text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
PROC. OF THE 7th EUR. CONF. ON PYTHON IN SCIENCE (EUROSCIPY 2014)
43
Py3DFreeHandUS: a library for voxel-array
reconstruction using Ultrasonography and attitude
sensors
Davide Monari∗† , Francesco Cenni† , Erwin Aertbeliën† , Kaat Desloovere†
arXiv:1412.6391v1 [cs.CV] 19 Dec 2014
F
Abstract—In medical imaging, there is a growing interest to provide real-time
images with good quality for large anatomical structures. To cope with this
issue, we developed a library that allows to replace, for some specific clinical
applications, more robust systems such as Computer Tomography (CT) and
Magnetic Resonance Imaging (MRI). Our python library Py3DFreeHandUS is
a package for processing data acquired simultaneously by ultra-sonographic
systems (US) and marker-based optoelectronic systems. In particular, US data
enables to visualize subcutaneous body structures, whereas the optoelectronic
system is able to collect the 3D position in space for reflective objects, that are
called markers. By combining these two measurement devices, it is possible
to reconstruct the real 3D morphology of body structures such as muscles, for
relevant clinical implications. In the present research work, the different steps
which allow to obtain a relevant 3D data set as well as the procedures for
calibrating the systems and for determining the quality of the reconstruction.
Index Terms—medical imaging, free-hand ultrasonography, optoelectronic systems, compounding
1
I NTRODUCTION
In medical imaging, 3D data sets are an essential tool to
explore anatomical volumes and to extract clinical features,
which can describe a particular condition of the patient. These
data are usually recorded by CT or MRI for identifying hard
or soft tissue, respectively, and provide a high image quality
together with a large field of view. On the other hand, these
systems are very expensive (espacially MRI ones) and time
consuming both for operators and patients. Plus, radioations
from CT are an issue. Therefore, for some clinical applications, it could be interesting to replace these systems with
others that can allow to provide 3D data sets quickly, although
without the same high image quality. Ultrasonography (US)
devices are systems largely used to collect medical images.
For example, it is very common to examine pregnant women.
This system, compared to other medical imaging systems, has
several advantages: real-time images, portability, no ionizing
radiation. However, one of the major drawbacks in US is
the limited field of view and the lack of spatial information
among different images acquired. Therefore a technique called
3D Freehand Ultrasound (3DUS) was originally proposed in
the 90s [Rankin93], [Prager99] with the aim of reconstructing
large 3D anatomical parts. The idea is to combine US images
and the corresponding position and orientation (POS) of the
US transducer; by simultaneously scanning a series of 2D
images and recording spatial information it is possible to
perform the relevant reconstruction and then the visualization
of the entire volume acquired. The aim of the present work
is to customize the 3DUS implementation by pushing on
vectorization in NumPy / SciPy along with memory waste
avoidance, for speeding up the processing phase as much as
possible. These aspects are essential in this context, since
for commodity hardware: i) memory resources are relatively
limited and 3D volumes involved here can quickly reach large
dimensions, ii) computation time can become unrealistic if
very large for- or while- loops are used in Python. In addition
the few existing applications for applying this technique have
at least one of the following disadvantages: i) not open-source;
ii) only supporting data streams from a limited number of
US/POS sensors; iii) they are written in low-level languages
such as C++, making rapid development and prototyping
more difficult. We developed a pure Python library called
Py3DFreeHandUS that solves all the above issues.
2
R EQUIREMENTS
Py3DFreeHandUS was developed in Python 2.7 (Python 3 not
yet supported), and uses the following libraries:
•
•
•
•
•
•
•
* Corresponding author: [email protected]
† KULeuven
c 2014 Davide Monari et al. This is an open-access artiCopyright ○
cle distributed under the terms of the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction
in any medium, provided the original author and source are credited.
http://creativecommons.org/licenses/by/3.0/
•
•
NumPy
SciPy (0.11.0+)
matplotlib
SymPy
pydicom
b-tk (Biomechanical ToolKit) [Barre14]
VTK
OpenCV (2.4.9+)
Cython + gcc (optional, we are "cythonizing" bottlenecks
but leaving pure Python implementation available)
We used the Python distribution Python(x,y) for development and testing, since it already includes all libraries but
b-tk.
44
3
PROC. OF THE 7th EUR. CONF. ON PYTHON IN SCIENCE (EUROSCIPY 2014)
D ESCRIPTION OF THE PACKAGE
The present package is able to process synchronized data
by US and POS, being as input DICOM and C3D files,
respectively. The operations flowchart is composed by: US
probe temporal and spatial calibration and 3D voxel array
reconstruction.
3.1
US probe temporal calibration
The aim of the temporal calibration is to estimate the time
delay between US and the POS devices. This procedure is
foundamental whenever it is not possible to hardware-trigger
US and POS devices, so data needs to be time-shifted later.
Time delay resolution cannot be lower than the inverse of the
lower frequency (normally US). Briefly, we moved vertically
up and down the US probe (rigidly connected to the POS
sensor) in a water-filled tank and generated two curves: the first
one being the vertical coordinate of the the POS sensor, the
second one being the vertical coordinate (in the US image) of
the center of the line representing the edge between water and
tank. These two sine-like signal were demeaned, normalized
to be inside the range [-1,+1] and cross-correlated with the
function matplotlib.pyplot.xcorr. The time of the
first peak for the cross-correlation estimates the time delay.
3.2
US probe spatial calibration
The probe spatial calibration is an essential procedure for
image reconstruction which allows to determine the pose
(position and orientation) of the US images with respect to
the POS device. The corresponding results take the form of
six parameters, three for position and three for orientation. The
quality of this step mainly influences the reconstruction quality
of the anatomical shape. To perform the probe calibration
we used two different steps. First we applied an established
procedure already published in the literature [Prager98] and
later we tuned the results by using an image compounding
algorithm [Wein08]. The established procedure was proposed
by Prager et al. [Prager98] and improved by Hsu [Hsu06], with
the idea of scanning the floor of a water tank by covering all
the degrees of freedom (see Figure 1); this scanning modality
produces clear and consistent edge lines (between water and
tank bottom) in the US images (B-scans). All the pixels lying
on the visible line in the B-scan should satisfy equations that
come from the different spatial transformations, which leave
to solve 11 identifiable parameters. Each B-scan can be used
to write 2 equations. The overdetermined set of equations is
solved using the Levenberg-Marquardt algorithm. We found
that it is essential to move the US transducers following the
sequence of movements suggested in [Prager98], in order to
have reasonable results. The equation that a pixel with image
coordinates (u, v) must satisfy (see [Prager98] for details) is
as follows:
0
sx u
0
= C TT T TR R TP sy v,
0
0
1
1
Fig. 1: The aim of the US probe spatial calibration is to find the
roto-translation matrix R TP from the image reference frame (P) to
the transducer reference frame (R). The other two roto-translation
matrices T TR and C TT (respectively, from transducer to optoelectronic
system and from optoelectronic system to calibration phantom) are
known for every time frame of the calibration acquisition.
where sx and sy are conversion factors from pixel to mm.
This is the code snippet for the equation creation:
from sympy import Matrix, Symbol, var
from sympy import cos as c, sin as s
# Pp
sx = Symbol(’sx’)
sy = Symbol(’sy’)
u = Symbol(’u’)
v = Symbol(’v’)
Pp = Matrix(([sx * u],\
[sy * v],\
[0],\
[1]\
))
# rTp
rTp, syms = creatCalibMatrix()
[x1, y1, z1, alpha1, beta1, gamma1] = syms
# tTr
tTr = MatrixOfMatrixSymbol(’tTr’, 4, 4)
tTr[3, 0:4] = np.array([0,0,0,1])
# cTt
x2 = Symbol(’x2’)
y2 = Symbol(’y2’)
z2 = Symbol(’z2’)
alpha2 = Symbol(’alpha2’)
beta2 = Symbol(’beta2’)
gamma2 = Symbol(’gamma2’)
cTt = Matrix(([c(alpha2)*c(beta2), ...
[s(alpha2)*c(beta2), ...
[-s(beta2), c(beta2)*s(gamma2), ...
[0, 0, 0, 1]\
)) # see [Prager98] for full expressions
# Calculate full equations
Pc = cTt * tTr * rTp * Pp
Pc = Pc[0:3,:]
# Calculate full Jacobians
x = Matrix([sx, sy, x1, y1, z1, alpha1, beta1,
gamma1, x2, y2, z2, alpha2, beta2, gamma2])
J = Pc.jacobian(x)
PY3DFREEHANDUS: A LIBRARY FOR VOXEL-ARRAY RECONSTRUCTION USING ULTRASONOGRAPHY AND ATTITUDE SENSORS
45
The equations system was solved by using the function
scipy.optimize.root with method=’lm’.
To validate the solution, the calibration part in this package
allows to visualize the corresponding covariance matrix; this
can be exploited to understand if some variable is not well
constrained. In addition, since in each B-scan it is necessary
to have the position for at least two pixels that belong to
the edge line, we developed an automatic tool for extracting
the corresponding lines in each image, based on the Hough
transform:
import cv2
# Threshold image
maxVal = np.iinfo(I.dtype).max
th, bw = cv2.threshold(I,np.round(thI*maxVal),
maxVal,cv2.THRESH_BINARY)
# Detect edges
edges = cv2.Canny(bw,thCan1,thCan2,
apertureSize=kerSizeCan)
# Dilate edges
kernel = np.ones(kerSizeDil,I.dtype)
dilate = cv2.dilate(edges, kernel, iterations=1)
# Find longest line
lines = cv2.HoughLinesP(dilate,1,np.pi/180,thHou,
minLineLength,maxLineGap)
maxL = 0
if lines == None:
a, b = np.nan, np.nan
else:
for x1,y1,x2,y2 in lines[0]:
L = np.linalg.norm((x1-x2,y1-y2))
if L > maxL:
maxL = L
a = float(y1 - y2) / (x1 - x2)
b = y1 - a * x1
# a, b being line parameters: y = a * x + b
Since we experienced unsatisfactory calibration results (in
terms of later reconstruction compounding) at this stage,
we passed those through an image compounding algorithm
which allows to achieve a good tuning. This is an image
based method which uses as input 2 perpendicular sweeps, at
approximately 90 degrees, for the same 3D volume [Wein08].
Briefly, a similarity measure (Normalized Cross Correlation,
NCC) between the two sweeps was applied to maximize this
measure with the final aim to find the calibration parameters
relative to the best overlapping between the images. The
initial values of this iterative method are the results of the
equations-based approach. A calibration quality assessment
was also implemented in terms of precision and accuracy
of the calibration parameters obtained. Precision gives an
indication of the dispersion of measures around their mean,
whereas the accuracy gives an indication of the difference
between the mean of the measures and the real value [Hsu06].
For example, this measure can be the known position of a point
in space (Point accuracy) or the known dimension of an object
(Distance accuracy).
Fig. 2: V’ is the smallest 3D voxel array parallelepipedon able to
contain all the US images. Others can be created, such as V, but they
are bigger, occupy more memory and contain more empty voxels.
The first step is to import the images (DICOM file, standard
format for medical imaging) and the synchronized kinematics
files (C3D format) containing pose data. A 3D voxel array
is then initialized. The 3D voxel array (a parallelepipedon)
should be the smallest one containing the sequence of all
the repositioned scans, as seen in Figure 2, in order to
avoid RAM waste. To face this issue, in the present package
two options are presented: reorienting manually the global
reference frame in order to be approximately aligned with
the scan direction during the acquisition; on the other hand,
by using the Principal Component Analysis (PCA), it is also
possible to find the scan direction and thereby realigning the
voxel array according to this direction.
The grey values of the original pixels in the 2D slices
are then copied in the new corresponding 3D position. This
procedure is performed by using an algorithm called Pixel
Nearest Neighbor (PNN) which runs through each pixel in
every image and fills the nearest voxel with the value of that
pixel; in case of multiple contributions to the same voxel, the
values are averaged. Below the code to perform this is shown.
Each 2D scan is positioned in the 3D volume in a vectorized
way.
# x, y, z: arrays for 3D coordinates of
# the pixels in image I
# idxV: unique ID for each voxel of the
# 3D voxel array
# V: 1D array containing grey values for the
# 3D voxel-array
# contV: 1D array containing current number of
# contributions for voxels
# I: 2D array containing US slice grey values
3.3
3D voxel array reconstruction
The 3D reconstruction is performed by positioning the 2D
US scans in the 3D space by using the corresponding pose.
idxV = xyz2idx(x, y, z, xl, yl, zl).astype(np.int32)
V[idxV] = (contV[idxV] * V[idxV]) / (contV[idxV] + 1)
+ I.ravel() / (contV[idxV] + 1) # iterative avg
46
PROC. OF THE 7th EUR. CONF. ON PYTHON IN SCIENCE (EUROSCIPY 2014)
Fig. 4: Three transversal and one longitudinal section of a reconstructed 3D voxel array (human calf scanning, about 90M voxels,
10mm3 each).
Fig. 3: Considering 2 US images consecutive in time, the convex hull
is the smaller object able to contain them. An easier shape can be
created, such as the parallelepipedon, but this is always bigger in
volume.
Only 2 outer loops exist, one for the DICOM file number
and one for the scan number. After all the scans are correctly
positioned in the 3D space, gaps can occur in the voxel array
when the voxel size is small compared to the distance between
the acquired images (e.g. scanning velocity significantly different from 0). Therefore interpolation methods are applied
for filling these empty voxels. For optimizing this process,
a robust method was also used, i.e. convex hull (see Figure
3), for restricting the gap filling operation only to the voxels
contained between 2 consecutive slices:
The quick-and-dirty way, known as VNN (Voxel Nearest
Neighbour), consists of filling a gap by using the closest voxel
having an assigned grey value. We also implemented another
(average cube) solution which consist of the following steps:
• Create a cube with side 3 voxels, centered around the
gap;
• Search the minimum percentage of non-gaps inside the
cube (100% = number of voxels in the cube);
• If that percentage is found, a non-gap voxels average
(weighted by the Euclidean distances) is performed into
the cube;
• If that percentage is not found, the cube size in incremented by 2 voxels (e.g. 5);
• If cube size is lesser or equal than a maximum size, start
again from point 2. Otherwise, stop and don’t fill the gap.
The entire voxel array can be subdivided in N parallelepipedal blocks, and the gap filling is performed on each
one at a time, to spare some of the RAM. The bigger the
number of blocks, the bigger the number of iterations to
go, but the smaller the block size, the RAM used and the
time spent per iteration. Finally, both the voxel array scans
silhouette (previously created with the wrapping convex hulls)
and the grey scale data voxel array are exported to VTI
files, after being converted to vtk.vtkImageData. These
can be opened with software like MeVisLab or Paraview for
visualization and further processing.
3.4
Preliminary results
The calibration quality assessments were 1.9 mm and 3.9
mm for the distance accuracy and reconstruction precision,
respectively. The average data processing time (calibration +
reconstruction + gap filling) over 3 trials on a human calf,
shown in Figure 4, was 5.9 min, on a 16 GB RAM Intel i7
2.7 GHz machine.
R EFERENCES
[Prager99] Prager RW, Gee AH, Berman L. Stradx: Real-time acquisition
and visualisation of freehand 3D ultrasound. Med Image Analysis
1999; 3(2):129-140.
[Rankin93] Rankin, R. N., Fenster, A., Downey, D. B., Munk, P. L., Levin,
M. F. and Vellet, A. D. (1993) Three-dimensional sonographic
reconstruction: techniques and diagnostic applications. Am. J.
Roentgenol., 161, 695-702.
[Prager98] Prager, R. W., Rohling, R. N., Gee, A. H. and Berman, L. (1998)
Rapid calibration for 3-D freehand ultrasound. Ultrasound Med.
Biol., 24, 855-869.
[Wein08] Wolfgang Wein and Ali Khamene. Image-Based Method for InVivo Freehand Ultrasound Calibration. Medical Imaging 2008:
Ultrasonic Imaging and Signal Processing, edited by Stephen
A. McAleavey, Jan D’hooge, Proc. of SPIE Vol. 6920, 69200K,
(2008).
[Hsu06]
Po-Wei Hsu, Richard W. Prager, Andrew H. Gee, and Graham
M. Treece. Rapid, easy and reliable calibration for freehand 3d
ultrasound. Ultrasound in Med. & Biol., Vol. 32, No. 6, pp. 823835, 2006.
[Barre14] Arnaud Barre, Stéphane Armand, Biomechanical ToolKit: Opensource framework to visualize and process biomechanical data,
Computer Methods and Programs in Biomedicine, Volume 114,
Issue 1, April 2014, Pages 80-87, ISSN 0169-2607.
| 5cs.CE
|
1
Anonymous Hedonic Game for Task Allocation in a
Large-Scale Multiple Agent System
arXiv:1711.06871v1 [cs.MA] 18 Nov 2017
Inmo Jang, Hyo-Sang Shin, and Antonios Tsourdos
Abstract—This paper proposes a novel game-theoretical autonomous decision-making framework to address a task allocation problem for a swarm of multiple agents. We consider
cooperation of self-interested agents and show that agents who
have social inhibition can converge to a Nash stable partition (i.e.,
social agreement) using our proposed decentralised algorithm
within polynomial time. The algorithm is simple and executable
based on local interactions with neighbour agents under a
strongly-connected communication network and even in asynchronous environments. We analytically present a mathematical
formulation for computing the lower bound of a converged
solution’s suboptimality and additionally show that 50 % of
suboptimality can be minimally guaranteed if social utilities
are non-decreasing functions with respect to the number of coworking agents. Through numerical experiments, it is confirmed
that the proposed framework is scalable, fast adaptable against
dynamical environments, and robust even in a realistic situation
where some of the agents temporarily somehow do not operate
during a mission.
Index Terms—Distributed robot systems, Networked robots,
Task allocation, Game theory, Self-organising systems
subgroups to each task. In the problem, it is assumed that
each agent can be assigned to at most one task, whereas each
task may require multiple agents: this case falls into ST-MR
(single-task robot and multi-robot task) category [10], [11].
According to [4], [5], [12]–[14], decision-making frameworks for such a large-scale multiple autonomous agent system
should be
•
•
•
•
I. I NTRODUCTION
Cooperation of a large number of possibly small-sized
robots, called robotic swarms, will play a significant role in
complex missions that existing operational concepts using a
few large robots could not deal with [1]. Even if every single
robot (or called agent) in a swarm is incapable of accomplishing a task alone, their cooperation will lead to successful
outcomes because the system is robust against failure or loss
of some individual agents [2]–[5]. The possible applications
include environmental monitoring [6], ad-hoc network relay
[7], disaster management [8], cooperative radar jamming [9],
to name a few.
Due to the large cardinality of a swarm robot system,
however, it is infeasible for human operators to supervise each
of them directly, but needed to entrust the swarm with certain
levels of decision-makings (e.g., task allocation, path planning,
and individual control). Thereby, only remained is to provide
a high-level mission description, which is manageable for a
few or even a single human operator. Nevertheless, there still
exist various challenges in the autonomous decision-making of
robotic swarms. Amongst them, this paper mainly addresses a
task allocation problem for a large-scale multiple agent system
where the number of agents is higher than that of tasks: how
to partition a set of agents into subgroups and assign the
Inmo Jang, Hyo-Sang Shin, and Antonios Tsourdos are with Centre for
Autonomous and Cyber-Physical Systems, Cranfield University, MK43 0AL,
United Kingdom (e-mail: [email protected]; [email protected];
[email protected]).
•
•
Decentralised: The desired collective behaviour can be
achieved not by any central control unit but by allowing
individual agents to locally make decisions based on local
information or local interactions with neighbour agents.
Scalable: The framework is operable for a wide range of
the system size (e.g., the number of agents).
Predictable: Human operators can estimate the quality
of an outcome obtained by the framework such as its
suboptimality. Preferably, a certain level of suboptimality
should be minimally guaranteed.
Flexible (or adaptable): The framework can quickly adapt
to any dynamic environments such as elimination or
addition of some agents or tasks.
Robust in asynchronous environments : Due to the large
cardinality of the system and its decentralisation, it is very
challenging for all the given agents to execute decisionmaking procedures synchronously. For synchronisation
in practice, “artificial delays and extra communication
must be built into the framework” [14], which may
cause considerable inefficiency on the system. Hence,
the framework can be operable even under asynchronous
environments.
Able to accommodate different interests of agents: Individual agents may have different levels of interest. Such
situations happen if the agents are “designed, owned, or
operated by several individuals or organisations that may
have different goals” [15].
In this paper, we propose a novel decision-making framework based on hedonic games [16]–[18], which model the
task allocation problem considered as a coalition-formation
game where self-interest agents are willing to form coalitions
to improve their own interests. The objective of this game is
to find a Nash stable partition, which is a social agreement
where all the agents agree with the current task assignment.
Despite any possible conflicts between the agents, this paper
shows that if they have social inhibition, then a Nash stable
partition can always be determined within polynomial times
in the proposed framework and all the desirable characteristics
mentioned above can be achieved. Furthermore, we analyse the
lower bound of the outcome’s suboptimality and show that 50
2
% of suboptimality is minimally guaranteed for a particular
case. Various settings of numerical experiments validate that
the proposed framework is scalable, flexible, and robust even
in asynchronous environments.
This paper is organised as follows. Section II reviews
existing literature on decentralised task allocation schemes for
robotic swarms, and introduces a recent finding in hedonic
game domains that inspires this study. Section III proposes
a decision-making framework based on anonymous hedonic
games, named GRAPE, and analytically proves the existence
of and the polynomial-time convergence to a Nash stable
partition. Section IV discusses the proposed framework’s
algorithmic complexity (i.e., scalability), suboptimality, and
adaptability. In Section V, we show that the framework can
also address a task allocation problem in which each task may
need a certain number of agents for completion. Numerical
simulations in Section VI confirm that the proposed framework
holds all the desirable characteristics. Finally, concluding
remarks are followed in Section VII.
II. R ELATED W ORK
A. Decentralised Coordination of Robotic Swarms
Existing approaches for task allocation problems can be
categorised into two branches, depending on how agents eventually reach a converged outcome: orchestrated and (fully) selforganised approaches [19]. In the former, additional mechanism such as negotiation and voting model is imposed so
that some agents can be worse off if a specific condition
is met (e.g., the global utility is better off). Alternatively,
in self-organised approaches, each agent simply makes a
decision without negotiating with other agents. The latter
generally induce less resource consumption in communication
and computation [20], and hence they are preferable in terms
of scalability. On the other hand, the former usually provide
a better quality of solutions with respect to the global utility,
and a certain level of suboptimality could be guaranteed [21]–
[23]. A comparison result between them [20] presents that
as the available information to agents becomes local, the
latter becomes to outperform the former. In the following,
we particularly review existing literature on self-organised
approaches because, for large-scale multiple agent systems,
scalability is a minimally essential feature and it is realistic
to regard that the agents know not the global information but
only each one’s local information.
Self-organised approaches can be categorised into top-down
approaches and bottom-up approaches according to which
level (i.e., between an ensemble and individuals) are mainly
focused on. Top-down approaches emphasise developing the
macroscopic model for a whole system. For instances, population fractions associated with given tasks are represented
as states, and the dynamics of the population fractions are
modelled by Markov chains [12], [24]–[26] or differential
equations [27]–[31]. Given a desired fraction distribution over
the given tasks, agents can converge to the desired status by
following local decision policies (e.g., the associated rows or
columns of the current Markov matrix). One advantage of
using top-down approaches is that, by simulation, a system’s
emergent behaviour can be averagely predicted regarding
convergence speed and the quality of a stable outcome (i.e.,
how well the agents converge to the desired fraction distribution) However, as top-down generated control policies
regulate agents, it may be difficult to accommodate each
agent’s individual preference. Also, each agent may have to
physically move around according to its local policy during
the entire decision-making process, and this fact may cause
unnecessary time and energy costs for the transitioning.
Bottom-up approaches focus on designing of a set of
individual rules (i.e., microscopic models) for every agent
that eventually lead to a certain emergent behaviour. Possible
behaviours of a single agent can be modelled by a finite state
machine [32], and change of behaviours occurs according to
a probabilistic threshold model [33]. A threshold value in
the model, which suggests the decision boundary between
motion A to motion B, is adjustable based on an agent’s past
experiences such as time spent for working a task [19], [34],
the success/failure rates [32], [35], and direct communication
from a central unit [33]. This feature can improve system
adaptability, and may have a potential to incorporate each
agent’s individual interest if required. However, it has been
shown in [35]–[41] that, in order to predict or evaluate an
emergent performance of a swarm utilising bottom-up approaches, a macroscopic model for the swarm eventually has
to be developed by abstracting the microscopic models.
B. Hedonic Games
Hedonic games [16]–[18] model a conflict situation where
self-interest agents are willing to form coalitions to improve
their own interests. Nash stability [18], which is inspired by
Nash equilibria in other game theories, plays a key role since
it yields a social agreement amongst the agents even without
having any negotiation between them. Many researchers have
investigated conditions under which a Nash stable partition
is guaranteed to exist and to be determined [18], [42]–[44].
Amongst them, the works in [43], [44] mainly addressed an
anonymous hedonic game, in which each agent considers the
size of a coalition to which it belongs instead of the identities
of the members. Recently, Darmann [44] showed that selfish
agents who have social inhibition (i.e., preference toward a
coalition with a fewer number of members) could converge to
a Nash stable partition in an anonymous hedonic game. The
author also proposed a centralised recursive algorithm that can
find a Nash stable partition within O(n2a · nt ) of iterations.
Here, na is the number of agents and nt is that of tasks.
C. Main Contributions
Inspired by the recent breakthrough of [44], we propose a
novel decentralised framework that models the task allocation
problem considered as an anonymous hedonic game. The
proposed framework is a self-organised approach in that agents
make decisions according to its local policies (i.e., individual
preferences). Unlike top-down or bottom-up approaches reviewed in the previous section, which primarily concentrate
on designing agents’ decision-making policies either macroscopically or microscopically, our work instead focuses on investigating and exploiting advantages from socially-inhibitive
3
agents, while simply letting them greedily behave according to
their individual preferences. Explicitly, the main contributions
of this paper are as follows:
1) This paper shows that selfish agents with social inhibition, which we refer to as SPAO preference (Definition
4), can reach a Nash stable partition within less algorithmic complexity compared with [44]: O(n2a ) of iterations
are required1 .
2) We provide a decentralised algorithm for each agent,
which is executable under a strongly-connected communication network of agents and even in asynchronous
environments. Depending on the network assumed, the
algorithmic complexity may be additionally increased
by O(dG ), where dG < na is the graph diameter of the
network.
3) This paper analyses the suboptimality of a Nash-stable
partition in term of the global utility. We firstly present a
mathematical formulation to compute the suboptimality
lower bound by using the information of a Nash stable
partition and agents’ individual utilities. Furthermore,
we additionally show that 50 % of the suboptimality
can be at least guaranteed if the social utility for each
coalition is defined as a non-decreasing function with
respect to the number of members in the coalition.
4) Our framework can accommodate different agents with
different interests as long as their individual preferences
hold SPAO.
5) Through various numerical experiments, it is confirmed
that the proposed framework is scalable, fast adaptable
to environmental changes, and robust even in a realistic
situation where some of agents are temporarily unable to
proceed a decision-making procedure and communicate
with other agents during a mission.
III. GROUP AGENT PARTITIONING AND PLACING E VENT
A. Problem Formulation
Firstly, we introduce the multi-robot task allocation problem
considered in this paper and underlying assumptions regarding
agents and tasks.
Problem 1. Suppose that there exist a set of na agents A =
{a1 , a2 , ..., ana } and a set of tasks T = T ∗ ∪ {tφ }, where
T ∗ = {t1 , t2 , ..., tnt } is a set of nt tasks and tφ is the void
task (i.e., not to perform any task). Each agent ai has the
individual utility ui : T × |A| → R, which is a function of
the task to which the agent is assigned and the number of coworking agents for the task (including itself) p ∈ {1, 2, ..., na }.
Since every agent is considered to have limited capabilities to
finish a task alone, the agent can be assigned to at most one
task. The objective of this task allocation problem is to find an
assignment that maximises the global utility, i.e., the sum of
individual utilities of the entire agents. The problem described
above is defined as follows:
X X
max
ui (tj , p)xij
(1)
{xij }
∀ai ∈A ∀tj ∈T
1 Note that the definition of iteration is described in Definition 5. This
comparison assumes the fully-connected communication network because the
algorithm in [44] is centralised.
TABLE I
N OMENCLATURE
Symbol
Description
A
ai
T∗
tj
tφ
T
(tj , p)
X
a set of na agents
the i-th agent
a set of nt tasks
the j-th task
the void task (i.e., not to work any task)
a set of tasks, T = T ∗ ∪ {tφ }
a task-coalition pair (i.e. to do task tj with p participants)
the set of task-coalition pairs, X = X ∗ ∪ {tφ },
where X ∗ = T ∗ × {1, 2, ..., n}
agent ai ’s preference relation over X
the strong preference of agent ai
the indifferent preference of agent ai
the weak preference of agent ai
a partition: a disjoint set that partitions the agent set A,
Π = {S1 , S2 , ..., Sm , Sφ }
the task-specific coalition for tj
the index of the task to which agent ai is assigned given Π
the graph diameter of the agent communication network
Pi
i
∼i
i
Π
Sj
Π(i)
dG
subject to
X
xij ≤ 1
∀ai ∈ A
(2)
∀tj ∈T
xij ∈ {0, 1}
∀ai ∈ A, ∀tj ∈ T
(3)
where xij is a binary decision variable that indicates whether
or not task tj is assigned to agent ai .
This paper uses the term social utility to indicate the sum
of total individual utilities within any agent group.
Assumption 1 (Homogeneous agents with limited capabilities). This paper considers a large-scale multi-robot system
that consists of physically homogeneous agents. This can be
justified because the realisation of a swarm can be in general
achieved through mass production [4]. Hence, each individual
utility ui is concerned with not the combinations of the agents
working for the task, but their cardinality. Despite that, it
is worth noting that agents in this paper may have different
preferences with respect to the given tasks (e.g., for an agent,
a spatially closer task is more preferred, whereas this may
not be the case for another agent). Besides, noting that “mass
production favours robots with fewer and cheaper components,
resulting in lower cost but also reduced capabilities [45]”, we
also assume that each agent can be only assigned to perform
at most a single task. According to [10], such a robot is called
a single-task (ST) robot.
Assumption 2 (Agents’ communication). The communication
network of the entire agents is at least strongly-connected.
Given a network, Ni denotes a set of neighbour agents for
agent ai .
Assumption 3 (Multi-robot-required tasks). Every task is a
multi-robot (MR) task, meaning that the task can require
multiple robots [10]. For now, we assume that each task can be
performed even by a single agent although it may take a long
4
time. However, in Section V, we will also address a particular
case in which some of the tasks need a certain number of
agents for completion.
Assumption 4 (Agents’ pre-known information). Every agent
ai only knows its own individual utility ui (tj , p) with regard to
every task tj , while not being aware of those of other agents.
Through communication, however, they can notice which
agent currently choses which task, i.e., partition (Definition
2). Note that the agents do not necessarily have to know the
true partition information at all the time. Each agent owns its
locally-known partition information.
Remark 1 (An advantage of Nash stability: low communication burden on agents). The rationale for the use of Nash
stability amongst various stable solution concepts in hedonic
games [16], [46]–[48] is that it can reduce communication
burden between agents required to reach an agreed assignment.
In the process of converging to a Nash stable partition, an
agent does not need to get any permission from other agents
when it is willing to deviate. This property may not be the case
for the other solution concepts. Therefore, each agent is only
required to notify its altered decision without any negotiation
with other agents. This fact can reduce communication burden
between the agents for collective decision-making in the
proposed approach.
B. Proposed Game-theoretical Approach: GRAPE
Let us transform Problem 1 into an anonymous hedonic
game event where every agent selfishly tends to join a coalition
according to its preference.
Definition 1 (GRAPE). An instance of GRoup Agent Partitioning and placing Event (GRAPE) is a tuple (A, T , P)
that consists of (1) A = {a1 , a2 , ..., ana }, a set of na
agents; (2) T = T ∗ ∪ {tφ }, a set of tasks; and (3)
P = (P1 , P2 , ..., Pna ), an na -tuple of preference relations of
the agents. For agent ai , Pi describes its preference relation
over the set of task-coalition pairs X = X ∗ ∪ {tφ }, where
X ∗ = T ∗ × {1, 2, ..., na }; a task-coalition pair (tj , p) is interpreted as “to do task tj with p participants”. For any
task-coalition pairs x1 , x2 ∈ X , x1 i x2 implies that agent
ai strongly prefers x1 to x2 , and x1 ∼i x2 means that the
preference regarding x1 and x2 is indifferent. Likewise, i
indicates the weak preference of agent ai .
Note that agent ai ’s preference relation can be derived
from its individual utility ui (tj , p) in Problem 1. For instances, given that ui (t1 , p1 ) > ui (t2 , p2 ), it can be said that
(t1 , p1 ) i (t2 , p2 ).
Definition 2 (Partition). Given an instance (A, T , P)
of GRAPE, a partition is defined as a set Π =
{S1 , S2 , ..., Sm , Sφ } that disjointly partitions the agent set A.
Here, Sj ⊆ A is the (task-specific) coalition for executing
task tj such that ∪m
j=0 Sj = A and Sj ∩ Sk = ∅ for j 6= k.
Sφ is the set of agents who choose the void task tφ . Note
that this paper interchangeably uses S0 to indicate Sφ . Given
a partition Π, Π(i) indicates the index of the task to which
agent ai is assigned. For examples, SΠ(i) is the coalition that
the agent belongs to, i.e., SΠ(i) = {Sj ∈ Π | ai ∈ Sj }.
The objective of GRAPE is to determine a stable partition
that all the agents agree on. In this paper, we seek for a Nash
stable partition, which is defined as follows:
Definition 3 (Nash stable). A partition Π is said to
be Nash stable if, for every agent ai ∈ A, it holds that
(tΠ(i) , |SΠ(i) |) i (tj , |Sj ∪ {ai }|), ∀Sj ∈ Π.
In other words, in a Nash stable partition, every agent
prefers its current coalition to joining any of the other coalitions. Thus, every agent does not have any conflict within this
partition, and no agent will not unilaterally deviate from its
current decision.
C. SPAO Preference: Social Inhibition
This section introduces the key condition, called SPAO, that
enables our proposed approach to provide all the desirable
properties described in Section I, and then explains its implications.
Definition 4 (SPAO). Given an instance (A, T , P) of GRAPE,
it is said that the preference relation of agent ai with
respect to task tj is SPAO (Single-Peaked-At-One) if it
holds that, for every (tj , p) ∈ X ∗ , (tj , p1 ) i (tj , p2 ) for any
p1 , p2 ∈ {1, ..., |A|} such that p1 < p2 . Besides, we say that
an instance (A, T , P) of GRAPE is SPAO if the preference
relation of every agent in A with respect to every task in T ∗
is SPAO.
For an example, suppose that Pi is such that
(t1 , 1) i (t1 , 2) i (t1 , 3) i (t2 , 1) ∼i (t1 , 4) i (t2 , 2).
This preference relation indicates that agent ai has (t1 , 1) i
(t1 , 2) i (t1 , 3) i (t1 , 4) for task t1 , and (t2 , 1) i (t2 , 2)
for task t2 . According to Definition 4, the preference relation
for each of the tasks holds SPAO. For another example, given
that
(t1 , 1) i (t1 , 2) i (t1 , 3) i (t2 , 2) ∼i (t1 , 4) i (t2 , 1),
the preference relation regarding task t1 holds SPAO, whereas
this is not the case for task t2 because of (t2 , 2) i (t2 , 1).
This paper only considers the case in which every agent
in an instance of GRAPE has SPAO preference relations
regarding all the given tasks. In this case, it is said that the
instance is SPAO, as mentioned in Definition 4. Agents under
this condition prefer to execute a task with smaller number of
collaborators, namely, they have social inhibition.
Remark 2 (Implications of SPAO). SPAO implies that an
agent’s individual utility should be a monotonically decreasing
function with respect to the size of a task-specific coalition. In
practice, SPAO can often emerge. For instance, experimental
and simulation results in [49, Figures 3 and 4] show that the
total work capacity resulted from a cooperation of multiple
robots does not proportionally increase due to interferences of
the robots. In such a non-superadditive environment [50], assuming that an agent’s individual work efficiency is considered
as its individual utility, the individual utility monotonically
drops as the number of collaborators enlarges even though
5
the social utility is increased. For another example, SPAO
also arises when individual utilities are related with sharedresources. As more agents use the same resource simultaneously, their individual productivities become diminished
(e.g., traffic affects travel times [51] [52, Example 3]). As
the authors in [50] pointed out, a non-superadditive case
is more realistic than a superadditive case: agents in a superadditive environment always attempt to form the grand
coalition whereas those in a non-superadditive case are willing
to reduce unnecessary costs. Note that social utility functions
are not restricted so that they can be either monotonic or nonmonotonic.
Remark 3 (Cooperation of selfish agents). The proposed
framework can accommodate selfish agents who greedily follow their individual preferences as long as the preferences hold
SPAO. This implies that the framework may be utilised for a
combination of swarm systems from different organisations.
D. Existence of and Convergence to a Nash Stable Partition
Let us prove that if an instance of GRAPE holds SPAO,
there always exists a Nash stable partition and it can be found
within polynomial time.
Definition 5 (Iteration). This paper uses the term iteration to
represent an iterative stage in which an arbitrary-chosen agent
compares the set of selectable task-coalition pairs given an
existing partition, and then determines whether or not to join
another coalition including the void task one.
Assumption 5 (Mutual exclusion algorithm). We assume that,
at each iteration, a single agent exclusively makes a decision
and updates the partition Π if necessary. This paper refers to
this agent as the deciding agent at the iteration. Based on the
resultant partition, another deciding agent also performs the
same process at the next iteration, and this process continues
until every agent does not deviate from a specific partition,
which is, in fact, a Nash stable partition. To implement this
algorithmic process in practice, the agents need a mutual
exclusion (or called mutex) algorithm to choose the deciding
agent at each iteration. In this section, for simplicity of
description, we assume that all the agents are fully-connected,
by which they somehow select and know the deciding agent.
However, in Section III-E, we will present a distributed mutex
algorithm that enables the proposed approach to be executed
under a strongly-connected communication network even in
an asynchronous manner.
as:
∆Π(i) := min{∆ | (tΠ(i) , |SΠ(i) | + ∆) i (tj , |Sj ∪ {ai }|),
∀Sj ∈ Π \ {SΠ(i) }, ∆ ∈ Z}.
(4)
Due to the SPAO preference relations, this value satisfies the
following characteristics: (a) if Π is Nash stable, for every
agent ai , it holds that ∆Π(i) ≥ 0; (b) if ∆Π(i) < 0, then
agent ai is willing to deviate to the most preferred coalition
at a next iteration; and (c) for the agent ai who deviated at
the last iteration and updated the partition as Π0 , it holds that
∆Π0 (i) ≥ 0.
From Definition 4, it is clear that the new instance (Ã, T , P)
still holds SPAO. Let Π0 denote a Nash stable partition in the
original instance (A, T , P). When a new agent ar ∈
/ A decides
to execute one of tasks in T and creates a new partition Π1 , it
holds that ∆Π1 (r) ≥ 0, as shown in (c). If there is no existing
agent aq ∈ A whose ∆Π1 (q) < 0, then the new partition Π1
is Nash stable.
Suppose that there exists at least an agent aq whose
∆Π1 (q) < 0. Then, the agent must be one of members in
the task-specific coalition that agent ar selected in the last
iteration. As agent aq moves to another coalition and creates
a new partition Π2 , for the previously-deviated agent ar , it
holds that ∆Π2 (r) ≥ 1. In other words, an agent who deviates
to a coalition and expels one of the existing agents in that
coalition will not deviate again, even if another agent joins
the coalition in a next iteration. This implies that at most |Ã|
of iterations are required to hold ∆Π̃(i) ≥ 0 for every agent
ai ∈ Ã, where the partition Π̃ is Nash stable.
Lemma 1 is an essential property not only for the existence
of and convergence to a Nash stable partition, but also for fast
adaptability to dynamic environments.
Theorem 1 (Existence). If (A, T , P) is an instance of GRAPE
holding SPAO, then a Nash stable partition always exists.
Lemma 1. Given an instance (A, T , P) of GRAPE that is
SPAO, suppose that a new agent ar ∈
/ A holding SPAO preference relations with regard to every task in T joins (A, T , P)
in which a Nash stable partition is already established. Then,
the new instance (Ã, T , P), where à = A ∪ {ar }, also (1)
satisfies SPAO; (2) contains a Nash stable partition; and (3)
the maximum number iterations required to re-converge to a
Nash stable partition is |Ã|.
Proof. This theorem will be proved by induction. Let M (n)
be the following mathematical statement: For |A| = n, if an
instance (A, T , P) of GRAPE is SPAO, then there exists a
Nash stable partition.
Base case : When n = 1, there is only one agent in an
instance. This agent is allowed to participate in its most
preferred coalition, and the resultant partition is Nash stable.
Therefore, M (1) is true.
Induction hypothesis : Assume that M (k) is true for a
positive integer k such that |A| = k. Let Π denote a Nash
stable partition in an instance (A, T , P).
Induction step : Suppose that a new agent ai ∈
/ A whose
preference relation regarding every task in T is SPAO joins
the instance (A, T , P). This induces a new instance (Ã, T , P)
where à = A ∪ {ai } and |Ã| = k + 1. From Lemma 1, it is
clear that the new instance also satisfies SPAO and has a
Nash stable partition Π̃. Consequently, M (k + 1) is true. By
mathematical induction, M (n) is true for all positive integers
n ≥ 1.
Proof. Given a partition Π, the number of agents that are
additionally tolerated by agent ai to her coalition is defined
Theorem 2 (Convergence). If (A, T , P) is an instance of
GRAPE holding SPAO, then the number of iterations re-
6
quired to determine a Nash stable partition is at most
|A| · (|A| + 1)/2.
Proof. Suppose that, given a Nash stable partition in an
instance where there exists only one agent, we add another
arbitrary agent and find a Nash stable partition for this new
instance, and repeat the procedure until all the agents in
A are included. From Lemma 1, if a new agent joins an
instance in which the current partition is Nash stable, then the
maximum number of iterations required to find a new Nash
stable partition is the number of the existing agents plus 1.
Therefore, it is trivial that the maximum number of iterations
to find a Nash stable partition of an instance (A, T , P) is given
as
|A|
X
k = |A| · (|A| + 1)/2.
(5)
k=1
Note that this is also the case even if the agents are
initialised into a random partition. Suppose that we have the
following setting: the entire agents A are firstly not movable
from the existing partition, except a set of free agents A0 ⊆ A;
whenever the agents A0 find a Nash stable partition Π0 , one
arbitrary agent in ar ∈ A \ A0 additionally becomes liberated
and deviates from the current coalition SΠ0 (r) to another
coalition in Π0 . In this setting, for the viewpoint of the agents
in A0 \SΠ0 (r) , the newly liberated agent is considered as a new
agent as that in Lemma 1. Accordingly, we can still utilise
Lemma 1 for the agents in A0 \ SΠ0 (r) ∪ {ar }. The agents
also can find a Nash stable partition if one of them moves
to SΠ0 (r) during the process, because, due to ar , it became
∆Π0 (i) ≥ 1 for every agent ai ∈ SΠ0 (r) \ {ar }. In a nutshell,
the agents A ∪ {ar } can converge to a Nash stable partition
within |A0 ∪{ar }|, as shown in Lemma 1, and hence Theorem
1 and this theorem are also valid for the case when the initial
partition of the agents are randomly given.
E. Decentralised Algorithm
In the previous section, it is assumed that only one agent
is somehow chosen to make a decision at each iteration.
On the contrary, in this section, we propose a decentralised
algorithm, as shown in Algorithm 1, in which every agent
does decision making and affects its neighbour agents simultaneously. Despite that, we show that the agents can converge
to a Nash stable partition thanks to our proposed distributed
mutex subroutine shown in Algorithm 2. The details of the
decentralised main algorithm are as follows.
Each agent ai has local variables such as Πi , satisfied,
i
r , and si (Line 1–2). Here, Πi is the agent’s locally-known
partition; satisfied is a binary variable that indicates whether or
not the agent satisfies with Πi ; ri ∈ Z+ is an integer variable
to represent how many times Πi has evolved (i.e., the number
of iterations happened for updating Πi until that moment);
and si ∈ [0, 1] is a uniform-random variable that is generated
when Πi is newly updated (i.e., a random time stamp). Given
Πi , agent ai examines which coalition is the most preferred
amongst others, assuming that other agents remain at the
existing partition (Line 5). Then, the agent joins the newly
found coalition if it is strongly preferred than the existing
Algorithm 1 Decision-making algorithm for each agent ai
// Initialisation
1: satisfied ← f alse; r i ← 0; si ← 0
2: Πi ← {Sφ = A, Sj = φ ∀tj ∈ T }
// Decision-making process begins
3: while true do
// Make a new decision if necessary
4:
if satisfied = f alse then
5:
(tj∗ , |Sj∗ |) ← arg max∀Sj ∈Πi (tj , |Sj ∪ {ai }|)
6:
if (tj∗ , |Sj∗ |) i (tΠi (i) , |SΠi (i) |) then
7:
Join Sj∗ and update Πi
8:
ri ← ri + 1
9:
si ∈ unif[0, 1]
10:
end if
11:
satisfied = true
12:
end if
// Broadcast the local information to neighbour agents
13:
Broadcast M i = {ri , si , Πi } and receive M k from its
neighbours ∀ak ∈ Ni
// Select the valid partition from all the received messages
14:
Construct Mircv = {M i , ∀M k }
15:
{ri , si , Πi }, satisfied ← D-M UTEX(Mircv )
16: end while
coalition. In this case, the agent updates Πi to reflect its new
decision, increases ri , and generates a new random time stamp
si (Line 6–10). In any case, since the agent ascertained that
the currently-selected coalition is the most preferred, the agent
becomes satisfied with Πi (Line 11). Then, agent ai generates
a message M i := {ri , si , Πi } and sends it to other neighbour
agents, and vice versa (Line 13).
Since every agent locally updates its locally-known partition
simultaneously, one of the partitions should be regarded as if it
were the partition updated by a deciding agent at the previous
iteration. We refer to this partition as the valid partition at the
iteration. The distributed mutex subroutine in Algorithm 2)
enables the agents to recognise the valid partition amongst
all the locally-known current partitions Πi ∀i even under
a strongly-connected network and in asynchronous environments. Before executing this subroutine, each agent ai collects
all the messages received from its neighbour agents in Ni
(including M i ) as Mircv = {M i , ∀M k }, where k : ak ∈ Ni .
Using this message set, the agent examines whether or not its
own partition Πi is valid. If there exists any other partition
Πk such that rk > ri , then the agent considers Πk more valid
than Πi . This also happens if sk > si and rk = ri . Since
Πk is considered as more valid, agent ai needs to re-examine
if there is a more preferred coalition given Πk in the next
iteration. Thus, the agent sets satisfied as f alse. (Line 3–10
in Algorithm 2). After that, depending on satisfied, each agent
proceeds the decision-making process again (i.e., Line 4–12 in
Algorithm 1) and/or just broadcasts the existing locally-known
partition to its neighbour agents (Line 13 in Algorithm 1).
In a nutshell, the distributed mutex algorithm makes sure
that there is only one valid partition that dominates (or will
finally dominate depending on the communication network)
any other partitions. In other words, regardless of the status
7
Algorithm 2 Distributed Mutex Subroutine
1: function D-M UTEX (Mircv )
2:
satisfied ← true
3:
for each message Mk ∈ Mircv do
4:
if (ri < rk ) or (ri = ri & si < sk ) then
5:
ri ← rk
6:
si ← sk
7:
Πi ← Πk
8:
satisfied ← f alse
9:
end if
10:
end for
11:
return {ri , si , Πi }, satisfied
12: end function
of the communication network and synchronisation amongst
agents, multiple partitions locally evolve and some of them
only eventually can survive at every main loop of Algorithm
1. In an extreme case, we may encounter multiple Nash stable
partitions in the very last (i.e., at n2a /2 of iterations as shown
in Theorem 2). Nevertheless, thanks to the mutex algorithm,
one of them can be distributedly selected by the agents. This
implies that agents using Algorithm 1 can find a Nash stable
partition in a decentralised manner.
IV. A NALYSIS
Remark 4 (The number of required iterations in practice).
Algorithm 1 allows the entire agents in A to involve the
decision-making process, whereas, in the proof for Theorem
2, a new agent can be involved after a Nash stable partition
of existing agents is found. Since agents using Algorithm 1
do not need to find every Nash stable partition for each subset
of the agents, unnecessary iterations can be reduced. Hence,
the number of required iterations in practice may become less
than that shown in Theorem 2, which is also supported by the
experimental results in Section VI.
B. Suboptimality
This section investigates the suboptimality lower bound (or
can be called approximation ratio) of the proposed framework
in terms of the global utility, i.e., the objective function in
Equation (1). Given a partition Π, the global utility value can
be equivalently rewritten as
X
ui (tΠ(i) , |SΠ(i) |).
(6)
J=
∀ai ∈A
Note that we can simply derive {xij } for Equation (1) from
Π for Equation (6), and vice versa. Let JGRAP E and JOP T
represent the global utility of a Nash stable partition obtained
by the proposed framework and the optimal value, respectively.
This paper refers to the fraction of JGRAP E with respect to
JOP T as the suboptimality of GRAPE, denoted by α, i.e.,
A. Algorithmic Complexity (Scalability)
Firstly, let us discuss about the running time for the proposed framework to find a Nash stable partition. This paper
refers to a unit time required for each agent to proceed the
main loop of Algorithm 1 (Line 4-15) as a time step. Depending on the communication network considered, especially if
it is not fully-connected, it may be possible that all the given
agents have to execute this loop to just propagate the valid
partition information over the agents without affecting ri as
Line 8. Because this process also spends a unit time step,
we call it as dummy iteration to distinguish from a (normal)
iteration, which increases ri .
Notice that such dummy iterations only sequentially happen
at most dG times before a normal iteration occurs, where dG
is the graph diameter of the communication network. Hence,
thanks to Theorem 2, the total required time steps until finding
a Nash stable partition is O(dG n2a ). For the fully-connected
network case, it becomes O(n2a ) because of dG = 1. Note that
this algorithmic complexity is less than that of the centralised
algorithm, i.e., O(n2a · nt ), in [44].
Every agent at each iteration investigates nt + 1 of selectable task-coalition pairs including (tφ , Sφ ) given a locallyknown valid partition (as shown in Line 5 in Algorithm 1).
Therefore, the computational overhead for an agent is O(nt )
per iteration. With consideration of the total required time
steps, the running time of the proposed approach for an agent
can be bounded by O(dG nt n2a ). Note that the running time
in practice can be much less than the bound since Theorem
2 was conservatively analysed, as described in the following
remark.
α := JGRAP E /JOP T .
(7)
The lower bound of the suboptimality can be determined
by the following theorem.
Theorem 3 (Suboptimality lower bound: general case). Given
a Nash stable partition Π obtained by GRAPE, its suboptimality in terms of the global utility is lower bounded as follows:
α ≥ JGRAP E /(JGRAP E + λ),
(8)
where
λ≡
X
∀Sj ∈Π
max
ai ∈A,p≤|A|
p· ui (tj , p)−ui (tj , |Sj ∪{ai }|)
(9)
Proof. Let Π∗ denote the optimal partition for the objective
function in Equation (6). Given a Nash stable partition Π, from
Definition 3, it holds that ∀ai ∈ A
ui (tΠ(i) , |SΠ(i) |) ≥ ui (t∗j←i , |Sj ∪ {ai }|),
(10)
where t∗j←i indicates task tj ∈ T to which agent ai should
have joined according to the optimal partition Π∗ ; and Sj ∈ Π
is the coalition for task tj whose participants follow the Nash
stable partition Π.
The right-hand side of the inequality in Equation (10) can
be rewritten as
ui (t∗j←i , |Sj ∪ {ai }|) = ui (t∗j←i , |Sj∗ |)−
ui (t∗j←i , |Sj∗ |) − ui (t∗j←i , |Sj ∪ {ai }|) ,
(11)
where Sj∗ ∈ Π∗ is the ideal coalition of task t∗j←i that maximises the objective function.
8
By summing the individual utilities of all the agents, the
inequality in Equation (10) can be said that
X
ui (tΠ (i), |SΠ (i)|)
∀ai ∈A
X
≥
∀ai ∈A
ui (t∗j←i , |Sj∗ |)
∀ai ∈A
X
−
{ui (t∗j←i , |Sj∗ |) − ui (t∗j←i , |Sj ∪ {ai }|)}.
∀ai ∈A
(12)
The left-hand side of the inequality in Equation (12) represents the objective function value of the Nash stable partition
Π, i.e., JGRAP E , and the first term of the right-hand side is the
optimal objective function value, i.e., JOP T . The second term
in the right-hand side can be interpreted as the summation of
the utility lost of each agent caused by the belated decision
to its optimal task, provided that other agents still follow the
Nash stable partition.
The upper bound of the second term is given by
|T \{tφ }|
X
|Sj∗ |
j=1
·
max {ui (t∗j←i , |Sj∗ |)
ai ∈Sj∗
−
ui (t∗j←i , |Sj
∪ {ai }|)}.
(13)
This is at most
X
∀Sj ∈Π
max
ai ∈A,p≤|A|
Lij [p] ≡ λ,
(14)
where Lij [p] = p · (ui (tj , p) − ui (tj , |Sj ∪ {ai }|)).
Hence, the inequality in Eqn (12) can be rewritten as
JGRAP E ≥ JOP T − λ.
Dividing the both sides by JGRAP E and rearranging them
yield the suboptimality lower bound of the Nash stable partition, as given by Equation (8).
Although Theorem 3 does not provide a fixed-value lower
bound, it can be determined as long as a Nash stable partition
and agents’ individual utility functions are given. Nevertheless,
as a special case, if the social utility for any coalition is
non-decreasing (or monotonically increasing) in terms of the
number of co-working agents, then we can obtain a fixed-value
lower bound for the suboptimality of a Nash stable partition.
Theorem 4 (Suboptimality lower bound: a special case).
Given an instance (A, T , P) of GRAPE, if (i) the social utility
for any coalition is non-decreasing with regard to the number
of participants, i.e., for any Sj ⊆ A and al ∈ A \ Sj , it holds
that
X
X
ui (tj , |Sj |) ≤
ui (tj , |Sj ∪ {al }|),
∀ai ∈Sj
Proof. Firstly, we introduce some definitions and notations
that facilitate to describe this proof. Given a partition Π of
an instance (A, T , P), the global utility is denoted by
X
V (Π) :=
ui (tΠ(i) , |SΠ(i) |).
(15)
∀ai ∈Sj ∪{al }
and (ii) all the individual utilities can derive SPAO preference
relations, then a Nash stable partition Π obtained by GRAPE
guarantees to provide 50% of suboptimality in terms of the
global utility.
We use the operator ⊕ as follows. Given any two partitions
ΠA = {S0A , ..., SnAt } and ΠB = {S0B , ..., SnBt },
ΠA ⊕ ΠB := {S0A ∪ S0B , S1A ∪ S1B , ..., SnAt ∪ SnBt }.
A
Note that, in this proof, we consider that ∪m
=
j=0 Sj
m
B
∪j=0 Sj = A. Thus, there may exist the same agent ai even
in two different coalitions in ΠA ⊕ ΠB . For instance, suppose
that ΠA = {{a1 }, {a2 }, {a3 }} and ΠB = {∅, {a1 , a3 }, {a2 }}.
Then, ΠA ⊕ ΠB = {{a1 }, {a1 , a2 , a3 }, {a2 , a3 }}. We regard
such an agent as two different agents in ΠA ⊕ΠB . Accordingly,
the operation ⊕ may increase the number of total agents in
the resultant partition.
Using the definitions described above, the condition (i)
implies that
V (ΠA ) ≤ V (ΠA ⊕ ΠB ).
(16)
From now on, we will show that 21 V (Π∗ ) ≤ V (Π̂),
where Π∗ = {S0∗ , S1∗ , ..., Sn∗t } is an optimal partition and
Π̂ = {Ŝ0 , Ŝ1 , ..., Ŝnt } is a Nash stable partition. By doing so,
this theorem can be proved. From the definition in Equation
(15), it can be said that
X
∗
V (Π̂ ⊕ Π∗ ) =
ui (tΠ̂(i) , |ŜΠ̂(i) ∪ SΠ̂(i)
|)
∀ai ∈A
X
+
∗
ui (tΠ∗ (i) , |ŜΠ∗ (i) ∪ SΠ
∗ (i) |),
(17)
∀ai ∈A−
−
where A is the set of agents whose decisions follow not the
Nash stable partition Π̂ but the optimal partition Π∗ .
Due to the condition (ii), the first term of the right-hand
side in Equation (17) is no more than
X
ui (tΠ̂(i) , |ŜΠ̂(i) |) ≡ V (Π̂).
(18)
∀ai ∈A
Likewise, the second term is also at most
X
ui (tΠ∗ (i) , |ŜΠ∗ (i) ∪ {ai }|).
(19)
∀ai ∈A−
By the definition of Nash stability (i.e., for every agent ai ∈ A,
ui (tΠ̂(i) , |ŜΠ̂(i) |) ≥ ui (tj , |Ŝj ∪ {ai }|), ∀Ŝj ∈ Π̂), the above
equation is at most
X
ui (tΠ̂(i) , |ŜΠ̂(i) |),
(20)
∀ai ∈A−
which is also no more than, because of A− ⊆ A,
X
ui (tΠ̂(i) , |ŜΠ̂(i) |) ≡ V (Π̂).
(21)
∀ai ∈A
In a nutshell, the left-hand side of Equation (17) holds the
following inequality:
V (Π̂ ⊕ Π∗ ) ≤ 2V (Π̂).
Thanks to Equation (16), it follows that
V (Π∗ ) ≤ V (Π̂ ⊕ Π∗ ).
Therefore, V (Π∗ ) ≤ 2V (Π̂), which completes this proof.
(22)
9
C. Adaptability
Our proposed framework is also expected to be adaptable
in dynamic environments, owing to its fast convergence to a
Nash stable partition. Thanks to Lemma 1, if a new agent
additionally joins an ongoing mission in which an agreed
assignment was already determined, the number of iterations
required for converging to a new Nash stable partition is
at most the number of the total agents. Responding to any
environmental change, the framework is able to establish a
new agreed task assignment within polynomial time.
V. GRAPE WITH M INIMUM R EQUIREMENTS
This section addresses another task allocation problem
where each task may require a certain number of agents for
completion. This problem can be defined as follows.
Problem 2. Given a set of agents A and a set of tasks T , the
objective is to find an assignment such that
X X
max
ui (tj , p)xij
(23)
{xij }
∀ai ∈A ∀tj ∈T
subject to
X
xij ≥ Rj
∀tj ∈ T
(24)
xij ≤ 1
∀ai ∈ A
(25)
∀ai ∈A
X
∀tj ∈T
xij ∈ {0, 1}
∀(ai , tj ) ∈ A × T
(26)
where Rj ∈ N is the number of minimum required agents for
task tj , and all the other variables are identically defined as
those in Problem 1. Here, it is considered that, for ∀ai ∈ A
and ∀tj ∈ T ,
ui (tj , p) = 0 if p < Rj
(27)
because task tj cannot be completed in this case.
Even if ui (tj , p) is monotonically decreasing at p ≥ Rj ,
because of (27), the individual utility can not be simply
transformed to a preference relation holding SPAO. Thus, we
need to modify the utility function to yield alternative values
for the case when p < Rj . We refer to the modified utility as
auxiliary individual utility ũi , which is defined as
(
u0i (tj , p) if p ≤ Rj
ũi (tj , p) =
(28)
ui (tj , p) otherwise
where u0i (tj , p) is the dummy utility of agent ai with regard
to task tj when p ≤ Rj .
The dummy utility is intentionally used also for the case
when p = Rj . This is because we desire the auxiliary
individual utility to support the agents to find an assignment
that holds Equation (24). For this, the auxiliary individual
utility should satisfy the following condition:
Condition 1. For every agent ai ∈ A, its preference relation
Pi holds that, for any two tasks ∀tj , tk ∈ T ,
(tj , Rj ) i (tk , Rk + 1).
This condition enables every agent to prefer a task for which
the number of co-working agents is less than its minimum
requirement, over any other tasks whose requirements are
already fulfilled. UnderPthis condition, as long as the agent set
A is such that |A| ≥ ∀tj ∈T Rj and a Nash stable partition
is found, the resultant assignment satisfies Equation (24).
Proposition 1. Given an instance of Problem 2 where ui (tj , p)
∀i ∀j is a monotonically decreasing function with regard to
∀p ≥ Rj , if the dummy utilities u0i (tj , p) ∀i ∀j in (28) are
set to satisfy Condition 1 and SPAO for ∀p ≤ Rj , then all
the resultant auxiliary individual utilities ũi (tj , p) ∀i ∀j ∀p
can be transformed to a na -tuple of preference relations P
that hold Condition 1 as well as SPAO for ∀p ∈ {1, ..., na }.
In the corresponding instance of GRAPE (A, T , P), a Nash
stable partition can be determined within polynomial times as
shown in Theorems 1 and 2 because of SPAO, and the resultant
partition can satisfy Equation (24) due to Condition 1.
Let us give an example. Suppose that there exist 100 agents
A, and 3 tasks T = {t1 , t2 , t3 } where only t3 has its minimum
requirement R3 = 5; for every agent ai ∈ A, individual
utilities for t1 and t2 , i.e., ui (t1 , p) and ui (t1 , p), are much
higher than that for t3 in ∀p ∈ {1, ..., 100}. We can find
a Nash stable partition for this example, as described in
Proposition 1, by setting that all u0i (tj , p), ∀j, ∀p ≥ Rj to
max∀tj {ui (tj , Rj )} for each agent ai .
After a Nash stable partition is found, in order to compute
the objective function value in (23), the original individual
utility function ui should be used instead of the auxiliary one
ũi .
Proposition 2. As similar to Theorem 3, the suboptimality
bound α for a Nash stable partition Π obtained by implementing Proposition 1 is such that
α≥
JGRAP E
JGRAP E
.
·
JGRAP E + λ̃ JGRAP E + δ
(29)
Here, δ ≡ J˜GRAP E −JGRAP E , where J˜GRAP E and JGRAP E
is the objective function value in (23) using ũi and ui given
the Nash stable partition, respectively. Likewise, λ̃ is the value
in (9) using ũi . In addition to this, if every ũi satisfies the
conditions for Theorem 4, then
α≥
1
JGRAP E
·
.
2 JGRAP E + δ
(30)
Proof. Since the Nash stable partition Π is obtained by using
ũi , it can be said from Equations (7) and (8) that
J˜GRAP E
J˜GRAP E
≥
.
J˜OP T
J˜GRAP E + λ̃
(31)
Due to the fact that ũi (tj , p) ≥ ui (tj , p) for ∀i, j, p, it is clear
that J˜GRAP E ≥ JGRAP E and J˜OP T ≥ JOP T . By letting
that δ ≡ J˜GRAP E − JGRAP E , the left term in (31) is at
most (JGRAP E + δ)/JOP T . Besides, the right term in (31) is
a monotonically-increasing function with regard to J˜GRAP E ,
and thus, it is lower bounded by JGRAP E /(JGRAP E − λ̃).
From this, Equation (31) can be rewritten as Equation (29) by
multiplying JGRAP E /(JGRAP E + δ).
10
Likewise, for the case when every ũi satisfies the conditions
for Theorem 4, it can be said that J˜GRAP E ≥ 1/2 · J˜OP T ,
which can be transformed into Equation (30) as shown above.
Peaked-reward
Submodular-reward
Individual Utility
40
20
20
40
60
(a) Social utility
A. Mission Scenario and Settings
1) Utility functions: Firstly, we introduce the social and
individual utilities used in this numerical experiment. We
consider that if multiple robots execute a task together as a
coalition, they are given a certain level of reward for the task.
The amount of the reward varies depending on the number of
the co-working agents. The reward is shared with the agents,
and each agent’s individual utility is considered as the shared
reward minus the cost required to personally spend for the
task (e.g., fuel consumption for movement). In this experiment,
the equal fair allocation rule [53], [54] is adopted. Under the
rule, a task’s reward is equally shared amongst the members.
Therefore, the individual utility of agent ai executing task tj
with a coalition Sj is defined as
ui (tj , |Sj |) = r(tj , |Sj |)/|Sj | − ci (tj ),
(32)
where r(tj , |Sj |) is the reward from task tj when it is executed
by Sj together, and ci (tj ) is the cost that agent ai needs to
pay for the task. Here, we simply set the cost as a function of
the distance from agent ai to task tj . We set that if ui (tj , |Sj |)
is not positive, agent ai prefers to join Sφ over Sj .
This experiment considers two types of tasks. For the
first type, a task’s reward becomes higher as the number of
participants gets close to a specific desired number. We refer
to such a task as a peaked-reward task, and its reward can be
defined as
rjmax · |Sj | −|Sj |/nd +1
j
·e
,
(33)
r(tj , |Sj |) =
ndj
where ndj represents the desired number, and rjmax is the
peaked reward in case that ndj of agents are involved in.
Consequently, the individual utility for agent ai with regard
to task tj becomes the following equation:
ui (tj , |Sj |) =
rjmax
ndj
d
· e−|Sj |/nj +1 − ci (tj ).
(34)
For the second type, a task’s reward becomes higher as
more agents are involved, but the corresponding marginal
gain decreases. This type of tasks is said to be (monotone)
submodular-reward, and the reward can be defined as
r(tj , |Sj |) = rjmin · logj (|Sj | + j − 1),
(35)
6
4
0
0
# Co-working Agents
This section validates the performances of the proposed
framework with respect to its scalability, suboptimality, adaptability against dynamic environments, and robustness in asynchronous environments.
Peaked-reward
Submodular-reward
8
2
0
VI. S IMULATION AND R ESULTS
10
60
Social Utility
Notice that if δ = 0 for the Nash stable partition in Proposition 2, then the suboptimality bounds become equivalent to
those in Theorems 3 and 4.
80
80
100
0
20
40
60
80
100
# Co-working Agents
(b) Individual utility
Fig. 1. Examples of the social utility for a coalition and an agent’s individual
utility
where rjmin indicates the reward obtained if there is only one
agent involved, and j > 1 is the design parameter regarding
the diminishing marginal gain. The resultant individual utility
becomes as follows:
ui (tj , |Sj |) = rjmin · logj (|Sj | + j − 1)/|Sj | − ci (tj ). (36)
Figure 1 illustrates examples of the social utilities and
individual utilities for the task types introduced above. For
simplification, agents’ costs are ignored in the figure. We set
rjmax , ndj , rjmin and j to be 60, 15, 10, and 2, respectively.
Notice that the individual utilities are monotonically decreasing in the both cases, as depicted in Figure 1(b). Therefore,
given a mission that entails these task types, we can generate
an instance (A, T , P) of GRAPE that holds SPAO.
2) Parameters generation: In the following sections, we
will mainly utilise Monte Carlo simulations. At each run,
nt tasks and na agents are uniform-randomly located in a
1000 m × 1000 m arena and a 250 m × 250 m arena within
there, respectively. For a scenario including peaked-reward
tasks, rjmax is randomly generated from a uniform distribution
over [1000, 2000]
and ndj is selected by the rounded
P× na /nt ;max
max
value of (rj / ∀tk ∈T ∗ rk )×na . For a scenario including
submodular-reward tasks, j is set as 2, and rjmin is uniformrandomly generated over [1000, 2000] × 1/ logj (na /nt + 1).
3) Communication network: Given a set of agents, we assume that their communication network is strongly-connected.
Furthermore, we also consider the fully-connected network in
some experiments in order to examine the influence of the
network. The communication network is randomly generated
at each instance, and is assumed to be sustained during a
mission except the robustness test simulations in Section VI-E.
B. Scalability
To investigate the effectiveness of nt and na upon the
scalability of the proposed approach, we conduct a Monte
Carlo simulation with 100 runs for the scenarios introduced
in Section VI-A with a fixed nt = 20 and various na ∈
{80, 160, 240, 320} and for those with na = 160 and nt ∈
{5, 10, 15, 20}. Figure 2 shows the statistical results using boxand-whisker plots, where the green boxes indicate the results
from the scenarios with the peaked-reward tasks and the magenta boxes are those with the submodular-reward tasks. The
blue and red lines connecting the boxes represent the average
11
mean values of the graph diameter dG for the instances with
na ∈ {80, 160, 240, 320} are 36, 58, 75 and 92, respectively,
the results show that the amount of dummy iterations happened
is much less than the bound value, which is dG as pointed out
in Section IV-A. On the contrary, under the fully-connected
network there is no need of such a dummy iteration, and thus
the required number of iterations and that of time steps are
the same.
(a) na ∈ {80, 160, 240, 320} with fixed nt = 20
(b) nt ∈ {5, 10, 15, 20} with fixed na = 160
Fig. 2. The number of iterations and that of time steps required for converging
into a Nash stable partition relative to the number of agents, depending on
communication networks (i.e., Strongly-connected vs. Fully-connected) and
utility function types (i.e., Peaked-reward vs. Submodular-reward)
value for each test case (na , nt ) under a strongly-connected
network and the fully-connected network, respectively.
The left subfigure in Figure 2(a) shows that the ratio of
the number of required iterations to that of agents linearly
increases as more agents are involved. This implies that the
number of iterations has a quadratic complexity (i.e., C1 n2a ),
as stated in Theorem 2, but with C1 being much less than
1
2 , which is the value from the theorem. Even, C1 can
become very low (e.g., 5 × 10−4 ) under the fully-connected
network. Such a C1 being smaller than 12 may be explained
by Remark 4: the algorithmic efficiency of Algorithm 1 can
reduce unnecessary iterations that may be induced in the
procedure of the proof for Theorem 2. On the other hand,
it is shown in the left subfigure in Figure 2(b) that the number
of required iterations decreases with regard to the number of
tasks. This behaviour may be caused by the fact that more
selectable options provided to a fixed number of agents can
reduce possible conflicts between the agents. Furthermore,
the trends regarding either na or nt have a higher slope
under a strongly-connected network than that under the fullyconnected network. This is because the former condition is
more sensitive to conflicts between agents, and thus causes
additional iterations. For example, agents at the middle nodes
of the network may change their decisions (and thus increase
the number of iterations) while the local partition information
of the agent at one end node is being propagated to agents at
the other end nodes. Such unnecessary iterations would not
have occured if the agents at the end nodes were directly
connected.
The right subfigures in Figure 2(a) and (b) indicate that
approximately 3–4 times of dummy iterations, compared with
the required number of normal iterations, are additionally
needed under a strongly-connected network. Noting that the
C. Suboptimality
This section examines the suboptimality of the proposed
framework by using Monte Carlo simulations with 100 instances. In each instance, there are nt = 3 of tasks and
na = 12 of agents that are connected by a strongly-connected
network. Figure 3 presents the true optimality for each instance, which is the ratio of the global utility obtained by
the proposed framework to that by a brute-force search, i.e.,
JGRAP E /JOP T , and the lower bound given by Theorem 3.
A blue circle and a red cross in the figure indicate the true
suboptimality and the lower bound, respectively. The results
show that the framework provides near-optimal solutions in
almost all cases and the suboptimality of each Nash stable
partition is enclosed by the corresponding lower bound.
The suboptimality may be improved if the agents are let
to investigate a larger search space, for examples, possible
coalitions caused by co-deviation of multiple agents. However,
this strategy in return may increase communication transactions amongst the agents because they have to notice each
other’s willingness unless their individual utility functions are
known, which is in contrast to Assumption 4. Besides, the
computational overhead for an agent at each iteration also
becomes more expensive than O(nt ), which is, as shown in
Section IV-A, the complexity bound for searching only the
currently-selectable coalitions. Hence, the resultant algorithm
complexity may hinder its applicability to a large-scale multiple agent system.
Figure 4 depicts the suboptimality lower bounds for the
large-size problems that were previously addressed in Section
VI-B. Firstly, it is clearly shown that the agent communication
network does not make any effect on the suboptimality lower
bound of a Nash stable partition. Secondly, although there is
no universal trend of the suboptimality with regard to na and
nt in both utility types, it is suggested that the features of the
lower bound given by Theorem 3 can be influenced by the
utility functions considered.
D. Adaptability
This section discusses the adaptability of our proposed
framework in response to dynamic environments such as the
inclusion or loss of agents and tasks. Suppose that there are
10 tasks and 160 agents in a mission, and a Nash stable
partition was already found as a baseline. During the mission,
the number of agents (or tasks) changes; the range of the
change is from losing 50% of the existing agents (or tasks)
to additionally including new ones as much as 50% of them.
For each dynamical environment, a new set of Monte Carlo
simulations with 100 instances are performed by randomly
12
100%
Suboptimality (%)
Suboptimality (%)
100%
80%
60%
40%
20%
80%
60%
40%
20%
0
50
Monte Carlo Case ID
(a) Peaked-reward task
100
0
50
100
Monte Carlo Case ID
(b) Submodular-reward task
(a) Dynamic Agents
Fig. 3. True suboptimality obtained by GRAPE (blue circle) and its lower
bound provided by Theorem 3 (red cross) under a strongly-connected communication network
(b) Dynamic Tasks
(a)
(b)
Fig. 4. The suboptimality lower bound, given by Theorem 3, of a Nash stable
partition (Left: fixed nt = 20 with varying na ∈ {80, 160, 240, 320}; Right:
fixed na = 160 with varying nt = {5, 10, 15, 20})
including or excluding a subset of the corresponding number
of agents or tasks. Here, we consider a strongly-connected
communication network.
Figure 5(a) illustrates that the more the additional agents
are involved, the more the number of iterations is required
for converging to a new Nash stable partition. This is because
the inclusion of a new agent may lead to additional iterations
at most as much as the number of total agents including
the newly-joined agent (Lemma 1). On the other hand, the
loss of existing agents does not seem to have any apparent
relation with the number of iterations. A possible explanation
is that the exclusion of an existing agent is favourable to other
agents due to SPAO preferences. This stimulates only a limited
number of agents who are preferred to move to the taskspecific coalition where the excluded agent was. This feature
induces fewer additional iterations to reach a new Nash stable
partition, compared with the case of adding a new agent.
Figure 5(b) shows that to eliminate existing tasks causes
more iterations than to include new tasks. This can be explained from the fact that removing any task releases the agents
performing the task free and it results in extra iterations as
much as the number of the freed agents. On the other hand,
adding new tasks induces relatively fewer additional iterations,
because only some of existing agents are attracted to those
tasks.
In summary, as the ratio of the number of agents to that of
tasks increases, the number of additional iterations to converge
a new Nash stable partition also increases. This actually
corresponds to the trend described in Section VI-B, i.e., the
Fig. 5. The number of additional iterations required for re-converging a Nash
stable partition relative to the number of agents (Baseline: nt = 10, na =
160). Negative values in the x-axis indicate that the corresponding number of
existing agents or tasks are lost; positive values indicate that the corresponding
number of new agents or tasks are included in an ongoing mission; a stronglyconnected communication network is used.
left subfigures in Figure 2(a) and (b). In all the cases of this
experiment, the number of additionally induced iterations still
remains the same order of the number of the given agents,
which implies that the proposed framework provides excellent
adaptability.
E. Robustness in Asynchronous Environments
This section investigates the robustness of the proposed
framework in asynchronous environments. This scenario assumes that a certain fraction of the given agents, which
are randomly chosen at each iteration, somehow can not
execute Algorithm 1 and even can not communicate with
other normally-working neighbour agents. We refer to such
agents as non-operating agents. Given that nt = 5 and
na = 40, the fractions of the non-operating agents are set as
{0, 0.2, 0.4, 0.6, 0.8}. In each case, we conduct 100 instances
of Monte Carlo experiments, in which the submodular-reward
utilities are used.
Figure 6(a) presents that the number of iterations required
for converging a Nash stable partition remains the same
level regardless of the fraction of the non-operating agents.
Despite that, the practically required time steps increase as
more agents become non-operating, as shown in Figure 6(b).
Note that time steps growth rate means the ratio of the total
required time steps to those for the case without non-operating
agents. These findings indicate that, due to communicational
discontinuity caused by non-operating agents at each iteration,
the framework may take more time to wait for these agents
to operate again and then to disseminate a valid partition
13
(a)
coloured agents are assigned to the same coloured task, for
example, yellow agents belong to a team for executing the
yellow task. The square size of a task indicates the reward
of the task, and the cost for an agent to execute the task is
considered as a function of the distance from the agent to
the task. The allocation results seem to be reasonable with
consideration of the task rewards and the costs. The number of
iterations required to find a Nash stable partition is 1355, 1380,
and 1295 for Scenario #1, #2, and #3, respectively. Since the
communication networks considered are more connected than
a strongly-connected one, the number of dummy iterations
happened is 20–30% of that of the iterations.
VII. C ONCLUSION
(b)
(c)
Fig. 6. Robustness test in asynchronous environments: the effectiveness of
non-operating agents (Baseline: nt = 10, na = 160, and the submodularreward tasks are used).
information over the entire agents. Although such dummy
iterations may increase in asynchronous environments, the
proposed framework is still able to find a Nash stable partition.
Furthermore, the resultant Nash stable partition’s lower bound
obtained by Theorem 3 is not affected, as presented in Figure
6(c).
F. Visualisation
We have na = 320 agents and nt = 5 tasks. The initial
locations of the given agents are randomly generated, and the
overall formation shape is different in each test scenario such
as circle, skewed circle, and square (denoted Scenario #1, #2,
and #3, respectively). The tasks are also randomly located
outside the agents. In this simulation, each agent is able to
communicate with its neighbour agents distant upto 50 m.
Here, the submodular-reward utilities are used.
Figure 7 shows the visualised task allocation results where
the circles and the squares indicate the positions of the agents
and the tasks, respectively. The lines between the circles
represent the communication networks of the agents. The
This paper proposed a novel game-theoretical framework
that addresses a task allocation problem for a robotic swarm
consisting of self-interested agents. We showed that selfish
agents whose individual interests are transformable to SPAO
preferences can converge to a Nash stable partition by using the proposed simple decentralised algorithm, which is
executable even in asynchronous environments and under a
strongly-connected communication network. We analytically
and experimentally presented that the proposed framework
provides scalability, a certain level of guaranteed suboptimality, adaptability, robustness, and an ability to accommodate
different interests of agents.
As this framework can be considered as a new sub-branch
of self-organised approaches, one of our ongoing works is to
compare it with one of the existing methods. Defining a fair
scenario for both methods is non-trivial and requires careful
consideration; otherwise, a resultant unsuitable scenario may
provide biased results. Secondly, another natural progression
of this study is to relax anonymity of agents and thus to consider a combination of the agents’ identities. Experimentally,
we have often observed that heterogeneous agents with social
inhibition also can converge to a Nash stable partition. More
research would be needed to analyse the quality of a Nash
stable partition obtained by the proposed framework in terms
of min max because our various experiments showed that the
outcome provides individual utilities to agents in a balanced
manner.
ACKNOWLEDGMENT
The authors gratefully acknowledge that this research was
supported by International Joint Research Programme with
Chungnam National University (No. EFA3004Z)
R EFERENCES
[1] H.-S. Shin and P. Segui-Gasco, “UAV Swarms: Decision-Making
Paradigms,” in Encyclopedia of Aerospace Engineering. John Wiley &
Sons, 2014, pp. 1–13.
[2] A. Khamis, A. Hussein, and A. Elmogy, “Multi-robot Task Allocation:
A Review of the State-of-the-Art,” in Cooperative Robots and Sensor
Networks 2015. Cham: Springer International Publishing, 2015, vol.
604, pp. 31–51.
[3] A. Jevtić, Á. Gutierrez, D. Andina, and M. Jamshidi, “Distributed Bees
Algorithm for Task Allocation in Swarm of Robots,” IEEE Systems
Journal, vol. 6, no. 2, pp. 296–304, 2012.
[4] E. Sahin, “Swarm Robotics: From Sources of Inspiration to Domains of
Application,” in Swarm Robotics. Berlin: Springer, 2005, pp. 10–20.
14
(a) Scenario #1
(b) Scenario #2
(c) Scenario #3
Fig. 7. Visualisated task allocation results (nt = 5, na = 320).
[5] M. Dorigo, M. Birattari, and M. Brambilla, “Swarm robotics,” p. 1463,
2014.
[6] K. Barton and D. Kingston, “Systematic Surveillance for UAVs: A
Feedforward Iterative Learning Control Approach,” in American Control
Conference, Washington, DC, USA, 2013, pp. 5917–5922.
[7] I. Bekmezci, O. K. Sahingoz, and S. Temel, “Flying Ad-Hoc Networks
(FANETs): A Survey,” Ad Hoc Networks, vol. 11, no. 3, pp. 1254–1270,
2013.
[8] M. Erdelj, E. Natalizio, K. R. Chowdhury, and I. F. Akyildiz, “Help from
the Sky: Leveraging UAVs for Disaster Management,” IEEE Pervasive
Computing, vol. 16, no. 1, pp. 24–32, 2017.
[9] I. Jang, J. Jeong, H.-S. Shin, S. Kim, A. Tsourdos, and J. Suk,
“Cooperative Control for a Flight Array of UAVs and an Application in
Radar Jamming,” IFAC-PapersOnLine, vol. 50, no. 1, pp. 8011–8018,
2017.
[10] B. P. Gerkey and M. J. Matarić, “A Formal Analysis and Taxonomy
of Task Allocation in Multi-robot Systems,” International Journal of
Robotics Research, vol. 23, no. 9, pp. 939–954, 2004.
[11] G. A. Korsah, A. Stentz, and M. B. Dias, “A Comprehensive Taxonomy
for Multi-robot Task Allocation,” The International Journal of Robotics
Research, vol. 32, no. 12, pp. 1495–1512, 2013.
[12] S. Bandyopadhyay, S.-J. Chung, and F. Y. Hadaegh, “Probabilistic and
Distributed Control of a Large-Scale Swarm of Autonomous Agents,”
IEEE Transactions on Robotics, vol. 33, no. 5, pp. 1103–1123, 2017.
[13] M. Brambilla, E. Ferrante, M. Birattari, and M. Dorigo, “Swarm
Robotics: a Review from the Swarm Engineering Perspective,” Swarm
Intelligence, vol. 7, no. 1, pp. 1–41, 2013.
[14] L. Johnson, S. Ponda, H.-L. Choi, and J. How, “Asynchronous
Decentralized Task Allocation for Dynamic Environments,” in Infotech@Aerospace 2011, St.Louis, MO, USA, 2011.
[15] C. M. Clark, R. Morton, and G. A. Bekey, “Altruistic relationships
for optimizing task fulfillment in robot communities,” Distributed Autonomous Robotic Systems 8, pp. 261–270, 2009.
[16] J. H. Drèze and J. Greenberg, “Hedonic Coalitions: Optimality and
Stability,” Econometrica, vol. 48, no. 4, pp. 987–1003, 1980.
[17] S. Banerjee, H. Konishi, and T. Sönmez, “Core in a Simple Coalition
Formation Game,” Social Choice and Welfare, vol. 18, no. 1, pp. 135–
153, 2001.
[18] A. Bogomolnaia and M. O. Jackson, “The Stability of Hedonic Coalition
Structures,” Games and Economic Behavior, vol. 38, no. 2, pp. 201–230,
2002.
[19] A. Brutschy, G. Pini, C. Pinciroli, M. Birattari, and M. Dorigo, “Selforganized Task Allocation to Sequentially Interdependent Tasks in
Swarm Robotics,” Autonomous Agents and Multi-Agent Systems, vol. 28,
no. 1, pp. 101–125, 2014.
[20] N. Kalra and A. Martinoli, “A Comparative Study of Market-Based and
Threshold-Based Task Allocation,” in Distributed Autonomous Robotic
Systems 7. Tokyo: Springer Japan, 2006, pp. 91–101.
[21] Y. Zhang and L. E. Parker, “Considering Inter-task Resource Constraints
in Task Allocation,” Autonomous Agents and Multi-Agent Systems,
vol. 26, no. 3, pp. 389–419, 2013.
[22] H. L. Choi, L. Brunet, and J. P. How, “Consensus-based Decentralized
Auctions for Robust Task Allocation,” IEEE Transactions on Robotics,
vol. 25, no. 4, pp. 912–926, 2009.
[23] P. Segui-Gasco, H.-S. Shin, A. Tsourdos, and V. J. Segui, “Decentralised
Submodular Multi-Robot Task Allocation,” in IEEE/RSJ International
Conference on Intelligent Robots and Systems, Hamburg, Germany,
2015, pp. 2829–2834.
[24] B. Acikmese and D. S. Bayard, “Markov Chain Approach to Probabilistic Guidance for Swarms of Autonomous Agents,” Asian Journal of
Control, vol. 17, no. 4, pp. 1105–1124, 2015.
[25] I. Chattopadhyay and A. Ray, “Supervised Self-Organization of Homogeneous Swarms Using Ergodic Projections of Markov Chains,” IEEE
Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics),
vol. 39, no. 6, pp. 1505–1515, 2009.
[26] N. Demir and B. Acikmese, “Probabilistic Density Control for Swarm
of Decentralized ON-OFF Agents with Safety Constraints,” in American
Control Conference, Chicago, IL, USA, 2015, pp. 5238–5244.
[27] S. Berman, A. Halasz, M. A. Hsieh, and V. Kumar, “Optimized
Stochastic Policies for Task Allocation in Swarms of Robots,” IEEE
Transactions on Robotics, vol. 25, no. 4, pp. 927–937, 2009.
[28] A. Halasz, M. A. Hsieh, S. Berman, and V. Kumar, “Dynamic Redistribution of a Swarm of Robots among Multiple Sites,” in IEEE/RSJ
International Conference on Intelligent Robots and Systems, San Diego,
CA, USA, 2007, pp. 2320–2325.
[29] M. A. Hsieh, A. Halasz, S. Berman, and V. Kumar, “Biologically
Inspired Redistribution of a Swarm of Robots among Multiple Sites,”
Swarm Intelligence, vol. 2, no. 2-4, pp. 121–141, 2008.
[30] T. W. Mather and M. A. Hsieh, “Macroscopic Modeling of Stochastic
Deployment Policies with Time Delays for Robot Ensembles,” The
International Journal of Robotics Research, vol. 30, no. 5, pp. 590–
600, 2011.
15
[31] A. Prorok, M. A. Hsieh, and V. Kumar, “The Impact of Diversity
on Optimal Control Policies for Heterogeneous Robot Swarms,” IEEE
Transactions on Robotics, vol. 33, no. 2, pp. 346–358, 2017.
[32] T. H. Labella, M. Dorigo, and J.-L. Deneubourg, “Division of Labor
in a Group of Robots Inspired by Ants’ Foraging Behavior,” ACM
Transactions on Autonomous and Adaptive Systems, vol. 1, no. 1, pp.
4–25, 2006.
[33] E. Castello, T. Yamamoto, Y. Nakamura, and H. Ishiguro, “Foraging Optimization in Swarm Robotic Systems Based on an Adaptive Response
Threshold Model,” Advanced Robotics, vol. 28, no. 20, pp. 1343–1356,
2014.
[34] H. Kurdi, J. How, and G. Bautista, “Bio-Inspired Algorithm for Task Allocation in Multi-UAV Search and Rescue Missions,” in AIAA Guidance,
Navigation, and Control Conference, San Diego, CA, USA, 2016.
[35] W. Liu, A. F. T. Winfield, J. Sa, J. Chen, and L. Dou, “Towards
Energy Optimization: Emergent Task Allocation in a Swarm of Foraging
Robots,” Adaptive Behavior, vol. 15, no. 3, pp. 289–305, 2007.
[36] W. Liu and A. F. T. Winfield, “Modeling and Optimization of Adaptive
Foraging in Swarm Robotic Systems,” The International Journal of
Robotics Research, vol. 29, no. 14, pp. 1743–1760, 2010.
[37] A. Martinoli, K. Easton, and W. Agassounon, “Modeling Swarm Robotic
Systems: a Case Study in Collaborative Distributed Manipulation,” The
International Journal of Robotics Research, vol. 23, no. 4, pp. 415–436,
2004.
[38] K. Lerman, A. Martinoli, and A. Galstyan, “A Review of Probabilistic
Macroscopic Models for Swarm Robotic Systems,” in Swarm Robotics.
Berlin: Springer, 2005, pp. 143–152.
[39] N. Correll and A. Martinoli, “System Identification of Self-Organizing
Robotic Swarms,” in Distributed Autonomous Robotic Systems 7.
Tokyo: Springer Japan, 2006, pp. 31–40.
[40] A. Prorok, N. Correll, and A. Martinoli, “Multi-level Spatial Modeling
for Stochastic Distributed Robotic Systems,” The International Journal
of Robotics Research, vol. 30, no. 5, pp. 574–589, 2011.
[41] A. Kanakia, J. Klingner, and N. Correll, “A Response Threshold
Sigmoid Function Model for Swarm Robot Collaboration,” Distributed
Autonomous Robotic Systems, pp. 193–206, 2016.
[42] D. Dimitrov and S. C. Sung, “Top responstiveness and Nash stability in
coalition formation games,” Kybernetika, vol. 42, no. 4, pp. 453–460,
2006.
[43] A. Darmann, E. Elkind, S. Kurz, and J. Lang, Internet and Network
Economics, ser. Lecture Notes in Computer Science, P. W. Goldberg,
Ed. Berlin, Heidelberg: Springer Berlin Heidelberg, 2012, vol. 7695.
[44] A. Darmann, “Group Activity Selection from Ordinal Preferences,” in
Algorithmic Decision Theory. ADT 2015. Lecture Notes in Computer
Science, ser. Lecture Notes in Computer Science, T. Walsh, Ed. Berlin,
Heidelberg: Springer Cham, 2015, vol. 9346, pp. 35–51.
[45] M. Rubenstein, A. Cornejo, and R. Nagpal, “Programmable Selfassembly in a Thousand-robot Swarm,” Science, vol. 345, no. 6198,
pp. 795–799, 2014.
[46] S. C. Sung and D. Dimitrov, “On Myopic Stability Concepts for Hedonic
Games,” Theory and Decision, vol. 62, no. 1, pp. 31–45, 2007.
[47] M. Karakaya, “Hedonic Coalition Formation Games: A New Stability
Notion,” Mathematical Social Sciences, vol. 61, no. 3, pp. 157–165,
2011.
[48] H. Aziz and F. Brandl, “Existence of Stability in Hedonic Coalition
Formation Games,” in Proceedings of the 11th International Conference
on Autonomous Agents and Multiagent Systems, Valencia, Spain, 2012,
pp. 763–770.
[49] J. Guerrero and G. Oliver, “Multi-robot Coalition Formation in Realtime Scenarios,” Robotics and Autonomous Systems, vol. 60, no. 10, pp.
1295–1307, 2012.
[50] O. Shehory and S. Kraus, “Feasible Formation of Coalitions Among
Autonomous Agents in Nonsuperadditive Environments,” Computational
Intelligence, vol. 15, no. 3, pp. 218–251, 1999.
[51] C. Nam and D. A. Shell, “Assignment Algorithms for Modeling Resource Contention in Multirobot Task Allocation,” IEEE Transactions
on Automation Science and Engineering, vol. 12, no. 3, pp. 889–900,
2015.
[52] L. B. Johnson, H.-L. Choi, and J. P. How, “The Role of Information Assumptions in Decentralized Task Allocation: A Tutorial,” IEEE Control
Systems, vol. 36, no. 4, pp. 45–58, 2016.
[53] W. Saad, Z. Han, M. Debbah, and A. Hjørungnes, “A Distributed
Coalition Formation Framework for Fair User Cooperation in Wireless
Networks,” IEEE Transactions on Wireless Communications, vol. 8,
no. 9, pp. 4580–4593, 2009.
[54] W. Saad, Z. Han, T. Basar, M. Debbah, and A. Hjørungnes, “Hedonic
Coalition Formation for Distributed Task Allocation among Wireless
Agents,” IEEE Transactions on Mobile Computing, vol. 10, no. 9, pp.
1327–1344, 2011.
| 2cs.AI
|
Leveraging Parallel Data Processing Frameworks with
Verified Lifting
Maaz Bin Safeer Ahmad
Alvin Cheung
Computer Science & Engineering
University of Washington
[email protected]
Computer Science & Engineering
University of Washington
[email protected]
http://casper.uwplse.org
Many parallel data frameworks have been proposed in recent years that let sequential programs access
parallel processing. To capitalize on the benefits of such frameworks, existing code must often be
rewritten to the domain-specific languages that each framework supports. This rewriting—tedious
and error-prone—also requires developers to choose the framework that best optimizes performance
given a specific workload.
This paper describes C ASPER, a novel compiler that automatically retargets sequential Java code
for execution on Hadoop, a parallel data processing framework that implements the MapReduce
paradigm. Given a sequential code fragment, C ASPER uses verified lifting to infer a high-level summary expressed in our program specification language that is then compiled for execution on Hadoop.
We demonstrate that C ASPER automatically translates Java benchmarks into Hadoop. The translated
results execute on average 3.3× faster than the sequential implementations and scale better, as well,
to larger datasets.
1
Introduction
As computing becomes increasingly ubiquitous, storage cheaper, and data collection tools more sophisticated, more data is being collected today than ever before. Data-driven advances are increasingly prevalent in various scientific domains. As such, effectively analyzing and processing huge datasets poses a
grand computational challenge.
Many parallel data processing frameworks have been developed to handle very large datasets [2,
5, 6, 9, 12], and new ones continue to be frequently released [1, 12, 23]. Most parallel data processing
frameworks come with domain-specific optimizations that are exposed either via library APIs [1, 2, 5, 6,
9, 23] or high-level, domain-specific languages (DSLs) for users to express their computations [12, 16].
Computations expressible using such API calls or DSLs are more efficient thanks to the frameworks’
domain-specific optimizations [3, 16, 20, 22].
However, the many issues with this approach often make domain-specific frameworks inaccessible
to non-experts such as researchers studying physical or social sciences. First, domain-specific optimizations for different workloads require an expert to decide up front the most appropriate framework for a
given piece of code. Second, end users must often learn new APIs or DSLs [1, 2, 5, 6, 9, 23] and rewrite
existing code to leverage the benefits provided by these frameworks. Doing so requires not only significant time and resource but also risks introducing new bugs into the application. Moreover, even users
willing to rewrite their applications must first understand the intent of the code which might have been
written by others. And manually written, low-level optimizations in the code often obscure high-level
intent. Finally, even after learning new APIs and rewriting code, newly emerging frameworks often turn
freshly rewritten code into legacy applications. Users must then repeat this process to keep pace with
Dimitrova, Piskac (Eds.): Fifth Workshop on
Synthesis (SYNT 2016)
EPTCS 229, 2016, pp. 67–83, doi:10.4204/EPTCS.229.7
© M. B. S. Ahmad & A. Cheung
This work is licensed under the
Creative Commons Attribution License.
68
Leveraging Parallel Data Processing Frameworks with Verified Lifting
new advances, requiring significant time investments that could be better spent in advancing scientific
discovery.
One way to improve the accessibility of these parallel data processing frameworks involves building
compilers that automatically convert applications written in common general-purpose languages (such
as Java or Python) to high-performance parallel processing frameworks, such as Hadoop or Spark. Such
compilers let users write their applications in familiar general-purpose languages and let the compiler
retarget portions of their code to high-performance DSLs [7, 11, 15]. The applications can then leverage the performance of these specialized frameworks without the overhead of learning how to program
individual DSLs. But such compilers don’t always exist, and building one can prove highly complex.
This paper demonstrates the application of verified lifting to automatically convert sequential Java
code fragments to MapReduce. As input, verified lifting takes program fragments written in a generalpurpose language and uses program synthesis to automatically find provably correct code summaries.
These summaries—expressed in our program specification language—encode the semantics of the input
code fragment. The found summaries are then used to translate the original input code to the target
high-performance DSL.
The concept of verified lifting has been previously applied to database applications [7] and stencil
computations [11]. This paper applies verified lifting to the conversion of sequential data processing Java
code to leverage the parallel data processing frameworks Apache Hadoop. The problem statement remains familiar and was first proposed in the MOLD compiler [15], which translates sequential Java code
for execution on Apache Spark. MOLD uses pre-defined rewrite rules to search the space of equivalent
Apache Spark implementations. It scans the input code for patterns that trigger such rewrite rules, an
approach fraught with many limitations. For instance, it requires the a priori definition of complicated
rewrite rules, which can be extremely brittle to code pattern changes. In comparison, our approach analyzes program semantics rather than program syntax, making it robust to code pattern changes. We also
do not rely on pre-defined translation rules and can thus discover new solutions and optimizations that
the user never knew existed.
We implemented our approach described above in a compiler called C ASPER. By converting sequential code fragments to Hadoop, C ASPER parallelizes computation at crucial program points where
input data collections are being processed. We used C ASPER to convert five benchmark programs with
encouraging results. This paper thus makes the following contributions:
• We describe the use of verified lifting to retarget sequential Java applications to Hadoop by converting code fragments within the application to Hadoop MapReduce tasks.
• We design a new program specification language to express the intent of Java code fragments using
the MapReduce paradigm.
• We employ static program analysis techniques to intelligently restrict the search space of all possible summaries that can be expressed in our specification language and use inductive synthesis to
find provably correct summaries for each input code fragment.
• We present encouraging preliminary results from using C ASPER to identify and optimize code
fragments written in sequential Java. To show the potential of our approach, we evaluate our
system on five MapReduce benchmarks used in prior work [17] to demonstrate its capabilities and
limitations.
In the following, we describe C ASPER’s design and illustrate its use to convert sequential Java programs into Hadoop tasks in §2. In §3 we explain verified lifting and describe how we implemented each
of its steps in C ASPER. §4 evaluates how C ASPER performs using varied benchmarks and shares our
preliminary results. §5 describes related work, and we conclude in §6.
M. B. S. Ahmad & A. Cheung
69
Figure 1: C ASPER system architecture diagram. Sequential code fragments (highlighted green) in the
input source file are translated to equivalent Hadoop tasks (highlighted orange).
2
System Overview
This section describes the architecture of the C ASPER compiler. C ASPER automatically identifies and
converts sequential Java source code fragments into semantically equivalent MapReduce tasks implemented using Hadoop. It generates new, optimized version of the input source code where original code
fragments are replaced by invocations to the generated MapReduce tasks. Figure 1 shows C ASPER’s
different components and how they interact in the compilation pipeline.
Before explaining each component in detail, we should generally note that by statically analyzing
Java input source code, C ASPER extracts code fragments that can potentially be translated. It then generates a high-level summary of each extracted code fragment. Expressed in our high level program
specification language (see §3.1), the summary is inferred by a program synthesizer. To quickly traverse
the large space of possible summaries, C ASPER bounds the search space considered by the synthesizer
and uses a bounded model-checking procedure to locate any candidate candidate summaries for a given
code fragment.1 The Hadoop code generator module uses the verified summary to produce code for
Hadoop MapReduce tasks. Lastly, the code generator module prepares a new version of the input source
1 While
not yet implemented in the current C ASPER prototype, any candidate summary that passes the bounded model
checking phase will be forwarded to a theorem prover, which verifies that the synthesizer-generated summary is semantically
equivalent to the original code.
70
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
Leveraging Parallel Data Processing Frameworks with Verified Lifting
int[][] histogram(int[] data) {
int[] hR = new int[256];
int[] hG = new int[256];
int[] hB = new int[256];
for (int i = 0; i < data.length; i += 3){
int r = data[i];
int g = data[i + 1];
int b = data[i + 2];
hR[r]++;
hG[g]++;
hB[b]++;
}
int[][] result = new int[3][];
result[0] = hR;
result[1] = hG;
result[2] = hB;
return result;
}
(a) Input source code
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
Map kvPairs = HistogramHadoop.execute();
hR = kvPairs.get(0);
hG = kvPairs.get(1);
hB = kvPairs.get(2);
(b) C ASPER wrapper to replace the loop in (a)
18
19
20
21
public class HistogramHadoop{
class HistogramMapper extends Mapper {
void map(int key, int[] value){
for (int i=0; i<value.length; i+=1) {
if(i%3==0) emit((0, value[i]), 1);
if(i%3==1) emit((1, value[i]), 1);
if(i%3==2) emit((2, value[i]), 1);
}}}
class HistogramReducer extends Reducer {
void reduce(Tuple key, int[] values) {
int value = 0;
for (int val:values) {value=value+val;}
emit(key, value);
}}
static Map execute() {
Job = Job.getInstance();
job.setMapper(HistogramMapper);
job.setReducer(HistogramReducer);
return job.execute();
}
}
(c) C ASPER-generated Hadoop task
Figure 2: C ASPER translation of the 3D Histogram benchmark.
code by replacing the original code fragments with invocations to the translated Hadoop tasks.
Throughout the remainder of this paper, we use, as a running example, a benchmark from the Phoenix
suite of benchmarks [17] that generates 3D histograms from image data stored in a file. Figure 2 shows
C ASPER’s translation of this 3D Histogram benchmark. In Figure 2a, the original program sequentially
iterates over an array of integers representing the intensity values of colors red, green, and blue for each
pixel. It then counts the number of times each value occurs for each color from Lines 9 to 11. The parallel
program C ASPER generates, on the other hand, emits a key-value pair with the tuple (color, intensity)
as key and 1 as value (lines 5 to 7 in Figure 2c). The generated pairs are then grouped by key and the
frequency of each key is calculated by adding all the 1’s in the reducer phase (line 12). Figure 2b shows
the final code generated by C ASPER that replaces the original loop, where it invokes the Hadoop task
shown in Figure 2c and uses the output from the Hadoop task to update program state accordingly. In our
evaluation, the translated version of the benchmark performed over 1.5× faster for a dataset of 50GB.
We now discuss the three essential modules that make up C ASPER’s compilation pipeline.
2.1
Program Analyzer
The program analyzer, the first component in C ASPER’s compilation pipeline, has two goals. First, it
identifies all code fragments that are candidates for conversion. Second, it generates synthesis specifications for each identified candidate code fragment. The program analyzer operations are grouped
into three sub-components: the code fragment identifier, the verification conditions generator, and the
grammar generator.
M. B. S. Ahmad & A. Cheung
71
In the current prototype, C ASPER’s code fragment identifier finds loops from the input program
and extracts them as candidates for conversion. C ASPER currently does not consider non-looping code
fragments (such as recursive functions) as candidates for conversion. In addition, C ASPER also ignores
loops containing calls to external library methods that are unrecognized by the C ASPER compiler. In
§3.4, we formally list the criteria needed for a code fragment to be extracted as a candidate for conversion,
highlighting some current limitations of C ASPER’s implementation.
The second program analyzer sub-component is the grammar generator, which aims to confine the
space of synthesizable summaries. Without doing so, the space of all possible summaries expressible
in our program specification language would be too large. The grammar generator takes as input the
code fragments extracted by the code fragment identifier and statically analyzes each one to extract
semantic information. It then uses the extracted information to generate a program grammar for every
code fragment. The challenge here is to generate a grammar expressive enough to express the correct
summary, but not so expressive as to make the problem of summary search intractable. In §3.3.2, we
explain how the grammar generator leverages static program analysis to construct a grammar for each
code fragment.
The third program analyzer sub-component is the verification conditions generator. This component
uses Hoare logic [10] and static program analysis to generate verification conditions for each code fragment. Verification conditions are logical statements that describe what must be true for a given summary
to be semantically equivalent to the original code fragment. We explain how C ASPER uses Hoare style
program verification to verify program equivalence in §3.2.
The output of the verification conditions generator is a search template for the summary, with: (i) the
search space specified by the grammar generator, and (ii) the verification conditions generator producing
the logical assertions that must be satisfied given a candidate summary. The summary generator uses this
template to search for a valid summary of the input code fragment, as we next explore.
2.2
Summary Generator
Using the program analyzer’s specifications, the summary generator traverses the search space to find a
summary that satisfies the verification conditions. It consists of two modules: the program synthesizer
and the theorem prover. The synthesizer takes the search space description and verification conditions
previously generated and searches for a code summary that satisfies the verification conditions. To make
the search problem tractable, it uses a bounded model checking procedure: the synthesizer checks for
correctness only over a small sub-domain. When a promising candidate for the summary is found,
viz., one that satisfies the verification conditions in the sub-domain, it is passed onto the formal theorem
prover 2 , which checks it for correctness over the entire domain of inputs. Candidate solutions that fail the
formal verification step are eliminated from the search space, and the search restarts for a new candidate
solution. Using this two-step verification process helps C ASPER quickly discard bad candidates. The
more computationally expensive process of formal verification is reserved only for promising candidate
solutions. As output, the summary generator emits a verifiably correct summary for each code fragment
that can then be translated to Hadoop. Note that the summary generator may not always find a solution
that can be proven correct: some fragments are impossible to translate to MapReduce while others might
have complex solutions that C ASPER currently can not generate. In such cases, C ASPER gives up on
translating the code fragment.
2 See
footnote 1.
72
2.3
Leveraging Parallel Data Processing Frameworks with Verified Lifting
Hadoop Code Generator
The summaries synthesized by the summary generator are expressed in our high-level program specification language. Generating Hadoop implementations from these high level program specifications is
straightforward and is achieved in C ASPER through syntax-directed translation rules. The code generator
module also outputs the code required to embed these Hadoop tasks into the original program. Essentially, C ASPER generates a new (source) version of the input code, where each code fragment that was
successfully translated is replaced by code that first invokes the corresponding Hadoop task and then
uses the output generated by the Hadoop task to update the state of the program. Figure 2b shows such
generated wrapper code for the 3D Histogram example. We present more details about C ASPER’s code
generation module in §3.5.
3
Converting Code Fragments
This section explains how C ASPER uses verified lifting to convert sequential Java code fragments to
MapReduce tasks. We review the concept of verified lifting in §3.1 and describe the program specification language C ASPER uses to express program summaries. In §3.2, we explain how C ASPER verifies
that the identified summaries preserve program semantics of the original code fragment. §3.3 discusses
the search process C ASPER uses to find program summaries, while §3.4 explains how C ASPER selects
suitable code fragments for translation. Finally, §3.5 explains code generation after the program summary has been inferred.
3.1
Verified Lifting
Verified lifting [7,11] is a general technique that infers the semantics of code written in a general-purpose
language by “lifting” it to summaries expressed using a high-level language. C ASPER specifies code
fragment summaries in our program specification language in the form of postconditions that describe
the effects of the code fragment on its output variables, i.e., variables that are modified within the code
fragment. The goals of our program specification language are:
• To generate summaries that C ASPER can translate to the target platform DSL. This excludes valid
summaries that cannot be translated. Therefore, the language should omit constructs that cannot
be translated easily to the target.
• To generate non-trivial summaries that exhibit parallel data processing. Obviously, this excludes
summaries that execute the computation sequentially. §4.2.2 discusses the sources of parallelism
in MapReduce and how C ASPER generates solutions that exploits them.
With these goals in mind, C ASPER’s inferred summaries must be of the form:
∀ v ∈ out putVariables . v = reduce(map(data, fm ), fr )[idv ]
(1)
where data is the iterable input data collection. The map function iterates over the data while calling
the fm function on each element. fm takes as input an element from data and generates potentially
multiple key-value pairs. map then collects and returns key-value pairs generated by invocations of fm .
The reduce function takes these key-value pairs, groups them by key, and calls fr for each key and all
values that correspond to that key. Function fr aggregates all values for the given key and emits a single
key-value pair. Like map, reduce collects all aggregated key-value pairs and returns an associative array
M. B. S. Ahmad & A. Cheung
73
that maps each variable’s ID to its final value. The variable ID is a unique identifier that C ASPER assigns
to every output variable. C ASPER requires that summaries (i.e., postconditions) be of the form described
in Eqn. (1) for easy translation to Hadoop tasks.
In the preceding discussion, fm and fr remain unspecified. Verified lifting seeks a definition of fm and
fr that makes a valid inferred summary, viz., one that preserves the semantics of the input code fragment.
To do this in C ASPER, the synthesizer generates the implementation of these two functions (see §3.3)
using the verification conditions computed by the program analyzer for each code fragment (see §3.4).
3.2
Verifying Equivalence
The summaries C ASPER infers must be semantically equivalent to the input code fragment. C ASPER establishes the validity of the inferred postconditions using Hoare-style verification conditions [10], which
represent the weakest preconditions of a code fragment that must be true to establish the postcondition of
the same code fragment under all possible executions. Generating verification conditions for simple assignment statements and conditionals is straightforward. For example, consider the imperative program
statement x := y + 3. To show that the candidate postcondition x > 10 is a valid postcondition, we
must prove that y + 3 > 10 is true before the statement is executes. In this case, y + 3 > 10 is called
the verification condition for this postcondition. Computing verification condition is easy for simple
statements. For a loop, however, computing verification conditions becomes more difficult since a loop
invariant is needed. The loop invariant is a hypothesis that asserts that the postcondition is true regardless
of how many times the loop iterates. Hoare logic states that the following three statements must hold for
the loop invariant (and postcondition) to be valid:
1. ∀σ . preCondition(σ ) → loopInvariant(σ )
2. ∀σ . loopInvariant(σ ) ∧ loopCondition(σ ) → loopInvariant(body(σ ))
3. ∀σ . loopInvariant(σ ) ∧ ¬loopCondition(σ ) → postCondition(σ )
Statement 1 asserts that the loop invariant must be true when the precondition is true for all program
states (σ ), i.e., the loop invariant must be true before entering the loop. Statement 2 asserts that for all
possible program states σ —assuming that the loop invariant is true and that the loop continues—the loop
invariant remains true after one more execution of the loop body; (here, body(σ ) returns a new program
state after executing the loop body at σ ). Statement 3 asserts that if the loop invariant is true and if loop
terminates, then the postcondition must be true for all possible program states.
Two challenges affect the identification of postconditions (and hence summaries) for code fragments
that involve loops. First, both the loop invariants and postcondition must be synthesized. Unlike prior
work on searching for invariants [8, 21], however, C ASPER needs to find loop invariants that are only
logically strong enough to establish the soundness of the postcondition, i.e., those that satisfy statement
3. This is made easier thanks to the specific form of the postcondition that C ASPER looks for. In addition, establishing the validity of the found invariants and postconditions requires checking all possible
program states, complicating the synthesis problem. We discuss how C ASPER makes the search problem
manageable in §3.3.3.
3.3
Searching for summaries
C ASPER seeks to infer a summary for each code fragment, where each summary is a postcondition of the
form explained in §3.1. This section describes how C ASPER uses synthesis to search for postconditions
and the loop invariants they require to prove the postconditions correct.
74
Leveraging Parallel Data Processing Frameworks with Verified Lifting
preCondition(hR, hG, hB, i) ≡
hR = [0..0] ∧ hG = [0..0] ∧ hB = [0..0] ∧ i = 0
postCondition(data, hR, hG, hB) ≡
∀ 0 ≤ j < hR.length. hR[ j] = reduce(map(data, fm ), fr )[(0, j)] ∧
∀ 0 ≤ j < hG.length. hG[ j] = reduce(map(data, fm ), fr )[(1, j)] ∧
∀ 0 ≤ j < hB.length. hB[ j] = reduce(map(data, fm ), fr )[(2, j)]
loopInvariant(data, hR, hG, hB, i) ≡
LoopCounterExp ∧
∀ 0 ≤ j < hR.length. hR[ j] = reduce(map(data[0 : i], fm ), fr )[(0, j)] ∧
∀ 0 ≤ j < hG.length. hG[ j] = reduce(map(data[0 : i], fm ), fr )[(1, j)] ∧
∀ 0 ≤ j < hB.length. hB[ j] = reduce(map(data[0 : i], fm ), fr )[(2, j)]
Figure 3: Definitions of precondition, postcondition and loop invariant for the 3D Histogram example.
3.3.1
Generating Verification Conditions
In §3.2, we explained the three verification conditions that must be satisfied by the synthesized summary. These verification conditions involve a precondition, postcondition, and loop invariant for the
code fragment. Preconditions are generated by extracting, through static program analysis, the program
state (values of input and output variables) just before the loop starts executing. When the value of a
variable before the loop starts cannot be determined, C ASPER generates a new variable to represent the
initial value. The loop invariant has a form similar to the postcondition (see §3.1); unlike the postcondition, however, which calls map and reduce on the entire data collection, the loop invariant calls map
and reduce only on the subset of the collection that has so far been traversed by the loop. Also, the loop
invariant includes an expression that describes the behavior of the loop counters.
Figure 3 shows the precondition, postcondition and loop invariant generated for the 3D Histogram
benchmark. The postcondition and loop invariant functions describe the behavior that must be true for
the bodies of fm and fr to be correct. For example, the postcondition states that for each index j of hR,
the value of hR[ j] must equal to the output of map and reduce functions for key (0, j).
3.3.2
Specifying Search Space
This section describes how C ASPER generates the grammar that the synthesizer uses to construct bodies
of fm and fr . By dynamically generating a grammar for each code fragment, C ASPER restricts the space
of summaries through which the synthesizer must search.
Recall that the function fm takes as parameters the input data collection and an index into the collection and returns a set of key-value pairs. C ASPER constructs the body of fm using emit statements
and conditionals. The current C ASPER prototype does not generate implementations of fm that involve
loops. Based on our experiments, we found that using the same number of emit statements as output
variables in the code fragment works well as a starting point. The number of emit statements can then
be increased if a solution cannot be found. In general, however, C ASPER takes a conservative approach
to avoid implementations with redundant emit statements since they generate unnecessary shuffle data,
consequently hurting performance. Each emit statement produces a key-value pair; the key and value
can be any expression generated by one of our expression grammars or tuples of such expressions.
The fr function reduces all values emitted by map for a given key into a single value. The body of
M. B. S. Ahmad & A. Cheung
fm
75
::= {EmitMap; EmitMap; EmitMap; }
EmitMap ::= emit(Exp, Exp) | if(BoolExp){ emit(Exp, Exp) }
Exp
::= IntExp | BoolExp | (Exp, Exp)
IntExp ::= IntTerm | data[IntExp] | IntExp + IntExp | IntExp % IntExp
IntTerm ::= intLiteral | loopCounter
BoolExp ::= true | false | IntExp == IntExp | BoolExp ∧ BoolExp
|
fr
FoldExp
BoolExp ∨ BoolExp
::= {value = IntLiteral; for(v in values){ value = FoldExp } emit(key, value); }
::= FoldTerm | FoldExp + FoldExp
FoldTerm ::= intLiteral | value | v
LoopCounterExp ::= LoopTerm <= LoopTerm <= LoopTerm
LoopTerm
::= loopCounter | intLiteral | data.length
Figure 4: Grammar generated for 3D Histogram example.
fr implements the folding operation. C ASPER uses the synthesizer to generate the folding expression
that reduces two values into one. It also generates an expression grammar to synthesize the folding
expression.
C ASPER generates expression grammars for each primitive data type found in the code fragment.
Each grammar can be used to generate expressions that evaluate to a value of its type. The expressions are
formulated using the operators and function calls from the original code fragment. Input variables, loop
counters, and literals from the code fragment are used as terminals. For arithmetic types, C ASPER lets
the synthesizer generate new constants as well. Furthermore, C ASPER generates an expression grammar
to construct the folding expression in fr and the loop counter expression in the loop invariant.
All expression grammars generated by C ASPER are bounded to a set level of recursion which the user
can specify. The recursive bound of a grammar controls the amount of times the synthesizer is allowed to
expand the non-terminals while formulating an expression. If the synthesizer cannot find a solution, the
expression grammars can be incrementally expanded by either introducing new operators and functions
that were not found in the code fragment or increasing the recursive bound on the grammar. The order in
which new constructs are added to the grammar is guided by priority values that we have encoded into
C ASPER.
Figure 4 shows the grammar generated for the 3D Histogram benchmark after 2 iterations of grammar
expansion. It is easy to see how to the solution presented in Figure 2 can be generated from this grammar.
3.3.3
Search Procedure
Despite all the search space constraints already discussed, the space of possible summaries remains large.
Therefore, to accelerate the search, C ASPER splits the verification process into two parts: it first uses a
bounded-checking procedure to find candidate invariants and postconditions. For candidate invariants
and postconditions that pass the bounded-checking procedure, it then uses a theorem prover to establish
soundness for all input program states. If the theorem prover fails (via a timeout) or returns unsat, the
synthesizer continues to search for a new candidate summary in the same search space. When it finds no
more candidate summaries, the synthesizer expands the grammar to increase the search space. It does
76
Leveraging Parallel Data Processing Frameworks with Verified Lifting
this by either adding new non-terminals, increasing the recursive bound for the grammar, or increasing
the number of emits made by fm , as discussed earlier. Configuration parameters specified by the user
control this iterative expansion of the search space. Eventually, the synthesizer either finds a verifiably
correct summary or halts efforts to convert the code fragment.
C ASPER also decouples the synthesis procedure from formal verification and uses off-the-shelf tools
for each of the two sub-problems. This methodology works well in practice to reduce the synthesis time.
3.4
Initial Code Extraction
The current C ASPER prototype parses the abstract syntax tree (AST) of the input program source code
to extract loops as individual fragments. C ASPER then analyzes each fragment’s AST to ensure it meets
the following criteria:
• The code fragment contains no unsupported library function calls. To synthesize summaries,
C ASPER must identify input and output variables (see §3.4.1), and the lack of library source code
makes this impossible unless models that describe library function semantics are encoded into the
compiler. C ASPER currently supports commonly used library functions, such as methods of the
java.lang.{String,Integer} and java.util.{ArrayList,Map} classes.
• Each loop contains no unstructured control flow. C ASPER’s current implementation cannot extract
necessary semantics from such loops, such as the premature terminations and loop stride.
• The code fragment contains no nested loops. C ASPER does not currently process nested loops. If
any is found, C ASPER attempts to optimize only the innermost loop.
• The code fragment contains no assignment statements that can create an alias. Moreover, C ASPER
does not currently perform any alias analysis and assumes that none of the input variables in the
code fragment is aliased. Thus, user defined objects cannot be assigned. Fields of these objects can
be modified as long as they are a primitive type. Similarly, array indexes can be modified—if array
is of an immutable type—but not the pointer to an array. Support for assigning common immutable
data structures, such as java.lang.{Integer,String}, has been built into the compiler.
C ASPER overlooks code fragments that do not satisfy these criteria. Once a loop has been marked for
conversion, it is normalized to a simpler form before further analysis. The normalization breaks down
large instructions into smaller, simpler ones (such as breaking down all expressions into binary ones)
and converts all loop constructs into while(true){...} loops. All of which are standard compiler
transformations.
3.4.1
Extracting Input and Output Variables
C ASPER makes additional passes on the normalized AST to extract input and output variables. It examines each assignment statement inside the code fragment in isolation and extracts assignment targets as
output variables. Similarly, all variables in the source of an assignment are extracted as input variables
(this may also include some output variables). Local variables declared inside a loop body are considered
neither input nor output variables. To determine whether a function call parameter is an input or output
variable, C ASPER must analyze the function’s source code. For library functions, this information must
be encoded into C ASPER beforehand. If a constant index of an array is accessed, then a separate input
variable is created for the array element. However, if any dynamic accesses are made, then the entire
array is considered an input variable.
M. B. S. Ahmad & A. Cheung
77
For the 3D Histogram example shown in Figure 2, arrays hR, hG and hB are labeled as output variables, and the data array is identified as an input variable. Variables i, r, g and b—all declared inside
the loop body—are considered to be neither input nor output variables.
3.5
Code Generation
After C ASPER finds a summary for each input code fragment, the last step is to convert each such
summary into a Hadoop task. The class encapsulating the Hadoop task has an execute method, which
takes as parameters all input variables in the code fragment. This method invokes the Hadoop task and
returns an associative array that maps each variable identifier to its final value as computed by the Hadoop
task. The associative array is then used to update the output variables before the remaining program is
executed. Translation of fm and fr to concrete Hadoop syntax is done using syntax-driven translation
rules. Since the postcondition is already in the MapReduce form, the rules to translate them into the
concrete syntax of Hadoop are straightforward and omitted here for brevity.
Figure 2c shows the final output code for the 3D Histogram example. HistogramHadoop is the class
generated by C ASPER, and the execute method invokes the Hadoop runtime with the generated map
and reduce classes. The resulting values—hR, hG, and hB—are compiled and returned by execute and
assigned to the original program’s corresponding output variables as shown in Figure 2b. The code that
reconstructs the arrays from key-value pairs is not shown for brevity.
4
Evaluation
We now describe our prototype implementation of C ASPER and present the results derived from applying
C ASPER to varied benchmarks.
4.1
Implementation
C ASPER’s program analysis and code generation modules are implemented by extending the open source
Java compiler Polyglot [14]. For synthesis, C ASPER uses an off-the-shelf synthesizer called SKETCH [18].
SKETCH uses counter-example guided inductive synthesis as its core algorithm. The program analyzer
encodes the verification conditions and search space in the SKETCH language. We implemented the
functions and data structures required to model the semantics of MapReduce programs in SKETCH as
well. In addition, C ASPER automatically models in SKETCH all program-specific user-defined data
types. SKETCH performs bounded model-checking to generate a summary, which we then use to generate the Hadoop Code. We have not yet implemented C ASPER’s formal verification component in
C ASPER and therefore rely solely on bounded model-checking to verify correctness.
4.1.1
Platform for Evaluation
We used our C ASPER prototype to translate sequential Java benchmarks into Hadoop tasks. We measured
the performance of both the original and the generated implementations on a 10 node cluster of Amazon
AWS m3.xlarge instances. Each m3.xlarge node was equipped with High Frequency Intel Xeon E52670 v2 (Ivy Bridge) 2.5 GHz processors, 15 GB of memory, and 80 GB of SSD storage. The cluster ran
Ubuntu Linux 14.04 LTS, Hadoop 2.7.2 and Spark 1.6.1. We used HDFS for input data storage in both
sequential and MapReduce implementations.
78
Leveraging Parallel Data Processing Frameworks with Verified Lifting
4.1.2
Benchmarks
We evaluated the performance of C ASPER on the following five benchmarks. These benchmarks were
taken from the Phoenix suite of benchmarks [17] and represent traditional problems that can be parallelized by rewriting using the MapReduce paradigm.
• The Summation benchmark sums all integer values in a list.
• The Word Count benchmark counts the frequency of each word in a body of text by iterating
through each word in the input file.
• The String Match benchmark determines whether a set of two strings is contained in a body of
text. It returns a Boolean value for each string as output. Like Word Count, this benchmark also
iterates through each word in the input file.
• The 3D Histogram benchmark generates a three-dimensional histogram that tallies the frequency
of each RGB color component in an image (Figure 2a). The output is an array for each color
component that holds the frequency of each intensity value.
• The Linear Regression benchmark iterates over a collection of cartesian points (x, y) and computes a number of coefficients for linear regression: namely, x, y, x ∗ x, x ∗ y, y ∗ y.
All benchmarks read input data from a text file saved on HDFS. For the generated Hadoop solutions,
class org.apache.hadoop.mapred.FileInputFormat is used to read and split data across multiple
mappers.
4.2
Compilation Performance
This section reports the time that C ASPER takes to generate Hadoop implementations and discusses the
quality of these implementations.
4.2.1
Scalability
Table 1 shows the average time (over 5 runs) required to synthesize a summary for each of the five
benchmarks. C ASPER synthesized Hadoop implementations for all benchmarks within an hour. Simpler
benchmarks, such as Summation and Word Count, were converted in under a minute and required only
one iteration of grammar generation. No benchmark required more than two iterations to successfully
synthesize an implementation.
Benchmark
Program Analysis
Synthesis and BMC
# of Grammar Iterations
Summation
Word Count
String Match
3D Histogram
Linear Regression
< 1s
< 1s
< 1s
< 1s
< 1s
13s
44s
1406s
2355s
1801s
1
1
2
2
2
Table 1: Average time for C ASPER to synthesize each benchmark.
M. B. S. Ahmad & A. Cheung
4.2.2
79
Sources of Parallelism
A MapReduce program has two primary sources of parallelism. First, processing can be parallelized in
the map phase by partitioning the input data and spawning multiple mappers to process each partition
simultaneously. Second, the reduce phase can be executed in parallel by grouping data to separate keys
and aggregating for each key simultaneously. Hadoop also supports the use of combiners. Before the
shuffle phase, combiners—if used—aggregate data locally on every node to offer additional parallelism
and decrease the amount of data that needs to be shuffled.
We now discuss the implementations C ASPER generated and how each leveraged both map and
reduce side parallelism.
The Summation benchmark produces as output a single integer variable. All data must be aggregated
together and cannot be split to multiple keys. The translated solution emits a key-value pair (0, number)
for each number in the input dataset during the map phase. These key-value pairs are aggregated locally
on each node in parallel before being sent to the reducer. Note that key 0 is the unique ID for the output
variable.
The C ASPER-generated implementation of the Word Count benchmark emits (word, 1) for each
word encountered. The reducer then sums the values for each key. All nodes aggregate data locally
(using a combiner) to compute word counts for the assigned data partition before the reducer aggregates
intermediate results. In addition, C ASPER uses the words as keys. Therefore, the aggregation for different
words is performed in parallel.
The generated String Match benchmark implementation parallelizes the search process. Each mapper iterates its assigned partition of text and emits (key,true) whenever a key being searched is encountered. The data is locally aggregated by doing a disjunction of all values for a given key. Reduce side
parallelism is leveraged as each key is aggregated in parallel.
The 3D Histogram benchmark resembles the word count problem. Hence, the C ASPER generated
implementation iterates over each pixel and emits ((color, intensity), 1), where the key is a tuple of
color and the intensity value. Data is aggregated in parallel in the reduce phase for each index of each
histogram, for a total of 255×3 keys. As with the preceding benchmarks, data is locally aggregated
before shuffling.
Linear Regression resembles the summation benchmark. All coefficients for a given point (x, y,
x ∗ x, y ∗ y, and x ∗ y) are calculated and emitted by the mapper, with a different key corresponding to
each coefficient. For each key, the values are aggregated (by summation) locally before being globally
reduced.
As is evident from all these benchmarks, C ASPER generated non-trivial implementations. C ASPER
leveraged reduce side parallelism, reducing each output variable in parallel by assigning to each variable
a unique ID and reducing data for each variable ID in parallel. For arrays, even greater parallelism was
achieved by reducing each index of the array in parallel. C ASPER also exploited map side parallelism
by evaluating expressions before they are emitted by the mapper (e.g., as in Linear Regression). Lastly,
C ASPER used the reduce class as a combiner to locally aggregate data whenever the reduce input and
output key-value pairs were of the same type.
To evaluate the quality of optimization C ASPER achieved, we compared the runtime performance of
the original sequential implementations to the Hadoop implementations generated. We also examined
the performance when synthesized summaries were manually translated to the Spark framework. Finally,
to add context, we compared the performance of Spark implementations generated by MOLD. Figure 5
graphs the results of all five benchmarks against different dataset sizes.
Leveraging Parallel Data Processing Frameworks with Verified Lifting
1200
800
2400
Sequential
Hadoop
Spark
Mold
Execution time (s)
Execution time (s)
1600
400
0
1800
1200
1000
Sequential
Hadoop
Spark
Mold
Execution time (s)
80
600
20
30
40
Input Size (GB)
50
10
(a) Summation
1200
800
20
30
40
Input Size (GB)
250
50
10
(b) Word Count
2000
Sequential
Hadoop
Spark
Mold
Execution time (s)
Execution time (s)
1600
500
0
0
10
750
Sequential
Hadoop
Spark
Mold
400
0
1500
1000
20
30
40
Input Size (GB)
50
(c) String Match
Sequential
Hadoop
Spark
Mold
500
0
10
20
30
40
Input Size (GB)
50
(d) Histogram
10
20
30
40
Input Size (GB)
50
(e) Linear Regression
Figure 5: Performance comparison of original implementations (blue) vs C ASPER optimized Hadoop
(orange) and Spark (Green) implementations. Performance of implementations when optimized using
MOLD added for reference (yellow).
4.2.3
Alternate Implementations
As discussed, C ASPER generates non-trivial implementations that effectively leverage the parallelism
offered by Hadoop MapReduce. However, these implementations may not be the most efficient ones.
For the 3D Histogram benchmark, an alternative Hadoop implementation would be to emit for each
pixel in the input data key-value pairs of the form (intensity, color). Hadoop would then group the data
by the 256 intensity values. Aggregation would involve simply counting the number of times each color
(Red, Green, or Blue) appears for a given key. Whether C ASPER generates this implementation or the
one discussed earlier in the paper depends upon which implementation the synthesizer discovers first. An
important opportunity for future work is to enable C ASPER to use heuristics to reason about the optimal
implementation.
4.3
Performance of the Generated Benchmarks
In all five benchmarks, the generated Hadoop implementations were not only faster than their sequential
counterparts but they also scaled better. Even for our smallest dataset (10GB), the Hadoop implementations outperformed the original implementations. The average speed up for Hadoop implementations
across all benchmarks was 3.3× with a maximum speedup of 4.5× in the case of String Match.
Translating the summaries synthesized by C ASPER into Spark yielded even higher speedups (up to
8.1×) since Spark uses cluster memory much more efficiently and minimizes disk I/O between different MapReduce stages. Extending C ASPER to automatically generate Spark code from the synthesized
summary is currently a work in progress.
M. B. S. Ahmad & A. Cheung
5
81
Related Work
MapReduce DSLs. MapReduce is a popular programming model. It scales elastically, integrates well
with distributed file systems, and abstracts away from the user low-level synchronization details. As
such, many systems have been built that compile code down to MapReduce [3–5]. However, these systems provide their own high-level DSLs in which the users must use to express their computation. In
contrast, C ASPER works with native Java programs and infers rewrites automatically.
Source-to-Source Compilers. Many efforts translate programs directly from low-level languages into
high-level DSLs. MOLD [15], a source-to-source compiler, relies on syntax-directed rules to convert
native Java programs to Apache Spark. Unlike MOLD, we translate on the basis of program semantics. This eliminates the need for rewrite rules, which are difficult to generate and brittle to code pattern
changes. Many source-to-source compilers have been built in other domains for similar purposes. For
instance, [13] evaluates numerous tools for C to CUDA transformations. However, these compilers often
require manual efforts to annotate the original source code. Our methodology works with code without
any user annotation.
Synthesizing Efficient Implementations. Extensive literature describes the use of synthesis to generate efficient implementations and optimizing programs. [19] is the most recent research that attempts to
synthesize MapReduce solutions with user-provided input and output examples. QBS [7] and STNG [11]
both use verified lifting and synthesis to convert low-level languages to specialized high-level DSLs for
database applications and stencil computations respectively.
6
Conclusion
This paper presented C ASPER, a compiler that automatically re-targets native Java code to execute on
Hadoop. C ASPER uses verified lifting to convert code fragments in the original program to a high-level
representation that can then be translated to generate equivalent Hadoop tasks for distributed data processing. We implemented a prototype of C ASPER and evaluated its performance on several MapReduce
benchmarks. Our experiments show that C ASPER can translate all input benchmarks, and the generated
programs can run on average 3.3× faster compared to their sequential counterparts.
7
Acknowledgment
The authors are grateful for the support of NSF grants CNS-1563788 and IIS-1546083 as well as DARPA
award FA8750-16-2-0032, and DOE award DE-SC0016260.
References
[1] Tyler Akidau, Robert Bradshaw, Craig Chambers, Slava Chernyak, Rafael J. Fernandez-Moctezuma,
Reuven Lax, Sam McVeety, Daniel Mills, Frances Perry, Eric Schmidt & Sam Whittle (2015): The
Dataflow Model: A Practical Approach to Balancing Correctness, Latency, and Cost in Massive-Scale,
Unbounded, Out-of-Order Data Processing. Proceedings of the VLDB Endowment 8, pp. 1792–1803,
doi:10.14778/2824032.2824076.
[2] Apache Hadoop. http://hadoop.apache.org. Accessed: 2016-04-19.
82
Leveraging Parallel Data Processing Frameworks with Verified Lifting
[3] Apache Hive. http://hive.apache.org. Accessed: 2016-04-20.
[4] Apache Pig. http://tensorflow.org/. Accessed: 2016-05-01.
[5] Apache Spark. https://spark.apache.org. Accessed: 2016-04-19.
[6] Apache Storm. http://storm.apache.org. Accessed: 2016-04-19.
[7] Alvin Cheung, Armando Solar-Lezama & Samuel Madden (2013): Optimizing Database-backed Applications with Query Synthesis. In: Proceedings of the 34th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’13, ACM, New York, NY, USA, pp. 3–14,
doi:10.1145/2491956.2462180.
[8] Michael D. Ernst, Jeff H. Perkins, Philip J. Guo, Stephen McCamant, Carlos Pacheco, Matthew S. Tschantz
& Chen Xiao (2007): The Daikon System for Dynamic Detection of Likely Invariants. Sci. Comput. Program.
69(1-3), pp. 35–45, doi:10.1016/j.scico.2007.01.015.
[9] GraphLab Create. https://dato.com/. Accessed: 2016-04-20.
[10] C. A. R. Hoare (1969): An Axiomatic Basis for Computer Programming. Communications of the ACM
12(10), pp. 576–580, doi:10.1145/363235.363259.
[11] Shoaib Kamil, Alvin Cheung, Shachar Itzhaky & Armando Solar-Lezama (2016): Verified Lifting of Stencil
Computations. SIGPLAN Not. 51(6), pp. 711–726, doi:10.1145/2980983.2908117.
[12] MongoDB 3.2. https://www.mongodb.org. Accessed: 2016-04-19.
[13] Cedric Nugteren & Henk Corporaal (2012): Introducing ’Bones’: A Parallelizing Source-to-source Compiler Based on Algorithmic Skeletons. In: Proceedings of the 5th Annual Workshop on General Purpose Processing with Graphics Processing Units, GPGPU-5, ACM, New York, NY, USA, pp. 1–10,
doi:10.1145/2159430.2159431.
[14] Polyglot. http://www.cs.cornell.edu/Projects/polyglot/. Accessed: 2016-05-01.
[15] Cosmin Radoi, Stephen J. Fink, Rodric Rabbah & Manu Sridharan (2014): Translating Imperative Code
to MapReduce. In: Proceedings of the 2014 ACM International Conference on Object Oriented Programming Systems Languages & Applications, OOPSLA ’14, ACM, New York, NY, USA, pp. 909–927,
doi:10.1145/2660193.2660228.
[16] Jonathan Ragan-Kelley, Connelly Barnes, Andrew Adams, Sylvain Paris, Frédo Durand & Saman Amarasinghe (2013): Halide: A Language and Compiler for Optimizing Parallelism, Locality, and Recomputation in Image Processing Pipelines. In: Proceedings of the 34th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’13, ACM, New York, NY, USA, pp. 519–530,
doi:10.1145/2491956.2462176.
[17] Colby Ranger, Ramanan Raghuraman, Arun Penmetsa, Gary Bradski & Christos Kozyrakis (2007): Evaluating MapReduce for Multi-core and Multiprocessor Systems. In: Proceedings of the 2007 IEEE 13th
International Symposium on High Performance Computer Architecture, HPCA ’07, IEEE Computer Society,
Washington, DC, USA, pp. 13–24, doi:10.1109/HPCA.2007.346181.
[18] SKETCH. https://people.csail.mit.edu/asolar/. Accessed: 2016-05-01.
[19] Calvin Smith & Aws Albarghouthi (2016): MapReduce Program Synthesis. SIGPLAN Not. 51(6), pp. 326–
340, doi:10.1145/2980983.2908102.
[20] Armando Solar-Lezama, Gilad Arnold, Liviu Tancau, Rastislav Bodik, Vijay Saraswat & Sanjit Seshia (2007): Sketching Stencils. In: Proceedings of the 28th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’07, ACM, New York, NY, USA, pp. 167–178,
doi:10.1145/1273442.1250754.
[21] Saurabh Srivastava & Sumit Gulwani (2009): Program Verification Using Templates over Predicate Abstraction. In: Proceedings of the 30th ACM SIGPLAN Conference on Programming Language Design and
Implementation, PLDI ’09, ACM, New York, NY, USA, pp. 223–234, doi:10.1145/1542476.1542501.
M. B. S. Ahmad & A. Cheung
83
[22] Arvind K. Sujeeth, Kevin J. Brown, Hyoukjoong Lee, Tiark Rompf, Hassan Chafi, Martin Odersky & Kunle
Olukotun (2014): Delite: A Compiler Architecture for Performance-Oriented Embedded Domain-Specific
Languages. ACM Trans. Embed. Comput. Syst. 13(4s), pp. 134:1–134:25, doi:10.1145/2584665.
[23] TensorFlow. http://tensorflow.org/. Accessed: 2016-04-20.
| 6cs.PL
|
Interpretable Apprenticeship Learning with
Temporal Logic Specifications
arXiv:1710.10532v1 [cs.SY] 28 Oct 2017
Daniel Kasenberg and Matthias Scheutz
Abstract— Recent work has addressed using formulas in
linear temporal logic (LTL) as specifications for agents planning in Markov Decision Processes (MDPs). We consider the
inverse problem: inferring an LTL specification from demonstrated behavior trajectories in MDPs. We formulate this as a
multiobjective optimization problem, and describe state-based
(“what actually happened”) and action-based (“what the agent
expected to happen”) objective functions based on a notion of
“violation cost”. We demonstrate the efficacy of the approach
by employing genetic programming to solve this problem in
two simple domains.
I. I NTRODUCTION
Apprenticeship learning, or learning behavior by observing expert demonstrations, allows artificial agents to learn
to perform tasks without requiring the system designer to
explicitly specify reward functions or objectives in advance.
Apprenticeship learning has been accomplished in agents
in stochastic domains, such as Markov Decision Processes
(MDPs), by means of inverse reinforcement learning (IRL),
in which agents infer some reward function presumed to
underlie the observed behavior. IRL has recently been criticized, especially in learning ethical behavior [2], because the
resulting reward functions (1) may not be easily explained,
and (2) cannot represent complex temporal objectives.
Recent work (e.g., [5], [8], [25]) has proposed using linear
temporal logic (LTL) as a specification language for agents
in MDPs. An agent in a stochastic domain may be provided
a formula in LTL, which it must satisfy with maximal probability. These approaches require the LTL specification to be
specified a priori (e.g., by the system designer, although [6]
construct specifications from natural language instruction).
This paper proposes combining the virtues of these approaches by inferring LTL formulas from observed behavior
trajectories. Specifically, this inference problem can be formulated as multiobjective optimization over the space of LTL
formulas. The two objective functions represent (1) the extent
to which the given formula explains the observed behavior,
and (2) the complexity of the given formula. The resulting
specifications are interpretable, and can be subsequently
applied to new problems, but do not need to be specified
in advance by the system designer.
The key contributions of this work are (1) the introduction
of this problem and its formulation as an optimization
problem; and (2) the notion of violation cost, and the stateand action-based objectives based on this notion.
Authors are with the Department of Computer Science, Tufts University, Medford, MA 02155, USA. The corresponding author is
[email protected]
In the remainder of the paper, we first discuss related
work; we then describe our formulation of this problem as
multiobjective optimization, defining a notion of “violation
cost” and then describing state-based and action-based objectives, corresponding to inferring a specification from “what
actually happened” and “what the demonstrator expected
to happen” respectively. We demonstrate the usefulness of
the formulation by using genetic programming to optimize
these objectives in two domains, called SlimChance and
CleaningWorld. We discuss issues pertaining to our approach
and directions for future work, and summarize our results.
II. R ELATED W ORK
The proposed problem draws primarily upon ideas from
apprenticeship learning (particularly, inverse reinforcement
learning), stochastic planning with temporal logic specifications, and inferring temporal logic descriptions of systems.
A. Apprenticeship Learning
Apprenticeship learning, the problem of learning correct
behavior by observing the policies or behavioral trajectories
of one or more experts, has predominantly been accomplished by inverse reinforcement learning (IRL) [19], [1].
IRL algorithms generally compute a reward function that
“explains” the observed trajectories (typically, by maximally
differentiating them from random behavior). Complete discussion of the many types of IRL algorithms is beyond the
scope of this paper.
The proposed approach bears some resemblance to IRL,
particularly in its inputs (sets of finite behavioral trajectories). Instead of computing a reward function based on
the observed trajectories, however, the proposed approach
computes a formula in linear temporal logic that optimally
“explains” the data. This addresses the criticisms of [2],
who claim that IRL is insufficient in morally and socially
important domains because (1) reward functions can be
difficult for human instructors to understand and correct,
and (2) some moral and social goals may be too temporally
complex to be representable using reward functions.
B. Stochastic Planning with Temporal Logic Specifications
There has been a wealth of work in recent years on
providing agents in stochastic domains (namely, Markov
Decision Processes) with specifications in linear temporal
logic (LTL). The most straightforward approach is [5], which
we describe further in section III-C. The problem is to
compute some policy which satisfies some LTL formula with
maximal probability.
More sophisticated approaches consider the same problem
in the face of uncertain transition dynamics [25], [8], partial observability [23], [22], and multi-agent domains [16],
[11]. Also relevant to the proposed approach is the idea of
“weighted skipping” that appears (in deterministic domains)
in [21], [24], [15].
The problem of inferring LTL specifications from behavior
trajectories is complementary to the problem of stochastic
planning with LTL specifications, much as IRL is complementary to “traditional” reinforcement learning (RL). Specifications learned using the proposed approach may be used
for planning, and trajectories generated from planning agents
may be used to infer the underlying LTL specification.
C. Inferring Temporal Logic Rules from Agent Behavior
The task of generating temporal logic rules that describe
data is not a new one. Automatic identification of temporal
logic rules describing the behavior of software programs (in
the category of “specification mining”) has been attempted
in, e.g., [9], [10], [17]. Lemieux et al’s Texada [17] allows
users to enter custom templates for formulas and retrieves all
formulas satisfied by the observed traces up to user-defined
support and confidence thresholds; this differs from the work
of Gabel and Su, who decompose complex specifications
into combinations of predefined templates. Specifications in
a temporal logic (rPSTL) have also been inferred from data
in continuous control systems in [13]. Each approach deals
with (deterministic) program traces.
The proposed approach is most strongly influenced by [4],
which casts the task of inferring temporal logic specifications
for finite state machines as a multiobjective optimization
problem amenable to genetic programming. Much of our
approach follows from this work; our novel contribution is
introducing the problem of applying such methods to agent
behavior in stochastic domains, and in particular our notion
of the violation cost as an objective function.
III. P RELIMINARIES
In this section we provide formal definitions of Markov
Decision Processes (MDPs) and linear temporal logic (LTL);
we then outline the approach taken in [5] for planning to
satisfy (with maximum probability) LTL formulas in MDPs.
A. Markov Decision Processes
The proposed approach pertains to agents in Markov Decision Processes (MDPs) augmented with a set, Π, of atomic
propositions. Since reward functions are not important to this
problem, we omit them. All notation and references to MDPs
in this paper assume this construction.
Formally, a Markov Decision Process is a tuple
M = hS, U, A, P, s0 , Π, Li
where
• S is a (finite) set of states;
• U is a (finite) set of actions;
U
• A : S → 2
specifies which actions are available in
each state;
•
•
•
•
P : S × U × S → [0, 1] is a transition function, with
P (s, a, s′ ) = 0 if a ∈
/ A(s), so that P (s, a, s′ ) is the
probability of transitioning to s′ by beginning in s and
taking action a;
s0 is an initial state;
Π is a set of atomic propositions; and
L : S → 2Π is the labeling function, so that L(s) is the
set of propositions that are true in state s.
A trajectory in an MDP specifies the path of an agent
through the state space. A finite trajectory is a finite sequence of state-action pairs followed by a final state (e.g.,
τ = (s0 , a0 ), · · · , (sT −1 , aT −1 ), sT ); an infinite trajectory
takes T → ∞, and is an infinite sequence of state-action
pairs (e.g., τ = (s0 , a0 ), (s1 , a1 ), · · · ). A sequence (finite or
infinite) is only a trajectory if P (st , at , st+1 ) > 0 for all
t ∈ {0, · · · , T − 1}. We will denote by TrajM the set of
all finite trajectories in an MDP M, and by ITrajM the set
of all infinite trajectories in M. We will denote by τ |T the
T -time step truncation (s0 , a0 ), · · · , (sT −1 , aT −1 ), sT of an
infinite trajectory r = (s0 , a0 ), (s1 , a1 ), · · · .
A policy M : TrajM × U → [0, 1] is a probability
distribution over an agent’s next action, given its previous
(finite) trajectory. A policy is said to be deterministic if,
for each trajectory, the returned distribution allots nonzero
probability for only one action; we write M : TrajM → U .
A policy is said to be stationary if the returned distribution
depends only on the last state of the trajectory; we write
π : S × U → [0, 1].
We denote ITrajM
M the set of all infinite trajectories that
may occur under a given policy M . More formally,
ITrajM
M = {τ = (s0 , a0 ), (s1 , a1 ), · · · ∈ ITrajM :
M (τ |T , aT ) > 0 for all T }
B. Linear Temporal Logic
Linear temporal logic (LTL) [20] is a multimodal logic
over propositions that linearly encodes time. Its syntax is as
follows:
φ ::=⊤ | ⊥ | p, where p ∈ Π | ¬φ | φ1 ∧ φ2 | φ1 ∨ φ2 |
φ1 → φ2 | Xφ | Gφ | Fφ | φ1 U φ2
Here Xφ means “in the next time step, φ”; Gφ means “in
all present and future time steps, φ”; Fφ means “in some
present or future time step, φ”; and φ1 U φ2 means “φ1 will
be true until φ2 holds”.
The truth-value of an LTL formula is evaluated over an
infinite sequence of valuations σ0 , σ1 , · · · , where for all i,
σi ⊆ Π. We say σ0 , σ1 , · · · φ if φ is true given the infinite
sequence of valuations σ0 , σ1 , · · · .
There is thus a clear mapping between infinite trajectories
and LTL formulas. We abuse notation slightly and define
L((s0 , a0 ), (s1 , a1 ), · · · ) = L(s0 ), L(s1 ), · · ·
We abuse notation further and say that for any τ ∈ ITrajM ,
τ φ if L(τ ) φ.
We define the probability that a given policy satisfies an
LTL formula φ by
M
PrM
M (φ) = Pr{τ ∈ ITrajM : τ φ}
That is, the probability that an infinite trajectory under M
will satisfy φ.
Each LTL formula can be translated into a deterministic
Rabin automaton (DRA), a finite automaton over infinite
words. DRAs are the standard approach to model checking
for LTL. A DRA is a tuple
D = hQ, Σ, δ, q0 , F i
where
• Q is a finite set of states;
Π
• Σ is an alphabet (in this case, Σ = 2 , so words are
infinite sequences of valuations);
• δ : Q × Σ → Q is a (deterministic) transition function;
• q0 is an initial state; and
• F = {(Fin1 , Inf 1 ), · · · , (Fink , Inf k )}, where Fin ⊆ Q,
Inf ⊆ Q for all (Inf, Fin) ∈ F specifies the acceptance
conditions.
A run r = q0 , q1 , · · · of a DRA is an infinite sequence of
DRA states such that there is some word σ0 σ1 · · · such that
δ(qi , σi ) = qi+1 for all i. A run r is considered accepting if
there exists some (Fin, Inf) ∈ F such that for all q ∈ Fin, q
is visited only finitely often in r, and Inf is visited infinitely
often in r.
C. Stochastic Planning with LTL Specifications
Planning to satisfy a given LTL formula φ within an
MDP M with maximum probability generally follows the
approach of [5].
The planning agent runs the DRA for φ alongside M by
constructing a product MDP M× which augments the state
space to include information about the current DRA state.
Formally, the product of an MDP M
=
hS, U, A, T, s0 , Π, Li and a DRA D = hQ, 2Π , δ, q0 , F i is
an MDP
×
×
M× = hS × , U × , A× , P × , s×
0 ,Π ,L i
where
×
• S = S × Q;
×
• U
= U ; A× = A;
×
′ ′
• P ((s, q), a, (s , q )) =
(
P (s, a, s′ ) if q ′ = δ(q, L(s′ ))
0
otherwise
s×
0 = (s0 , δ(q0 , L(s0 )))
×
×
• Π = Π; and L = L.
The agent constructs the product MDP M× , and then
computes its accepting maximal end components (AMECs).
An end component E of an MDP M× is a set of states
SE ⊂ S × and an action restriction (mapping from states to
sets of actions) AE : SE → 2U such that (1) any agent in SE
that performs only actions as specified by AE will remain
in SE ; and (2) any agent with a policy assigning nonzero
probability to all actions in AE is guaranteed to eventually
visit each state in AE infinitely often.
An end component thus specifies a set of states SE such
that with an appropriate choice in policy, the agent can
guarantee that it will remain in SE forever, and that it will
reach every state in SE infinitely often. An end component
is maximal if it is not a proper subset of another end
component. An end component is accepting if there is some
(Fin, Inf) ∈ F such that (1) if q ∈ Fin, then (s, q) ∈
/ SE
for all s ∈ S; and (2) there exists some q ∈ Inf, s ∈ S
such that (s, q) ∈ SE . In this case, by entering SE and
choosing an appropriate policy (for instance, a uniformly
random policy over AE ), the agent guarantees that the DRA
run will be accepting. A method for computing the AMECs
of the product MDP is found in [3].
The problem of satisfying φ with maximal probability
is thus reduced to the problem of reaching, with maximal
probability, any state in any AMEC. [5] shows how this can
be solved using linear programming.
IV. O PTIMIZATION P ROBLEM
Suppose that an agent is given some set of finite behavior trajectories τ 1 , · · · , τ m ∈ TrajM , where τ i =
(si0 , ai0 ), · · · , (siTi −1 , aiTi −1 ), siTi for all i ∈ {1, · · · , m}.
We refer to the agent whose trajectories are observed as
the demonstrator, and the agent that observes the trajectories
as the apprentice. There may be several demonstrators satisfying the same objectives; this does not affect the proposed
approach.
The proposed problem is to infer an LTL specification that
well (and succinctly) explains the observed trajectories. This
can be cast as a multiobjective optimization problem with
two objective functions:
1) An objective function representing how well a candidate LTL formula explains the observed trajectories
(and distinguishes them from random behavior); and
2) An objective function representing the complexity of
a candidate LTL formula.
This section proceeds by describing a notion of “violation
cost” (and defining the violation cost of infinite trajectories
and policies) and using it to define two alternate objective
functions representing (a) how well a candidate formula
explains the actual observed state sequence (a “state-based”
objective function), and (b) how well a candidate formula
explains the actions of the demonstrator in each state (an
“action-based” objective function). We then describe the
simple notion of formula complexity we will utilize, and
formulate the optimization problem.
•
A. Violation Cost
We are interested in computing LTL formulas that well explain the demonstrator’s trajectories. These formulas should
be satisfied by the observed behavior, but not by random
behavior within the same MDP (since, for example, the trivial
formula G ⊤ will be satisfied by the observed behavior,
but also by random behavior). Ideally we could assign a
“cost” either to trajectories (finite or infinite) or to policies
(and, particularly, to the uniformly random policy in M),
where the cost of a trajectory or policy corresponds to
its adherence to or deviance from the specification. Given
such
function C,the objective would be to minimize
P a cost
i
C(τ
)
−
C(πrand ) , where πrand : S × U → [0, 1] is
i
the uniformly-random (stationary) policy over M:
(
1
if a ∈ A(s)
πrand (s, a) = |A(s)|
0
otherwise
Π⊗ = Π, L⊗ = L
The state s−1 and action a−1 are added so that the agent
may choose to “skip” time step t = 0. This is necessary for
the case that s0 violates the specification.
Note that the transition dynamics of M⊗ are such that N
(the set of “skipped” time step indices) can be defined as
The obvious choice of such a cost function (over infinite
trajectories τ ) would be the indicator function 1τ 2φ which
returns 0 if τ φ and 1 otherwise; this function may be
extended to general policies M by 1 − PrM
M (φ). This function, however, cannot distinguish between small and large
deviances from the specification. For example, given the
specification G p, this function cannot differentiate between
τ such that p is almost always true and τ such that p is never
true. We thus propose a more sophisticated cost function.
For τ ∈ ITrajM , N a set of nonnegative integers, we
define τ \N to be the subsequence of τ omitting the stateaction pairs with time step indices in N . For example,
(s0 , a0 ), (s1 , a1 ), (s2 , a2 ), , · · · \{1} = (s0 , a0 ), (s2 , a2 ), · · · .
Each time step with an index in N is said to be “skipped”.
We define the violation cost of an infinite trajectory
τ ∈ ITrajM subject to the formula φ as the (discounted)
minimum number of time steps that must be skipped in order
for the agent to satisfy the formula:
T C(s⊗ , (a, ã), s⊗ ) = 1ã=susp
Violφ (τ ) = min
∞
X
N ⊆N0
τ \N φ t=0
γ t 1t∈N
(1)
Note that if τ φ, then Violφ (τ ) = 0.
In order to define a similar measure for policies, we
must construct an augmented product MDP M⊗ , which is
similar to M× as described in section III-C, but allows
an agent to “skip” states by performing at each time step
(simultaneously with their normal actions), a “DRA action”
ã ∈ {keep, susp}, where keep causes the DRA to transition
as usual, and susp causes the DRA to not update in response
to the new state.
Formally, given an MDP M = hS, U, A, T, s0 , Π, Li
and a DRA D = hQ, Σ, δ, q0 , F i corresponding to the
specification φ, we may construct a product MDP M⊗ =
⊗
⊗
hS ⊗ , U ⊗ , A⊗ , T ⊗ , s⊗
−1 , Π , L i as follows:
⊗
• S = (S ∪ {s−1 }) × Q
⊗
• U
= (U ∪ {a(−1 }) × Ũ , where Ũ = {keep, susp}
{a−1 } × Ũ if s = s−1
⊗
• A ((s, q)) =
A(s) × Ũ
otherwise
⊗
• s−1 = (s−1 , q0 )
⊗ ⊗
• P (s−1 , (a−1 , keep), (s0 , δ(q0 , L(s0 )))) = 1
⊗ ⊗
• P (s−1 , (a−1 , susp), (s0 , q0 )) = 1
⊗
′ ′
• Otherwise, P ((s, q), (a, ã), (s , q )) =
′
′
′
P (s, a, s ) if q = δ(q, L(s )) and ã = keep
P (s, a, s′ ) if q ′ = q and ã = susp
0
otherwise
•
N = {t ∈ N0 : ãt−1 = susp}
(2)
′
Define the transition cost s⊗ , (a, ã), s⊗ in M⊗ as
′
(3)
The violation cost of a (non-product) trajectory τ can then
be rewritten as a discounted sum of the transition costs at
each stage, minimized over the DRA actions ã−1 , ã0 , ã1 , · · · ,
subject to the constraint that the DRA run from carrying out
τ and the DRA actions must be accepting. This indicates
that the violation cost of a policy π may be thought of as
the state-value function for the policy π with respect to T C.
Indeed, we will define the violation cost of a policy this way.
We define a product policy to be a stationary policy π ⊗ :
⊗
S × (U ∪ {a−1 }) → [0, 1]. When we consider the violation
cost of a policy, we will assume a product policy of this
form.
There are two reasons for this. First, when evaluating a
candidate specification, we wish to assume the demonstrator
had knowledge of that specification (or else we would
be unable to notice complex temporal patterns in agent
behavior), and thus that the demonstrator’s policy is over
product states. Second, we wish to allow the demonstrator
to observe the new (non-product) state st before deciding
whether to “skip” time step t. That is, st should be observed
before ãt−1 is chosen, which is inconsistent with the typical
policy π : S ⊗ × U ⊗ → [0, 1] over the product space.
We can easily construct a product policy from the uni⊗
formly random policy on M. We define πrand
((s, q), a) =
πrand (s, a) for all s ∈ S, a ∈ A.
Upon constructing the product MDP M⊗ , we compute
its
S
SE i ,
AMECs (as in section III-C). Then let Sgood =
i∈{1,··· ,p}
and let Sbad be the set of states in the product space
from which no state in Sgood can be reached; these can be
determined by breadth-first search.
We can use a form of the Bellman update equation to
perform policy evaluation on a product policy π ⊗ . For
each state s⊗ ∈ Sbad , we initialize the cost of this state
1
, and we do not
to the maximum discounted cost, 1−γ
update these costs. This is done to enforce the constraint
that the minimization should be over accepting DRA runs.
Otherwise, the violation cost will always be trivially zero
(since ã = keep will always be picked). The update equation
has the following form:
X
X
π ⊗ ((s, q), a)
Viol(k+1) ((s, q)) ←
P (s, a, s′ )
a∈A(s)
s′ ∈S
min{1 + γViol(k) (s′ , q), γViol(k) (s′ , δ(q, L(s′ )))}
(4)
Algorithm 1 Best DRA state sequence for finite state
sequence s0 , · · · , sT
π⊗
G ET R ABIN S TATE S EQUENCE(V iolφrand , M⊗ ,
Sbad , s0 , · · · , sT )
Ct [s⊗ ] ← ∞ for all t ∈ {−1, 0, · · · , T }, s⊗ ∈ S ⊗
R−1 = {s⊗
−1 }
C−1 [s⊗
−1 ] ← 0
seq−1 [s⊗
−1 ] ← q0
for t ∈ {0, · · · , T } do
Rt = ∅
for (s, q) ∈ Rt−1 do
q ′ ← δ(q, L(st ))
Rt ← Rt ∪ {(st , q), (st , q ′ )}
if Ct−1 [(s, q)] + γ t < Ct [(st , q)] then
Ct [(st , q)] ← Ct−1 [(s, q)] + γ t
seqt [(st , q)] ← (seqt−1 [(s, q)], q)
end if
if Ct−1 [(s, q)] < Ct−1 [(st , q ′ )] then
Ct [(st , q ′ )] ← Ct−1 [(s, q)]
seqt [(st , q ′ )] ← (seqt−1 [(s, q)], q ′ )
end if
end for
end for
π⊗
argmin CT [s⊗ ] + γ T +1 Violφrand (s⊗ ) return
s⊗
T ←
1: function
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
s⊗ ∈RT \Sbad
π⊗
seqT [s⊗
T ],
and 16). The sequence seqt [(st , qt+1 )] = q0 , · · · , qT +1 that
achieves this minimal cost is also computed (lines 5, 13, and
17).
The apprentice then assumes that the demonstrator acted
randomly from time step T + 1 onward. Although this
assumption is probably incorrect, it is not entirely unreasonable, since it avoids the assumption that the demonstrator
attempted to satisfy the formula after time step T + 1,
which would artificially drive the net violation cost down;
this allows the apprentice to reuse values that are already
computed in order to evaluate the random policy.
Employing this assumption, the apprentice determines the
optimal product-space interpretation as seqT [s⊗
T ], where
s⊗
T =
π⊗
argmin CT [s⊗ ] + γ T +1 Violφrand (s⊗ )
1) State-based objective function: We first consider an
approach to estimating the violation cost of a finite trajectory that considers only the states visited in the trajectory,
ignoring the demonstrator’s actions.
The state-based violation cost is the minimand of (6),
which is the second value returned by Algorithm 1:
π⊗
T +1
CT [s⊗
Violφrand (s⊗
T]+γ
T)
22: end function
T +1
ViolSφ (τ ) = CT [s⊗
Violφrand (s⊗
T]+γ
T)
1
The min{·} in (4) is where the optimization over ã
(implicitly) occurs. Choosing ã = susp incurs a transition
cost of 1 and then causes the DRA to remain in state q;
choosing ã = keep incurs no transition cost, but causes the
DRA to transition to state δ(q, L(s′ )). The ability for the
demonstrator to optimize over ã after observing the new state
s′ corresponds to the location of the min{·} in the Bellman
update.
We define the violation cost of a policy as the function that
results when running this update equation to convergence:
⊗
Violπφ ((s, q)) = lim Viol(k) ((s, q))
k→∞
(5)
We now consider state-based (“what actually happened”)
and action-based (“what the agent expected to happen”)
objective functions, for explaining sets of finite trajectories.
The crux of both the state- and action-based objective
functions is Algorithm 1. Given a finite sequence of states
s0 , · · · , sT , Algorithm 1 determines the “optimal productspace interpretation” of s0 , · · · , sT . We define a product
space interpretation of a sequence of states s0 , · · · , sT in
an MDP M as a sequence of DRA states q0 , · · · , qT +1
such that, for all i ∈ {1, · · · , T + 1}, either qi = qi−1 , or
qi = δ(qi−1 , L(si−1 )). That is, a product-space interpretation
specifies a possible trajectory in M⊗ that is consistent with
the observed trajectory in M.
Algorithm 1 uses dynamic programming to determine, for
each time step t, the set of states Rt of DRA states that the
demonstrator could be in at time t (lines 3,7, and 10), as well
as the minimal violation cost Ct [qt+1 ] that would need to be
accrued in order to be in each such state qt+1 (lines 2,4, 12,
(6)
s⊗ ∈RT \Sbad
(7)
m
Thus the state-based objective function for τ , · · · , τ is the
sum of the estimated violation costs of all observed finite
trajectories, less m times the expected violation cost of the
random policy from the initial state:
!
m
X
π⊗
S
S i
Violφ (τ ) − mViolφrand (s⊗
(8)
Obj (φ) =
−1 )
i=1
The main drawback of the state-based approach is that
by ignoring the observed actions, the apprentice neglects
a crucial detail: that what the demonstrator “expected” or
“intended” to satisfy may differ from what actually was
satisfied. The fact that p did not occur does not mean that
the demonstrator was not attempting to make p occur with
maximal probability, particularly if p is a very rare event. To
solve this problem, we consider an action-based approach.
2) Action-based objective function: We now consider
estimating the violation cost of a finite trajectory τ by using
the observed state-action pairs to compute a partial policy
over the product MDP M⊗ .
To compute the action-based violation cost of a set of
trajectories τ 1 , · · · , τ m (Algorithm 2), the apprentice first
runs Algorithm 1 to determine the optimal product-space interpretation q0i , · · · , qTi +1 for each trajectory τ i (line 4), and
uses this to compute the resulting product-space sequence
⊗i
⊗i
i i
s−1
, · · · , s⊗i
T where st = (st , qt+1 ).
The assumption that for each i ∈ {1, · · · , m}, t ∈
{0, · · · , Ti − 1}, the demonstrator performed ait when in the
inferred product MDP state s⊗i
t , induces an action restriction
(lines 6 and 11) A∗ : S ⊗ → 2U where
S
S
6 ∅
{ait } =
{ait } if
i,t:
i,t:
∗ ⊗
⊗i
⊗
A (s ) = s⊗ =s⊗i
s =st
⊗ t⊗
A (s )
otherwise
Algorithm 2 Action-based violation cost of set of finite
set of finite state-action trajectories τ 1 , · · · , τ m where τ i =
(si0 , ai0 ), · · · , (siTi −1 , aiTi −1 ), siTi
π rand
1: function ACTION B ASEDV IOLATION C OST (V iolφM
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
, M⊗ ,
m
1
Sbad , τ , · · · , τ )
A∗ [s⊗ ] ← ∅ for s⊗ ∈ S ⊗
for i ∈ {1, · · · , m} do
q0i , · · · , qTi i +1 , V
←
G ET R ABIN S TATE S E i
i
QUENCE (s0, · · · , sTi )
for t ∈ {−1, 0, · · · , Ti − 1} do
i
i
A∗ [(sit , qt+1
)] ← A∗ [(sit , qt+1
)] ∪ {ait }
end for
end for
for s⊗ = (s, q) ∈ S ⊗ do
if A∗ [s⊗ ] = ∅ then
A∗ [s⊗ ] = A(s)
end if
end for
π ⊗∗
Compute ViolφA using (4)
π ⊗∗
15: return ViolφA (s⊗
−1 )
16: end function
The apprentice may then compute, using the Bellman update
rand ⊗
(4), the violation cost of the policy πA
(s ) that uniformly∗
randomly chooses an action from A∗ (s⊗ ) at each state s⊗
(line 14):
(
1
if a ∈ A∗ (s⊗ )
⊗
⊗
|A∗ (s⊗ )|
πA
∗ (s , a) =
0
otherwise
The action-based objective function is then
π⊗
π⊗
rand
ObjA (φ) = ViolφA∗ (s⊗
(s⊗
−1 ) − Violφ
−1 )
(9)
B. Formula Complexity
Given two formulas that equally distinguish between the
observed behavior and random behavior, we wish to select
the less complex of the two. Here it suffices to simply
minimize the number of nodes in the parse tree for the
LTL formula (that is, the total number of symbols in the
formula). There are also more sophisticated ways to evaluate
formula complexity (such as that used in [4]), but they are
not necessary for our purposes.
C. Multiobjective Optimization Problem
Given some set of finite trajectories τ 1 , · · · , τ m , we thus
frame the problem of inferring some LTL formula φ that
describes τ 1 , · · · , τ m as
min (Obj(φ), F C(φ))
φ∈LTL
where Obj is either ObjS , as described in (8), or ObjA , as
described in (9); F C is formula complexity (in this case,
the number of nodes in the formula) as specified in section
IV-B.
V. E XAMPLES
To demonstrate the effectiveness of the proposed objective
functions, we employed genetic programming to evolve a set
of LTL formulas (where formulas are represented by their
parse trees) in two domains. A summary of the domains used
is in Table I. In all demonstrations, we used MOEAFramework [12] for genetic programming, using standard tree
crossover and mutation operations [14]. We consider (separately) the state-based and action-based objectives. NSGAII over each set of objectives was run for 50 generations
with a population size of 100. This process was repeated 20
times. We employed BURLAP [18] for MDP planning, and
Rabinizer 3 [7] for converting LTL formulas to DRAs. In
each case, we restricted search to formulas of the form G φ.
The tables in this section show formulas that are Pareto
efficient in at least two NSGA-II runs - that is, there were
no solutions within those runs that outperformed them on
both objectives. For any Pareto inefficient formula φ, there
is some formula φ′ which both (1) better explains the
demonstrated trajectories (as measured by the violation-cost
objective function) and (2) is simpler. Thus it is reasonable
to restrict consideration to only Pareto efficient solutions.
A. SlimChance domain
The SlimChance domain consists of two states: sGOOD ,
a “good” state, and sBAD , a “bad” state. The agent has two
actions: try, and notry. If the agent performs notry, the
next state is always sBAD ; if the agent performs try, the
next state is sGOOD with small probability ǫ = 0.01 and
sBAD otherwise. Thus, performing the try action is “trying”
to make the good state occur, but will rarely succeed.
The set Π of atomic propositions for this problem consists
of a single proposition good, which is true in sGOOD but
false in sBAD . We then suppose that the agent is attempting
to satisfy the simple LTL formula G good.
A demonstrator attempting to minimize violation cost
generated three trajectories of 10 time steps each. This
resulted in the following trajectories (note that τ 1 = τ 3 ,
which occurred randomly):
τ 1 , τ 3 =(sBAD , try), (sBAD , try), (sBAD , try),
(sBAD , try), (sBAD , try), (sBAD , try), (sBAD , try),
(sBAD , try), (sBAD , try), (sBAD , try), sBAD
τ 2 =(sBAD , try), (sGOOD , try), (sBAD , try), (sBAD , try),
(sBAD , try), (sBAD , try), (sBAD , try), (sBAD , try),
(sBAD , try), (sBAD , try), sBAD
Tables II and III show all solutions that were Pareto efficient
in at least two runs, for ObjS and ObjA respectively. The
results emphasize the distinction between the two objective
functions. In Table II the correct formula G good is Pareto
efficient in two runs, but in most runs the obviously-incorrect
G ⊥ is the only Pareto efficient formula (and note that
ObjS (G ⊥) ≅ ObjS (G good)). In contrast, Table III shows
that when using ObjA , the true function G good is Pareto
efficient in all twenty runs.
TABLE I
S UMMARY OF EXAMPLE DOMAINS , WITH RUN TIMES FOR STATE - BASED / ACTION - BASED OBJECTIVES
# States
2
77
Domain
SlimChance
CleaningWorld
# Actions
2
5
“Actual” specification
G good
G (X(vacuum) U roomClean)
Time, state-based (s)
174.8 ± 18.0
19139.3 ± 671.2
Time, action-based (s)
372.0 ± 43.3
32932.3 ± 1755.0
TABLE II
TABLE IV
PARETO EFFICIENT SOLUTIONS IN STATE - BASED S LIM C HANCE
PARETO EFFICIENT SOLUTIONS IN STATE - BASED C LEANING W ORLD
Formula φ
G⊥
G good
ObjS (φ)
-0.3139852
-0.3139852
F C(φ)
2
2
# Runs
18
2
TABLE III
PARETO EFFICIENT SOLUTIONS IN ACTION - BASED S LIM C HANCE
Formula φ
G good
G(good U (X good))
G(good ∨ X good)
G((X good) U good)
G((X good) ∨ good)
ObjA (φ)
-0.4623490
-0.4939355
-0.9400473
-0.9400473
-0.9400424
F C(φ)
2
5
5
5
5
# Runs
20
5
5
3
2
Formula φ
G roomClean
G(F roomClean)
G((X roomClean) ∨ vacuum)
G((G ⊤) U roomClean)
G(F(undock U roomClean))
ObjS (φ)
-208.69876
-216.91139
-217.40816
-216.91169
-216.91170
F C(φ)
2
3
5
5
5
# Runs
20
20
2
2
2
TABLE V
PARETO EFFICIENT SOLUTIONS IN ACTION - BASED C LEANING W ORLD
Formula φ
G(roomClean)
G(F roomClean)
G(vacuum ∨ F roomClean)
G(F(roomClean ∨ dock))
G((F roomClean) ∨ dock)
G((XroomClean) ∨ vacuum)
ObjA (φ)
-72.74240
-75.15686
-75.15832
-75.15782
-75.15832
-75.64639
F C(φ)
2
3
5
5
5
5
# Runs
20
20
3
3
2
2
B. CleaningWorld domain
In the CleaningWorld domain, the agent is a vacuum
cleaning robot in a dirty room. The room is characterized
by some initial amount dirt ∈ N0 of dirt; the agent has
some battery level battery ∈ N0 . The actions available to
the agent are: vacuum, which reduces both dirt and battery
by one; dock, which plugs the robot into a charger, allowing
it to increment battery for each time step it remains docked;
undock, which unplugs the robot from the charger; wait,
which allows the robot to remain docked if it is currently
docked, but otherwise simply decrements battery. If the
robot’s battery dies (battery = 0), the robot may only
perform the dummy action beDead. The domain has two
propositions batteryDead, which is true iff battery = 0,
and roomClean, which is true iff dirt = 0. There are also
propositions corresponding to each action (where, e.g., the
proposition vacuum is true whenever the agent’s last action
was to vacuum). The agent is to satisfy the LTL objective
G ((X vacuum) U roomClean).
An agent attempting to minimize violation cost for this
specification produced three demonstration trajectories of 10
time steps each. Because CleaningWorld is deterministic, all
three trajectories were identical. Here we represent each state
s by (d, b) where d is the amount of dirt still in the room in
state s and b is the robot’s current battery level.
formulas (in particular, G roomClean) arguably better describe agent behavior than the “actual” specification φact =
G((X vacuum) U roomClean): they are simpler than φact
while generating identical trajectories. This is reflected by
the fact that φact was generated by the algorithm for both
state- and action-based runs, but ObjS (φact ) = −215.78773,
ObjA (φact ) = −75.10621, and F C(φact ) = 5, which is
Pareto dominated by G(F roomClean) when considering
either ObjS or ObjA . Perhaps because of this, the actual
formula is never recovered (although similar formulas occasionally are, such as G((XroomClean) ∨ vacuum)).
VI. D ISCUSSION
While for demonstration purposes we chose to use NSGAII for optimization, in principle any algorithm that can
optimize over LTL formulas should suffice. Exploring other
algorithms is a topic for future work. In particular, the
genetic programming methods employed operate entirely on
the syntax of LTL; a method that can make some use of LTL
semantics may find optimal solutions more efficiently.
Optimizing over the space of all LTL formulas is difficult
because of the combinatorial nature of this space. Since the
number of LTL formulas of length ℓ increases exponentially
in ℓ, optimization algorithms like NSGA-II are likely to
recover simple formulas that explain the demonstrator’s behavior reasonably well, but are less likely to recover complex
1 2 3
τ , τ , τ =((5, 3), vacuum), ((4, 2), vacuum),
formulas that better explain the behavior.
((3, 1), dock), ((3, 1), wait), ((3, 3), undock),
We do not specify how to select between Pareto efficient
((3, 3), vacuum), ((2, 2), vacuum), ((1, 1), dock), solutions; this depends on the relative degree to which
system designers value simplicity versus explanatory power.
((1, 1), wait), ((1, 3), undock), (1, 3)
In practice, system designers with clear preferences could
convert the given problem into a single-objective problem
Tables IV and V show all solutions that were Pareto with objective f (Obj(φ), F C(φ)) where f is some nondeefficient in at least two runs, for ObjS and ObjA respectively. creasing function encoding these preferences.
The major drawback of the proposed approach is its scalThe formulas G roomClean and G(F roomClean) are
generated in all 20 runs by both ObjS and ObjA . These ability. Table I indicates that evaluation on CleaningWorld
with the action-based objective took, on average, roughly
9h 9m. For problems with much larger state and action
spaces, this approach is certainly intractable. Theoretically,
a single iteration in the computation of Violπφ takes time in
O(|S|2 |Q||U |). Run time for objective function evaluation
also scales linearly in the total number of demonstration
time steps. Identifying approaches with better theoretical and
practical run times is a topic for future work.
This paper also assumes that the demonstrator is operating
in an environment with complete information (e.g., an MDP),
no other agents, and known transition dynamics. Extensions
to unknown transition dynamics, POMDPs, and multi-agent
domains are a topic for future work.
In both given examples, the “true” specification can be
modeled using a reward function: in SlimChance, give high
reward if and only if the agent is in sGOOD ; in CleaningWorld, give high reward only when roomClean is true. IRL
may easily recover these reward functions, and would likely
converge more quickly than our approach. These examples
are meant more to show the viability of the proposed
approach than its superiority to IRL in these domains.
While the given problem assumes that the apprentice passively observes the demonstrator’s trajectories, future work
could consider an active learning approach, in which the
apprentice (for example) poses new MDPs involving the
same predicates (or perturbs the given MDP), and ‘asks’ the
demonstrator to generate trajectories in the posed MDPs.
VII. C ONCLUSION
In this paper, we introduced the problem of inferring linear
temporal logic (LTL) specifications from agent behavior in
Markov Decision Processes as a road to interpretable apprenticeship learning, combining the representational power and
interpretability of temporal logic with the generalizability of
inverse reinforcement learning. We formulated this as a twoobjective optimization problem, and introduced objective
functions using a notion of “violation cost” to quantify the
ability of an LTL formula to explain demonstrated behavior.
We presented results using genetic programming to solve this
problem in the SlimChance and CleaningWorld domains.
VIII. ACKNOWLEDGEMENTS
This project was in part supported by ONR grant N0001416-1-2278.
R EFERENCES
[1] Pieter Abbeel and Andrew Y Ng. Apprenticeship learning via
inverse reinforcement learning. Proc. 21st International Conference
on Machine Learning (ICML), pages 1–8, 2004.
[2] Thomas Arnold, Daniel Kasenberg, and Matthias Scheutz. Value
alignment or misalignment–what will keep systems accountable? In
3rd International Workshop on AI, Ethics, and Society, 2017.
[3] Christel Baier and Joost-Pieter Katoen. Principles of Model Checking.
The MIT Press, 2008.
[4] Daniil Chivilikhin, Ilya Ivanov, and Anatoly Shalyto. Inferring
Temporal Properties of Finite-State Machine Models with Genetic
Programming. In Proc. 2015 Annual Conference on Genetic and
Evolutionary Computation, pages 1185–1188, 2015.
[5] Xu Chu Ding, Stephen L. Smith, Calin Belta, and Daniela Rus.
LTL control in uncertain environments with probabilistic satisfaction
guarantees. In Proceedings - IFAC World Congress, volume 18, pages
3515–3520, 2011.
[6] Juraj Dzifcak, Matthias Scheutz, Chitta Baral, and Paul Schermerhorn.
What to do and how to do it: Translating natural language directives
into temporal and dynamic logic representation for goal management
and action execution. In Proceedings - IEEE International Conference
on Robotics and Automation, pages 4163–4168, 2009.
[7] Javier Esparza and Jan Ketı́nský. From LTL to deterministic automata:
A safraless compositional approach. In Lecture Notes in Computer
Science (including subseries Lecture Notes in Artificial Intelligence
and Lecture Notes in Bioinformatics), volume 8559, pages 192–208.
Springer International Publishing, 2014.
[8] Jie Fu and Ufuk Topcu. Probably Approximately Correct MDP
Learning and Control With Temporal Logic Constraints. In Robotics:
Science and Systems X, 2014.
[9] Mark Gabel and Zhendong Su. Symbolic mining of temporal specifications. In Proc. 30th International Conference on Software Engineering,
ICSE ’08, pages 51–60, New York, NY, USA, 2008. ACM.
[10] Mark Gabel and Zhendong Su. Online inference and enforcement
of temporal properties. In Proceedings of the 32Nd ACM/IEEE
International Conference on Software Engineering - Volume 1, ICSE
’10, pages 15–24, New York, NY, USA, 2010. ACM.
[11] M. Guo and D. V. Dimarogonas. Multi-agent plan reconfiguration
under local LTL specifications. The International Journal of Robotics
Research, 34(2):218–235, 2014.
[12] David Hadka. Moea framework: a free and open source java framework for multiobjective optimization, 2012.
[13] Zhaodan Kong, Austin Jones, Ana Medina Ayala, Ebru Aydin Gol,
and Calin Belta. Temporal Logic Inference for Classification and Prediction from Data. Proceedings of the 17th International Conference
on Hybrid Systems: Computation and Control, pages 273–282, 2014.
[14] John R Koza. Genetic programming: on the programming of computers by means of natural selection, volume 1. MIT press, 1992.
[15] Morteza Lahijanian, Shaull Almagor, Dror Fried, Lydia E Kavraki,
and Moshe Y Vardi. This Time the Robot Settles for a Cost:
A Quantitative Approach to Temporal Logic Planning with Partial
Satisfaction. In Proceedings of the AAAI Conference on Artificial
Intelligence, volume 29, pages 3664–3671, 2015.
[16] Kevin Leahy, Austin Jones, Mac Schwager, and Calin Belta. Distributed Information Gathering Policies under Temporal Logic Constraints. In IEEE Conference on Decision and Control (CDC),
volume 54, pages 6803–6808, 2015.
[17] Caroline Lemieux, Dennis Park, and Ivan Beschastnikh. General ltl
specification mining. In Automated Software Engineering (ASE), 30th
IEEE/ACM International Conference on, pages 81–92. IEEE, 2015.
[18] James MacGlashan. Brown-UMBC Reinforcement Learning and
Planning (BURLAP), 2016.
[19] Andrew Ng and Stuart Russell. Algorithms for inverse reinforcement
learning. In Proc. Seventeenth International Conference on Machine
Learning, volume 0, pages 663–670, 2000.
[20] Amir Pnueli. The temporal logic of programs. In 18th Annual
Symposium on Foundations of Computer Science, pages 46–57, 1977.
[21] Luis I. Reyes Castro, Pratik Chaudhari, Jana Tümová, Sertac Karaman, Emilio Frazzoli, and Daniela Rus. Incremental sampling-based
algorithm for minimum-violation motion planning. In Proc. IEEE
Conference on Decision and Control, pages 3217–3224, 2013.
[22] Rangoli Sharan and Joel Burdick. Finite state control of POMDPs
with LTL specifications. In Proceedings of the American Control
Conference, pages 501–508, 2014.
[23] Mária Svoreňová, Martin Chmelı́k, Kevin Leahy, Hasan Ferit Eniser,
Krishnendu Chatterjee, Ivana Černá, and Calin Belta. Temporal logic
motion planning using POMDPs with parity objectives. In Proceedings
of the 18th International Conference on Hybrid Systems Computation
and Control, pages 233–238, 2015.
[24] Jana Tumova, Gavin C Hall, Sertac Karaman, Emilio Frazzoli, and
Daniela Rus. Least-violating control strategy synthesis with safety
rules. In Proceedings of the 16th International Conference on Hybrid
Systems: Computation and Control, pages 1–10, 2013.
[25] Eric M. Wolff, Ufuk Topcu, and Richard M. Murray. Robust
control of uncertain Markov Decision Processes with temporal logic
specifications. In IEEE Conference on Decision and Control (CDC),
volume 51, pages 3372–3379, 2012.
| 3cs.SY
|
Sensitivity analysis based on Cramér von Mises distance
Fabrice Gamboa∗
Thierry Klein∗†
Agnès Lagnoux∗
December 1, 2017
arXiv:1506.04133v2 [math.PR] 30 Nov 2017
Abstract
In this paper, we first study a sensitivity index that is based on higher moments and generalizes
the so-called Sobol one. Further, following an idea of Borgonovo (see [3]), we define and study a new
sensitivity index based on the Cramér von Mises distance. This index appears to be more general
than the Sobol one as it takes into account the whole distribution of the random variable and not
only the variance. Furthermore, we study the statistical properties of its Pick and Freeze estimator.
Keywords: Sensitivity analysis, Cramér von Mises distance, Pick and Freeze method, functional delta-method.
1
Introduction
A very classical problem in the study of computer code experiments (see [26]) is the evaluation of the
relative influence of the input variables on some numerical result obtained by a computer code. In this
context, a sensitivity analysis is performed. Such a topic has been widely studied in the last decades and
is still challenging nowadays (see for example [27, 25, 13] and references therein). More precisely, the
result of the numerical code Y is seen as a function of the vector of the distributed input (Xi )i=1,··· ,d
(d ∈ N∗ ). Statistically speaking, we are dealing here with the following unnoisy non parametric model
Y = f (X1 , . . . , Xd ),
where f is a regular unknown numerical function on the state space E1 × E2 × . . . × Ed on which the
distributed variables (X1 , . . . , Xd ) are living. Generally, the random inputs are assumed to be independent
and a sensitivity analysis is performed using the so-called Hoeffding decomposition (see [29, 1]). In this
functional decomposition, f is expanded as an L2 -sum of uncorrelated functions involving only a part of
the random inputs. This leads, for any subset v of Id := {1, . . . , d}, to an index called the Sobol index
([27]) that measures the amount of randomness (more precisely, the part of the variance) of Y due to the
subset of input variables (Xi )i∈v . Since nothing has been assumed on the nature of the inputs, one can
consider the vector (Xi )i∈v as a single input. Without loss of generality, we then consider the case where
v reduces to a singleton. More precisely, the numerator Hv2 of the Sobol index related to the input Xv is
Hv2 := Var (E [Y |Xv ])
while the denominator of the index is nothing more than the variance of Y . Notice that we also have:
h
i
h
i
2
2
Hv2 = E (E[Y |Xv ] − E[Y ]) = Var(Y ) − E (E[Y ] − E [Y |Xv ])
(1)
In order to estimate Hv2 , Sobol in [27] proposed to rewrite the variance of the conditional expectation as
a covariance (see equation (3)). Further, a well tailored design of experiment called the Pick and Freeze
scheme is considered [19]. More precisely, let X v be the random vector such that Xvv = Xv and Xiv = Xi0
if i 6= v where Xi0 is an independent copy of Xi . Then, setting
Y v := f (X v ),
∗ Institut de Mathématiques de Toulouse, 118 Route de Narbonne 31062 Toulouse Cedex 9.
[email protected]
† ENAC - Ecole Nationale de l’Aviation Civile , Université de Toulouse, France
1
(2)
France.
an obvious computation leads to the following relationship (see, e.g., [19])
Var(E[Y |Xv ]) = Cov (Y, Y v ) .
(3)
The last equality leads to a natural Monte-Carlo estimator, the so-called Pick and Freeze estimator,
2
N
N
X
X
1
1
v
TN,Cl
=
Yj Yjv −
(Yj + Yjv )
N j=1
2N j=1
where for j = 1, · · · , N , Yj (resp. Yjv ) are independent copies of Y (resp. Y v ). The sharp statistical
properties and some functional extensions of the Pick and Freeze method are considered in [19, 18, 12].
Notice that the Sobol indices and their Monte-Carlo estimation are order two methods since they derive
from the L2 -Hoeffding functional decomposition. This is their main drawback. As an illustration consider
the following example. Let X1 and X2 be two independent
standardized random variables having the
same third and fourth moments with E X15 6= E X25 . Let us consider the following model
Y = X1 + X2 + X12 X22 .
One gets
Var (E [Y |X1 ]) = Var(X1 + X12 ) = Var(X2 + X22 ) = Var (E [Y |X2 ]) .
Y is an exchangeable function of the inputs but X1 and X2 do not share the same distribution. So
that, X1 and X2 should not have the same importance. That shows the need to introduce a sensitivity
index that takes into account all the distribution and not only the second order behavior. As pointed out
before, Sobol indices are based on an L2 decomposition. As a matter of fact, they are well adapted to
measure the contribution of an input on the deviation around the mean of Y . Nevertheless, it seems very
intuitive that the sensitivity of an extreme quantile of Y could depend on sets of variables that cannot
be captured using only the variances. Thus the same index should not be used for any task and we need
to define more general indices.
There are several ways to generalize the Sobol indices. For example, one can define new indices through
contrast functions based on the quantity of interest (see [16]). Unfortunately the Monte-Carlo estimators
of these indices are computationally very expensive. In [11], Da Veiga presents a way to define moment
independent measures through dissimilarity distances. These measures define a unified framework that
encompasses some known sensitivity indices. They are efficiently estimated in low dimensions but as
claimed by the author “it is well known that density estimation suffers from the curse of dimensionality”.
Now, as pointed out in [3, 5, 6, 24, 23], there are situations where higher order methods give a sharper
analysis on the relative influence of the input and allow finer screening procedures. Borgonovo et al.
propose and study an index based on the total variation distance (see [3, 5, 6]); while Owen et al. suggest
to use procedures based on higher moments (see [24, 23]).
Our paper follows these tracks. We will first revisit the work of Owen et al. by studying the asymptotic
properties of the multiple Pick and Freeze estimation. Further, we propose a new natural index based on
the Cramér von Mises distance between the distribution of the output Y and its conditional law when
an input is fixed. We will show that this approach leads to natural self-normalized indices. Indeed,
as for Sobol indices, the sum of all first order indices is uniformly bounded. Notice that these indices
take into account the whole output distribution instead of only the order two moments and contrary to
most of the other known indices, they are naturally defined for multivariate outputs. As a consequence,
they are well-tailored to perform a sensitivity analysis for any vectorial output. Furthermore, we show
that surprisingly a Pick and Freeze scheme is also available to estimate this new index. This scheme is
not really expensive and easy to implement. The sample size required for the estimation is of the same
order as the size needed for the classical Sobol index estimation allowing its use in concrete situations.
As a consequence, considering a sample with the appropriate size, one can provide simultaneously the
Cramér von Mises indices and the Sobol indices. Other advantage of the Cramér von Mises index with
respect to the general ones presented in [11] is that the theoretical statistical properties of its estimation
can be derived. Indeed, we prove a Central Limit Theorem for the estimator that allows one to exhibit
confidence intervals.
The paper is organized as follows. In the next section, we will study the statistical properties of the
multiple Pick and Freeze method proposed earlier by Owen et al ([24, 23]). Section 3 is devoted to the
new index built on the Cramér von Mises distance. In the last section, we give some numerical simulations
that illustrate the interest of the new index. Moreover, we revisit a real data example introduced in [10]
and studied in [15, 7].
2
2
Higher-moment indices
In the sequel, for any integer k, the notation Ik stands for the set {1, . . . , k}. Following [24, 23], we
generalize the numerator of the Sobol index defined in (1) by considering higher order moments: for any
integer q > 2, and singleton v ∈ Id , we set
q
Hvq := E [(E[Y |Xv ] − E[Y ]) ] .
Properties Obviously, Hvq is non negative only for even q and
q
|Hvq | 6 E [|Y − E[Y ]| ] .
Further, Hvq is invariant by any translation of the output.
Estimation procedure In view of the estimation of Hvq , we first notice that
"
Hvq
=E
q
Y
#
Y
v,i
− E[Y ]
i=1
" l
#
q
X
Y
q
q−l
q−l
v,i
=
(−1) E [Y ] E
Y
l
i=1
l=0
Q0
q
with the usual convention i=1 Y v,i = 1 and l = q!/l!(q − l)!. Here, Y v,1 = Y and for i = 2, . . . , q, Y v,i
is constructed independently as Y v defined in Equation (2).
Second, we use
scheme and consider the following Pick and Freeze design constituted by
a Monte-Carlo
v,i
a N -sample Yj
of Y v,1 , . . . , Y v,q . The Monte-Carlo estimator is then
(i,j)∈Iq ×IN
v
Hq,N
=
q
v q−l v
X
q
(−1)q−l P 1
Pl
l
l=0
where for any N ∈ N∗ , j ∈ IN and l ∈ Iq , we have defined
v
Pl,j
−1
q
=
l
X
l
Y
k1 <...<kl ∈Iq
i=1
!
Yjv,ki
and
v
Pl
N
1 X v
P .
=
N j=1 l,j
Notice that we generalize the estimation procedure of [18] and use all the available information by
considering the means over the set of indices k1 , . . . , kl ∈ Id , kn 6= km .
v
Asymptotic properties of Hq,N
v
Theorem 2.1. Hq,N
is strongly consistent and asymptotically Gaussian:
√
v
N Hq,N
− Hqv
L
→ N 0, σ 2
N →∞
where
q
X
σ = q Var(Y ) + (q − 1)Cov(Y, Y v,2 )
al bl
2
!2
,
l=1
al =
b1 = (−1)
q−1
l
E[Y ]l−1 ,
q
q−1
q(q − 1)E[Y ]
+
l = 1, . . . , q
q−1
X
q
l=2
l
"
(−1)
q−l
q−l−1
(q − l)E[Y ]
i=1
and
q
bl =
(−1)q−l E[Y ]q−l ,
l
3
E
l
Y
l = 1, . . . , q.
#
Y
v,i
Interpretation and comments The collection of all indices (Hvq )q is much more informative than the
classical Sobol index with respect to v. Nevertheless, it has several drawbacks. First, these indices are
moment-based and it is well known that they are not stable when the moment order increases. Second,
q
they may be negative when q is odd. To overcome this fact, one may introduce E [|E[Y |Xv ] − E[Y ]| ] but
the Pick and Freeze estimation procedure is then lost. Third, the Pick and Freeze estimation procedure
is computationally expensive and may be unstable: it requires a q × N -sample of the output Y . In order
to have a good idea of the influence of an input on the law of the output, we need to estimate the first
K − 1 indices Hvq : Hv2 , . . . , HvK . Hence, we need to run the code K × N times.
In a nutshell, these indices are not attractive in a practical point of view. In the next section, we then
introduce a new sensitivity index that is based on the conditional distribution of the output and requires
only 3 × N runs. Concretely, it compares the output distribution to the conditional one whereas the q
higher-order moment indices only compare the q-th output moment to the conditional one.
3
The Cramér von Mises index
In this section, the code will be denoted by Z = f (X1 , . . . , Xd ) ∈ Rk . It is worth noticing that here we
can consider multivariate outputs unlike in Section 2 and [7], e.g., Let F be the distribution function of
Z:
F (t) = P (Z 6 t) = E 1 {Z6t} , for t = (t1 , . . . , tk ) ∈ Rk
and F v be the conditional distribution function of Z conditionally on Xv :
F v (t) = P (Z 6 t|Xv ) = E 1 {Z6t} |Xv , for t = (t1 , . . . , tk ) ∈ Rk .
Notice that {Z 6 t} means that {Z1 6 t1 , . . . , Zk 6 tk }. Obviously, E [F v (t)] = F (t). Now, we define
Y (t) = 1 {Z6t} and take p = 2. Since for any fixed t ∈ Rk , Y (t) is a real-valued random variable, we
apply the framework presented in Section 2. More precisely, for any v ∈ Ip let ∼ v be Ip \ {v} and we
first perform the Hoeffding decomposition of Y (t):
Y (t) = 1 {Z6t} = E[Y (t)] + (E[Y (t)|Xv ] − E[Y (t)]) + (E[Y (t)|X∼v ] − E[Y (t)]) + R(t, v)
(4)
where
R(t, v) = Y (t) − E[Y (t)] − (E[Y (t)|Xv ] − E[Y (t)]) − (E[Y (t)|X∼v ] − E[Y (t)]) .
As done usually, we compute the variance of both sides of (4) that leads to
Var(Y (t)) = F (t)(1 − F (t))
= Var (E[Y (t)|Xv ] − E[Y (t)]) + Var (E[Y (t)|X∼v ] − E[Y (t)]) + Var(R(t, v))
= Var (F v (t)) + Var (F ∼v (t)) + Var(R(t, v))
h
i
h
i
2
2
= E (F v (t) − F (t)) + E (F ∼v (t) − F (t)) + Var(R(t, v))
(5)
by the decorrelation of the different terms involved in the Hoeffding decomposition.
Remark 3.1. A straightforward application of the results of Section 2 provides for any fixed t ∈ Rk a
consistent and asymptotically normal procedure for the estimation of
h
i
h
i
2
2
E (F v (t) − F (t)) = Var (F v (t)) and E (F ∼v (t) − F (t)) = Var (F ∼v (t)) .
Now we integrate the terms in (5) in t ∈ Rk with respect to the distribution of Z:
Z
F (t)(1 − F (t))dF (t)
k
ZR h
Z
Z
i
h
i
2
2
v
∼v
=
E (F (t) − F (t)) dF (t) +
E (F (t) − F (t)) dF (t) +
Var(R(t, v))dF (t)
Rk
Rk
(6)
Rk
This integration has to be understood in the Riemmann-Stieltjes sense (see, e.g., [28]). Notice that the
first term in the right hand side of (6) represents a Cramér Von Mises type distance of order 2 between
4
the distribution L (Z) of Z and the distribution L (Z|Xv ) of Z given Xv .
Following the classical way of defining Sobol indices, we normalize the previous equation by
Z
F (t)(1 − F (t))dF (t)
Rk
leading to
R
1=
h
i
h
i
R
R
2
2
∼v
E (F v (t) − F (t)) dF (t)
E
(F
(t)
−
F
(t))
dF (t)
k
R
k Var(R(t, v))dF (t)
R
R
+
+R R
F (t)(1 − F (t))dF (t)
F (t)(1 − F (t))dF (t)
F (t)(1 − F (t))dF (t)
Rk
Rk
Rk
Rk
(7)
Then we define the Cramér Von Mises indices with respect to v and ∼ v by
h
i
h
i
R
R
2
2
v
∼v
dF (t)
(t)) dF (t)
k E (F (t) − F (t))
k E (F (t) − F
R
R
v
∼v
R
R
S2,CV
and S2,CV
.
M :=
M :=
F
(t)(1
−
F
(t))dF
(t)
F (t)(1 − F (t))dF (t)
k
R
Rk
Properties 3.2. These new indices are naturally adapted to multivariate outputs and they share the
same properties as the classical Sobol index. Namely,
1. as seen in (7), the different contributions sum to 1.
2. they are invariant by translation, by any isometry and by any non degenerated scaling of the components of Y .
Remark 3.3.
1. We could have defined the following indices instead
h
i
h
i
Z E (F (t) − F ∼v (t))2
Z E (F (t) − F v (t))2
dF (t) and
dF (t).
F (t)(1 − F (t))
F (t)(1 − F (t))
Rk
Rk
normalizing by F (t)(1 − F (t)) (like in the Anderson-Darling statistic) before the integration phase.
Nevertheless, the previous integrals might be not defined. Moreover, even if the integrals are well
defined, one may encounter numerical explosion during the estimation procedure that might be
produced for small and large values of t since the normalizing factor then cancels.
2. In this paper, we only consider first-order sensitivity indices as well for the classical Sobol indices
and for the Cramér von Mises indices. Anyway, as well as for the Sobol indices, one may define
higher-order and total Cramér von Mises indices. The construction of the former is straightforward
taking v no longer a singleton. For example, if one is interested in the second-order Cramer von
Mises index with respect to the first and second inputs, it suffices to take v = {1, 2}. Concerning
T ot,v
the latter, the total Cramér von Mises index S2,CV
M with repect to v is defined by
h
i
2
∼v
E
(F
(t)
−
F
(t))
dF (t)
Rk
R
=1−
.
F (t)(1 − F (t))dF (t)
Rk
R
T ot,v
∼v
S2,CV
M := 1 − S2,CV M
3. To use the Hoeffding decomposition, the inputs are required to be independent. Anyway, one can
compute the Cramér von Mises index when the inputs are dependent. Nevertheless, there are then
difficult to interpret.
3.1
General comments on the Cramér von Mises indices
Cramér von Mises indices versus Sobol indices
Cramér von Mises and Sobol indices are both based on the Hoeffding decomposition and sum to 1. Nevertheless, the former are based on the whole distribution of the output, in contrast with the latter that
are only based on the order-two moments. Notice that two variables that have a different influence on
the output may have the same Sobol indices (just as two random variables with different distribution can
have the same variance). This point represents one limitation of Sobol indices and does not occur with
5
the Cramér von Mises indices as one can see in Section 4.1.
In addition, remark that a null value for a Sobol index does not imply that the input is unimportant
whereas a null value for a Cramér von Mises index means that the input is unimportant. Moreover,
by definition, a large Cramér von Mises index means that the input variable under concern has a great
influence on the output in regions where the output has a large distribution mass. That is why we advice
the practitioner to use them in a general context. Nevertheless, when one is interested in the mean output
behavior, the Sobol indices are more adapted. Indeed, as noted in [16], the Sobol indices minimize the
contrast associated to the mean. In the same spirit, if one is interested in specific feature of the output
(for example an α-quantile), one should use the index based on the associated contrast. See [16] for more
details on the notion of contrast and the results therein.
In contrast, the indices based on the whole distribution partially get rid of such limitations and pathological patterns. However, one can build an example based, e.g., on two input variables that leads to the
1
2
same indices S2,CV
M and S2,CV M once the integration with respect to t has been done.
Cramér von Mises indices versus moment independent indices
There already exists several moment-independent indices: some of them have been introduced by Borgonovo et al. (density-based indices [5], cumulative distribution function based indices [9]). See also [4]
for other indices and references therein. More recently, Da Veiga [11] shows that those indices are special
cases of a class of sensitivity indices based on the Csizár f -divergence. A lot of classical “distances” between probability measures as, e.g., the Kullback-Leibler divergence, the Hellinger distance and the total
variation distance belong to this family of divergences. Other dissimilarity measures exist to compare
probability distributions: in particular, integral probability metrics [20].
In comparison with the indices defined in Equation (17) in [8], we can notice that the integration is done
with respect to the distribution of the output in the former indices while the integration is done with
respect to the Lebesgue measure in the latter indices. Our method represents at least two advantages:
(i) the index always exist whatever the output distribution (ii) such an integration weights the support
of the output distribution.
Since the space of the probability measures on Rk is of infinite dimension, the different distances on this
space are not equivalent; hence they are very difficult to compare. Each index is constructed on a specific
distance and has its own interest. Despite the fact that the Cramér von Mises indices have no clear
dual formulation, they present the following remarkable advantages. As we will see in the next sections,
one can easily estimate them with a low simulation cost that does not depend on the dimension of the
output. The sample required for their estimation also provide Sobol√indicies estimation. In addition,
these estimators are asymptotically normal and converge at the rate N which allows the practitioner
to build confidence intervals.
v
∼v
The rest of the section is dedicated to the estimation of S2,CV
M (and S2,CV M ). One has to estimate both
the numerator and the denominator of the indices. Nevertheless, when the output Z has independent
coordinates that are absolutely continuous with respect to the Lebesgue measure, we have
Z
F (t)(1 − F (t))dF (t) = E[F (Z)(1 − F (Z)] =
Rk
1
1
− k.
2k
3
Thus the normalizing factor reduces to 21k − 31k . As a consequence, we propose two versions of Central
Limit Theorems: the first one deals with the numerator’s estimator and can be applied when the output
Z has independent coordinates that are absolutely continuous with respect to the Lebesgue measure
whereas the second one concerns the general estimator and may apply in any other cases.
6
3.2
Numerator estimation and its asymptotic properties
v
v
We denote the numerator of S2,CV
M by N2,CV M . Notice that it can be rewritten as
2
v
v
N2,CV
=
E
E
F
(
Z̃)
−
F
(
Z̃)
Xv
M
Z̃
where Z̃ is an independent copy of Z.
v
Then we proceed to a double Monte-Carlo scheme for the estimation of N2,CV
M and consider the following
design of experiment consisting in:
1. The classical Pick and Freeze sample, that is two N -samples of Z: (Zjv,1 , Zjv,2 ), 1 6 j 6 N ;
2. A third N -sample of Z independent of (Zjv,1 , Zjv,2 )16j6N : Wk , 1 6 k 6 N .
v
The empirical estimator of N2,CV
M is then given by
2
N
N
N
1 X 1 X
1 X
v
b2,CV
v,1
v,2
v,1
v,2
N
=
1
1
−
1
+
1
.
M
{Zj 6Wk } {Zj 6Wk }
{Zj 6Wk }
N
N
2N j=1 {Zj 6Wk }
j=1
k=1
(8)
bv
Now we established the consistency of N
2,CV M that follows directly from an auxiliary lemma (see Section
6).
bv
Corollary 3.4. N
2,CV M is strongly consistent as N goes to infinity.
bv
Now we turn to the asymptotic normality of N
2,CV M . We follow van der Vaart [29] to establish the
following proposition (more precisely Theorems 20.8 and 20.9, Lemma 20.10 and Example 20.11).
v
bv
Theorem 3.5. The sequence of estimators N
2,CV M is asymptotically Gaussian in estimating N2,CV M .
√ v
v
b
That is, N N
converges in distribution towards the centered Gaussian law with a
2,CV M − N2,CV M
limiting variance ξ 2 whose explicit expression can be found in the proof.
Remark 3.6. Thanks to Theorem 3.5, we are now able to provide
√ asymptotic confidence intervals for
v
bv
the estimation of N2,CV
M . They are of the form (N2,CV M ± zα ξ/ N ), where zα is the 1 − α/2 quantile
of a standard normal distribution. Unfortunately, the variance ξ 2 is unknown but thanks to its explicit
form it is easy to replace it by a consistent estimator ξb and use Slutsky’s Lemma to have an asymptotic
confidence interval.
3.3
Estimation of the general index and its asymptotic properties
v
In order to estimate the general index S2,CV
M , we first estimate its numerator as in Subsection 3.2 and
v
then its denominator that we denote D2,CV M . Notice that it can be rewritten as
v
D2,CV
M = E [F (Z)(1 − F (Z))]
and estimated using the design of experiment already introduced for the estimation of the numerator by
2
N
N
N
X
X
X
1
1
1
v
b
D2,CV M =
1 {Z v,1 6Wk } + 1 {Z v,2 6Wk } −
1 {Z v,1 6Wk } + 1 {Z v,2 6Wk }
.
j
j
j
j
N
2N j=1
2N j=1
k=1
(9)
Proceeding as in Subsection 3.2, we have
v
Corollary 3.7. Sb2,CV
M is strongly consistent as N goes to infinity.
The following Central Limit Theorem comes from the functional Delta method.
v
v
Theorem 3.8. The sequence of estimators Sb2,CV
M is asymptotically Gaussian in estimating S2,CV M .
√ v
v
converges in distribution towards the centered Gaussian law with a
That is, N Sb2,CV M − S2,CV
M
limiting variance that can be computed.
7
3.4
Practical advices
In a general setting, for all the nice properties of the Cramér von Mises indices and their efficient estimation
easy to implement, we recommend to use the Cramér von Mises indices. As a consequence, considering
a sample with the appropriate size, one can estimate once at a time the Cramér von Mises indices and
the Sobol indices. More precisely, if one wants to estimate p Sobol indices a sample size of (p + 1)N
is required. With only N more output evaluations, we get both the p Sobol indices and the Cramér
von Mises ones. Furthermore, the theoretical theorems provides confidence intervals that controlled the
accuracy of the estimations. Anyway, when the practitioner is interested in a specific feature (e.g., mean
behavior or quantile) of the output, he should use more suited indices (e.g., the classical Sobol indices
for the mean or the indices introduced in [16] for the quantile).
4
4.1
Numerical applications
A flavor of the method applied on a toy model
Let us consider the quite simple linear model
Y = αX1 + X2 , α > 0,
where X1 has a Bernoulli distribution with success probability 0 < p < 1 and X1 , X2 are independent.
Assume further that X2 has a continuous distribution F2 on R such that E[X2 ] = αp and with finite
variance Var(X2 ) = α2 p(1 − p). With these choices, the random variables αX1 and X2 share the same
expectation and the same variance. Thus X1 and X2 have the same first order Sobol indices all equal to
1/2.
We present a general closed formula to compute our new indices and show that in some particular cases
an exact formula is available. Then we perform a simulation study in order to illustrate the Central Limit
Theorem and analyse the practical behaviour of our estimators.
4.1.1
General closed formula
On one hand, the distribution of Y given X1 = 0 and the distribution of Y given X1 = 1 are given by
(
L(Y |X1 = 0) = L(X2 )
L(Y |X1 = 1) = L(X2 + α).
On the other hand, the conditional distribution of Y given X2 is
P (Y = α + X2 | X2 ) = 1 − P (Y = X2 |X2 ) = p.
Hence, the distribution function of Y is the mixture pF2 (· − α) + (1 − p)F2 (·). Tedious computations lead
to
Z
1
S2,CV
=
6p(1
−
p)
(F2 (t) − F2 (t − α))2 [(1 − p)dF2 (t) + pdF2 (t − α)]
M
R
and
2
S2,CV
M
1
−
= 1 − 6p(1 − p)
2
Z
F2 (t − α)dF2 (t)
R
(the normalizing factor being 1/6 as explained before).
1
2
As p goes to 0 (and α goes to infinity), (S2,CV
M , S2,CV M ) goes to (0, 1) while the two classical Sobol
indices remain equal to 1/2. Our new indices shed lights on the fact that, for small p, X2 has much more
influence on Y than X1 which follows the intuition. This fact is not detected by the classical Sobol indices.
Similarly we can compute the indices of order q (q > 2):
H1q = αq [p(1 − p)q + (−p)q (1 − p)]
8
and H2q = E[(X2 − µ)q ].
Particular cases (i) if X2 is a centered Gaussian random variable with variance Var(X2 ) = α2 p(1 − p),
one can easily derive an explicit formula for H2q :
q!
1q∈2N∗ .
· (q/2)!
p
(ii) if X2 is uniformly distributed on [0, b] with b = 2α 3p(1 − p), one can easily derive an explicit
formula for the indices introduced before:
2α
1
α 2
1
1−
S2,CV M = 6p(1 − p) ×
1α6b + 1α>b
b
3b
3
!
2
b−α
2
S2,CV
1α6b .
M = 1 − 3p(1 − p) 1 −
b
H2q = E[(X2 − m)q ] =
2q/2
Moreover, H2q = E[(X2 − µ)q ] = (b/2)q /(q + 1)1q∈2N∗ .
(iii) if X2 is exponentially distributed with mean 1/λ = α
formula for the indices introduced before:
1
−λα 2
S2,CV
)
M = 2p(1 − p)(1 − e
p
p(1 − p), one can easily derive an explicit
2
−λα
and S2,CV
).
M = 1 − 3p(1 − p)(1 − e
Moreover, H2q = E[(X2 − µ)q ] = q!λ−q /2.
4.1.2
Simulation study
A numerical illustration with sample sizes N =100 and 500 is presented in Figures 1 and 2 (remind that in
order to estimate both indices we compute 4N values of the output function). We consider the case where
the random variable X2 is uniformly distributed (for the other cases the simulations provide the same
kind of results). We estimate the Cramér von Mises indices thanks to Equation (8) and renormalize it
by the factor 1/6 since the output has a continuous distribution. Then we estimate the limiting variance
in (12) in order to provide asymptotic confidence intervals. In Figures 1 and 2, the blue line represents
2
1
the true value of index D2,CV
M (first row) or D2,CV M (second row). The red dashed line (resp. the red
dashed line with +) represents the index estimation based on (8) (resp. the confidence interval). In the
left column, α is fixed to 1/2 and p varies while in the right one, p is fixed to 1/4 and α varies.
4.2
A non linear model
Now, let us consider the following non linear model
Y = exp{X1 + 2X2 },
(10)
where X1 and X2 are independent standard Gaussian random variables. The distribution of Y is lognormal and we can derive both its density and its distribution functions:
2
1
ln y
fY (y) = √
e−(ln y) /10 1 R+ (y) and FY (y) = Φ √
10πy
5
where Φ stands for the distribution function of the standard Gaussian random variable. Its density
function will be denoted by f in the sequel. Then tedious computations lead to the Cramér von Mises
1
2
indices S2,CV
M and S2,CV M .
Proposition 4.1. Assume that Y is defined by Equation (10) then
1
S2,CV
M =
6
arctan 2 − 2 ≈ 0.1145
π
and
2
S2,CV
M =
√
6
arctan 19 − 2 ≈ 0.5693.
π
9
Uniform distribution and N=100
0.8
0.5
0.7
0.4
0.6
0.3
0.5
First index
First index
Uniform distribution and N=100
0.6
0.2
0.1
0.4
0.3
0
0.2
−0.1
0.1
−0.2
0.1
0.2
0.3
0.4
0.5
0.6
Value of p for fixed α=1/2
0.7
0.8
0
0.2
0.9
0.4
0.6
Uniform distribution and N=100
0.9
1.2
0.8
1.8
2
1.6
1.8
2
0.7
Second index
1
Second index
1.6
Uniform distribution and N=100
1.4
0.8
0.6
0.4
0.6
0.5
0.4
0.3
0.2
0
0.1
0.8
1
1.2
1.4
Value of α for fixed p=1/4
0.2
0.2
0.3
0.4
0.5
0.6
Value of p for fixed α=1/2
0.7
0.8
0.9
0.1
0.2
0.4
0.6
0.8
1
1.2
1.4
Value of α for fixed p=1/4
Figure 1: Example 1 - X2 uniformly distributed and N =100.
10
Uniform distribution and N=500
Uniform distribution and N=500
0.4
0.4
0.35
0.35
First index
First index
0.3
0.25
0.2
0.3
0.25
0.15
0.2
0.1
0.05
0.1
0.2
0.3
0.4
0.5
0.6
Value of p for fixed α=1/2
0.7
0.8
0.9
0.2
0.4
0.6
0.65
0.9
0.6
0.8
0.55
0.7
0.6
0.4
0.35
0.3
0.4
0.5
0.6
Value of p for fixed α=1/2
2
1.6
1.8
2
0.45
0.4
0.2
1.8
0.5
0.5
0.1
1.6
Uniform distribution and N=500
1
Second index
Second index
Uniform distribution and N=500
0.8
1
1.2
1.4
Value of α for fixed p=1/4
0.7
0.8
0.9
0.2
0.4
0.6
0.8
1
1.2
1.4
Value of α for fixed p=1/4
Figure 2: Example 1 - X2 uniformly distributed and N =500.
11
Remark 4.2. In this simple example, one can compute the indices of order q (q > 2):
i
h
i
h
and H2q = E (e2X1 +1/2 − e5/2 )q .
H1q = E (eX1 +2 − e5/2 )q
The Sobol indices and their estimation based on the Pick-Freeze scheme with a sample of size N are
computed using equation (6) in [19]. We also compute the Cramér von Mises indices and their estimation
based on (8). Moreover, we estimate the limiting variances in both cases (see equation (12) for the Cramér
von Mises indices and equation (12) in [19] for the Sobol indices) in order to provide confidence intervals.
The results are presented in Table 1.
N = 102
N = 103
N = 104
True values
Est. values
CI 5%
Est. values
CI 5%
Est. values
CI 5%
Cramér von Mises
1
2
D2,CV
D2,CV
M
M
0.1145
0.5693
0.1287
0.6097
[-0.0601,0.3175]
[0.4692,0.7503]
0.1358
0.6007
[0.07861,0.19297] [0.54897,0.65242]
0.1166
0.5585
[0.09930,0.13382] [0.54150,0.57540]
Sobol indices
S1
S2
0.0118
0.3738
0.0425
0.1954
[0.0265,0.0585]
[0.0430,0.3477]
0.1198
0.2345
[-0.5633,0.8030]
[0.1343,0.3347]
0.01685
0.26252
[0.0010,0.0327] [-1.2744, 1.7994]
Table 1: Model (10). The Cramér von Mises and Sobol indices, their estimations based on (8) and (6)
in [19] and the associated 5%-confidence intervals.
As a conclusion, with only N = 103 , the statistical method provides a precise estimation of the different
indices. Moreover, in this example, the Sobol and Cramér von Mises indices give the same influence
ranking of the two random inputs. Nevertheless, the estimation of the Cramér von Mises indices seems
to be more efficient to give the true ranking.
4.3
Application: The Giant Cell Arthritis Problem
Context and goal
In this subsection, we consider the realistic problem of management of suspected giant cell arthritis
posed by Bunchbinder and Detsky in [10]. More recently, this problem was also studied by Felli and
Hazen [15] and Borgonovo et al. [7]. As explained in [10], “giant cell arthritis (GCA) is a vasculitis of
unknown etiology that affects large and medium sized vessels and occurs almost exclusively in patients
50 years or older”. This disease may lead to severe side effects (loss of visual accuity, fever, headache,...)
whereas the risks of not treating it include the threat of blindness and major vessels occlusion. A patient
with suspected GCA can receive a therapy based on Prednisone. Unfortunately, a treatment with high
Prednisone doses may cause severe complications. Thus when confronted to a patient with suspected
GCA, the clinician must adopt a strategy. There is a considerable literature on sensitivity analysis for
these sorts of models, based on the utility of learning a model input before choosing a treatment strategy
(see, e.g., [14] and [22]). In [10], the authors considered four different strategies:
A : Treat none of the patients;
B : Proceed to the biopsy and treat all the positive patients;
C : Proceed to the biopsy and treat all the patients whatever their result;
D : Treat all the patients.
The clinician wants to adopt the strategy optimizing the patient outcomes measured in terms of utility.
The reader is referred to [21] for more details on the concept of utility. The basic idea is that a patient
with perfect health is assigned a utility of 1 and the expected utility of the other patients (not perfectly
healthy) is calculated subtracting some “disutilities” from this perfect score of 1. These strategies are
represented in Figures 3 to 6 with the different inputs involved in the computation of the utilities.
For example in strategy A (see Figure 3), the utility of a patient having GCA and developing severe GCA
complications is given by 1 − ds − dugc − dudx . His entire sub-path is then
g × gc × (1 − ds − dugc − dudx ).
12
Figure 3: The decision tree for the treat none alternative
Figure 4: The decision tree for the biopsy and the treat positive alternative
The input parameters and the modelisation of the random ones
As seen in Figures 3 to 6, the different strategies involve input parameters like, e.g., the proportion g of
patients having GCA or the probability gc for a patient to develop severe GCA complications (fixed at
0.8 as done in [10]) or even the disutility associated to having GCA symptoms. Table 2 summarizes the
input parameters involved.
The values P[·] and D(·) refer respectively to the probability of an event and to the disutility associated
with an event. The minimum and maximum values m and M depict each parameter’s range for the
sensitivity analysis. The base values are provided by a clinician expertise. The utilities of the different
strategies when all the input parameters are set to their base value are summarized in Table 3.
The base value of some input parameters are reliable while the others are really uncertain that leads us to
consider them as random. As a consequence, if YA , YB , YC and YD represent the outcomes corresponding
to the four different strategies A to D, the clinician aims to determine
max{E[YA ], E[YB ], E[YC ], E[YD ]}
(11)
with the uncertain model input presented in Table 2. A sensitivity analysis is then performed to determine the most influential input variables on the outcome.
As done in [15, 7], all the random inputs will be independently based on Beta distributions. The Beta
density parameters corresponding to each random input are determined by fitting the base value as their
mean and capturing 95% of the probability mass in the range defined by the minimum and maximum.
13
Figure 5: The decision tree for the biopsy and the treat all alternative
Figure 6: The decision tree for the treat all alternative
The remaining 5% will be equally distributed to either side of this range if possible. Concretely, each
random input will be distributed as
Z1 m6Z<M + U 1 m>Z + V 1 Z>M
where Z, U and V are independent random variables. Z is Beta distributed with parameters (α, β). U
and V are uniform random variables on [0, m] and [M, 1] respectively.
14
Fixed parameters
P[having GCA]
D(having symptoms of GCA)
D(having a temporal artery biopsy)
D(not knowing the true diagnosis)
Symbols
g
dus
dub
dudx
Fixed value
0.8
0.12
0.005
0.025
Uncertain parameters
P[developing severe complications of GCA]
P[developing severe iatrogenic side effects]
Efficacy of high dose Prednisone
Sensitivity of temporal artery biopsy
D(major complication from GCA)
D(Prednisone therapy)
D(major iatrogenic side effect)
Symbols
gc
pc
e
sens
dugc
dup
dupc
Base
0.3
0.2
0.9
0.83
0.8
0.08
0.3
–
–
–
–
–
–
–
–
Min. m
0.05
0.05
0.8
0.6
0.3
0.03
0.2
Max. M
0.5
0.5
1
1
0.9
0.2
0.9
–
–
–
–
–
–
–
–
Beta(α,β)
α
β
4.179
11.011
2.647
10.589
27.787 3.087
7.554
1.547
27.454 6.864
4.555
52.380
15.291 35.680
Table 2: The data used by Buchbinder and Detsky [10] in their analysis
Treatment alternative
A Treat none
B Biopsy and treat positive
C Biopsy and treat all
D Treat all
Utilility
0.6870
0.7575
0.7398
0.7198
Expectation
0.6991
0.7570
0.7371
0.7171
Table 3: The utilities of the different strategies when all the input parameters are set to their base value
(second column) and their expectation when they are random (third column).
Sensitivity analysis
As already mentioned, the clinician wants to determine the highest utility. In [4], the authors then consider the highest utility as output and lead a sensitivity analysis to determine the input having the largest
influence on this output. Since we are able to treat multivariate outputs, we consider a more general
framework in this paper: the output is the four-dimensional random variable Y = (YA , YB , YC , YD ) where
YS represents the outcome corresponding to strategy S.
We compare three different methodologies and indices. First, we consider the Sobol indices introduced in
[17] (Multivariate). Second, we consider the indices constructed in this paper, based on the Cramér von
Mises distance and estimated by the ratios of the numerator estimator (8) and the denominator estimator
(9). Third, we consider the index presented in [4] and named β defined by
βi = E[sup{|FY (y) − FY |Xi (y)|}].
y∈Y
Then we use the estimator given in [7, Table 1] adapted to the multivariate case that is based on the
tedious and costly estimation of conditional expectations.
Results
Table 4 summarizes the sensitivity measures of the seven random inputs on the multivariate output with
the three different methodologies while Table 5 presents the associated ranks. It is worth mentioning
that the same total sample size has been used to compare properly the three methodologies.
As a conclusion, in this example, unlike the indices defined by Borgonovo et al., the multivariate sensitivity
indices and the Cramér von Mises indices provide the same ranking. The main advantage of the Cramér
von Mises sensitivity methodology with respect to the one of Borgonovo et al. is that one can use the
Pick and Freeze estimation scheme which provides an accurate estimation (see (8)) simple to implement.
Notice that in [7], the authors study a slightly different model that explains the numerical differences
between their results and the ones of the present paper. Furthermore, they perform a sensitivity analysis
on the best alternative with the greater mean instead of considering the multivariate output.
15
N = 102
N = 103
N = 104
Sensitivity meas.
Multivariate
Borgonovo et al.
Cramér von Mises
Multivariate
Borgonovo et al.
Cramér von Mises
Multivariate
Borgonovo et al.
Cramér von Mises
1
0.3690
0.1195
0.3496
0.4024
0.1788
0.3494
0.3828
0.3842
0.3494
2
0.0193
0.1047
0.0745
0.1201
0.1192
0.0750
0.1333
0.1572
0.0775
3
0.0105
0.1064
0.0206
0.0516
0.1009
0.0209
0.0618
0.1033
0.0232
4
-0.0821
0.1022
-0.0010
-0.0190
0.1007
-0.0008
-0.0016
0.0930
0.0011
5
-0.0617
0.1046
0.0084
-0.0043
0.1044
0.0086
0.0100
0.0986
0.0108
6
0.1150
0.1063
0.1042
0.2403
0.1195
0.1045
0.3182
0.1775
0.1056
7
-0.0751
0.1027
0.0105
0.0093
0.1028
0.0109
0.0217
0.1061
0.0124
Cputime
0.0624
1.5132
0.9048
0.0156
57.8452
10.1089
0.0312
5.1988 103
436.8028
Table 4: Sensitivity measures. The estimation of the Cramér von Mises indices is the ratio of (8) and
(9).
N = 102
N = 103
N = 104
Sensitivity meas.
Multivariate
Borgonovo et al.
Cramér von Mises
Multivariate
Borgonovo et al.
Cramér von Mises
Multivariate
Borgonovo et al.
Cramér von Mises
1
1
1
1
1
1
1
1
1
Ranking
62357
36257
62375
62375
62573
62375
62375
62735
62375
4
4
4
4
4
4
4
4
4
Table 5: Ranks. The estimation of the Cramér von Mises indices is the ratio of (8) and (9).
5
Conclusion
In this paper, we first study the asymptotic properties of the multiple Pick and Freeze scheme proposed
by Owen et al. for the estimation of higher order Sobol indices. This index has several drawbacks
that lead us to propose a new natural index based on the Cramér von Mises distance between the
distribution of the output Y and the conditional law when an input is fixed. This new index contains
all the distributional information, is naturally defined for multivariate outputs and provides a rigorous
sharper way for a fast screening of complex computer codes. Furthermore, our approach is generic and
may be extended and implemented for general outputs (vectorial, valued on a manifold, functional, ...).
Concerning its estimation, we show that surprisingly a Pick and Freeze scheme is also available for the
estimation procedure and prove that it is efficient in a theoretical point of view as well as in a practical
one. More precisely, we establish a Central Limit Theorem that confirms the good statistical properties
of our estimator and allows us to build confidence intervals. Furthermore, the estimation is well working
with moderate sample sizes as shown in toy examples. Finally, the performance of the method is proven
on a real data example.
Acknowledgement
The authors are greatly indebted to the referees for their fruitful and detailed suggestions or comments
which permit us to greatly improve our paper.
Part of this research was conducted within the frame of the Chair in Applied Mathematics OQUAIDO
and the ANR project PEPITO (ANR-14-CE23-0011).
16
6
6.1
Proofs
Proof of Theorem 2.1
Proof of Theorem 2.1. The consistency follows from a straightforward application of the strong law of
large numbers. The asymptotic normality is derived by two successive applications of the delta method
[29] .
(1) Let Wj1 := (Yjv,1 , . . . , Yjv,p )T (j = 1, . . . , N ) and g 1 be the mapping from Rp to Rp whose l-th
coordinate is given by
!
−1
l
X
Y
p
1
gl (x1 , . . . , xp ) =
xki .
l
i=1
k1 < . . . < kl
ki ∈ Ip , i = 1, . . . , l
Then (Wj1 )j=1,...,N is an i.i.d. sample distributed as W 1 := (Y v,1 , . . . , Y v,p )T .
Let Σ1 be the covariance matrix of Wj1 . Clearly, one has Σ1ii = Var(Y ) for i ∈ Ip while for i 6= j,
Σ1ij = Cov(Y v,i , Y v,j ) = Cov(Y, Y v,2 ). The multidimensional Central Limit Theorem gives that
N
X
√
1
L
N
Wj1 − m → Np 0, Σ1 ,
N →∞
N j=1
where m := (E[Y ], . . . , E[Y ])T . We then apply the so-called delta method to W 1 and g 1 so that
√ 1 1
L
T
N g W N − g1 E W 1
→ N 0, Jg1 E W 1 Σ1 Jg1 E W 1
N →∞
where Jg1
E W 1 is the Jacobian of g 1 at point E W 1 . Notice that for i ∈ Ip and k ∈ Ip ,
p−1
1
l
∂gl1
l−1
E W
= p ml−1 = E[Y ]l−1 =: al .
∂xk
p
l
T
Thus Σ2 := Jg1 E W 1 Σ1 Jg1 E W 1
is given by
Σ2ij = pai aj Σ111 + (p − 1)Σ112 .
(2) Now consider Wj2 := (Pjv,1 , . . . Pjv,p )T (j = 1, . . . , N ) and g 2 the mapping from Rp to R defined by
p
X
p
g (y1 , . . . , yp ) =
(−1)p−l y1p−l yl .
l
2
l=0
Then (Wj2 )j=1,...,N is an i.i.d. sample distributed as W 2 := (P v,1 , . . . P v,p )T .
We apply once again the delta method to W 2 so that
√ 2 2
L
T
N g W N − g2 E W 2
→ N 0, Jg2 E W 2 Σ2 Jg2 E W 2
N →∞
where Jg2
E W 2 is the Jacobian of g 2 at point E W 2 . Notice that for k ∈ Ip ,
∂g 2
E W 2 = (−1)p−1 p(p − 1)E[Y ]p−1
∂y1
" l
#
p−1
X
Y
p
p−l
p−l−1
v,i
+
(−1) (p − l)E[Y ]
E
Y
l
i=1
l=2
and
∂g 2
E W2 =
∂yl
p
(−1)p−l E[Y ]p−l .
l
17
Thus the limiting variance is
2
σ := Jg2
p
T
X
= p Σ111 + (p − 1)Σ112
E W 2 Σ2 J g 2 E W 2
ai bi
!2
,
i=1
where bi is the i-th coordinate of ∇g 2 E W 2 .
6.2
An auxiliary result and the proofs of the results of Section 3
Lemma 6.1. Let G and H be two measurable functions. Let (Uj )j∈IN and (Vk )k∈IN be two independent
samples of i.i.d. random variables. Assume that G(U1 , V1 ) and H(U1 , U2 , V1 ) are both integrable and
centered. We define SN and TN by
SN =
N
1 X
G(Uj , Vk )
N2
TN =
and
j,k=1
N
1 X
H(Ui , Uj , Vk ).
N3
i,j,k=1
Then SN and TN converge a.s. to 0 as N goes to infinity.
4
Proof of Lemma 6.1. Notice that if E[SN
]=O
0. Now,
4
E[SN
]=
1
N2
then by Borel-Cantelli lemma, SN converges a.s. to
1 X
E[G(Ui1 , Vj1 )G(Ui2 , Vj2 )G(Ui3 , Vj3 )G(Ui4 , Vj4 )]
N8
where the sum is taken over all the indices i1 , i2 , i3 , i4 , j1 , j2 , j3 , j4 from 1 to N . The only cases leading
to terms in O N1 or even in O (1) appear when we sum over indices that are all different except two i’s
or two j’s or over indices that are all different. Nevertheless, in those cases, at least one term of the form
E[G(Ui , Vj )] appears. Since the function G is centered, those cases are then discarded.
The proof of the result concerning TN follows the same tracks.
Proof of Corollary 3.4. The proof is based on Lemma 6.1. First, we define Zj = Zjv,1 , Zjv,2 ,
G(Zj , Wk ) = 1 {Z v,1 6Wk } 1 {Z v,2 6Wk } ,
j
j
1
F (Zj , Wk ) =
1 {Z v,1 6Wk } + 1 {Z v,2 6Wk } ,
j
j
2
H(Zi , Zj , Wk ) = F (Zi , Wk )F (Zj , Wk ).
18
Second, we proceed to the following decomposition
bv
N
2,CV M
2
N
N
N
X
X
X
1
1
1
=
1 {Z v,1 6Wk } 1 {Z v,2 6Wk } −
1 {Z v,1 6Wk } + 1 {Z v,2 6Wk }
j
j
j
j
N
2N j=1
N j=1
k=1
=
N
N
X
1
1 X
v,1
v,2
v,1
v,2
v,1
v,2
1
−
1
+
1
1
+
1
1
{Zj 6Wk } {Zj 6Wk }
{Zi 6Wk }
{Zi 6Wk }
{Zj 6Wk }
{Zj 6Wk }
N2
4N 3
j,k=1
=
=
1
N2
N
X
i,j,k=1
G(Zj , Wk ) −
j,k=1
1
N3
N
X
H(Zi , Zj , Wk )
i,j,k=1
N
N
1 X
1 X
{G(Z
,
W
)
−
E[G(Z
,
W
)]}
−
{H(Zi , Zj , Wk ) − E[H(Zi , Zj , Wk )]}
j
k
j
k
N2
N3
j,k=1
+
=
1
N2
N
X
i,j,k=1
E[G(Zj , Wk )] −
j,k=1
1
N3
N
X
E[H(Zi , Zj , Wk )]
i,j,k=1
N
N
1 X
1 X
{G(Z
,
W
)
−
E[G(Z
,
W
)]}
−
{H(Zi , Zj , Wk ) − E[H(Zi , Zj , Wk )]}
j
k
j
k
N2
N3
j,k=1
i,j,k=1
1
1
E[H(Z1 , Z2 , W1 )] − E[H(Z1 , Z1 , W1 )].
+ E[G(Z1 , W1 )] − 1 −
N
N
The two first sums converge almost surely to 0 by Lemma 6.1. The remaining term goes to E[G(Z1 , W1 )]−
E[H(Z1 , Z2 , W1 )] as N goes to infinity.
v
It remains to show that N2,CV
M = E[G(Z1 , W1 )] − E[H(Z1 , Z2 , W1 )]. On the one hand,
v
N2,CV
M
Z
=
E[(F (t) − F v (t))2 ]dF (t) = E[Hv2 (W1 )]
R
= E[Cov(1 {Z v,1 6W1 } , 1 {Z v,2 6W1 } )]
1
1
= EW [EZ [1 {Z v,1 6W1 } 1 {Z v,2 6W1 } ] − EZ [1 {Z v,1 6W1 } ]2 ].
1
1
1
On the other hand,
E[G(Z1 , W1 )] − E[H(Z1 , Z2 , W1 )]
1
= E[1 {Z v,1 6W1 } 1 {Z v,2 6W1 } ] − E[(1 {Z v,1 6W1 } + 1 {Z v,2 6W1 } )(1 {Z v,1 6W1 } + 1 {Z v,2 6W1 } )]
1
1
1
1
2
2
4
= EW [EZ [1 {Z v,1 6W1 } 1 {Z v,2 6W1 } ]] − E[1 {Z v,1 6W1 } 1 {Z v,2 6W1 } ]
1
1
1
2
= EW [EZ [1 {Z v,1 6W1 } 1 {Z v,2 6W1 } ]] − E[E[1 {Z v,1 6W1 } 1 {Z v,2 6W1 } |W1 ]]
1
1
1
2
= EW [EZ [1 {Z v,1 6W1 } 1 {Z v,2 6W1 } ]] − E[E[1 {Z v,1 6W1 } |W1 ]E[1 {Z v,2 6W1 } |W1 ]]
1
1
1
2
= EW [EZ [1 {Z v,1 6W1 } 1 {Z v,2 6W1 } ]] − E[E[1 {Z v,1 6W1 } |W1 ]]E[E[1 {Z v,2 6W1 } |W1 ]]
1
1
1
2
= EW [EZ [1 {Z v,1 6W1 } 1 {Z v,2 6W1 } ]] − E[1 {Z v,1 6W1 } ]E[1 {Z v,2 6W1 } ]
1
1
1
2
= EW [EZ [1 {Z v,1 6W1 } 1 {Z v,2 6W1 } ]] − E[1 {Z v,1 6W1 } ]2
1
1
1
= EW [EZ [1 {Z v,1 6W1 } 1 {Z v,2 6W1 } ] − EZ [1 {Z v,1 6W1 } ]2 ]
1
1
1
that completes the proof.
19
Proof of Theorem 3.5. We define for t ∈ R,
G1,2
N (t, t) =
N
1 X
1 v,1 1 v,2 ,
N j=1 {Zj 6t} {Zj 6t}
GiN (t) =
N
1 X
1 v,i , i = 1, 2,
N j=1 {Zj 6t}
FN (t) =
N
1 X
1 {Wk 6t}
N
k=1
bv
and we rewrite N
2,CV M as a regular function depending on the four empirical processes defined above:
bv
N
2,CV M =
1
2 #
Z "
GN + G2N
1,2
GN −
dFN .
2
By Donsker’s theorem,
√ 1,2
L
e G1N − F, G2N − F, FN − F
N GN − G,
→ G = (G1 , G2 , G3 , G4 )
N →∞
e = G(t, t) and G is a centered Gaussian process of dimension
where G(t, s) = P Z v,1 6 t, Z v,2 6 s , G(t)
4 with covariance function defined by
T
Π(t, s) = E At ATs − E (At ) E (As ) , for (t, s) ∈ R2
and At := 1 {Z v,1 6t} 1 {Z v,2 6t} , 1 {Z v,1 6t} , 1 {Z v,2 6t} , 1 {W 6t}
T
.
Since these processes are càd-làg functions of bounded variation, we introduce the maps ψ1 , ψ2 :
BV1 [−∞, +∞]2 7→ R and Ψ : BV1 [−∞, +∞]4 7→ R defined by
Z
F2 + F3
ψi (F1 , F2 ) = (F1 )i dF2 , i = 1, 2 and Ψ(F1 , F2 , F3 , F4 ) = ψ1 (F1 , F4 ) − ψ2
, F4 ,
2
where BVM [a, b] is the set of càd-làg functions of variation bounded by M . Hence,
1,2
v
1
2
b2,CV
N
=
Ψ
G
,
G
,
G
,
F
,
N
M
N
N
N
Now using the chain rule 20.9 and Lemma 20.10 in [29], the map Ψ is Hadamard-differentiable from the
domain BV1 [−∞, +∞]4 into R whose derivative is given by
h2 + h3
F2 + F3
, F4
, h4
(h1 , h2 , h3 , h4 ) 7→ Dψ1 (F1 , F4 )(h1 , h4 ) − Dψ2
2
2
where the derivative of ψi are given by Lemma 20.10:
Z
Z
(h1 , h2 ) 7→ h2 ϕi ◦ F1 |+∞
−
h
dϕ
◦
F
+
ϕ0i (F1 )h1 dF2
2−
i
1
−∞
with ϕi (x) = xi and h− is the left-continuous version of a càd-làg function h.
Applying the functional delta method 20.8 in [29] we get the weak convergence of
to the following limit distribution
Z
Z
Z
2
e
G4− d(F − G) + G1 dF − F (G2 + G3 )dF.
20
√
v
bv
N N
2,CV M − N2,CV M
Since the map Ψ is continuous on the whole space BV1 [−∞, +∞]4 , the delta method in its stronger form
20.8 in [29] implies that the limit variable is the limit in distribution of the sequence
√
2
e 1
e F, F, F )
N G1,2
DΨ(G,
N − G, GN − F, GN − F, FN − F
Z
Z
√
1,2
2
1
2
e
e
= N
(FN − F )− d F − G) +
GN − G − F GN + GN − 2F dF .
We define
Z
e = G(W
e
1 {W <t} d(F 2 (t) − G(t)
) − F (W )2 ,
U :=
Z
V :=
=
1 {Z v,1 6t} 1 {Z v,2 6t} − 1 {Z v,1 6t} + 1 {Z v,2 6t} F (t) dF (t)
1
F (Z v,1 )2 + F (Z v,2 )2 − F (Z v,1 ∨ Z v,2 ).
2
By independence, the limiting variance ξ 2 is
ξ 2 = VarU + VarV.
6.3
(12)
Proof of Proposition 4.1
Proof of Proposition 4.1. First of all, the distribution function of Y conditioned on X1 is given by
F (1) (t) = P(Y 6 t|X1 ) = Φ
ln t − X1
2
.
Then
1
N2,CV
M
Z
h
i
E (F (1) (t) − FY (t))2 fY (t)dt
R
"
2 #
Z
2
ln y
1
ln t − X1
√
−Φ √
e−(ln t) /10 dt
=
E Φ
2
+
5
10πt
R
!
!2
√
Z
2
5z
−
X
dz
1
=
E Φ
− Φ (z) e−z /10 √
2
2π
R
!!
√
2
5X2 − X1
Φ(X2 ) − Φ
=E
2
=
where X1 and X2 are independent standard Gaussian random variables. In the same way,
h
√
i
2
N2,CV
=
E
(Φ(X
)
−
Φ
5X
−
2X
)2 .
2
2
1
M
Thus we are lead to compute the bivariate function:
ϕ(α, β) := E (Φ(X2 ) − Φ (αX2 − βX1 ))2
√
√
for (α, β) = ( 5/2, 1/2) and (α, β) = ( 5, 2). The term E Φ(X2 )2 is
E Φ(X2 )2 =
Z
1
Φ(z) f (z)dz =
Φ(z)3
3
2
21
+∞
=
−∞
1
.
3
We introduce three independent
random variables
U , U 0 and V distributed as standard Gaussian random
h
i
2
variables. Then the term E Φ (αX2 − βX1 ) can be rewritten as
h
2
E Φ (αX2 − βX1 )
i
p
h
i2
2
2
2
√
=E Φ
α +β V
= E E 1 U 6 α2 +β 2 V |V
i h
ii
h h
ii
h h
= E E 1 U 6√α2 +β 2 V |V E 1 U 0 6√α2 +β 2 V |V = E E 1 U 6√α2 +β 2 V 1 U 0 6√α2 +β 2 V |V
h
i
p
p
= E 1 U 6√α2 +β 2 V 1 U 0 6√α2 +β 2 V = P U 6 α2 + β 2 V, U 0 6 α2 + β 2 V
p
=: G( α2 + β 2 ).
Integrating by parts, we have
Z
2
2
dz
G0 (a) = 2 zΦ(az)e−(a +1)z /2
2π
R
h
Z
i+∞
1
−(a2 +1)z 2 /2
−(a2 +1)z 2 /2
−
a
=−
Φ(az)e
f
(az)e
dz
π(a2 + 1)
−∞
R
a
1
√
.
=
π(a2 + 1) 2a2 + 1
Since G(1) = 1/3, we get
Z a
p
p
√
1
x
1
1
1
1
2 − arctan 3) =
√
G(a) = +
+
(arctan
1
+
2a
arctan
1 + 2a2
dx
=
2
3
3 π
π
2x2 + 1
1 π(x + 1)
and
h
i 1
p
p
√
1
1
2
E Φ (αX2 − βX1 ) = + (arctan 1 + 2(α2 + β 2 ) − arctan 3) = arctan 1 + 2(α2 + β 2 ).
3 π
π
In the same way, the last term E [Φ(X2 )Φ (αX2 − βX1 )] is given by
r
E [Φ(X2 )Φ (αX2 − βX1 )] = P U 6 V,
1 + β2 0
U 6V
α2
!
where U , U 0 and V are independent standard Gaussian random variables.
Remind that we only need to
q
√
√
1+β 2
consider (α, β) = ( 5/2, 1/2) and (α, β) = ( 5, 2) in which cases
α2 = 1. Thus the last term equals
1/3 in both cases. It remains to divide by the normalizing factor 1/6 to get the result.
Remark 6.2. In the previous proof, we establish that
G(a) = P (U 6 aV, U 0 6 aV )
√
is equal to π1 arctan 1 + 2a2 where U , U 0 and V are independent standard Gaussian random variables.
√
Actually, this result is
also a straightforward consequence of Lemma 4.3 in [2] with X = (aV −U )/ a2 + 1
√
and Y = (aV − U 0 )/ a2 + 1. Nevertheless, since our proof is different and elegant, we decide not to skip
it.
References
[1] A. Antoniadis. Analysis of variance on function spaces. Statistics: A Journal of Theoretical and
Applied Statistics, 15(1):59–71, 1984.
[2] J.M. Azaïs and M. Wschebor. Level sets and extrema of random processes and fields. John Wiley &
Sons, Inc., Hoboken, NJ, 2009.
22
[3] E. Borgonovo. A new uncertainty importance measure. Reliability Engineering & System Safety,
92(6):771–784, 2007.
[4] E. Borgonovo and M. Baucells. Invariant probabilistic sensitivity analysis. Management Science,
59(11):2536–2549, 2013.
[5] E. Borgonovo, W. Castaings, and S. Tarantola. Moment independent importance measures: New
results and analytical test cases. Risk Analysis, 31(3):404–428, 2011.
[6] E. Borgonovo, W. Castaings, and S. Tarantola. Model emulation and moment-independent sensitivity
analysis: An application to environmental modelling. Environmental Modelling & Software, 34:105–
115, 2012.
[7] E. Borgonovo, G. Hazen, and E. Plischke. Probabilistic sensitivity measures: Foundations and
estimation. Submitted, pages 1–24, 2014.
[8] E. Borgonovo and B. Iooss. Moment Independent Importance Measures and a Common Rationale.
Preprint, 2015.
[9] E. Borgonovo and B. Iooss. Moment-Independent and Reliability-Based Importance Measures, pages
1–23. Springer International Publishing, Cham, 2016.
[10] R. Buchbinder and A. S. Detsky. Management of suspected giant cell arteritis: A decision analysis.
J. Rheumatology, 19(9):1220–1228, 1992.
[11] S. Da Veiga. Global sensitivity analysis with dependence measures. J. Stat. Comput. Simul.,
85(7):1283–1305, 2015.
[12] Y. De Castro and A. Janon. Randomized pick-freeze for sparse Sobol indices estimation in high
dimension. ArXiv e-prints, March 2014.
[13] E. De Rocquigny, N. Devictor, and S. Tarantola. Uncertainty in industrial practice. Wiley Online
Library, 2008.
[14] J.C. Felli and G. Hazen. Sensitivity analysis and the expected value of perfect information. Med.
Decis. Making, 18(1):95–109, 1998.
[15] J.C. Felli and G. Hazen. Javelin diagrams: A graphical tool for probabilistic sensitivity analysis.
Decision Analysis, 1(2):93–107, 2004.
[16] J.-C. Fort, T. Klein, and N. Rachdi. New sensitivity analysis subordinated to a contrast. Communications in Statistics - Theory and Methods, 2015 to appear.
[17] F. Gamboa, A. Janon, T. Klein, and A. Lagnoux. Sensitivity analysis for multidimensional and
functional outputs. Electronic Journal of Statistics, 8:575–603, 2014.
[18] F. Gamboa, A. Janon, T. Klein, A. Lagnoux-Renaudie, and C. Prieur. Statistical inference for Sobol
pick freeze Monte Carlo method, March 2013.
[19] A. Janon, T. Klein, A. Lagnoux, M. Nodet, and C. Prieur. Asymptotic normality and efficiency of
two sobol index estimators. ESAIM: Probability and Statistics, 18:342–364, 1 2014.
[20] A. Müller. Integral probability metrics and their generating classes of functions. Adv. in Appl.
Probab., 29(2):429–443, 1997.
[21] J. von Neumann and O. Morgenstern. Theory of Games and Economic Behavior. Princeton, NJ.
Princeton University Press, 1953.
[22] J. E. Oakley. Decision-theoretic sensitivity analysis for complex computer models. Technometrics,
51(2):121–129, 2009.
[23] A.B. Owen. Variance components and generalized sobol’ indices. SIAM/ASA Journal on Uncertainty
Quantification, 1(1):19–41, 2013.
23
[24] A.B. Owen, J. Dick, and S. Chen. Higher order sobol’ indices. Information and Inference, 3(1):59–81,
2014.
[25] A. Saltelli, K. Chan, and E.M. Scott. Sensitivity analysis. Wiley Series in Probability and Statistics.
John Wiley & Sons, Ltd., Chichester, 2000.
[26] T. J. Santner, B. Williams, and W. Notz. The Design and Analysis of Computer Experiments.
Springer-Verlag, 2003.
[27] I. M. Sobol. Sensitivity estimates for nonlinear mathematical models. Math. Modeling Comput.
Experiment, 1(4):407–414 (1995), 1993.
[28] H. J. ter Horst. On Stieltjes integration in Euclidean space. J. Math. Anal. Appl., 114(1):57–74,
1986.
[29] A. W. van der Vaart. Asymptotic statistics, volume 3 of Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge, 1998.
24
| 10math.ST
|
PREFERENTIAL ATTACHMENT AND VERTEX ARRIVAL
TIMES
Benjamin Bloem-Reddy and Peter Orbanz
arXiv:1710.02159v1 [math.PR] 5 Oct 2017
University of Oxford and Columbia University
We study preferential attachment mechanisms in random graphs
that are parameterized by (i) a constant bias affecting the degreebiased distribution on the vertex set and (ii) the distribution of times
at which new vertices are created by the model. The class of random
graphs so defined admits a representation theorem reminiscent of
residual allocation, or “stick-breaking” schemes. We characterize how
the vertex arrival times affect the asymptotic degree distribution,
and relate the latter to neutral-to-the-left processes. Our random
graphs generate edges “one end at a time”, which sets up a one-toone correspondence between random graphs and random partitions
of natural numbers; via this map, our representation induces a result
on (not necessarily exchangeable) random partitions that generalizes
a theorem of Griffiths and Spanó. A number of examples clarify how
the class intersects with several known random graph models.
The term preferential attachment describes generative mechanisms for random
graph models that select the terminal vertices of a new edge with probability biased
by the vertex degrees. These models come in many shapes and guises [e.g. 4, 5,
24, 32], and are often motivated by their ability to generate (and hence explain)
power law distributions. Degree-biased selection is a form of size bias [3], and this
interplay between size-biasing and power laws is not confined to random graph
models, but also encountered in random partitions, which are used in population
genetics, machine learning, and other fields [e.g. 36, 16, 10]. In partition models,
power laws arise as heavy-tailed distributions of block sizes. Size-biased sampling as
such, however, need not result in a power laws: The most basic form of size-biased
sampling from a countable number of categories is a type of Pólya urn with an
unbounded number of colors, or, equivalently, the one-parameter Chinese restaurant
process [36]. It does not generate a power law. To obtain power laws, plain size-biased
sampling can be modified in two ways:
(i) By biasing the size-biased probability of each category downward.
(ii) By forcing new categories to arrive at a faster rate than that induced by plain
size-biased sampling.
An example is the two-parameter Chinese restaurant process with parameter (α, θ),
which modifies the Chinese restaurant process with a single parameter θ—a model
that corresponds to plain size-biased sampling—by effectively (i) damping the size
bias by a constant offset α, and (ii) increasing the rate at which new categories arrive.
1
2
An example of (ii) is the Barabási–Albert random graph model, in which vertices
arrive at fixed, constant time intervals; if these times were instead determined at
random by size-biased sampling, intervals would grow over time.
The premise of this work is to study preferential attachment mechanisms in random graph models by explicitly controlling these two effects:
(i) The attachment probability, proportional to the degree degk of each vertex k,
is biased by a constant offset as degk − α.
(ii) Vertex arrival times are taken into account, by explicitly conditioning the
generation process on a (random or non-random) sequence of given times.
The result is a class of random graphs parametrized by the offset α and a sequence t
of vertex arrival times. Each such (α, t)-graph can be generalized by randomizing the
arrival times, i.e. to an (α, T )-graph for a random sequence T . Preferential attachment models that constantly bias the attachment probability have been thoroughly
studied [30, 24]. We consider the range α ∈ (−∞, 1), and the case α ∈ [0, 1) turns
out to be of particular interest. The effects (i) and (ii) are not independent, and in
models with a suitable exchangeability property, the effect of α can equivalently be
induced by controlling the law of the arrival times. In this sense, (ii) can provide
more control over the model than (i).
Section 1 characterizes (α, T )-graphs by a representation in terms of independent beta random variables, reminiscent of stick-breaking constructions of random
partitions.
Section 2 considers implications for random partitions and urns. (α, T )-graphs
generate edges “one end at a time”, updating the vertex degrees after each step.
Although such a scheme differs from the usual preferential attachment model, it
is similar to so-called “sequential” versions considered by [5, 32]. This sets up a
one-to-one correspondence between multigraphs and partitions of natural numbers:
There is a bijection Φ such that
Φ(partition) = graph ,
which translates our results on graphs into statements about partitions. If G is an
(α, T )-graph, the random partition Φ−1 (G) may or may not be exchangeable. The
subclass of such partitions that are exchangeable are precisely the exchangeable
Gibbs partitions [21, 36]. Arrival times in such partitions, known as record indices,
have been studied by Griffiths and Spanò [23]. Broadly speaking, our results recover
those of Griffiths and Spanò if Φ−1 (G) is exchangeable, but show there is a larger
class of random partitions—either of Gibbs type, or not exchangeable—for which
similar results hold. Non-exchangeable examples include partitions defined by the
Yule–Simon process [42, 39]. Additionally, our representation result for graphs yields
an analogous representation for this class of partitions; it also relates the work
of Griffiths and Spanò [23] to that of Berger, Borgs, Chayes, and Saberi [5] on
Benjamini–Schramm limits of certain random preferential attachment graphs.
3
Section 3 studies degree asymptotics of (α, T )-graphs. Properly scaled degree
sequences of such graphs converge. The limiting degrees are neutral-to-the-left sequences of random variables that satisfy a number of distributional identities. We
characterize cases in which power laws emerge, and relate the behavior of the degree distribution to sparsity. The range of power laws achievable is constrained by
whether or not the average degree is bounded.
Section 4 discusses examples, and shows how the class of (α, T )-graphs overlaps
with several known models, such as the Barabási–Albert model [4], edge exchangeable graphs [14, 11, 28], and the preferential attachment model of Aiello, Chung,
and Lu [1, 2]. We obtain new results for some of these submodels. For preferential attachment graphs, for example, limiting degree sequences are known to satisfy
various distributional identities [32, 25, 26]. These results assume fixed intervals between vertex arrival times. We show similar results hold if arrival times are random.
Perhaps most closely related is the work of Peköz, Röllin, and Ross [33] on random
immigration times in a two-color Pólya urn, which corresponds to a certain (α, T )model with i.i.d. interarrival times. We use this correspondence to answer a question
posed in [33] about interarrival times with geometric distribution.
1. Preferential attachment and arrival times. Consider an undirected multigraph g, possibly with self-loops, with a countably infinite number of edges. The
graph models considered in the following insert edges into the graph one at a time.
It is hence convenient to represent g as a sequence of edges
(1)
g = (l1 , l2 ), (l3 , l4 ), . . .
where lj ∈ N for all j ∈ N .
Each pair (l2n−1 , l2n ) represents an undirected edge connecting the vertices l2n−1 and
l2n . The vertex set of g is V(g) := {l1∗ , l2∗ , . . .}, the set of all distinct values occurring
in the sequence. We assume vertices are enumerated in order of appearance, to wit
(2)
l1 = 1
and
lj+1 ≤ max {l1 , . . . , lj } + 1
for all j ∈ N .
Consequently, V(g) is either a consecutive finite set {1, . . . , m}, or the entire set N.
Let G be the set of multigraphs so defined, equipped with the topology inherited from
the product space N∞ , which makes it a standard Borel space. For our purposes, a
random graph is a random element
G = (L1 , L2 ), (L3 , L4 ), . . .)
of G, for N-valued random variables Ln . Note that the same setup can be used to
model directed multigraphs.
For a graph g, let gn := (lj )j≤2n denote the subgraph given by the first n edges,
and degk (n) the degree of vertex k in gn . The arrival time of vertex k is
tk := min {j ∈ N | lj = k} ,
4
with tk = ∞ if g has fewer than k vertices. The set of possible arrival time sequences
is T := {(1 = t1 < t2 < . . . ≤ ∞)}. Arrival times in T2 := {t ∈ T | tk even for k > 1}
is a sufficient, though not necessary, condition for g to be a connected graph; it is
necessary and sufficient for each gn to be connected. If T = (T1 , T2 , . . .) is a random
sequence of arrival times, the interarrival times
∆k := Tk − Tk−1
where T0 := 0
are random variables with values in N ∪ {∞}.
1.1. Degree-biased random graphs. The term preferential attachment describes
a degree bias: Recall that the degree-biased distribution on the vertices of a graph
gn with n edges is P (k ; gn ) := degk (n)/2n. We embed P into a one-parameter family
of laws
degk (n) − α
Pα (k ; gn ) :=
for α ∈ (−∞, 1) ,
2n − α|V(gn )|
the α-degree biased distributions. Both P and Pα are defined on a graph gn , in
which each edge is either completely present or completely absent. To permit edges
to be generated “one end at a time”, we observe Pα can be rewritten as
(3)
Pα (k ; l1 , . . . , lj ) =
|{i ≤ j|lj = k}| − α
j − α max {l1 , . . . , lj }
for k ≤ max {l1 , . . . , lj } ,
which is well-defined even if j is odd.
(α, t)-graph. Given are α ∈ (−∞, 1) and t ∈ T. Generate L1 , L2 , . . . as
Ln := k
if n = Tk
and
Ln ∼ Pα ( • ; L1 , . . . , Ln−1 )
otherwise .
Then G := ((L1 , L2 ), (L3 , L4 ), . . .) is a random graph, whose law we denote DB(α, t).
The sequence t may additionally be randomized: We call G an (α, T )-graph if its law
is DB(α, T ), for some random element T of T. Examples of multigraphs generated
using different distributions for T are shown in Fig. 1.
As a consequence of the family of laws (3), the finite-dimensional distributions of
(α, t)-graphs have a simple product form.
Proposition 1.
Let Gn be an (α, t)-graph with n/2 edges and k vertices. Then
Pα,t [Gn/2 = ((L1 , L2 ), . . . , (Ln−1 , Ln ))]
(4)
=
k
Y
Γ(Tj − jα)Γ(#{Li = k | i ≤ n} − α)
1
.
Γ(n − kα)
Γ(Tj − 1 − (j − 1)α + δ1 (j))Γ(1 − α)
j=1
/
5
1.2. Representation result. Fix α ∈ (−∞, 1) and t ∈ T. Let Ψ1 , Ψ2 , . . . be independent random variables with Ψ1 = 1,
(5)
Ψj ∼ Beta 1 − α, tj − 1 − (j − 1)α for j ≥ 2 ,
and define
(6)
Wj,k :=
j
X
i=1
Ψi
k
Y
(1 − Ψ` )
and
Ij,k := [Wj−1,k , Wj,k ) with W0,k = 0 .
`=i+1
Q
Note that Wj,k = k`=j+1 (1−Ψ` ) and Wk,k = 1. Hence, ∪kj=1 Ij,k = [0, 1). Generate a
random sequence U1 , U2 , . . . ∼iid Uniform[0, 1). For each n, let tk(n) be the preceding
arrival time, i.e. the largest tk with tk ≤ n, and set
(
k(n)
if n = tk(n)
(7)
Ln :=
.
j such that Un ∈ Ij,k(n) otherwise
Then H(α, t) := ((L1 , L2 ), . . .) is a random element of G.
d
Theorem 2. A random graph G is an (α, T )-graph if and only if G = H(α, T )
for some α ∈ (−∞, 1) and a random element T of T.
/
Products of the form (6), for the same sequence of beta variables Ψj , previously
have appeared in two separate contexts: Griffiths and Spanò [23] identify Ψj Wj,∞
as the limiting relative block sizes in exchangeable Gibbs partitions, conditioned on
the block arrival times. This corresponds to the special case of Theorem 2 where the
random variables (L1 , L2 , . . .) define an exchangeable Gibbs partition (see Sections 2
and 4.2). In work of Berger, Borgs, Chayes, and Saberi [5], a version of (5)–(7)
arises as the representation of the Benjamini–Schramm limit of certain preferential
attachment graphs (in which case all interarrival times are fixed to a single constant).
That two problems so distinct lead to the same (and arguably not entirely obvious)
distribution raises the question whether (5) can be understood in a more conceptual
way. One such way is by regarding the graph as a recursive sequence of Pólya urns:
Conditionally on an edge attaching to one of the first k vertices, it attaches to vertex
k with probability Ψk and one of the first k − 1 with probability 1 − Ψk , and so on
for k − 1, . . . , 2. A related interpretation is in terms of the special properties of beta
and gamma random variables. Let Ga and Ba,b generically denote a gamma random
variable with parameters (a, 1) and a beta variable with parameters (a, b). Beta
and gamma random variables satisfy a set of relationships sometimes referred to
collectively as the beta-gamma algebra [e.g. 38]. These relationships revolve around
the fact that, if Ga and Gb are independent, then
d
Ga
(8)
Ga+b , Ba,b = Ga + Gb ,
,
Ga + Gb
6
where the pair on the left is independent, and so is the pair on the right. In the
context of (α, T )-graphs, conditionally on the sequence ∆1 , ∆2 , . . . of interarrival
times, generate two sequences of gamma variables
G (1) , G (2) , . . . ∼iid Gamma(1 − α, 1)
G∆2 −1 , G∆3 −1 , . . . ,
and
all mutually independent given (∆k ). The variables Ψk can then be represented as
(9)
Ψj
d
=
G (j)
P
(i) +
i≤j G
i<j G∆i+1 −1
P
=:
Ψ0j .
Such recursive fractions are not generally independent, but as a consequence of
d
(8), equality in law holds even jointly, (Ψ1 , Ψ2 , . . .) = (Ψ01 , Ψ02 , . . .), recovering the
d
variables in Theorem 2. Identity (8) further implies Ba,b+c = Ba,b Ba+b,c , again with
independence on the right. Abbreviate
τj := tj − 1 + α(j − 1)
d
Ψj = B1−α,τj .
such that (5) becomes
The recursion (9) then implies
d
B1−α,τj = B1−α+τj−1 ,∆j −α B1−α,τj−1
hence
d
Ψj |Ψj−1 = Ψj−1 B1−α+τj−1 ,∆j −α ,
with independence on the right of both identities. Informally, one may think of G (k)
as an (unnormalized) propensity of vertex k to attract edges, of those edges attaching
to one of the first k vertices. The requisite normalization in (9) depends on propensities of previously created vertices (represented by the variables G (1) , . . . , G (k−1) ), and
contributions of the “head start” given to previously created vertices (represented
by the variables G∆j −1 ).
2. Graphs and urns. Any graph in G defines a partition of N, and vice versa.
This fact is used below to classify α-degree biased graphs according to the properties
of the random partition they define. More precisely, a partition of N is a sequence
π = (b1 , b2 , . . .) ⊂ N of subsets, called blocks, such that each n ∈ N belongs to one
and only one block. The set of all partitions is denoted P(N), and inherits the topology of N∞ . A partition can equivalently be represented as a sequence π = (l1 , l2 , . . .)
of block labels, where lj = k means j ∈ bk . There is hence a bijective map
Φ : P(N) → G
given by
(l1 , l2 , . . .) 7→ (l1 , l2 ), (l3 , l4 ), . . . ,
which is a homeomorphism of P(N) and G. It identifies blocks of π with vertices of
g = Φ(π). In population genetics, the smallest element of the kth block of a partition
π is known as a record index [23]. Thus, the kth arrival time in g is precisely the
kth record index of π.
The generative process of a random partition Π can be thought of as an urn: Start
with an empty urn, and add consecutively numbered balls one at a time, each colored
7
with a randomly chosen color. Colors may reoccur, and are enumerated in order of
first appearance. Let Bk (n) be the set of all balls sharing the kth color after n balls
have been added. For n → ∞, one obtains a random partition Π = (B1 , B2 , . . .) of N,
with blocks Bk := ∪n Bk (n). In analogy to the (α, t)-graphs above, we define:
(α, t)-urn. Given are α ∈ (−∞, 1) and t ∈ T.
• If n = t for some k, add a single ball of a new, distinct color to the urn.
k
• Otherwise, add a ball of a color already present in the urn, where the jth
color is chosen with probability proportional to |Bj (n)| − α.
A familiar special case of such an urn is the Pólya urn with m colors, obtained
for α = 0 and t = (1, 2, . . . , m, ∞, ∞, . . .). Another is the two-parameter Chinese
restaurant process [36], also known as the Blackwell–MacQueen urn [7, 34]: If t is
randomized by generating (T1 = 1, T2 , T3 , . . .) according to
(10)
P[Tk+1 = Tk + t | Tk ] = (θ + αk)
Γ(θ + Tk )Γ(Tk + t − 1 − αk)
,
Γ(θ + Tk + t)Γ(Tk − αk)
for some θ > −α, the partition has law CRP(α, θ).
In general, (α, t)-urns define a class of random partitions Π that are coherent, in
the sense that
P(Πn−1 = {B1 , . . . , Bk }) =
k+1
X
P(Πn = An→j (Πn−1 )) ,
j=1
where An→j (Πn−1 ) denotes the operation of appending n to block Bj in Πn−1 . Partitions for which these probabilities depend only on the sizes of the blocks, and which
are therefore invariant under permutations of the elements, are exchangeable random partitions [36]. That is, there is an exchangeable partition probability function
(EPPF) p(·), symmetric in its arguments, such that
p(|B1 |, . . . , |Bk |) = P(Πn = {B1 , . . . , Bk }) ,
which is invariant under the natural action of the symmetric group. A special subclass are the exchangeable partitions of Gibbs type, for which the EPPF has the
unique product form [21]
(11)
k
Y
Γ(|Bj | − α)
p(|B1 |, . . . , |Bk |) = Vn,k
,
Γ(1 − α)
j=1
for a suitable sequence of coefficients Vn,k satisfying the recursion
(12)
Vn,k = (n − αk)Vn+1,k + Vn+1,k+1 .
8
The distribution of the arrival times can be deduced from (11) and (12) as
(13)
P[Tk+1 = Tk + t | Tk ] =
Γ(Tk + t − 1 − αk) VTk +t,k+1
,
Γ(Tk − αk)
VTk ,k
of which (10) for the CRP is a special case. Denote the law of T1 , . . . , Tk generated
by (13) as Pα,V (T1 , . . . , Tk ).
Alternatively, consider the (α, T )-urn counterpart of the EPPF, given in Proposition 1,
(14) pα,T (|B1 |, . . . , |Bk |; T1 , . . . , Tk ) = P[Πn = {B1 , . . . , Bk } | T1 , . . . , Tk ]
k
Y
Γ(Tj − jα)Γ(|Bj | − α)
1
=
.
Γ(n − kα)
Γ(Tj − 1 − (j − 1)α + δ1 (j))Γ(1 − α)
j=1
Define
(15)
α,T
:=
Vn,k
k
Y
Γ(Tj − jα)
1
,
Γ(n − kα)
Γ(Tj − 1 − (j − 1)α + δ1 (j))
j=1
in which case (14) takes on the Gibbs-like form
pα,T (|B1 |, . . . , |Bk |; T1 , . . . , Tk ) =
α,T
Vn,k
k
Y
Γ(|Bj | − α)
.
Γ(1 − α)
j=1
This general formula holds for all (α, T )-urns. In the case that Π is exchangeable,
these relationships imply a further characterization of exchangeable Gibbs partitions: (11) is obtained by marginalizing the arrival times from (14) according to
Pα,V .
Proposition 3. Let Π be a random partition generated by an (α, T )-urn, with
finite-dimensional conditional distributions given by (14). Then Π is exchangeable
if and only if there exists some sequence of coefficients V = (Vn,k ) satisfying
Vn,k =
k
X Pα,V (T1 , . . . , Tk ) Y
Γ(Tj − jα)
α,T
= E[Vn,k
],
Γ(n − kα)
Γ(Tj − 1 − (j − 1)α + δ1 (j))
T1 ,...,Tk
Tk ≤n
j=1
for all k ≤ n, in which case (11) holds and Π is an exchangeable Gibbs partition.
/
It is straightforward to verify that Φ(Π) is an (α, t)-graph if and only if Π is an
(α, t)-urn. This correspondence is used in Section 4 to classify some (α, t)-graphs
according to the urns they define. It also allows us to translate properties of random
graphs into properties of random partitions, and vice versa. Theorem 2 implies the
following result on partitions, which gives a representation of exchangeable Gibbs
partitions.
9
Corollary 4. A random partition Π is an (α, t)-urn if and only if it is disd
tributed as Π = (L1 , L2 , . . .), for variables Ln generated according to (5)–(7).
/
3. Degree asymptotics. Let G be an (α, t)-graph, and Gn the subgraph given
by its first n edges. The degree sequence of Gn is the vector D(n) = (degk (n))k≥1 ,
where vertices are ordered by appearance. Denote by md (n) the number of vertices
in Gn with degree d. The empirical degree distribution
(pd (n))d≥1 := |V(gn )|−1 (md (n))d≥1
is the probability that a vertex sampled uniformly at random from Gn has degree d.
The degree sequence and the degree distribution as Gn grows large are characterized
by the scaling behavior induced by α and t, which yields power laws and related
properties.
3.1. Linear and sub-linear regimes. As will become clear in the next section, the
scaling behavior of (α, t)-graphs is the result of products of the form
(16)
k
Y
Wj,k =
(1 − Ψi )
as
k→∞,
i=j+1
where (Ψj )j>1 are as in (5). In particular, two regimes of distinct limiting behavior emerge. To which of the two regimes an (α, t)-graph belongs is determined by
whether or not Wj,k converges to a non-zero value as k → ∞.
We consider (α, t)-graphs that satisfy the following assumption:
(17)
a.s.
|V(Gn )|/nσ −−−→ µ−σ
σ
n→∞
for some
0<σ≤1
and
0 < µσ < ∞ .
Slower vertex arrival rates (e.g., logarithmic) result in graphs that are almost surely
dense (see Section 3.4), and as such exhibit less interesting structural properties. For
example, in order to generate power law distributions in (α, t)-graphs, the asymptotic arrival rate must be super-logarithmic, which follows from work on exchangeable random partitions and can be read from [36, Chapter 3].
For a growing graph sequence satisfying the assumption (17), consider the limiting
average degree,
lim d¯n = lim
n→∞
n→∞
2n
σ
µ−σ
σ n
= lim
n→∞
2
n
µ−σ
σ
1−σ
.
The average degree is almost surely bounded if σ = 1, which we call the linear
regime; for σ ∈ (0, 1), the sub-linear regime, it diverges. This is a consequence of
Proposition 5 below: For a graph Gn on k(n) vertices, the probability that the end
of edge n + 1 is attached to vertex j is equal to Ψj Wj,k(n) , which results in vertex
j participating in a constant proportion of edges if and only if Wj,k(n) is bounded
away from zero as n grows large.
10
Proposition 5. For fixed α ∈ (−∞, 1) and t ∈ T such that (17) is satisfied for
some σ ∈ (0, 1], let Wj,k be as in (16). Then for each j ≥ 1, Wj,k converges almost
surely as k → ∞ to some random variable Wj,∞ , which is non-zero if and only if
σ < 1.
/
Remark 6. For slower vertex arrivals (e.g.logarithmic) or when the limiting
number of vertices is finite, Wj,k(n) also converges to a non-zero value.
/
3.2. Limiting joint distributions of degree sequences for given arrival times. The
previous section suggests that the limit of the scaled degrees should depend on the
random variables (Ψj )j>1 . Indeed, for any (α, t)-graph G, it can be shown that for
any r ∈ N+ ,
(18)
n−1 degj (n)
a.s.
−−−→ ξj
1≤j≤r n→∞
1≤j≤r
where
d
ξj = Ψj
∞
Y
(1 − Ψi ) ,
i=j+1
and (Ψj )j>1 are as in (5). Griffiths and Spanò [23] showed that relative degrees
with such a limit uniquely characterize exchangeable Gibbs partitions among all exchangeable partitions; if the random partition Φ−1 (G) is exchangeable, that result
applies to G (see Section 4.2). For a general (α, t)-graph, Φ−1 (G) need not be exchangeable, and indeed there are examples for which n−1 D(n) converges to zero, in
which case Wj,k(n) does, as well. In such cases, one may ask more generally whether
a finite, non-zero limit
D∞ := lim n−1/γ D(n) ,
n→∞
exists for an appropriate scaling exponent γ. Theorem 7 establishes that this is true
for (α, t)-graphs.
Theorem 7. Let G be an (α, t)-graph for some α ∈ (−∞, 1) and t ∈ T. Then
(18) holds. If t is such that (17) holds with σ = 1, assume
lim
j→∞
tj
= µ ∈ (1, ∞) .
j
Then for every r ∈ N+ , there exists a positive, finite constant Mr (t), and positive
random variables ζ1 , . . . , ζr such that
a.s.
Mr (t)n−r/γ deg1 (n) · · · degr (n) −−−→ ζ1 · · · ζr
n→∞
where
γ=
µ−α
.
µ−1
The
Pr mixed moments also converge: For any p1 , . . . , pr > −(1 − α)/2 with p̄ =
j=1 pj , there exists some Mp̄ (t) ∈ (0, ∞) such that
(19)
Mp̄ (t)E lim n−p̄/γ deg1 (n)p1 · · · degr (n)pr = E ζ1p1 · · · ζrpr .
n→∞
11
Furthermore,
(20)
ζj
d
1≤j≤r
= Ψj
r
Y
(1 − Ψi ) 1≤j≤r ,
i=j+1
where Ψ1 = 1 and (Ψj )j>1 are as in (5).
/
In the sub-linear regime, (18) agrees with and generalizes the result of [23] for exchangeable Gibbs partitions (though the proof uses different methods). In the linear
regime, the mixed moments of the scaled degrees also converge to those of products of independent beta random variables. However, the result does not completely
describe the joint distributions, due to the presence of the unknown scaling terms
Mp̄ (t). These terms depend on the moments p1 , . . . , pr , and on t, and express the
randomness remaining, for large k(n), in Wj,k(n) after the part that scales with n is
removed; in particular, they result from early fluctuations of the process. Section 4
provides stronger results in several cases for which these terms are well-behaved.
3.3. Neutrality. It was noted in Section 2 that the map Φ−1 from graphs to
partitions translates results on graphs into results on partitions. Conversely, one
can transfer properties from partitions to graphs. A sequence (X1 , X2 , . . .) of random
variables is neutral-to-the-left (NTL) if the relative increments
X1 ,
Xj
X2
, . . . , Pj
,...
X1 + X2
i=1 Xi
are independent random variables in (0,1) [17, 23]. If Π is an exchangeable partition,
Griffiths and Spanò [23] show that the limiting relative block sizes of Π are NTL if
and only if Π is an exchangeable Gibbs partition. If so, the random graph Φ(Π) has a
limiting degree sequence D∞ that is NTL. Due to the representation in Theorem 2,
this property generalizes beyond the exchangeable case:
Corollary 8.
The limiting degree sequence D∞ of an (α, t)-graph is NTL.
/
3.4. Sparsity and power law degrees. Suppose Gn is the subgraph of an (α, T )graph G, given by the first n edges. Since Gn is finite, one can sample a vertex
uniformly at random from its vertex set and report its degree Dn . One can then
ask whether the sequence of random degrees Dn converges in distribution to some
limiting variable D. We show in this section that that is indeed the case for (α, T )graphs, under some regularity conditions. We also show how the degree distribution
is related to the sparsity, or, equivalently, the edge density, of (Gn ).
The sequence (Gn ) is defined to be ε-dense if
(21)
lim sup
n→∞
n
= cε > 0
|V(Gn )|ε
for some
ε≥1.
12
If ε < 2, the graph sequence is typically called sparse; when ε ≥ 2, the sequence is
dense. Note that ε > 2 is only possible for multigraphs. The level of sparsity follows
from σ: Graph models in the linear regime correspond to ε = 1 [9, 5, 2]; graph models
in the sub-linear regime with σ > 21 have appeared in the literature [12, 41, 14, 11],
with 1 < ε < 2. See Section 4 for examples.
For functions a and b, we use the notation
n↑∞
a(n) ∼ b(n)
:⇔
lim a(n)/b(n) → 1 .
n→∞
The sequence (Gn ) has power law degree distribution with exponent η > 1 if
pd (n) =
(22)
md (n)
d↑∞
−−−→ pd ∼ L(d)d−η
|V(Gn )| n→∞
for all large d ,
for some slowly varying function L(d), that is, limx→∞ L(rx)/L(x) = 1 for all r > 0
[6, 20].
In the sub-linear regime, the degree distribution follows from results due to Pitman and Hansen [36, Lemma 3.11], see also [22], on the limiting block sizes of
exchangeable random partitions (see Section 4.2 for more details). In particular, if
(17) is satisfied by an (α, t)-graph Gα = Φ(Πα ) for σ = α ∈ (0, 1), then there exist an
exchangeable random partition Π and a positive, finite random variable S such that
a.s.
|V(Φ(Π2n ))|/nα −−−→ S, Π = Πα and S = µα . The limiting degree distribution is
n→∞
pαd
(23)
=
α
Γ(d − α)
Γ(d + 1)Γ(1 − α)
α
d−(1+α) .
Γ(1 − α)
d↑∞
∼
In the linear regime, σ = 1, with limiting mean interarrival time µ1 . We show (see
Appendix A.4) that the resulting limiting degree distribution is a generalization of
the classical Yule–Simon distribution (which corresponds to α = 0) [42, 39, 19],
(24)
pγd
=
γ
Γ(d − α)Γ(1 − α + γ)
Γ(d + 1 − α + γ)Γ(1 − α)
d↑∞
∼
γ
Γ(1 − α + γ) −(1+γ)
d
,
Γ(1 − α)
where γ := µµ11−α
−1 , as in (19).
The tail behavior of the two distributions (23), (24) partition the range of possible
values of the power law exponent, as summarized by the following theorem.
Theorem 9.
Let G be a random (α, T )-graph for some α ∈ (−∞, 1) and T ∈ T.
If
a.s.
Tj j −1/σ −−−→ µσ
j→∞
for some
σ ∈ (0, 1]
and
1 < µσ < ∞ ,
then G has ε-density with ε = 1/σ. If σ = 1, assume that E[∆j ] = µ1 , Var(∆j ) < ∞
2 |i − j|−`∆ for all i, j > 1, some C 2 ≥ 0, and
for all j ∈ N+ , and |Cov(∆i , ∆j )| ≤ C∆
∆
13
some `∆ > 0. Then the degree distribution converges asymptotically,
(
pαd
if σ = α ∈ (0, 1)
md (n)
p
−−−→
,
|V(Gn )| n→∞ pγd
if σ = 1
which for large d follows a power law with exponent
(
1 + α ∈ (1, 2)
if σ = α ∈ (0, 1)
η=
.
1 + γ ∈ (2, ∞) if σ = 1
/
The distributions (23), (24) have the following representation, which is useful for
generating realizations from those distributions.
Corollary 10. Let G be a random (α, T )-graph for some α ∈ (−∞, 1) and
T ∈ T satisfying the conditions of Theorem 9. Then the degree Dn of a vertex
sampled uniformly at random from Gn converges in distribution to D0 , where D0 is
sampled as
(
Beta(α, 1 − α) if σ = α ∈ (0, 1)
0
D ∼ Geom(B) for B ∼
.
Beta(γ, 1 − α) if σ = 1
/
This representation can be refined further: The proof of Theorem 9 shows, by
extending techniques introduced by Berger, Borgs, Chayes, and Saberi [5], that the
neighborhood of a random vertex can be coupled to a Poisson point process on the
unit interval. That yields the following representation:
Corollary 11. Let G be a random (α, T )-graph for some α ∈ (−∞, 1) and
T ∈ T satisfying the conditions of Theorem 9. Then the degree Dn of a vertex
sampled uniformly at random from Gn converges in distribution to D0 , where D0 is
sampled as
(
Beta(α, 1) if σ = α ∈ (0, 1)
1−B
0
G1−α
for B ∼
.
D ∼ Poisson
B
Beta(γ, 1) if σ = 1
/
Remark 12. Based on the fact that (1 − B)/B is distributed as a so-called beta
prime random variable, additional distributional identities may be deduced. To give
one, let G1 , Gα , and Gγ be independent Gamma random variables. Then one can
replace (1 − B)/B above by G1 /Gα (for σ = α < 1) or G1 /Gγ (for σ = 1).
/
14
Fig 1. Examples of (α, T )-graphs generated using different arrival time distributions. Each graph
has 500 edges. Left: arrival times generated by CRP(α, θ), with α = 0.1, θ = 5. Middle: interarrival
times are i.i.d. Poisson+ (2), with α = 0.1. Right: interarrival times are i.i.d. Geom(0.25), with
α = 0.5.
3.5. A note on almost surely connected graphs. Suppose one requires each graph
in the evolving sequence (Gn ) drawn from an (α, T )-graph to be almost surely
connected. That holds if and only if T ∈ T2 , i.e. if each arrival time after T1 = 1 is
even. A simple way to generate T ∈ T2 is to sample ∆2 , ∆3 , . . . as in the generation
of general T ∈ T, and to set
(25)
T2 = 2∆2 ,
Tk = Tk−1 + 2∆k
for k > 2 .
In the sub-linear regime, doubling the interarrival times does not affect the degree
asymptotics. In the linear regime, the change has noticeable affect. For example,
suppose the variables ∆k above are drawn i.i.d. from some probability distribution
on N+ with mean µ. Then by Theorem 9, the limiting degree distribution has power
law exponent η2 = 1 + 2µ−α
2µ−1 . For T 6∈ T2 , the upper limit of η is ∞, no matter the
value of α; for T ∈ T2 , one has η2 < 3 − α. Hence, if α > 0, then η2 ∈ (2, 3), implying
that the limiting degree distribution has finite mean, but unbounded variance for
any µ.
4. Examples. We next discuss several subclasses of (α, T )-graphs. One is obtained by fixing all interarrival times to the same, constant value (Section 4.1). This
includes the Barabási–Albert random tree as a special case. Other subclasses can
be obtained by imposing exchangeability assumptions. One is that the vertex assignment variables Ln form an exchangeable sequence, and hence that the induced
random partition Φ−1 (G) is exchangeable (Section 4.2). This subclass overlaps with
the class of “edge exchangeable” graphs [14, 11, 28]. If the interarrival times ∆k
are exchangeable (Section 4.3), the induced partition is not exchangeable. This case
includes a version of the random graph model of [1, 2, 13].
15
4.1. Barabási–Albert trees and graphs with constant interarrival time. The basic
preferential attachment model popularized by [4] generates a random graph as
follows: With parameter d ∈ N+ , start with any finite connected graph. At each
step, select d vertices in the current graph independently from P0 in (3), add a new
vertex, and connect it to the d selected vertices (multiple connections are allowed).
The Barabási–Albert model can be expressed in terms of a sequence (L1 , L2 , . . .)
with given arrival times as follows: Start, say, with a graph consisting of a single vertex and d self-loops. Thus, tk = 1 and L1 = . . . = L2d = 1. Each new vertex requires 2d stubs, so tk+1 = tk + 2d. At time tk+1 − 1 = (2d)k, just before the
(k + 1)-st vertex arrives, the graph Gkd = (Ln )n≤2kd has k vertices and kd edges.
For tk ≤ n < tk+1 , we then set
Ln = k
if n odd
and
Ln ∼ P0 ( • ; G(k−1)d )
if n even.
The single vertex with self-loops is chosen as a seed graph here only to keep notation
reasonable. More generally, any graph with k vertices and n edges can be encoded
in the variables L1 , . . . , L2n and the first k arrival times.
When d = 1, the result is a tree (with a loop at the root). When d ≥ 2, the above
sampling scheme does not produce an (α, t)-graph. However, the following modified
sampling scheme produces an (α, t)-graph with ∆j = 2d for all j > 1. Start as before
with a graph consisting of a single vertex and d self-loops. When the k-th vertex
arrives at time tk = 2(k − 1)d + 1, set Ltk = k and for tk + 1 ≤ n < tk+1 , set
Ln ∼ Pα ( • ; (Li )i<n ) .
The modified sampling scheme differs from the basic preferential attachment model
in that it updates the degrees after each step, allows loops, and does not require that
each vertex begin with d edges. Although the connectivity properties of the resulting
graph may be substantially different from the Barabási–Albert model, the degree
properties are similar to modifications that have been considered by [5, 32]. In this
case, the results of Section 3.2 can be strengthened, showing that the scaled degrees
converge to random variables that satisfy distributional relationships generalizing
those of the beta-gamma algebra (8), as discussed by Dufresne [18]. (See also [27].)
These relationships emerge due to the behavior of Wj,k as k → ∞, which separates
into two pieces: A deterministic scaling factor, and the random variables that appear
below.
Proposition 13. Let G be an (α, t)-graph with α ∈ (−∞, 1) such that for some
d ∈ N+ , tj = 2d(j − 1) + 1 for all j ≥ 1. Then for every r ∈ N+ ,
n
2m
− 2d−1
2d−α
a.s.
(degj (n))1≤j≤r −−−→ (ξj )1≤j≤r .
n→∞
16
The vector of random variables (ξj )1≤j≤r satisfies the following distributional identities: Denote ᾱ := 2d − α, and define
Zr =
2d−1
Y
1/ᾱ
Gr+1−i/ᾱ
Zr0 =
and
i=1
2d−1
Y
1/ᾱ
G1−i/ᾱ
and
Zr00 = Zr0
i=1
r−1
Y
Bkᾱ,1−α ,
k=1
where all of the random variables are independent of each other and of (ξj )j . Then
with Ψ1 = 1, and (Ψj )j>1 ∼ Beta(1 − α, (j − 1)(2d − α)), the following distributional identities hold:
r
Y
d
(26)
Zr · ξj 1≤j≤r = Grᾱ · Ψj
(1 − Ψi ) 1≤j≤r
i=j+1
(27)
Zr0
· ξj
d
1≤j≤r
= Grᾱ
r
Y
Bkᾱ−2d+1,2d−1 · Ψj
(28)
· ξj
d
1≤j≤r
= Grᾱ−2d+1
r−1
Y
Bkᾱ−2d+1,2d−1 · Ψj
Zr00 · ξj
d
1≤j≤r
= G1−α · Ψj
r
Y
(1 − Ψi )
1≤j≤r
i=j+1
k=1
(29)
(1 − Ψi ) 1≤j≤r
i=j+1
k=1
Zr0
r
Y
r
Y
(1 − Ψi ) 1≤j≤r .
i=j+1
/
Remark 14. For a gamma random variable Ga , Gab has so-called generalized
gamma distribution, denoted GGa(a/b, 1/b). Hence, Zr above is equal to the product
of GGa((r + 1)ᾱ − i, ᾱ) random variables, and similarly for Zr0 . Generalized gamma
random variables also appear in the limits of the preferential attachment models in
[32, 37], and arise in a range of other applications [31].
/
Results on power law degree distributions in preferential attachment models are
numerous [e.g. 4, 1, 2, 5]. It is well-known that the degree distribution of the
Barabási–Albert tree exhibits a power law with exponent η = 3 [4, 8], which agrees
with the following implication of Theorem 9.
Corollary 15. For the constant interarrival time model considered above, the
degree distribution converges to (24) with γ = (2d − α)/(2d − 1). In particular, the
α-weighted Barabási–Albert tree has power law exponent η = 3 − α.
/
4.2. Graphs with exchangeable vertex assignments. Suppose G is a random graph
such that the random partition Π = Φ−1 (G) is exchangeable (see [36] for more on
exchangeable partitions). Equivalently, the vertex assignments L1 , L2 , . . . are exchangeable, and there is hence a random probability measure µ on N such that
(30)
L1 , L2 , . . . | µ ∼iid µ .
17
We first note an implication of Theorem 9. Recall that a random graph is sparse if
its density (21) is ε < 2.
Corollary 16. If a graph generated by an exchangeable partition is sparse and
has a power law degree distribution, then σ > 1/2, and hence η ∈ (3/2, 2).
/
Recall from Section 2 that an exchangeable Gibbs partition is an exchangeable
random partition Π such that the probability of any finite restriction Πn can be
written as
P(Πn = {deg1 (n), . . . , degk (n)}) = Vn,k
k
Y
Γ(degj (n) − α)
,
Γ(1 − α)
j=1
for a suitable sequence of weights Vn,k . Griffiths and Spanò [23] studied the block
arrival times (called, in that context, the record indices) of exchangeable Gibbs
partitions. For the random graph induced by such partitions, their results show that
(31)
∞
Y
1
a.s.
d
degj (n) −−−→ Pj = Ψj
(1 − Ψi ) ,
n→∞
n
i=j+1
where Ψj is distributed as in (5). (This result is also contained in Theorem 7.) They
prove that an exchangeable random partition is of Gibbs form if and only if the
sequence (Pj )j≥1 is NTL conditioned on (Tj )j≥1 ; this result has implications for
some recent network models.
Crane and Dempsey [14] and Cai, Campbell, and Broderick [11] call a random
graph ((L1 , L2 ), . . .) edge exchangeable if there is some random probability measure
ν on N2 such that
(L1 , L2 ), (L3 , L4 ), . . . | ν ∼iid ν .
Janson [28] refers to such a graph as being rank one if ν = µ ⊗ µ for some random
probability measure on N, which is just (30). Thus, rank one edge exchangeable
graphs are precisely those corresponding to exchangeable random partitions via Φ.
The intersection of edge exchangeable and (α, T )-graphs are precisely those (α, T )graphs that have exchangeable vertex assignments, in which case Π is an exchangeable Gibbs partition. That includes the case Π ∼ CRP(α, θ) above, for which [14]
call G = Φ(Π) the Hollywood model.
Proposition 17. Let G be a rank one edge exchangeable graph, and let D∞ be
the limiting degree proportions n−1 (deg1 (n), deg2 (n), . . . ). Then D∞ is NTL if and
only if G is distributed as a (α, T )-graph, where T distributed as in (13), in which
case D∞ is distributed as in (31).
/
The results of Section 3.2 specialize for G = Φ(Π) where Π has law CRP(α, θ). In
particular, consider conditioning on the first r arrival times, rather than all arrival
18
times. As Proposition 18 shows, the scaled degrees have the same basic structure as
in Section 3.2, but E[Wr,∞ ] is captured by a single beta random variable.
Proposition 18. Let G be an (α, T )-graph for fixed α ∈ (0, 1) such that T
are the arrival times induced by a CRP(α, θ) partition process (10). Then for every
r ∈ N+ , conditioned on T1 , . . . , Tr ,
a.s.
n−1 (degj (n))1≤j≤r | T1 , . . . , Tr −−−→ (ξ1 , . . . , ξr ) ,
n→∞
where
(32)
ξj
r
Y
d
1≤j≤r
= BTr −rα,θ+rα · Ψj
(1 − Ψi ) 1≤j≤r ,
i=j+1
with Ψ1 = 1, Ψj ∼ Beta(1−α, Tj −1−(j−1)α) and BTr −rα,θ+rα mutually independent
random variables for j ≥ 1. This implies the joint distributional identity
(33)
GTr +θ · ξj
d
1≤j≤r
= GTr −rα · Ψj
r
Y
(1 − Ψi )
1≤j≤r
.
i=j+1
and the marginal identities for all j > 1
d
ξj = B1−α,Tj −1+θ+α
d
ξj+1 | ξj , ∆j+1 = ξj BTj +θ,∆j+1 .
/
The random variable ξj is independent of j, given Tj . Among all (α, T )-graphs
derived from exchangeable Gibbs partitions, this property characterizes those derived from CRP(α, θ) partitions, and stems from the arrival time distribution (see
Section 4.4).
Proposition 19. Let G be an (α, T )-graph such that T are the arrival times
induced by an exchangeable Gibbs partition Π. For any j ≥ 1, the marginal distribution of ξj conditioned on jth arrival time depends only on Tj if and only if Π has
law CRP(α, θ).
/
4.3. Graphs with exchangeable interarrival times. We next consider (α, T )-graphs
for which the interarrival times ∆j = Tj − TP
j−1 are exchangeable. An immediate
consequence of exchangeability is that k −1 j≤k ∆j → µ almost surely for some
constant µ in [1, ∞]. Theorem 7 implies:
Corollary 20. If µ is finite, the limiting degrees scale as n1/γ in (19), where
γ = (µ − α)/(µ − 1). If Var(∆j ) is finite for all j, then the degree distribution converges to (24).
/
19
Stronger results hold when the interarrivals are i.i.d. geometric variables, corresponding to the Yule–Simon model [42, 39]. Recall that a positive random variable Mσ is said to have Mittag–Leffler distribution with parameter σ ∈ (0, 1) if
Mσ = Zσ−σ , where Zσ is a positive σ-stable random variable, characterized by the
σ
Laplace transform E[e−λZ ] = e−λ and density fσ (z). See [36, 25] for details. Define Zσ,θ for θ > −σ as a random variable with the polynomially tilted density
−σ
fσ,θ ∝ z −θ fσ (z), and let Mσ,θ = Zσ,θ
. We denote the law of Mσ,θ by ML(σ, θ),
which is known as the generalized Mittag–Leffler distribution [35, 25].
Proposition 21. Let G be an (α, T )-graph with α = 0, and T constructed from
i.i.d. Geom(β) interarrival times, for β ∈ (0, 1). Then for every r ∈ N+ , conditioned
on T1 , . . . , Tr ,
a.s.
n−(1−β) (degj (n))1≤j≤r | T1 , . . . , Tr
−−−→
(ξ1 , . . . , ξr ) ,
n→∞
where
ξj
(34)
d
1≤j≤r
= M1−β,Tr −1 BTr ,(Tr −1)
with M1−β,Tr −1 , BTr ,(Tr −1)
β
1−β
· Ψj
β
1−β
r
Y
(1 − Ψi )
,
1≤j≤r
i=j+1
, Ψ1 = 1 and Ψj ∼ Beta(1, Tj − 1) mutually indepen-
dent random variables for j ≥ 1. This implies the joint distributional identity
GT1−β
r
· ξj
d
1≤j≤r
= GTr · Ψj
r
Y
(1 − Ψi )
1≤j≤r
,
i=j+1
and the marginal identities for j > 1
(35)
1−β
ξj = M1−β B1,T
j −1
(36)
ξj+1 | ξj , ∆j+1 = ξj BT1−β
j ,∆j+1
(37)
d
d
d
d
ξj = M1−β,Tj −1 BTj ,(Tj −1)
β
1−β
Ψj = M1−β,Tj −1 B
1,
d
Tj −1
1−β
= M1−β,Tj B
1,
Tj −1+β
1−β
∆j+1
(38)
(39)
d
ξj+1 | ξj , ∆j+1 = ξj B
Tj ∆j+1
,
1−β 1−β
Y
i=1
B Tj −1+i−β
1−β
β
, 1−β
ξj GT1−β
= G1 .
j
d
/
Peköz, Röllin, and Ross [33] consider the following two-color Pólya P
urn: Let
∆1 , ∆2 , . . . be drawn i.i.d. from some distribution P∆ , and define Tj = ji=1 ∆j .
Starting with w white balls and b black balls, at each step n 6= Tj , a ball is drawn
and replaced along with another of the same color. On steps n = Tj , a black ball is
20
P{n + 1 is arrival time|Gn }
depends on
Φ−1 (G) is
(α, T )-urn
L(Φ−1 (G))
L(G)
L(∆k )
n
n, #vertices in Gn
n, #vertices in Gn , degrees
deterministic
independent
n + 1 − Tk(n)
yes
yes
no
yes
yes
yes
CRP(θ)
Gibbs partition
exch. partition
–
Yule–Simon process
(α, T )-urn
Hollywood model
⊂ rank one edge exch.
rank one edge exch.
PA tree
ACL [1, 2]
(α, T )-graph
(10)
(13)
–
δ2
Geom(β)
i.i.d.
Table 1
Classification of different models according to which statistics of Gn determine the probability that
a new vertex is observed at time n + 1.
added to the urn. Of interest is the distribution of the number of white balls in the
urn after n steps.
In the language of (α, T )-graphs, consider a seed graph Gw+b with kw+b < w + b
vertices and w + b edges arranged arbitrarily, the only constraint being that there
exists a bipartition Vw ∪ Vb = V(Gw+b ) so that the total degree of the vertices in
Vw is Dw (w + b) = w, and of those in Vb is Db (w + b) = b. For T constructed from
i.i.d. interarrivals, Dw (n) corresponds to the number of white balls after n steps.
For interarrivals drawn i.i.d. from the geometric distribution, the following result
characterizes the limiting distribution of Dw (n), which was left as an open question
by Peköz, Röllin, and Ross [33].
Proposition 22. Let Dw (n) be the number of white balls in the Pólya urn
with immigration from [33] starting with w white balls and b black balls, where the
immigration times have i.i.d. Geom(β) distribution. Then
n−(1−β) Dw (n)
a.s.
−−−→
n→∞
d
ξw,w+b = Bw,b Bw+b,(w+b−1)
β
1−β
M1−β,w+b−1 ,
which implies the distributional identities
(40)
d
ξw,w+b = Bw, (w−1)β+b M1−β,w+b−1
1−β
(41)
d
ξw,w+b = Bw, wβ+b M1−β,w+b
1−β
(42)
1−β
ξw,w+b Gw+b
d
= Gw .
/
4.4. Classification by arrival time probabilities. De Blasi, Favaro, Lijoi, Mena,
Prünster, and Ruggiero [16] classify exchangeable partitions according to the quantities on which the probability of observing a new block in draw n + 1 depends [16,
Proposition 1], conditionally on the partition observed up to time n. This classification can be translated to random graphs via the induced partition Φ−1 (G), and can
21
be extended further since partitions induced by (α, T )-graphs need not be exchangeable: See Table 1. One might also consider a sequence of interarrival distributions
indexed by the number of vertices, yielding a bespoke generalization of the last row,
where the probability of a new vertex depends on n + 1 − Tk(n) and the number of
vertices.
Acknowledgments. We are grateful to Nathan Ross for helpful comments on
the manuscript. BBR is supported by the European Research Council under the
European Union’s Seventh Framework Programme (FP7/2007–2013) / ERC grant
agreement no. 617071. PO was supported in part by grant FA9550-15-1-0074 of
AFOSR.
References.
[1] W. Aiello, F. Chung, and L. Lu. A random graph model for power law graphs. Experiment.
Math., 10(1):53–66, 2001.
[2] W. Aiello, F. Chung, and L. Lu. Random evolution in massive graphs. In J. Abello, P. M.
Pardalos, and M. G. C. Resende, editors, Handbook of Massive Data Sets, volume II, pages
97–122, 2002.
[3] R. Arratia, L. Goldstein, and F. Kochman.
Size bias for one and all.
08 2013.
https://arxiv.org/abs/1308.2729.
[4] A.-L. Barabási and R. Albert. Emergence of scaling in random networks. Science, 186(5439):
509–512, 1999.
[5] N. Berger, C. Borgs, J. T. Chayes, and A. Saberi. Asymptotic behavior and distributional
limits of preferential attachment graphs. Ann. Probab., 42(1):1–40, 01 2014.
[6] N. H. Bingham, C. M. Goldie, and J. L. Teugels. Regular Variation, volume 27 of Encyclopedia
of Mathematics and its Applications. Cambridge University Press, 07 1989.
[7] D. Blackwell and J. B. MacQueen. Ferguson distributions via pólya urn schemes. Ann. Statist.,
1(2):353–355, 03 1973.
[8] B. Bollobás, O. Riordan, J. Spencer, and G. Tusnády. The degree sequence of a scale-free
random graph process. Random Structures & Algorithms, 18(3):279–290, 2001.
[9] B. Bollobás, S. Janson, and O. Riordan. The phase transition in inhomogeneous random
graphs. Random Structures & Algorithms, 31(1):3–122, 2007.
[10] T. Broderick, M. I. Jordan, and J. Pitman. Beta processes, stick-breaking and power laws.
Bayesian Analysis, 7(2):439–476, 06 2012.
[11] D. Cai, T. Campbell, and T. Broderick. Edge-exchangeable graphs and sparsity. In D. D.
Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural
Information Processing Systems 29, pages 4242–4250. Curran Associates, Inc., 12 2016.
[12] F. Caron and E. B. Fox. Sparse graphs using exchangeable random measures. Journal of the
Royal Statistical Society: Series B (Statistical Methodology), 79(5):1–44, 2017.
[13] F. Chung and L. Lu. Complex Graphs and Networks. (CBMS Regional Conference Series in
Mathematics). American Mathematical Society, 2006.
[14] H. Crane and W. Dempsey. Edge exchangeable models for network data. 03 2016. URL
https://arxiv.org/abs/1603.04571.
[15] D. J. Daley and D. Vere-Jones. An Introduction to the Theory of Point Processes, volume II
of Probability and its Applications. Springer-Verlag New York, 2008.
[16] P. De Blasi, S. Favaro, A. Lijoi, R. H. Mena, I. Prünster, and M. Ruggiero. Are Gibbs-type
priors the most natural generalization of the Dirichlet process? IEEE Transactions on Pattern
Analysis and Machine Intelligence, 37(2):212–229, 02 2015.
[17] K. Doksum. Tailfree and neutral random probabilities and their posterior distributions. Ann.
Probab., 2(2):183–201, 04 1974.
22
[18] D. Dufresne. G distributions and the beta-gamma algebra. Electron. J. Probab., 15:2163–2199,
2010.
[19] R. Durrett. Random Graph Dynamics. Cambridge Series in Statistical and Probabilistic
Mathematics. Cambridge University Press, New York, NY, USA, 2006.
[20] W. Feller. An Introduction to Probability Theory and Its Applications, volume 2. Wiley, 2nd
edition, 1971.
[21] A. Gnedin and J. Pitman. Exchangeable Gibbs partitions and stirling triangles. Journal of
Mathematical Sciences, 138(3):5674–5685, 2006. ISSN 1573-8795.
[22] A. Gnedin, B. Hansen, and J. Pitman. Notes on the occupancy problem with infinitely many
boxes: general asymptotics and power laws. Probab. Surveys, 4:146–171, 2007.
[23] R. C. Griffiths and D. Spanò. Record indices and age-ordered frequencies in exchangeable
Gibbs partitions. Electron. J. Probab., 12:1101–1130, 2007.
[24] R. van der. Hofstad. Random Graphs and Complex Networks, volume 1 of Cambridge Series
in Statistical and Probabilistic Mathematics. Cambridge University Press, 2016.
[25] L. F. James. Generalized Mittag–Leffler distributions arising as limits in preferential attachment models. 09 2015. URL http://arxiv.org/abs/1509.07150.
[26] S. Janson. Limit theorems for triangular urn schemes. Probability Theory and Related Fields,
134(3):417–452, 2006.
[27] S. Janson. Moments of gamma type and the Brownian supremum process area. Probab.
Surveys, 7:1–52, 2010.
[28] S. Janson. On edge exchangeable random graphs. 02 2017. URL https://arxiv.org/abs/
1702.06396.
[29] S. V. Kerov. Coherent random allocations, and the Ewens–Pitman formula. Journal of Mathematical Sciences, 138(3):5699–5710, 2006.
[30] T. F. Móri. The maximum degree of the Barabási–Albért random tree. Comb. Probab. Comput., 14(3):339–348, May 2005.
[31] E. A. Peköz, A. Röllin, and N. Ross. Generalized gamma approximation with rates for urns,
walks and trees. Ann. Probab., 44(3):1776–1816, 05 2016.
[32] E. A. Peköz, A. Röllin, and N. Ross. Joint degree distributions of preferential attachment
random graphs. Advances in Applied Probability, 49(2):368–387, 2017.
[33] E. A. Peköz, A. Röllin, and N. Ross. Pólya urns with immigration at random times. Bernoulli,
forthcoming.
[34] J. Pitman. Some developments of the Blackwell-MacQueen urn scheme, volume Volume 30 of
Lecture Notes–Monograph Series, pages 245–267. Institute of Mathematical Statistics, Hayward, CA, 1996.
[35] J. Pitman. Poisson-Kingman partitions, volume Volume 40 of Lecture Notes–Monograph Series,
pages 1–34. Institute of Mathematical Statistics, Beachwood, OH, 2003.
[36] J. Pitman. Combinatorial Stochastic Processes, volume 1875 of Ecole d’Eté de Probabilités de
Saint-Flour. Springer-Verlag Berlin Heidelberg, 2006.
[37] J. Pitman and M. Racz. Beta-gamma tail asymptotics. Electron. Commun. Probab., 20:1–7,
2015.
[38] D. Revuz and M. Yor. Continuous martingales and Brownian motion. Springer Science &
Business Media, 3rd edition, 1999.
[39] H. A. Simon. On a class of skew distribution functions. Biometrika, 42(3-4):425–440, 1955.
[40] F. G. Tricomi and A. Erdélyi. The asymptotic expansion of a ratio of gamma functions. Pacific
J. Math., 1(1):133–142, 1951.
[41] V. Veitch and D. M. Roy. The class of random graphs arising from exchangeable random
measures. 12 2015. URL http://arxiv.org/abs/1512.03099.
[42] G. U. Yule. A Mathematical Theory of Evolution, Based on the Conclusions of Dr. J. C. Willis,
F.R.S. Philosophical Transactions of the Royal Society of London B: Biological Sciences, 213
(402-410):21–87, 1925. ISSN 0264-3960.
23
A. Proofs.
A.1. Proof of Theorem 2.
Proof. Let G(α, T ) = ((L1 , L2 ), (L3 , L4 ), . . . ) be an (α, T )-graph, and let kn
denote the number of vertices after n edge ends have been drawn. For notational
convenience, let Dj (n) := degj (n) be the degree of the jth vertex at step n. Then
from the sequence of laws Pα (k; l1 , . . . , lj ) in (3), the probability of a particular
sequence is
P[Gn/2 (α, T ) | T = t] =
kn
Y
Γ(tj − jα)Γ(Dj (n) − α)
1
.
Γ(n − kn α)
Γ(tj − 1 − (j − 1)α + δ1 (j))Γ(1 − α)
j=1
Now let Hn (α, T ) denote the first n elements of the sequence H(α, T ), so that
P[Hn/2 (α, T ), (Ψj )2≤j≤kn | T = t] =
D (n)−α−1
kn
Y
Γ(tj − jα)Ψj j
(1 − Ψj )D̄j−1 (n)−(j−1)α−1
Γ(1 − α)Γ(tj − 1 − (j − 1)α)
i=2
.
Hence, by marginalizing (Ψj )2≤j≤kn , it follows that
Z
P[Hn/2 (α, T ) | T = t] =
P[Hn/2 (α, T ), (Ψj )2≤j≤kn | T = t] dΨ2 · · · dΨkn
[0,1]kn
= P[Gn/2 (α, T ) | T = t] .
The equality is true for all n ∈ N+ , and t ∈ T, implying that P[G(α, T )] = P[H(α, T )].
A.2. Proof of Proposition 5. The proof will make repeated use of the expectation
of Wj,k as k → ∞. Observe that
− log E[Wj,k ] =
k
X
log 1 +
i=j+1
1−α
ti − 1 − (i − 1)α
k
X
=
i=j+1
1−α
+ O(i−2 )
ti − 1 − (i − 1)α
By assumption in the linear regime, ti − 1 − (i − 1)α is well-approximated by (i −
1)(µ1 −α) for i large enough, resulting in a finite error Cj,k that converges as k → ∞.
Hence,
k−1
1 − α X −1
1−α
k−1
−2
− log E[Wj,k ] = Cj,k +
i + O(i ) = Cj,k +
log
+ O(j −1 ) ,
µ1 − α
µ1 − α
j
i=j
and therefore for all j > 1,
(43)
E[Wj,k ] = Aj
j
k−1
1−α
µ1 −α
as
k→∞.
24
Clearly, this converges to zero for each j. Similarly, in the sub-linear regime,
k−1
X
1−α
+ O(i−2 )
1/σ − iα
µ
i
σ
i=j
1−σ
σ(1 − α)
µσ − α(k − 1)− σ
0
= Cj,k
+
log
+ O(j −1 ) .
− 1−σ
α(1 − σ)
µσ − αj σ
− log E[Wj,k ] =
Since
1−σ
σ
0
Cj,k
+
> 0,
(44) E[Wj,k ] =
A0j
1−
σ(1−α)
α
µσ j
1−σ
σ
− σ(1−α)
α
α(1−σ)
1−
µσ (k − 1)
α(1−σ)
as k → ∞ .
1−σ
σ
Clearly, this converges to something non-zero for each j.
Proof of Proposition 5. Define Mj,k := Wj,k /E[Wj,k ]. Since E[Mj,k+1 | Mj,k ] = Mj,k
and E[Mj,j+1 = 1], Mj,k is a nonnegative martingale with mean 1 for k > j; it therefore converges almost surely to a random variable Mj,∞ . Hence,
Wj,∞ := lim Wj,k = Mj,∞ lim E[Wj,k ] .
k→∞
k→∞
2 ] = 1 + O(k −1 ), and therefore M
Simple calculations show that E[Mj,k
j,k is bounded
in L2 and also in L1 . Hence, Wj,∞ exists if and only if limk→∞ E[Wj,k ] exists. By
(43) and (44), that is the case.
By Markov’s inequality, for any λ > 0,
(
− 1−α
1
O(k µ1 −α ) for σ = 1
P[Wj,k ≥ λ] ≤ E[Wj,k ] ≤
,
λ
O(λ−1 ) for σ < 1
indicating that Wj,∞ = 0 in the linear case. On the other hand,
−1
−1
P[Wj,k ≤ λ] = P[Wj,k
≥ λ−1 ] ≤ λE[Wj,k
] = λE[Wj,k ]−1
k
Y
i=j+1
(
1−α
O(k µ1 −α ) for σ = 1
≤
O(λ) for σ < 1
1
ti −iα
1
ti −1−(i−1)α
1−
1−
,
indicating that Wj,∞ > 0 in the sub-linear case.
A.3. Proofs for Section 3.2. The proof of Theorem 7 uses martingale methods
adapted and extended from those used by Móri [30]. See also [24, 19]. For notational
convenience, we let Dj,n = Dj (n). Fix r ∈ N+ and p = (p1 , p2 , . . . ) such that
25
pj > −(1 − α) for each 1 ≤ j ≤ r, and pj = 0 for j > r. Let p̄j =
p̄0 = 0 and p̄j = p̄r := p̄ for j > r. For fixed t ∈ T, define
Pj
i=1 pi ,
with
(45)
Zn (p, t) :=
kn
r
Y
Γ(Dj,n − α + pj ) Y
Γ(n − kn α)
Γ(tk − 1 − (j − 1)α)Γ(tk − jα + p̄k )
.
Γ(n − kn α + p̄kn )
Γ(Dj,n − α)
Γ(tk − 1 − (k − 1)α + p̄k )Γ(tk − kα)
j=1
k=r+1
The asymptotic behavior of Zn (p, t) is described by the following two lemmas, from
which Theorem 7 follows.
Lemma 23. Let Zn (p, t) be as in (45), with pj > −(1 − α) for each j ≥ 1. Then
Zn (p, t) is a nonnegative martingale with respect to (Fn )n≥Tr , the filtration generated
by the (α, t) sampling process, for n ≥ Tr , and therefore
a.s.
Zn (p, t) −−−→ ξ1p1 · · · ξrpr .
n→∞
Furthermore, if pj > −(1 − α)/2 for each j ≥ 1, then Zn (p, t) converges in L2 and
therefore also in L1 .
/
Proof. Zn (p, t) is nonnegative by construction. Furthermore,
kn+1
Y Γ(Dj,n+1 − α + pj )
Γ(n + 1 − kn+1 α)
| Fn := E[Rn+1 (p) | Fn ]
Γ(n + 1 − kn+1 α + p̄kn+1 )
Γ(Dj,n+1 − α)
j=1
Γ(n − kn α + p̄kn )Γ(n + 1 − (kn + 1)α)
1{tkn +1 = n + 1} ,
= Rn (p) 1{tkn +1 > n + 1} +
Γ(n + 1 − (kn + 1)α + p̄kn +1 )Γ(n − kn α)
E
from which it follows that E[Zn+1 (p, t) | Fn ] = Zn (p, t) for all n ≥ Tr . Furthermore,
E[Zn (p, t)] < ∞ and does not depend on n. To bound Zn (p, t) in L2 , observe that
by the properties of the gamma function [40],
(46)
cp̄ :=
Γ(n − kn α)
= n−p̄ (1 + O(n−1 ))
Γ(n − kn α + p̄kn )
as
n→∞,
which implies that c2p̄ /c2p̄ → 1. Using the fact that x 7→ Γ(x + y)/Γ(x) is increasing
in x for y ≥ 0, it follows that Zn (p, t)2 ≤ Zn (2p, t) (see [24, Section 8.7] for a similar
argument). Hence, E[Zn (p, t)2 ] ≤ E[Zn (2p, t)] < ∞ for pj > −(1 − α)/2.
Lemma 24. Let Zn (p, t) be as in (45), with pj > −(1 − α) for each j ≥ 1. Then
there exists a constant Mp̄ (t) such that
Zn (p, t) = Mp̄ (t)n−p̄/γ
r
Y
p
j
Dj,n
(1 + O(n−1 ))
as
n→∞.
j=1
/
26
Proof. The properties of the gamma function show that for each j ≤ r,
Γ(Dj,n − α + pj )
pj
= Dj,n
(1 + O(n−1 ))
Γ(Dj,n )
as
n→∞.
Hence, it suffices to examine the behavior of the product
kn
Γ(tj − 1 − (j − 1)α)Γ(tj − jα + p̄j )
Γ(1 − α + p1 ) Y
Γ(n − kn α)
Mn (p, t) :=
Γ(n − kn α + p̄kn ) Γ(1 − α)
Γ(tj − 1 − (j − 1)α + p̄j )Γ(tj − jα)
j=2
n−1
Y
m − km α
m − km α + p̄km 1{tkm +1 6= m + 1}
m=1
n−1
kY
n −1
Y
tj+1 − 1 − jα + p̄j
m − km α
=
:= Mn(1) (p, t)Mn(2) (p, t) .
m − km α + p̄km
tj+1 − 1 − jα
=
m=1
j=1
Taking the logarithm, the first term is
ln Mn(1) (p, t)
n−1
X
=
=
m=1
n−1
X
ln(m − km α) − ln(m − km α + p̄km )
ln 1 +
m=1
n−1
X
p̄km
p̄km
= C1 −
+ O(m−2 ) ,
m − km α
m − km α
m=1
where C1 captures the error in the approximation for terms where p̄km > m − km α,
which is finite because pj = 0 for j > r. By assumption when µ < ∞, for any > 0
there exists some finite K such that for all j ≥ K, |km − m/µ| ≤ . Therefore,
absorbing the additional error into C1 ,
ln Mn(1) (p, t) = C1 −
n−1
X
m=1
p̄
p̄µ
+ O(m−2 ) = C1 −
ln n + O(n−1 ) .
m(1 − α/µ)
µ−α
In the sub-linear regime (µ = ∞), the second term is −p̄ ln n. Similarly, for the
second term of Mn (p, t),
ln Mn(2) (p, t)
=
kX
n −1
j=1
kX
n −1
p̄j
p̄j
ln 1 +
= C2 +
+ O(j −2 )
tj+1 − 1 − jα
tj+1 − 1 − jα
= C2 +
j=1
kX
n −1
j=1
p̄j
p̄
+ O(j −2 ) = C2 +
ln n + O(n−1 ) .
j(µ − α)
µ−α
In the sub-linear regime, the second term is also O(n−1 ). Hence, letting Mp̄ (t) =
eC1 +C2 , the result follows.
27
Proof of Theorem 7. Equations (18) and (19) follow from Lemmas 23 and 24.
To establish (20), let p \ pj be p with the jth element set to zero, observe that
E[ζ1p1 · · · ζrpr ] = E[Ztr (p, t)]
Γ(1 − α + pr )Γ(tr − rα)Γ(tr − 1 − (r − 1)α + p̄r−1 )
= E[Ztr −1 (p \ pr , t)]
Γ(1 − α)Γ(tr − 1 − (r − 1)α)Γ(tr − rα + p̄r )
r
r
Y
Γ(1 − α + pj )Γ(tj − jα)Γ(tj − 1 − (j − 1)α + p̄j−1 ) Y
p
=
=
E[Ψj j (1 − Ψj )p̄j−1 ] .
Γ(1 − α)Γ(tj − 1 − (j − 1)α)Γ(tj − jα + p̄j )
j=2
j=2
A.4. Proofs for Section 3.4. In the sub-linear regime, the degree distribution
follows from results on exchangeable random partitions and can be read from [36,
Lemma 3.11]. The proof of Theorem 9 in the linear case relies on the following
lemma, which says that for large enough j, Wj,k is well-approximated by (j/k)1−1/γ
for all k > j, where γ = (µ − α)/(µ − 1).
Lemma 25. Let G be a random (α, T ) graph for some α ∈ (−∞, 1) and T ∈ T.
Assume that Var(∆j ) < V∆2 for all j ∈ N+ , that E[∆j ] = µ for all j > 1, and that
2 |i − j|−`∆ for all i, j > 1, some C 2 ≥ 0, and some ` > 0. Let
|Cov(∆i , ∆j )| = C∆
∆
∆
Wj,k be as in (6). Then for every λ ≥ 0, there exists some K < ∞ that does not
depend on n such that for all k ≥ K,
"
#
m 1−1/γ
P
sup
Wj,m
−1 ≤λ ≥1−λ.
k
k+1≤m≤n
/
Proof. Condition on T = t and define for fixed j
k
Y
Mj,k := Wj,k
j=k+1
Tj − jα
,
Tj − 1 − (j − 1)α
which is a martingale with mean 1 for k ≥ j + 1. By Doob’s maximal inequality [e.g.
38, Chapter II], for any c > 0 and j ≥ 1,
(47)
P
1
|Mj,m − 1| ≥ c | T ≤ 2 E[(Mj,k − 1)2 | T ] .
c
j+1≤m≤k
sup
The right-hand side is
(48)
1
1
E[(Mj,k − 1)2 | T ] = 2
2
c
c
2 | T]
E[Wj,k
E[Wj,k
1 1 1
−1 ≤ 2
+ O(j −2 ) .
2
| T]
c 1−αj
28
The product in Mj,k is
X
X
k
k
1−α
1−α
−2
exp
log 1 +
= exp
+ O(i )
Ti − 1 − (i − 1)α
Ti − 1 − (i − 1)α
i=j+1
i=j+1
1−1/γ
k
e−j,k 1 + O(j −1 )
=
j
where
k
X
j,k :=
i=j+1
(1 − α)
(1 − α)
−
.
(i − 1)(µ − α) Ti − 1 − (i − 1)α
Define ¯j,k to be sum of the absolute values of the terms in j,k . Because it is a
running sum of nonnegative terms, ¯j,k is a nonnegative submartingale. Therefore,
another application of Doob’s maximal inequality followed by Jensen’s inequality
shows that
P[
sup
|j,m | ≥ c ] ≤ P[
j+1≤m≤k
¯j,m ≥ c ]
sup
j+1≤m≤k
≤
k
X
E[¯
j,k ]
1
1
E[|Tj − (j − 1)µ − 1|]
≤
c
c (µ − α)
j2
i=j+1
≤
1
c (µ − α)
k
X
i=j+1
X
1/2
j
1
Var
(∆
−
µ)
i
j2
i=2
k
X
1
1
2 2−`∆ 1/2
(V 2 (j + C∆
j
))
c (µ − α)
j2 ∆
i=j+1
q
2
V∆2 + C∆
2/`+
2/`+
−(1+`+ /2)
≤
−
+
O(j
)
,
c (µ − α) j `+ /2 k `+ /2
≤
where `+ := min{`∆ , 1}. Let Aδj,k be the event that |j,m | < j −δ for each j + 1 ≤
m ≤ k and some fixed 0 < δ < `+ /2. Then
q
2
2 V∆2 + C∆
1
−δ
δ
−(1+`+ /2−δ)
P[Aj,k ] = P[ sup |j,m | < j ] ≥ 1 −
+ O(j
) .
`+ (µ − α) j `+ /2−δ
j+1≤m≤k
Hence, the probability of Aδj,k → 1 as j → ∞. Now, taking expectation with respect
to T of the left-hand side of (47) yields
m 1−1/γ
ET P
sup |Wj,k
(1 + j,m + O(j −1 )) − 1| ≥ c | T
j
k+1≤m≤k
m 1−1/γ
= P sup |Wj,k
− 1| ≥ c · · ·
j
j+1≤m≤k
1−1/γ
−1 )) − 1| ≥ c | T
P supj+1≤m≤k |Wj,k m
(1
+
+
O(j
j,m
j
···
× ET 1Aδ
1−1/γ
j,k
P supj+1≤m≤k |Wj,k m
−
1|
≥
c
j
m 1−1/γ
+ ET (1 − 1Aδ )P sup |Wj,k
(1 + j,m + O(j −1 )) − 1| ≥ c | T .
j,k
j
j+1≤m≤k
29
Observe that as j → ∞, the last term converges to zero by monotone convergence,
while the ratio inside the expectation converges to one on sets where Aδj,k obtains,
which have asymptotic measure equal to one. The result follows by monotone convergence and comparison with (48).
The following lemma gives an estimate based on the beta-gamma algebra, and
says that for large enough j, Ψj can be approximated by G1−α /((j − 1)µ − 1 − jα).
Lemma 26. For j < k, let Ψj and Ψk be independent beta random variables as
in (5), and let G (j) and G (k) be independent Gamma(1 − α, 1) random variables as
in (9). Then there exists some J < ∞ such that for every j ≥ J and λ > 0,
"
#
G (j) G (k)
P sup Ψj Ψk −
<λ ≥1−λ.
(j(µ − α) − (µ − 1))(k(µ − α) − (µ − 1))
j≤k≤n
/
Proof. Denote the quantity of interest as Ψ
j,k , and let Σj :=
Then using the distributional identity in (9),
Pj
i=1 G
Pj−1
(i) +
i=1
G∆i+1 −1 .
(j(µ − α) − (µ − 1))(k(µ − α) − (µ − 1)) − Σj Σk
(j(µ − α) − (µ − 1))(k(µ − α) − (µ − 1))
Ψj Ψk
=
···
(j(µ − α) − (µ − 1))(k(µ − α) − (µ − 1))
|Ψ
j,k | = Ψj Ψk
×
+
j X
k
X
(i) (m)
G G
2
− (1 − α)
i=1 m=1
j−1
k
XX
+
j X
k−1
X
G (i) G∆m+1 −1 − (1 − α)(µ − 1) · · ·
i=1 m=1
j−1 X
k−1
X
G∆j+1 −1 G∆m+1 −1 − (µ − 1)2 .
G∆j+1 −1 G (m) − (1 − α)(µ − 1) +
i=1 m=1
i=1 m=1
Ψ
Denote by ¯Ψ
j,k the term-wise absolute value of j,k . As a cumulative sum of nonnegative terms, ¯Ψ
j,k is a nonnegative submartingale for k > j, and hence by Doob’s
maximal inequality, for any c > 0,
1 Ψ
].
P[ sup |Ψ
¯Ψ
j,k | ≥ c] ≤ P[ sup
j,k ≥ c] ≤ E[¯
c j,n
j≤k≤n
j≤k≤n
The independence properties of the beta and gamma random variables, along with
simple but tedious calculations yields
E[¯
Ψ
j,n ] ≤
C(α, µ, V∆2 ) + O(j −1 )
,
(j(µ − α) − (µ − 1))(n(µ − α) − (µ − 1))
where C(α, µ, V∆2 ) is a constant that depends only on the model parameters. The
result follows.
30
Finally, the proof of Theorem 9 requires that for large enough j, the error term
rj,k
:=
k
X
r=j
1
2r2−2/γ
(∆r+1 − µ)
is small with high probability, which is established by the following lemma.
Lemma 27.
For any c > 0,
P[ sup |rj,k | < c] ≥ 1 −
j≤k≤n
V∆ 1
(1 + O(j −2−2/γ )) .
c j 1−2/γ
/
Proof. Denote by ¯rj,k the term-wise absolute value of rj,k . ¯rj,k is a nonnegative
submartingale for k > j, and an application of Doob’s maximal inequality yields
the result.
Proof of Theorem 9. Let U0 be a Uniform[0, 1] random variable, and define
j0 = dU0 kn e to be a vertex selected uniformly at random from Gn , which has kn
vertices. For simplicity, assume that either all Tj are even for j > 1 (which is the
case for graphs that are a.s. connected for all n), or that Tj has equal probability of
being odd or even for all j > 1. (These assumptions are not necessary, but greatly
simplify notation. More general assumptions may be accommodated.) Conditioned
on (Ψj )j>1 , the expected number of directed edges from vertex j0 to any vertex
j ∈ N+ is
(49)
kn
X
E[Nj0 →j,n ] =
kn
X
Pj0 →j,r =
r=j0 ∨j
Ψj0 Wj0 ,r Ψj Wj,r (∆r+1 /2 − 1)
r=j0 ∨j
+ 1{j0 ≥ j, Tj0 odd}Ψj Wj,j0 + 1{j0 < j, Tj even}Ψj0 Wj0 ,j−1 .
Define
P̂j0 →j,r
G (j0 ) G (j) (µ/2 − 1) j0 j 1−1/γ
.
:=
j0 j(µ − α)2
r r
By Lemmas 25 to 27, for n large enough, with probability at least 1 − 2λ,
(1 − λ)
kn
X
r=j0 ∨j
P̂j0 →j,r ≤
kn
X
r=j0 ∨j
Pj0 →j,r ≤ (1 + λ)
kn
X
r=j0 ∨j
P̂j0 →j,r .
31
Similar approximations hold for the final two terms of (49). Let j+ = j0 ∨ j. Now
summing over r yields
G (j0 ) G (j) (µ/2 − 1) 1
E[N̂j0 →j,n ] =
(µ − α)2 (1 − 2/γ) kn
j0 j
kn kn
−1/γ
kn
j+
1−2/γ
−1
1+
−1
O(j+
)
.
Pdkn ue
Defining for u ∈ (0, 1), E[N̂U0 ,n (u)] := j=1
E[N̂j0 →j,n ], it follows from standard
results on convergence to Poisson processes [15] that N̂U0 ,n (u) converges weakly to
a Poisson point process with intensity
(50)
1−2/γ
→
−
G1−α (1 − α)(µ/2 − 1)
1 −1/γ −(1−1/γ)
−1/γ 1 − u+
λ (u0 , u) =
(u0 u)
+ u− u+
,
1−2/γ
µ − α (µ − α)(1 − 2/γ)
2
u
+
where u+ = u0 ∨u and u− = u0 ∧u. Similarly, for incoming edges to j0 , the symmetry
←
−
→
−
of the sampling process yields λ (u0 , u) = λ (u0 , u).
It follows that the degree of a randomly sampled vertex is a Poisson random
variable with mean parameter
Z 1
←
−
→
−
−1/γ
(51)
Λ(G1−α , U0 ) =
λ (U0 , u) + λ (U0 , u)du = G1−α (U0
− 1) .
0
Hence, by the conjugacy relationship between the Poisson and Gamma distributions,
the probability that the degree of a randomly sampled vertex, conditioned on U0 , is
equal to d + 1 is
P[D = d + 1 | U0 ] = E[e−Λ(G1−α ,U0 ) Λ(G1−α , U0 )d | U0 ]/d!
Γ(d + 1 − α)
1/γ
(1−α)/γ
=
(1 − U0 )d U0
.
Γ(d + 1)Γ(1 − α)
Finally, taking the expectation with respect to the uniform random variable U0 ,
Z 1
Γ(d + 1 − α)
P[D = d + 1] =
(1 − u1/γ )d u(1−α)/γ du
Γ(d + 1)Γ(1 − α) 0
Γ(d + 1 − α)
Γ(d + 1)Γ(1 − α + γ)
=
γ
Γ(d + 1)Γ(1 − α) Γ(d + 1 + 1 − α + γ)
Γ(d + 1 − α)Γ(1 − α + γ)
=γ
,
Γ(d + 1 + 1 − α + γ)Γ(1 − α)
which is the stated result. Corollary 10 follows by checking moments. For σ = 1,
1/γ
Corollary 11 follows from (51) by observing that U0 is distributed as Beta(γ, 1);
for σ ∈ (0, 1), it follows from a similar integral identity.
32
A.5. Proofs for Section 4.
Proof of Proposition 13. Let m̄ := 2m−1 and ᾱ := 2m−α. When tj = 2m(j − 1) + 1,
(45) can be manipulated into the form
m̄
Γ(n − kn α)Γ(kn ᾱ − m̄ + p̄) Y Γ(kn − i/ᾱ)Γ(r + 1 + (p̄ − i)/ᾱ)
Zn (p, t) =
···
Γ(n − kn α + p̄)Γ(kn ᾱ − m̄)
Γ(kn + (p̄ − i)/ᾱ)Γ(r + 1 − i/ᾱ)
i=1
r
Γ(rᾱ) Y Γ(Dj,n − α + pj )
×
Γ(rᾱ + p̄)
Γ(Dj,n − α)
j=1
m̄
p̄−i
Γ(rᾱ) Y Γ(r + 1 + ᾱ )
= Xn (p, t)
,
Γ(rᾱ + p̄)
Γ(r + 1 − ᾱi )
i=1
where
Xn (p, t) :=
n
2m
−p̄ m̄ Y
r
ᾱ
p
j
Dj,n
(1 + O(n−1 )) .
j=1
Therefore, algebraic manipulations of Zn (p, t) show that for large n:
m̄
Y
p̄/ᾱ
Zn (p, t) = Xn (p, t)E[ Gr+1− i ]/E[Grp̄ᾱ ]
(52)
i=1
ᾱ
m̄
r
Y
Y
p̄/ᾱ
p̄
Zn (p, t) = Xn (p, t)E[ G1− i ]/E[Grᾱ
Bjp̄ᾱ−m̄,m̄ ]
(53)
ᾱ
i=1
m̄
Y
(54)
Zn (p, t) = Xn (p, t)E[
i=1
(55)
Zn (p, t) = Xn (p, t)E[
m̄
Y
i=1
j=1
p̄/ᾱ
G1− i ]/E[Grp̄ᾱ−m̄
ᾱ
p̄/ᾱ
G1− i ]E[
ᾱ
r
Y
r−1
Y
Bjp̄ᾱ−m̄,m̄ ]
j=1
p̄
p̄
B(j−1)
ᾱ,1−α ]/E[G1−α ] .
j=2
On the other hand, Lemma 23 establishes that Zn (p, t) converges almost surely to
the product of random variables ξ1p1 · · · ξrpr . Now, letting p \ pj denote p with the jth
element set to zero,
E[ξ1p1 · · · ξrpr ] = E[ lim Zn (p, t)] = E[Ztr (p, t)]
n→∞
Γ(1 − α + p̄)Γ(rᾱ − m̄)Γ((r − 1)ᾱ + p̄r−1 )
Γ(1 − α)Γ(rᾱ − m̄ + p̄r )Γ((r − 1)ᾱ)
r
r
Y
Y
Γ(1 − α + pj )Γ(j ᾱ − m̄)Γ((j − 1)ᾱ + p̄j−1 )
p
=
= E[
Ψj j (1 − Ψj )p̄j−1 ] .
Γ(1 − α)Γ((j − 1)ᾱ)Γ(j ᾱ − m̄ + p̄j )
= E[Ztr −1 (p \ pr , t)]
j=2
By equating (52)–(55) to E[Ztr (p, t)], the results follow.
j=2
33
Proof of Proposition 18. Define T1:r := (T1 , T2 , . . . , Tr ) and
r
Γ(n + θ) Y Γ(Dj,n − α + pj )
Γ(n + θ + p̄)
Γ(Dj,n − α)
Znα,θ (p, T1:r ) :=
(56)
j=1
= n−p̄
(57)
r
Y
p
j
Dj,n
(1 + O(n−1 )) .
j=1
α,θ
The fact that E[Zn+1
(p, T1:r ) | Fn ] = Znα,θ (p, T1:r ) shows that Znα,θ (p, T1:r ) is a nonnegative martingale for n ≥ Tr . Znα,θ (p, T1:r ) can be bounded in L2 following an
argument similar to that given in the proof of Lemma 23, and hence it converges
almost surely to ξ1p1 · · · ξrpr . Therefore,
E[ lim n
n→∞
=
−p̄
r
Y
p
j
Dj,n
| T1:r ] = E[ξ1p1 · · · ξrpr | T1:r ]
j=1
Γ(1 − α + pr )Γ(Tr + θ)Γ(Tr − 1 − (r − 1)α + p̄r−1 )
...
Γ(1 − α)Γ(Tr + θ + p1 )Γ(Tr − 1 − (r − 1)α)
r−1
Y Γ(1 − α + pj )Γ(Tj − jα)Γ(Tj − 1 − (j − 1)α + p̄j−1 )
×
Γ(1 − α)Γ(Tj − jα + p̄j )Γ(Tj − 1 − (j − 1)α)
j=2
= E[BTp̄ r −rα,θ+rα ·
r
Y
p
Ψj j (1 − Ψj )p̄j−1 ] ,
j=2
from which the result follows.
Proof of Proposition 19. For an exchangeable Gibbs partition with α ∈ (0, 1)
and a sequence of coefficients vn,k , one may construct a martingale similar to that
in (56). It is then straightforward to show that the scaled degrees converge almost
surely to random variables (ξj )1≤j≤r , conditionally on T1 , . . . , Tr . Conditioned on
Tj , ξj has marginal moments
E[ξjp | Tj ] =
VT +1,j −1
Γ(1 − α + p)
1+p j
.
Γ(1 − α)
VTj ,j
Kerov [29] showed that the only exchangeable Gibbs partitions with coefficients that
can be represented as a ratio Vn,k = vk /cn are those of the CRP(α, θ), in which case
the product in the previous equation becomes independent of j. See also [23, Lemma
4.1], which also implies the result.
34
Proof of Proposition 21. Define for p > −1,
Znβ (p, T1:r )
r
Y
Γ(Dj,n + pj )
Γ(n)
:=
Γ(n + p̄(1 − β))
Γ(Dj,n )
j=1
= n−p̄(1−β)
r
Y
p
j
Dj,n
(1 + O(n−1 )) ,
j=1
which is by construction a nonnegative martingale, bounded in L2 , for n ≥ Tr .
Therefore,
E[ lim n−p̄(1−β)
n→∞
r
Y
p
j
Dj,n
] = E[ξ1p1 · · · ξrpr ]
j=1
r
Y
Γ(1 + pj )Γ(Tj )Γ(Tj − 1 + p̄j−1 )
Γ(Tr + p̄)
=
Γ(Tr + p̄(1 − β))
Γ(Tj − 1)Γ(Tj + p̄j )
j=2
−p̄(1−β)
= E[M1−β,Tr −1 ]E[B p̄
β
Tr ,(Tr −1) 1−β
]E[
r
Y
p
Ψj j (1 − Ψj )p̄j−1 ] ,
j=2
from which the result follows. The marginal identities result from altering the above
martingale to contain only Dj,n , which is also a martingale. Similar treatment of
the moments, along with the beta-gamma algebra and the identity (1.3) in James
[25] yields the rest of the identities.
Proof of Proposition 22. In analogy to the previous proof, for every p > −1,
Znβ (p, w + b) =
Γ(Dw,n + p)
Γ(n)
Γ(n + p(1 − β)) Γ(Dw,n )
p
= n−p(1−β) Dw,n
(1 + O(n−1 ))
is a nonnegative martingale, bounded in L2 , for n ≥ w + b. Therefore, it converges
p
almost surely to a random variable ξw,w+b
, where
p
E[ξw,w+b
]=
Γ(w + b)Γ(w + p)
Γ(w + b + p(1 − β))Γ(w)
p(1−β)
= E[Gwp ]/E[Gw+b
p
] = E[Bw,b
Bp
β
w+b,(w+b−1) 1−β
Mp1−β,w+b−1 ] .
The other identities can be verified by checking moments; they also follow from the
beta-gamma algebra and from [25].
Department of Statistics
24–29 St. Giles’
Oxford OX1 3LB, UK
E-mail: [email protected]
Department of Statistics
1255 Amsterdam Avenue
New York, NY 10027, USA
E-mail: [email protected]
| 10math.ST
|
Model Reduction for Aperiodically
Sampled Data Systems ⋆
Mert Baştuğ ∗ Laurentiu Hetel ∗ Mihály Petreczky ∗
arXiv:1703.01990v1 [cs.SY] 6 Mar 2017
∗
Centre de Recherche en Informatique, Signal et Automatique de Lille
(CRIStAL), UMR CNRS 9189, Ecole Centrale de Lille, 59650
Villeneuve d’Ascq, France (e-mail: [email protected],
[email protected], [email protected]).
Abstract: Two approaches to moment matching based model reduction of aperiodically
sampled data systems are given. The term “aperiodic sampling” is used in the paper to indicate
that the time between two consecutive sampling instants can take its value from a pre-specified
finite set of allowed sampling intervals. Such systems can be represented by discrete-time linear
switched (LS) state space (SS) models. One of the approaches investigated in the paper is to
apply model reduction by moment matching on the linear time-invariant (LTI) plant model,
then compare the responses of the LS SS models acquired from the original and reduced order
LTI plants. The second approach is to apply a moment matching based model reduction method
on the LS SS model acquired from the original LTI plant; and then compare the responses of the
original and reduced LS SS models. It is proven that for both methods, as long as the original
LTI plant is stable, the resulting reduced order LS SS model of the sampled data system is
quadratically stable. The results from two approaches are compared with numerical examples.
Keywords: Model reduction, sampled data systems, quadratic stability, numerical algorithms,
linear systems theory.
1. INTRODUCTION
The topic of model reduction deals with computing simpler
approximation models for an original complex model [Antoulas (2005)]. For system classes which can be represented
by state-space (SS) models, the “complexity” of a model
refers usually to the state-space dimension of the corresponding model. Hence a “simpler model” is a model with
less number of states whose input-output behavior is close
to the one of the original system. In this paper, the term
model reduction is used in this sense, i.e., approximating
the input-output behavior of an original SS model with an
another SS model with less number of states.
Aperiodically sampled data systems appear commonly in
applications since they can be used for modeling various
phenomena encountered in the context of large scale
networked and embedded control systems [Hespanha et al.
(2007); Hristu-Varsakelis and (Editors); Zhang et al.
(2001); Hetel et al. (2017); Brockett (1997); Donkers et al.
(2011)]. In turn, the dimension of corresponding SS models
for such systems can be very big due to the interaction
of different subsystems in the network. Simulations for
control synthesis or performance specifications regarding
the output behavior can easily become intractable due
to the complexity of the original model. Hence, model
reduction approaches for such systems can be of great
importance.
⋆ This work was partially supported by ESTIREZ project of Region
Nord-Pas de Calais, France and by ANR project ROCC-SYS (ANR14-CE27-0008).
The paper states two model reduction procedures based
on moment matching for aperiodically sampled data systems. Model reduction for sampled data systems has been
considered previously on [Barb and Weiss (1993); Shieh
and Chang (1984)]. Both papers deal with the case of
periodical sampling and are valid only for the case when
the considered plant is stable. In contrast, in the present
paper the general aperiodic sampling case considered and
the considered plant is allowed to be unstable.
In this paper, the sampling interval is considered to be
time varying and assumed to be taking its values from
a finite set of possible sampling intervals. The inputoutput behavior at sampling instants of such sampled data
systems can be modeled by discrete-time linear switched
(LS) SS representations [Gu et al. (2003); Zhang (2001);
Donkers et al. (2009)]. One of the model reduction procedures considered in the paper can be summarized as
follows: Apply a classical moment matching algorithm to
the original continuous-time linear time-invariant (LTI)
plant to get a reduced order model, and then get an LS SS
representation to model the aperiodically sampled system.
The other approach given is to apply an analogous moment
matching based model reduction algorithm to the LS SS
representation which is computed from the original LTI
plant. Since the sampling interval at each time instant acts
as an additional control input for aperiodically sampled
data systems, intuitively, a method of approximating the
input-output behavior of such systems should make use
of the information of the allowed sampling interval set.
The second approach is given in accordance with this idea.
For both of the approaches, it is shown that the resulting
reduced order discrete-time LS SS representation of the
sampled data system will be quadratically stable as long
as the original continuous-time LTI model is stable.
The paper is organized as follows: In Section 2, we present
the procedure of modeling an aperiodically sampled data
system with an LS SS representation. In Section 3 we
present a brief overview of the concept of model reduction by moment matching for LTI SS representations and
present the first model reduction approach in detail. In
Section 4 we briefly review the concept of model reduction by moment matching for LS SS representations and
present the second model reduction approach in detail.
In Section 5 we show that the proposed methods preserve
stability. In Section 6 we illustrate the two approaches and
compare their performances with two numerical examples.
Notation 1. In the following, we will use Z, N and R+ to
denote respectively the set of integers; the set of natural
numbers including 0 and the set [0, +∞) of nonnegative
real numbers. We will use Ia to denote the a × a identity
matrix with a ∈ N\{0}.
2. MODELING OF SAMPLED DATA SYSTEMS
WITH LS SS REPRESENTATIONS
In this section we present briefly the process of modeling
an aperiodically sampled continuous-time LTI system with
a discrete-time LS SS model. We start with the formal
definition of continuous-time LTI state-space (SS) representations.
An LTI SS representation ΣLTI is a tuple ΣLTI = (A, B, C)
with A ∈ Rn×n , B ∈ Rn×m , C ∈ Rp×n . The state
x(t) ∈ Rn and the output y(t) ∈ Rp of the LTI system
ΣLTI at time t ≥ 0 is defined by 1
ẋ(t) = Ax(t) + Bu(t)
(1)
ΣLTI
y(t) = Cx(t), ∀t ∈ R+ .
In the following, dim(ΣLTI ) will be used to denote the
dimension n of the state-space of ΣLTI and the number
n will be called the order of ΣLTI .
Let ΣLTI = (A, B, C) be a continuous-time LTI SS representation of the form (1). Let the state x(tk ) and output
y(tk ) of ΣLTI be sampled in arbitrary time instants tk ,
k ∈ N such that t0 = 0 and tk+1 − tk ∈ H = {ĥ1 , . . . , ĥD },
ĥ1 , . . . , ĥD ∈ R+ for all k ∈ N to form the constant control
signal u(t) = uk for all t ∈ [tk , tk + 1), k ∈ N. Note that
the sequence tk , k ∈ N is monotonically increasing. The
resulting sampled data system ΣSD can be represented as
follows:
ẋ(t) = Ax(t) + Buk , t ∈ [tk , tk+1 ), k ∈ N
ΣSD
yk = Cx(tk )
(2)
tk+1 = tk + hk , hk ∈ H = {ĥ1 , . . . , ĥD }.
In (2), x(t) ∈ Rn is the state, uk ∈ Rm is the constant
input and y(t) ∈ Rp is the output at time t ∈ R+ ;
A ∈ Rn×n , B ∈ Rn×m and C ∈ Rp×n are the same
as the system parameters (A, B, C) of ΣLTI . We call the
set H = {ĥ1 , . . . , ĥD }, as the finite sampling interval
1 Unless stated otherwise, we take x(0) = x = 0 for all classes of
0
systems discussed in the paper for notational simplicity. Note that
the result of the paper can easily be extended to the case of non-zero
initial states.
set and the value hk ∈ H as the kth sampling interval.
We will use the shorthand notation ΣSD = (A, B, C, H)
for the sampled data system of the form (2). Note that
different from the model (1), model (2) has also hk ∈ H,
k ∈ N as the control parameter in addition to the input
u(t) = uk , t ∈ [tk , tk+1 ), k ∈ N.
The state xk = x(tk ) and output yk = y(tk ) of the sampled
data system ΣSD in (2) at sampling instants tk , k ∈ N can
be written by induction as
!
Z hk
xk+1 = x(tk+1 ) = eAhk xk +
eAs ds Buk , ∀k ∈ N,
0
yk = Cxk .
(3)
Let
Θ(hk ) =
Z
hk
eAs ds.
(4)
0
It is easy to see that the following holds:
eAhk = In + AΘ(hk ).
(5)
Replacing (5) in (3) and defining the matrix functions
Φ : H → Rn×n and Γ : H → Rn×m as Φ(hk ) = eAhk =
In + AΘ(hk ) and Γ(hk ) = Θ(hk )B, (3) can be rewritten
as
xk+1 = Φ(hk )xk + Γ(hk )uk ,
Σdisc
(6)
yk = Cxk , ∀hk ∈ H, k ∈ N.
With equation (6), the sampled LTI plant ΣSD in (2) is
modeled by a discrete-time, time-varying linear system
Σdisc whose read-out map (map represented by the matrix
C) is time-invariant. Here, the discrete time instants k ∈ N
of (6) corresponds to the time instants tk ∈ R+ , k ∈ N for
the original sampled data system ΣSD . In addition, the
state xk and the output yk of (6) corresponds to the state
x(tk ) and output y(tk ) of ΣSD at the sampling instants
tk ∈ R+ when u(t) = uk for t ∈ [tk , tk+1 ). Hence we have
built the relationship between the sampled data system
ΣSD = (A, B, C, H) and the corresponding discrete-time,
linear time-varying system representation Σdisc .
Since the sampling interval hk between any two consecutive sampling instants can take its values only from the
finite set H = {ĥ1 , . . . , ĥD } one approach to design control
for the model (6) is to create an LS SS model from (6).
The idea is that since the set H has D elements, Θ(hk ) can
only take D different values for all k ∈ N. In turn, (6) can
be used to create an LS SS representation with D discrete
modes. Below we summarize this procedure.
Notation 2. Let a, b ∈ N. In the following, we use Iba to
denote the set Iba = {c ∈ N | a ≤ c ≤ b}.
Let the matrices Â1 , . . . , ÂD ∈ Rn×n and B̂1 , . . . , B̂D ∈
Rn×m be defined by
Âi = In + AΘ(ĥi ), ∀i ∈ ID
1 ,
B̂i = Θ(ĥi )B, ∀i ∈ ID
1 .
(7)
Using (7), (6) can be rewritten as the following SS representation
ΣLS
xk+1 = Âσk xk + B̂σk uk
yk = Cxk , ∀k ∈ N.
(8)
where σk ∈ ID
1 is called the value of the switching sequence
at time k ∈ N.
Approach 1
Approach 2
Models of the form (8) are a subclass of discrete-time LS
SS representations where the read-out map represented
by the matrix C is constant and independent from the
value of the switching signal σk at each time instant
k ∈ N. Hence from now on, we will refer to the system
representations of the form (8) as LS SS representations
and formally define the tuple ΣLS = ({(Âi , B̂i , C)}D
i=1 )
p×n
with Âi ∈ Rn×n , B̂i ∈ Rn×m for all i ∈ ID
1 , C ∈ R
as an LS SS representation. We remark that the discretetime LS SS representation described by (8) completely
models the behavior of ΣSD in sampling instants. More
clearly, note that each linear mode (Âi , B̂i , C), i ∈ ID
1
corresponds to the ith element of the sampling interval
set H = {ĥ1 , . . . , ĥD }, i.e., if the kth sampling interval
hk , k ∈ N is chosen as hk = ĥi , i ∈ ID
1 , then the value
of the switching signal at time instant k is σk = i. In
the following, analogous to the LTI case, dim(ΣLS ) will be
used to denote the dimension n of the state-space of ΣLS
and the number n will be called the order of ΣLS .
ΣLTI
ΣLTI
Now we can state the problem considered in the paper as
follows.
Problem Let ΣLTI be a continuous-time LTI plant model
of order n, which is to be sampled aperiodically with
respect to the set H = {ĥ1 , . . . , ĥD } to form the sampled
data system ΣSD . Compute a discrete-time model Σ̄LS of
order r < n which is an approximation of the input-output
behavior of ΣSD in sampling instants.
Two intuitive approaches (see Figure 1) can be proposed
for the solution of this problem:
Approach 1 Let ΣLTI = (A, B, C) be the continuoustime LTI plant which is to be sampled aperiodically
with respect to H = {ĥ1 , . . . , ĥD }. Compute from ΣLTI
another LTI SS model Σ̄LTI = (Ā, B̄, C̄) of order r < n
who approximates the input-output behavior of ΣLTI .
Let Σ̄SD = (Ā, B̄, C̄, H) be the sampled data system
corresponding to Σ̄LTI . Compute from Σ̄SD the LS SS
ˆ , C̄)}D ) of order r < n of the
model Σ̄LS = ({(Āˆi , B̄
i
i=1
form (8).
Approach 2 Let ΣSD = (A, B, C, H) be the sampled
data system with H = {ĥ1 , . . . , ĥD } of the form (2)
corresponding to the continuous-time LTI plant ΣLTI =
(A, B, C) of the form (1). Let ΣLS = ({(Âi , B̂i , C)}D
i=1 ) of
the form (8) be the corresponding LS SS model for the
sampled data system ΣSD . Compute from ΣLS another LS
¯ ¯
SS model Σ̄LS = ({(Âi , B̂i , Ĉ)}D
i=1 ) of order r < n who
approximates the input-output behavior of ΣLS .
Remark 1. Note that the symbol¯over a system representation Σ is used for indicating that Σ̄ is an approximation
system for Σ; whereas the subscripts LTI, SD, LS are used
for referring to the particular class of system representations of the form (1), (2) and (8) respectively. The symbol
¯when used above a system matrix A, B or C of a system
Σ, indicates that Ā, B̄ or C̄ are the system parameters of
Σ̄. Finally, the symbol ˆ over a system parameter A or B
of ΣSD is used to indicate that Âi and B̂i for all i ∈ ID
1
are the matrices of the form (7) of the resulting LS SS
Model Reduction
Σ̄LTI
Sampling
8
2
Σ̄SD
LS Modeling
5
1
3
Σ̄LS
ΣSD
6
4
Sampling
LS Modeling
ΣLS
7
Model Reduction
Σ̄LS
Fig. 1. Overview of the two model reduction approaches.
representation corresponding to the sampled data system
with H = {ĥ1 , . . . , ĥD }. The symbols ˆ¯ and ¯ˆ used above
a system matrix in the definitions of Approach 1 and
Approach 2 respectively can be interpreted with respect
to this remark.
3. FIRST APPROACH OF MODEL REDUCTION
In this section, firstly we recall the concepts of Markov
parameters and moment matching for LTI SS representations. Then we present Approach 1 in detail.
3.1 Review: Moment Matching for LTI SS Representations
Notation 3. Let a ∈ N\{0}. The set of continuous and
absolutely continuous maps of the form R+ → Ra is
denoted by C(R+ , Ra ) and AC(R+ , Ra ) respectively; and
the set of Lebesgue measurable maps of the form R+ → Ra
which are integrable on any compact interval is denoted
by Lloc (R+ , Ra ).
We define the input-to-state map XΣx0LTI and input-to0
output map YΣxLTI
of a system ΣLTI = (A, B, C) of the
form (1) as the maps
XΣx0LTI : Lloc (R+ , Rm ) → AC(R+ , Rn ); u 7→ XΣx0LTI (u),
0
0
YΣxLTI
: Lloc (R+ , Rm ) → C(R+ , Rp ); u 7→ YΣxLTI
(u),
defined by letting t 7→ XΣx0LTI (u)(t) be the solution to
the first equation of (1) with x(0) = x0 , and letting
0
YΣxLTI
(u)(t) = CXΣx0LTI (u)(t) for all t ∈ R+ as in second
equation of (1).
A moment’s reflection lets us see that the kth Taylor series
coefficient Mk of YΣ0LTI around t = 0 for the unit impulse
input will be
Mk = CAk B, k ∈ N
(9)
0
where A defined to be In . The coefficients Mk , k ∈ N
are called the Markov parameters or the moments of the
system ΣLTI . Hence, it is possible to approximate the
input-output behavior of ΣLTI by another system Σ̄LTI
(possibly of reduced order), whose first some number of
Markov parameters are equal to the corresponding ones
of ΣLTI . If this number is chosen to be N ∈ N, we will
call such approximations as N -partial realizations of ΣLTI .
More precisely, a continuous-time LTI SS representation
Σ̄LTI = (Ā, B̄, C̄) is an N -partial realization of another
continuous-time LTI SS representation ΣLTI = (A, B, C)
if
CAk B = C̄ Āk B̄, k = 0, . . . , N.
The problem of model reduction of LTI systems by moment matching can now be stated as follows. Consider
an LTI system ΣLTI = (A, B, C) of the form (1) and fix
N ∈ N. Find another LTI system Σ̄LTI of order r strictly
less than n such that Σ̄LTI is an N -partial realization of
ΣLTI .
Below we recall a basic theorem on how to compute an
N -partial realization for the LTI case. For this purpose,
we define the N -partial reachability space of a continuoustime LTI realization ΣLTI = (A, B, C) as
N
RLTI
= im B AB · · · AN B .
(10)
for all N ∈ N with A0 := In .
Theorem 1. (Moment Matching for LTI SS Representations, [Antoulas (2005)]). Let ΣLTI = (A, B, C) be a
continuous-time LTI SS representation of the form (1),
N ∈ N and V ∈ Rn×r be a full column rank matrix such
that
N
RLTI
= im(V ).
If Σ̄LTI = (Ā, B̄, C̄) is an LTI SS representation such that
the matrices Ā, B̄, C̄ are defined as
Ā = V −1 AV , B̄ = V −1 B, C̄ = CV,
(11)
where V −1 is a left inverse of V , then Σ̄LTI is an N -partial
realization of ΣLTI .
3.2 Approach 1
Now the first approach for model reduction of aperiodically
sampled data systems can be stated in detail as follows:
Approach 1 Let ΣLTI = (A, B, C) be the continuous-time
LTI plant of order n which is to be sampled aperiodically
with respect to H = {ĥ1 , . . . , ĥD }. By using Theorem
1, compute an N -partial realization Σ̄LTI = (Ā, B̄, C̄) of
ΣLTI such that the order of Σ̄LTI is r < n. Let Σ̄SD =
(Ā, B̄, C̄, H) be the sampled data system corresponding
to Σ̄LTI . Compute from Σ̄SD the LS SS model Σ̄LS =
ˆ , C̄)}D ) of order r < n of the form (8), with
({(Āˆi , B̄
i
i=1
the procedure given in Section 2.
Corollary 1. (Theorem 1). The resulting reduced order LS
SS model Σ̄LS computed by Approach 1 corresponds to
the discrete-time, time varying model Σdisc
x̄k+1 = Φ̄(hk )x̄k + Γ̄(hk )uk ,
(12)
Σ̄disc
ȳk = C̄ x̄k , ∀hk ∈ H, ∀k ∈ N,
of the sampled data system. In (12)
!
Z hk
−1
Āhk
Ās
, Γ̄(hk ) =
x̄k = V xk , Φ̄(hk ) = e
e ds B̄.
0
4. SECOND APPROACH OF MODEL REDUCTION
In this section, we recall the concepts of Markov parameters and moment matching for LS SS representations
and the analogy between the LTI case. Then we present
Approach 2 in detail.
4.1 Review: Moment Matching for LS SS Representations
Notation 4. In the sequel, we use the following notation
and terminology: If s = s0 · · · sN is a sequence with N + 1
elements, N ∈ N, we denote the number N as |s| = N
and call |s| as the length of the sequence |s|. We use Q
to denote the set of finite sequences in Q = {1, . . . , D},
D ≥ 1, i.e., Q = {σ = σ0 · · · σN | σ0 , · · · , σN ∈ Q, N ∈
N}; U to denote the set of finite sequences in Rm , i.e.,
U = {u = u0 · · · uN | u0 , · · · , uN ∈ Rm , N ∈ N}; X
to denote the set of finite sequences in Rn , i.e., X =
{x = x0 · · · xN | x0 , · · · , xN ∈ Rn , N ∈ N} and Y to
denote the set of finite sequences in Rp , i.e., Y = {y =
y0 · · · yN | y0 , · · · , yN ∈ Rp , N ∈ N}. In addition, we will
write U × Q = {(u, σ) ∈ U × Q | |u| = |σ|}.
We define the input-to-state map XΣx0LS and input-to-output
0
map YΣxLS
of a system ΣLS of the form (8) as the maps
U × Q → X; (u, σ) 7→ XΣx0LS (u, σ) = x,
0
U × Q → Y; (u, σ) 7→ YΣxLS
(u, σ) = y,
defined by letting k 7→ XΣx0LS (u, σ)k be the solution to
the first equation of (8) with x(0) = x0 , and letting
0
YΣxLS
(u, σ)k = CXΣx0LS (u, σ)k for all k ∈ N as in second
equation of (8).
Using (8), one can see that the coefficients appearing in
the output of ΣLS for any pair of input and switching
sequences (u, σ) ∈ U × Q are of the form
C B̂j , j ∈ ID
1
(13)
and
C Âk1 · · · ÂkM B̂j ; k1 , · · · , kM , j ∈ ID
1 , M ∈ N\{0}. (14)
Analogously to the linear case we will call the coefficients
of the form (13) and (14) as the Markov parameters
of ΣLS = ({(Âi , B̂i , C)}D
i=1 ). Specifically, we will call
the Markov parameters of the form (13) as the Markov
parameters of length 0 and the Markov parameters of the
form (14) as the Markov parameters of length M for any
M ∈ N\{0}.
In [Bastug et al. (2014)] and [Bastug et al. (2016)] it
is shown that similarly to the LTI case, it is possible
to approximate the input-output behavior of ΣLS by
another LS SS representation Σ̄LS (possibly of reduced
order), whose Markov parameters up to a certain length
N ∈ N is equal with the corresponding ones of ΣLS
2
. Again, we will call such approximations as N -partial
realizations of ΣLS . More precisely, a discrete-time LS SS
¯ ¯
representation Σ̄LS = ({(Âi , B̂i , C̄)}D
i=1 ) is an N -partial
realization of another discrete-time LS SS representation
ΣLS = ({(Âi , B̂i , C)}D
i=1 ) if
¯
C B̂j = C̄ B̂j , j ∈ ID
1
2
Even though the results in [Bastug et al. (2014)] and [Bastug et al.
(2016)] are stated in the continuous-time context, the analogous results on N -partial realizations of discrete-time LS SS representations
are also valid. See [Bastug et al. (2015)] for an application of these
results for the model reduction of affine LPV systems in the discretetime context.
4.2 Approach 2
and
¯ ¯
¯
C Âk1 · · · ÂkM B̂j = C̄ Âk1 · · · ÂkM B̂j
ID
1
IN
1 .
for all k1 , · · · , kM , j ∈
and M ∈
Note that an N partial realization Σ̄LS of ΣLS will have the same output
with ΣLS for all time instants up to N , i.e., k ∈ IN
0 , for all
input and switching sequences. The reason why the output
of an N -partial realization is indeed an approximation for
the output of the original system model for also the time
instants k > N can be found in [Bastug et al. (2015)].
The problem of model reduction of LS systems by moment
matching can now be stated as follows: Consider an LS SS
model ΣLS = ({(Âi , B̂i , C)}D
i=1 ) of the form (8) of order n
and fix N ∈ N. Find another LS SS model Σ̄LS of order r
¯ ¯
strictly less than n such that Σ̄LS = ({(Âi , B̂i , C̄)}D
i=1 ) is
an N -partial realization of ΣLS .
Next, we recall a theorem of model reduction with N partial realizations for the LS case [Bastug et al. (2014)].
This theorem (Theorem 2) can be considered as the analogous of Theorem 1 for the LS case (or in other words
Theorem 1 is a special case of 2 when the LS system consists of only one LTI system). For this purpose, we define
inductively the N -partial reachability space of a discretetime LS SS representation ΣLS = ({(Âi , B̂i , C)}D
i=1 ) as
[
0
im(Bj ),
RLS
= span
j∈ID
1
N
0
RLS
= RLS
+
X
N −1
), N ≥ 1.
im(Ak RLS
(15)
k∈ID
1
for all N ∈ N. In (15) the summation operator must be
interpreted as the Minkowski sum of vector spaces.
Theorem 2. (Moment Matching for LS SS Representations, [Bastug et al. (2014)]). Let ΣLS = ({(Âi , B̂i , C)}D
i=1 )
be a discrete-time LS SS representation of the form (8),
N ∈ N and V ∈ Rn×r be a full column rank matrix such
that
N
RLS
= im(V ).
(16)
¯ ¯
If Σ̄LS = ({(Âi , B̂i , C̄)}D
i=1 ) is an LS SS representation such
¯ ¯
,
the
matrices Âi , B̂i , C̄ are defined as
that for each i ∈ ID
1
¯
¯
Âi = V −1 Âi V , B̂i = V −1 B̂i , C̄ = CV,
(17)
where V −1 is a left inverse of V , then Σ̄LS is an N -partial
realization of ΣLS .
Note that the key of model reduction lies in the number
of columns of the full column rank projection matrix
V ∈ Rn×r such that r < n. Choosing the number N small
enough such that the matrix V satisfies the condition (16)
and it has r < n columns, results in the reduced order
N -partial realization Σ̄LS of order r. A simple algorithm
with polynomial computational complexity to compute
the matrix V in Theorem 2 is given in [Bastug et al.
(2014)]. Note also that the counterpart of Theorem 2 can
be given dually, using matrix representations of the N unobservability space. These discussions are left out of this
paper for simplicity and they can be found in detail in
[Bastug et al. (2014)].
Now the second approach for model reduction of aperiodically sampled data systems can be stated in detail as
follows:
Approach 2 Let ΣSD = (A, B, C, H) be the sampled data
system with H = {ĥ1 , . . . , ĥD } of the form (2) corresponding to the continuous-time LTI plant ΣLTI = (A, B, C) of
the form (1) of order n. Let ΣLS = ({(Âi , B̂i , C)}D
i=1 ) of
the form (8) be the corresponding LS SS model for the
sampled data system ΣSD computed with the procedure
given in Section 2. By using Theorem 2 compute an N ¯ ¯
partial realization Σ̄LS = ({(Âi , B̂i , C̄)}D
i=1 ) of order r < n
for ΣLS .
Since we have proposed computing an N -partial realization based on the model (8) of an aperiodically sampled
system as a model reduction approach for it, it could be
helpful to relate the definition of N -partial realization to
the model (6) of the sampled data system. The following
corollary is the direct consequence of the definition of N partial realizations in the switched case and the representations (6) and (8).
Corollary 2. (Theorem 2). The N -partial realization Σ̄LS
of the original LS SS model ΣLS of the sampled data
system computed by Approach 2 corresponds to the
time-varying model
x̄k+1 = Φ̄(hk )x̄k + Γ̄(hk )uk ,
(18)
Σ̄disc
ȳk = C̄ x̄k , ∀hk ∈ H, ∀k ∈ N,
such that the outputs yk of Σdisc in (6) and ȳk of Σ̄disc in
(18) for any input u ∈ U will be the same for all k ∈ IN
0 .
In other words,
CΦ(hk ) · · · Φ(hi )Γ(hi ) = C̄ Φ̄(hk ) · · · Φ̄(hi )Γ̄(hi ),
k
k
for all k ∈ IN
0 and i ∈ I0 , where hl ∈ H for all l ∈ Ii . The
relationship between Σdisc of (6) and Σ̄disc of (18) can be
constructed by stating that
x̄k = V −1 xk , Φ̄(hk ) = V −1 Φ(hk )V , Γ̄(hk ) = V −1 Γ(hk )
for all hk ∈ H and k ∈ N.
5. CONSERVATION OF STABILITY
In this section, we build the relationship between the
stability of the original continuous-time LTI system ΣLTI
and the quadratic stability of the reduced order discretetime LS SS model Σ̄LS computed with both approaches.
More specifically, we will show that as long as the original
continuous-time LTI system ΣLTI is stable, the reduced
order discrete-time LS SS representation Σ̄LS modeling
the sampled data system computed by Approach 1 or
Approach 2 will be quadratically stable. As the final
result of this section, we will extend this conservation
of stability argument for the representations of the form
(6) of aperiodically sampled data systems. We will start
with presenting two technical lemmas for the purpose of
presenting this result.
5.1 Technical Lemmas
In the sequel, we denote the fact that a matrix G is positive
definite (resp. positive semi-definite, negative definite,
negative semi-definite) with G > 0 (resp. G ≥ 0, G < 0,
G ≤ 0).
Definition 1. (Quadratic stability).
Let ΣLS =
({(Âi , B̂i , C)}D
)
be
an
LS
SS
representation
of the form
i=1
(8). The LS SS representation ΣLS is quadratically stable
if and only if there exists a symmetric positive definite
P ∈ Rn×n such that
(19)
ÂT
∀i ∈ ID
i P Âi − P < 0,
1 .
Lemma 1. Let ΣLTI = (A, B, C) be an LTI SS representation. For any H = {ĥ1 , . . . , ĥD } with D ∈ N\{0}, the LS
SS model ΣLS = ({(Âi , B̂i , C)}D
i=1 ) of the sampled data
system ΣSD = (A, B, C, H) is quadratically stable if ΣLTI
is stable.
Proof. The stability of ΣLTI = (A, B, C) implies the
stability of the autonomous system Σaut
LTI = (A, 0, 0). Then
there exists a P > 0, P T = P and V (x(t)) = x(t)T P x(t)
such that
V (x(t)) < V (x(0)), ∀x(0) ∈ Rn , x(0) 6= 0, ∀t ∈ R+ \{0}.
Then for all x(0) ∈ Rn , x(0) 6= 0 and for all t ∈ R+ \{0}
x(t)T P x(t) − x(0)T P x(0) < 0.
(20)
At
Replacing x(t) with e x(0) in (20) yields that for all
x(0) ∈ Rn , x(0) 6= 0 and for all t ∈ R+ \{0}
T
(21)
x(0)T eA t P eAt − P x(0) < 0.
In turn (21) implies
T
eA ĥi P eAĥi − P < 0, ∀ĥi ∈ H.
Using (6) we conclude that
ΦT (ĥi )P Φ(ĥi ) − P < 0, ∀ĥi ∈ H.
The proof of the statement follows by noticing that
Φ(ĥi ) = Âi , for all i ∈ ID
1 .
Lemma 1 establishes the connection between the stability
of ΣLTI and the quadratic stability of ΣLS . Hence the proof
of conservation of stability in the directions of 4 and 8
on Figure 1 is done. The following lemma establishes the
remaining part of the conservation of stability argument
stated in the beginning of this section (namely, in the
directions of 1 and 7 on Figure 1).
Lemma 2. Let ΣLS = ({(Âi , B̂i , C)}D
i=1 ) be a quadratically stable LS SS representation of the form (8) and
P > 0 be a solution of (19). If the left inverse V −1
of the matrix V ∈ Rn×r in Theorem 2 is chosen as
¯ ¯
V −1 = (V T P V )−1 V T P , then Σ̄LS = ({(Âi , B̂i , C̄)}D
i=1 )
in Theorem 2 is also quadratically stable.
Proof. See Appendix A.
Remark 2. Note that even though Lemma 2 is presented in
this paper as a step on proving the stability of the reduced
order models for the sampled data systems computed
with Approach 1 or Approach 2, on the condition of
stability of the original plant; it can also be considered
as an independent stability result for model reduction of
discrete-time LS SS representations.
With Lemma 2 we have established the connection between the quadratic stability of ΣLS and Σ̄LS in the direction of 7 in Figure 1. Using this proof, proving the
counterpart argument in the direction of 1 in Figure 1 is
trivial 3 . Therefore, the first statement in the beginning
of the section has been proven.
5.2 Main Stability Result
With the following theorem, we can relate the conservation
of stability argument to the discrete-time, time varying
representations of the reduced order aperiodically sampled
data systems of the form (12) and (18) respectively.
Theorem 3. Let ΣLTI = (A, B, C) be stable. Then
(i) The model Σ̄disc of the form (12) corresponding to Σ̄LS
computed with Approach 1 is quadratically stable.
(ii) The model Σ̄disc of the form (18) corresponding to Σ̄LS
computed with Approach 2 is quadratically stable.
Proof. The proof of part (i) follows directly from the
subsequent application of Lemma 2 and 1 to ΣLTI (note
that to apply Lemma 2 and 1 to ΣLTI = (A, B, C) in this
order, one simply considers ΣLTI as an LS SS representation consisting only of one LTI system (A, B, C)).
The proof of part (ii) follows directly from the subsequent
application of Lemma 1 and 2 to ΣLTI .
6. NUMERICAL EXAMPLES
In this section, two generic numerical examples are presented to illustrate and compare the two proposed model
reduction procedures 4 .
Example 1
In the first example, the two approaches are applied
to get a reduced order model for a single-input singleoutput (SISO), stable system ΣLTI = (A, B, C) of order
50, sampled to form the sampled data system ΣSD =
(A, B, C, H) with H = {ĥ1 , ĥ2 , ĥ3 , ĥ4 } = {1, 1.5, 2, 3}.
For Approach 1, firstly a reduced order continuous-time
LTI 17-partial realization Σ̄LTI of ΣLTI with order 18 is
computed. Then this model is used to get the reduced
order (of order 18) LS SS model Σ̄LS of the sampled data
system ΣSD . For simulation, the output sequence yk of the
original sampled data system ΣSD and ȳk of the reduced
order LS model Σ̄LS are acquired for k = IK
0 where K + 1
is the number of sampling instants of the simulation; by
applying the same white Gaussian noise input sequence
u = u0 · · · uK , uk ∈ N (0, 1) for all k ∈ IK
0 and sampling
sequence h = h0 · · · hK , hk ∈ H for all k ∈ IK
0 . For this
example, the total time horizon is chosen as [0, 50]. For
each simulation, the distance of the values ȳk to the values
yk , k ∈ IK
0 are compared with the best fit rate (BFR)
[Ljung (1999)] which is defined as
3
After this statement, one must also remark that numerical issues
related to the particular implementation of the moment matching
algorithm can cause instability in the reduced order model even in
the linear case and even when the original system is stable, [Antoulas
(2005)].
4 The implementation of the two approaches in Matlab
is freely available (with the examples) for experimentation
from https://sites.google.com/site/mertbastugpersonal/. The original system parameters used and reduced order system parameters
computed can be obtained from the same site.
Response of the Different System Models
20
4.5
×104
Response of the Different System Models
Continuous Output
Output at the Sampling Instants
Approach 1: Output of LSS Model Based on Reduced LTI Plant
Approach 2: Output of the reduced LSS model
4
15
3.5
10
3
2.5
y(t)
y(t)
5
0
2
1.5
-5
1
-10
Continuous Output
Output at the Sampling Instants
Approach 1: Output of LSS Model Based on Reduced LTI Plant
Approach 2: Output of the reduced LSS model
-15
-20
0.5
0
-0.5
0
5
10
15
20
25
30
35
40
45
50
0
0.5
1
t
2
2.5
3
3.5
4
4.5
5
t
Fig. 2. Example 1: The outputs resulting from the two
approaches compared with the original output. For
this simulation, the BFR of Approach 1 is 53.8214%
and the one of Approach 2 is 98.3700%.
1.5
qP
K
2
k=0 kyk − ȳk k2
BFR = 100% max 1 − qP
K
2
k=0 kyk − ym k2
, 0
(22)
where ym is the mean of the sequence {yk }K
k=0 . The mean
of the BFRs for 200 such different simulations is acquired
as 53.8303% whereas the best BFR is acquired as 73.9027%
and the worst as 23.0697%. The output sequence {ȳk }K
k=0
of the simulation giving the closest value to the mean of
the BFRs over this 200 simulations is illustrated in Figure
2 together with the original output sequence {yk }K
k=0 .
Then Approach 2 is applied to the same example.
Firstly the original LTI SS representation ΣLTI is used
to construct to LS SS representation ΣLS which models
the behavior of the sampled data system with respect to
the sampling interval set H. The model ΣLS is then used
to get the reduced order LS SS representation Σ̄LS using
Theorem 2. The reduced order LS SS representation Σ̄LS
in this case is a 2-partial realization of order 18. When
the same 200 simulations is done with this model with the
specifications given for Approach 1 of this example, the
mean of the BFRs over these simulations is 98.4222%, best
99.6449%, worst 95.5082%. Again, the output sequence
giving the closest value to the mean of the BFRs over this
200 simulations is illustrated in Figure 2 together with the
original output sequence.
Example 2
In the second example, the two approaches are applied
to get a reduced order model for a SISO unstable system ΣLTI = (A, B, C) of order 10 sampled to form the
sampled data system ΣSD = (A, B, C, H) with H =
{ĥ1 , ĥ2 , ĥ3 , ĥ4 } = {0.1, 0.15, 0.2, 0.3}. For Approach 1,
a reduced order continuous-time LTI 3-partial realization
Σ̄LTI of ΣLTI with order 4 is computed. Then this model is
used to get the reduced order (of order 4) LS SS model
Σ̄LS of the sampled data system ΣSD . The simulations
are done with the input and sampling sequences with the
specifications analagous to the ones given for Example 1.
Fig. 3. Example 2: The outputs resulting from the two
approaches compared with the original output. For
this simulation, the BFR of Approach 1 is 91.4432%
and the one of Approach 2 is 96.4294%.
Table 1. The mean of BFRs over 200 simulations for both of the examples using Approach 1 and Approach 2
Ex. / App.
Example 1
Example 2
Approach 1
53.8303%
91.5237%
Approach 2
98.4222%
96.1276%
For this example, the total time horizon is chosen as [0, 5].
The mean of the BFRs for 200 simulations is acquired as
91.5237% whereas the best BFR is acquired as 94.7476%
and the worst as 76.8753%. The output sequence {ȳk }K
k=0
of the simulation giving the closest value to the mean of
the BFRs over this 200 simulations is illustrated in Figure
3 together with the original output sequence.
Then Approach 2 is applied to the same example.
Firstly the original LTI SS representation ΣLTI is used
to construct to LS SS representation ΣLS which models
the behavior of the sampled data system with respect to
the sampling interval set H. The model ΣLS is then used
to get the reduced order LS SS representation Σ̄LS using
Theorem 2. The reduced order LS SS representation Σ̄LS
in this case is a 0-partial realization of order 4. When
the same 200 simulations is done with this model with
the corresponding same input and sampling sequences
used for Approach 1, the mean of the BFRs is acquired
96.1276%, best 97.8198%, worst 91.2306%. Again, the
output sequence giving the closest value to the mean of
the BFRs over this 200 simulations is illustrated in Figure
3 together with the original output sequence.
The results of the two simulations for both of the examples
are summarized in Table 1 for Approach 1 and Approach 2. Intuitively, the reason why Approach 2 is superior to Approach 1 for these examples can be explained
as follows: Approach 1 is based on model reduction of
the original LTI plant where the sampling behavior is not
considered at all. Whereas Approach 2 applies the model
reduction on the model which already takes into account
the specific set of allowed sampling intervals. It should be
noted that this statement may change depending on the
specific example, since for the moment, no formal proof of
comparison for the two methods can be given.
7. CONCLUSIONS
Two approaches for model reduction of sampled data
systems by moment matching is proposed. One approach
relies on applying a classical model reduction by moment
matching algorithm to the original LTI plant whereas
the other relies on computing a reduced order model
from the LS SS model of the sampled data system. With
some numerical examples, the use of two approaches are
illustrated and compared. For both approaches, it is shown
that the stability of the original continuous-time LTI plant
guarantees the quadratic stability of the resulting reduced
order discrete-time LS SS model with respect to any finite
allowed sampling interval set.
Hristu-Varsakelis, D. and (Editors), W.L. (2005). Handbook of networked and embedded control systems.
Boston: Birkhauser.
Lin, H. and Antsaklis, P.J. (2009). Stability and stabilizability of switched linear systems: A survey of recent results. IEEE Transactions on Automatic Control, 54(2),
308 – 322.
Ljung, L. (1999). System Identification, Theory for the
User. Prentice Hall, Englewood Cliffs, NJ.
Shieh, L.S. and Chang, Y.F. (1984). Model simplification
and digital design of multivariable sampled-data control
systems via a dominant-data matching method. Applied
Mathematical Modelling, 8(10), 355–364.
Zhang, W. (2001). Stability analysis of networked control
systems. Ph.D. thesis, CASE Western Reserve University, Cleveland, OH.
Zhang, W., Branicky, M., and Phillips, S. (2001). Stability
of networked control systems. IEEE Control Systems
Magazine, 21(1), 84–99.
REFERENCES
Antoulas, A.C. (2005). Approximation of Large-Scale
Dynamical Systems. SIAM, Philadelphia, PA.
Barb, F.D. and Weiss, M. (1993). Model reduction techniques for sampled-data systems. Numerical Algorithms,
4, 47–64.
Bastug, M., Petreczky, M., Toth, R., Wisniewski, R., Leth,
J., and Efimov, D. (2015). Moment matching based
model reduction for LPV state-space models. In Proc.
of the 54th IEEE Conference on Decision and Control,
5334 – 5339. Osaka.
Bastug, M., Petreczky, M., Wisniewski, R., and Leth, J.
(2014). Model reduction by moment matching for linear
switched systems. In Proc. of the American Control
Conference (ACC), 3942 – 3947. Portland, OR, USA.
Bastug, M., Petreczky, M., Wisniewski, R., and Leth, J.
(2016). Model reduction by nice selections for linear
switched systems. IEEE Transactions on Automatic
Control, PP, 1–1. doi:10.1109/TAC.2016.2518023.
Brockett, R.W. (1997). Minimum attention control. In
Proc. of the 36th IEEE Conference on Decision and
Control (CDC), 2628 – 2632. San Diego, CA, USA.
Donkers, M.C.F., Hetel, L., Heemels, W.P.M.H., van de
Wouw, N., and Steinbuch, M. (2009). Hybrid Systems:
Computation and Control, chapter Stability Analysis of
Networked Control Systems Using a Switched Linear
Systems Approach, 150–164. Springer, Berlin Heidelberg.
Donkers, M., Tabuada, P., and Heemels, W. (2011). On
the minimum attention control problem for linear systems: A linear programming approach. In Proc. of the
50th IEEE Conference on Decision and Control and
European Control Conference (CDC-ECC), 4717 – 4722.
Orlando, FL, USA.
Gu, K., Kharitonov, V.L., and Chen, J. (2003). Stability
of Time-Delay Systems. Springer US, New York, NY.
Hespanha, J., Naghshtabrizi, P., and Xu, Y. (2007). A
survey of recent results in networked control systems.
IEEE Special Issue on Technology of Networked Control
Systems, 95(1), 138–162.
Hetel, L., Fiter, C., Omran, H., Seuret, A., Fridman,
E., Richard, J.P., and Niculescu, S.I. (2017). Recent
developments on the stability of systems with aperiodic
sampling: an overview. Automatica, accepted.
Appendix A. PROOF OF LEMMA 2
Below, we will use the following simple claims:
(C1) If S ∈ Rn×n is symmetric negative (respectively
positive) definite, then Ŝ = V T SV is also symmetric
negative (respectively positive) definite.
(C2) V −1 = (V T P V )−1 V T P
(V T P V )−1 .
=⇒ V −1 P −1 (V −1 )T =
(CS) (Schur Complement Lemma for positive/negative
definiteness). Let S ∈ Rn×n be a symmetric positive
definite matrix and G ∈ Rn×n . Then GT SG − S < 0 ⇐⇒
GS −1 GT − S −1 < 0.
Note that, by the assumption of the theorem, (19) holds.
Multiplying (19) by V T from left and V from right for all
i ∈ ID
1 and using (C1) yields
T
V T ÂT
i P Âi V − V P V < 0,
By (CS) it follows that
∀i ∈ ID
1 .
−1
Âi V (V T P V )−1 V T ÂT
< 0, ∀i ∈ ID
(A.1)
i −P
1 .
In turn, multiplying (A.1) by V −1 from left and (V −1 )T
from right for all i ∈ ID
1 and using (C1) yields
−1 T
V −1 Âi V (V T P V )−1 V T ÂT
) − V −1 P −1 (V −1 )T < 0,
i (V
(A.2)
T
for all i ∈ ID
1 . Using (C2) and choosing P̄ = V P V , the
inequality (A.2) can be rewritten as
¯
¯
−1
< 0, ∀i ∈ ID
(A.3)
Âi P̄ −1 ÂT
i − P̄
1 .
Finally, using (CS) one more time for (A.3) yields
¯
¯
∀i ∈ ID
(A.4)
ÂT
1 .
i P̄ Âi − P̄ < 0,
T
Since P̄ = V P V is symmetric and positive definite
by (C1), (A.4) proves the quadratic stability of Σ̄LS =
¯ ¯
({(Âi , B̂i , C̄)}D
i=1 ).
| 3cs.SY
|
2016 1st International Conference on New Research Achievements in Electrical and Computer Engineering (ICNRAECE)
May 13, 2016
(AmirKabir University of Technology) ـــTehran, Iran
Multi-focus image fusion using VOL and EOL in
DCT domain
Mostafa Amin-Naji
Ali Aghagolzadeh
Faculty of Electrical and Computer Engineering
Babol Noshirvani University of Technology
Babol, Iran
[email protected]
Faculty of Electrical and Computer Engineering
Babol Noshirvani University of Technology
Babol, Iran
[email protected]
and temperature sensors. Also there are some limitations such
as limited band width, energy consumption and processing
time, which leads us to process the local input images for
decreasing the amount of transmission data [3]. Therefore
many researchers are seeking efficient methods for multi-focus
image fusion.
Abstract—The purpose of multi-focus image fusion is
gathering the essential information and the focused parts from
the input multi-focus images into a single image. These multifocused images are captured with different depths of focus of
cameras. Multi-focus image fusion is very time-saving and
appropriate in discrete cosine transform (DCT) domain,
especially when JPEG images are used in visual sensor networks
(VSN). The previous works in DCT domain have some errors in
selection of the suitable divided blocks according to their
criterion for measurement of the block contrast. In this paper, we
used variance of Laplacian (VOL) and energy of Laplacian
(EOL) as criterion to measure the contrast of image. Also in this
paper, the EOL and VOL calculations directly in DCT domain
are prepared using vector processing. We developed four
matrices which calculate the Laplacian of block easily in DCT
domain. Our works greatly reduce error due to unsuitable block
selection. The results of the proposed algorithms are compared
with the previous algorithms in order to demonstrate the
superiority of the output image quality in the proposed methods.
The several JPEG multi-focus images are used in experiments
and their fused image by our proposed methods and the other
algorithms are compared with different measurement criteria.
Several multi-focus image fusion researches have been
done in the spatial domain [4-14]. The simplest methods in the
spatial domain used a weighted arithmetic mean of the source
image pixels intensity [14]. This method is associated with an
image blurring and decreasing the image contrast. Multi-scale
image fusion methods are very common and convenient. The
Laplacian pyramid transform [15], gradient pyramid-based
transform [16] and the premier ones, discrete wavelet
transform (DWT) [17] and shift-invariant wavelet transform
(SIDWT) [18] are some examples of image fusion methods
based on multi-scale transform. Because of computational
complexity in multi-scale methods, they need more processing
time and energy consumption. Also image fusion methods
based on discrete wavelet transform need a large number of
convolution operations. In addition, due to creating ringing
phenomena in edge place of image, quality of the output image
is reduced.
Keywords—multi-focus; image fusion; VSN; DCT; energy;
variance; laplacian
Because of aforementioned problems in multi-scale
transform based image fusion methods, the researchers tended
to multi-focus images fusion in DCT domain. The DCT based
methods have simple calculation and are suitable for
implementation in real-time applications when images are
compressed in joint photographic experts group format (JPEG)
[19-21]. Tang [22] proposed two image fusions techniques
DCT+Average and DCT+Contrast. These methods have
undesirable side effects on the output images like blurring and
blocking artifact which reduces the quality of the output image.
Haghighat et al. in [19] introduced the multi-focus image
fusion method based on variance in DCT domain
(DCT+Variance). This method divides the input images into
8×8 blocks and then creates a merged output image by
selecting the corresponding blocks which have larger variance
value as a criterion for evaluation of image contrast. In [23]
Phamila introduced a method called DCT+AC-Max. This
I. INTRODUCTION
Image fusion is used to gather special and necessary
information from multiple images into fewer images or
preferably a single image. The fused image includes all
important information from the input images, ideally, and it’s
more accurate explanation of the scene than each input images.
Due to limited depth of focus in cameras, it is difficult to
capture an image that all components be obvious in it. This
issue occurs in optical lenses of CCD/CMOS cameras [1].
Therefore, some parts of the captured images with camera
sensors in visual sensor network (VSN) are blurred. In VSN
there is capability to record images with different depth of
focuses using several cameras [2]. The camera generate large
amount of data compared to the other sensors such as pressure
728
bxx + byy = −b(x − 1, y − 1) − 4 × b(x − 1, y)
− b(x − 1, y + 1) − 4 × b(x, y − 1) + 20 × b(x, y)
−4 × b(x, y + 1) − b(x + 1, y − 1)
−4 × b(x + 1, y) − b(x + 1, y + 1)
(5)
method selects the blocks which has more number of higher
values AC coefficient in DCT domain. This method can’t
always choose the blocks properly, because the number of
higher valued AC coefficients as a fusion criterion is invalid
when the majority of AC coefficients are zero. So it makes a
mistake in selection of the proper focused block. The DCT+SF
method in DCT domain is introduced by Cao et al [24]. This
method selects the block with higher value of the spatial
frequency which is computed for each block. Although these
methods (DCT+Variance, DCT+AC-Max and DCT+SF) have
advantages over the previous methods, due to selection the
unsuitable block from input images, the quality of the output
image is reduced.
bxx + byy can be computed with convolving mask (6) on
8×8 block. The size of mask (6) is 3×3 and the output matrix
from convolution of the mask on the block is a 6×6 matrix.
Also the EOL value is computed by the sum of the squares of
6×6 matrix elements.
In order to improve the quality of the output fused image,
new image fusion methods in DCT domain are presented in
this paper. These efficient methods are based on energy of
Laplacian (EOL) and variance of Laplacian (VOL). In multifocus image fusion process, EOL and VOL are suitable
criterions for measurement the contrast of image. The multifocus image fusion using EOL is introduced in the spatial
domain [7, 9]. But multi-focus image fusion methods based on
DCT domain are more favorable and more useful for VSN and
real-time applications. So in this article, we calculate the EOL
and VOL in DCT domain using vector processing and they are
used as a criterion for measure the contrast of the given image.
The proposed methods increase the quality of the output image
substantially.
1
0
0
0
𝑚=
0
0
0
(0
−1
−4
−1
0
𝑛=
0
0
0
(0
Two-dimensional DCT transform of N×N blocks of the
image b(m, n) and its inverse DCT using vector processing are
given as (1) and (2), respectively.
(1)
(2)
b = C . B. C
(3)
−4
−20
−4
0
𝑒=
0
0
0
( 0
and
A. EOL Calculation In DCT Domain
Energy of Laplacian (EOL) measures the image border
sharpness and it's calculated by (4) in the spatial domain [9].
EOL = ∑ ∑(bxx + byy )
x
2
-4
+20
-4
-1
-4
-1
(4)
y
where
729
0
1
0
0
0
0
0
0
0
−1
−4
−1
0
0
0
0
0
0
0
0
𝑑=
0
0
0
(0
where C and C t are orthogonal matrix consisted of the cosine
coefficients and the transpose coefficients, respectively. B is
the DCT coefficients for image's matrix of b. For C, we have:
C −1 = C t
-1
(6)
We are defining m, n, d and e matrices as below:
II. PROPOSED METHOD
t
-4
The 6×6 output matrix can be equivalent with the definition
of four matrices which their multiplication on 8×8 block results
in relation (5).
This article is organized as follows: In section 2, the EOL
and VOL calculations in DCT domain are introduced. The
proposed methods are explained in section 3 and it is compared
with the previous algorithms using different experiments in
section 4. Finally conclusions are presented in section 5.
B = C. b. C t
-1
0
−4
−20
−4
0
0
0
0
1
0
1
0
0
0
0
0
0
0
−1
−4
−1
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
−4
−20
−4
0
0
0
0
1
0
1
0
0
0
0
0
0
1
0
1
0
0
0
0
0
0
−1
−4
−1
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
1
0
0
0
0
0
0
−1
−4
−1
0
0
0
0
1
0
0
0
0
0
0
0
−4
−20
−4
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0)8×8
0
0
0
0
0
−1
−4
−1
0
0
0
0
1
0
0
0
0
0
0
0
−4
−20
−4
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0)8×8
0
0
0
0
0
0
0
0)8×8
0
0
0
0
0
−4
−20
−4
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0)8×8
q=m.b.n , p=d.b.e
N−1 N−1
(7)
1
σ = 2 ∑ ∑ Laplacian2 (k, l) − μ2
N
2
Energy of Laplacian (4), as the sum of entry-wise products
of the elements, can be rewritten by the trace of a product by
(8).
where μ is mean value of Laplacian of the block.
2
Haghighat et.al in [19] computed the variance of 8×8 block
in DCT domain is as (20):
EOL = ∑ ∑(bxx + byy ) = ∑ ∑(q + p)2
x
y
x
y
= trace((q + p)(q + p)t )
7
(8)
7
σ 2 DCT = ∑ ∑
where m, n, d and e are matrices that result in (5).
k=0 l=0
Defining:
d2 (k, l)
− d(0,0)
64
(20)
where d(k, l) is the DCT representation of the block.
P = C t . p. C , Q = C t . q. C
t
(9)
t
M = C . m. C , N = C . n. C
(10)
D = C t . d. C , E = C t . e. C
(11)
For calculate variance of image Laplacian directly in DCT
domain we should replace d(k,l) in (20) with achieved
LaplacianDCT in (15). Finally the variance of Laplacian (VOL)
in DCT domain is obtained as (21):
p and q can be defined as:
t
(19)
k=0 l=0
t
7
t
t
q=C. Q. C = C. M. C . C. B. C . C. N. C = C. M. B. N. C
t
p=C. P. C t = C. D. C t . C. B. C t . C. E. C t = C. D. B. E. C t
VOLDCT = ∑ ∑
(12)
k=0 l=0
(13)
(15)
When b is a matrix of the block values and B is its DCT,
we have:
trace(b. bt ) = trace(B. B t )
(16)
The region of the focused image has more information and
high contrast. Subsequently this region has more raised and
evident edges. The amount and intensity of edges in image is
used as criterion to specify the image quality and contrast.
Energy of Laplacian and variance of Laplacian are the
appropriate measures to show the amount of edges in image.
Therefore, the block of image which comes from focused area
has higher EOL or VOL value than the block of the unfocused
area. The proposed method divides the input images into 8×8
blocks. In next step, DCT coefficients of each block are
calculated. The first input block and the second one are named
as imA and imB, respectively. Then EOL or VOL value is
computed for every block in DCT domain. Block with higher
EOL or VOL value is represented as focused area and it is
selected for the output fused image. The proposed algorithm
makes a decision map X(i,j) as (22) by comparing the EOL or
VOL value of the corresponding blocks.
The EOL in DCT domain can be written as (17) using (8)
and (16).
EOLDCT = trace((Q + P)(Q + P)t ) =
trace((Q + P)(Qt + P t ))
(17)
where P and Q, according to (9), are DCT of q and p,
respectively.
Finally, EOL in DCT domain for B (DCT representation of
8×8 block) is calculated by combining (14) and (17):
EOLDCT =
trace((M. B. N + D. B. E) × ((M. B. N)t + (M. B. N)t )
(21)
C. Block Selection
For a simple description of the proposed algorithm, two
images A and B are considered. The proposed image fusion
process could be extended for more than two images. It is
assumed the input images were aligned by an image
registration method before performing the image fusion
process. The general structure of the proposed method for two
images fusion is shown in Fig.1.
(14)
In other word, the Laplacian of block in DCT domain is
computed as (15):
LaplacianDCT = Q + P
LaplacianDCT (k, l)
64
−LaplacianDCT (0,0)
From (8), (9), (10) and (11) we can find Q and P in DCT
domain as (14):
Q=M.B.N , P=D.B.E
7
(18)
B. VOL Calculation In DCT Domain
Variance of Laplacian (VOL) of image is calculated in
spatial domain as (19):
X(i, j) =
1 if (EOL or VOL)DCT (imA) > (EOL or VOL)DCT (imB)
{−1 if (EOL or VOL)DCT (imA) < (EOL or VOL)DCT (imB)
0
otherwise
(22)
730
Fig. 1. General structure of proposed method.
For X(i, j) = 1, the block of imA is selected for output
fused image. Subsequently for X(i, j) = −1, the block of imB
is selected.
of image fusion are used. The structural similarity (SSIM) [28]
D. Consistency Verification (CV)
Suppose the central block of an area among selected
blocks for fused image comes from image B, but the majority
of neighboring blocks comes from image A. It means that the
central block should belong to image A. Li et al. used the
majority filter for consistency verification [17]. The central
block replaced by the corresponding block from image A
using a majority filter which is applied on decision map
M(m,n). Therefore the consistency verification (CV) is applied
as a post-processing after image fusion process to improve the
quality of the output image and reduce the error due to
unsuitable block selection.
Fig. 2. Standard test images used for simulation
and Mean-squared error (MSE) [29] are used as metrics that
needs the ground truth image for the referenced images. The
total information transferred from source images to the fused
image (QAB/F ), the total loss of information (LAB/F), noise or
artifacts added in fused image due to fusion process (NAB/F)
which provided by Petrovic [30], and feature mutual
information (FMI) [31] are used for the non-reference images
which their ground truth image is not available.
III. EXPERIMENTAL RESULTS AND ANALYSIS
This section discusses the performance of the proposed
method and exhibits the results of simulation for the proposed
method and also the other methods for comparison. The results
of simulation for the proposed method compared with the
results of the previous methods such as the methods which are
based on multi-scale transform like DWT [17] and SIDWT
[18], and the methods based on DCT domain like
DCT+Average [21], DCT+Variance [19], DCT+ AC_Max [23]
and DCT+SF [24]. The used test images in simulations shown
in Fig.2 obtained from the online database [25] as the
referenced images. Also the famous non-referenced multifocus image “Disk” is used for second experiment obtained
from the online database [26]. The DCT+Variance algorithm
code of MATLAB simulation is taken from the online database
[27]. For the wavelet based methods, DWT with DBSS (2,2)
and the SIDWT for Haar basis with three levels of
decomposition are considered. Simulations of these methods
are done by “Image Fusion Toolbox” [26]. All of the
DCT+Average, DCT+ AC_Max and DCT+SF methods are
simulated by the authors in MATLAB. For the majority filter
used in CV, an averaging mask of size 5×5 is considered for all
experiments.
The first experiment conducted for applying the proposed
methods and the other methods on 12 pairs of artificial multifocus images generated from six images depicted in Fig.2. The
non-focused conditions of each pairs produced by artificial
blurring using two averaging masks of size 5×5 and 9×9.
Blurring process is performed on right or left half of the
images. The average values of SSIM and MSE for the
proposed and the other algorithms are listed in TABLE I. Our
proposed method shows the best results among the other
methods. The second experiments conducted for assessment of
the proposed algorithm and the others on real multi-focus
image (“Disk”) which it was captured with different depth of
focuses in camera. Evaluation performance metrics QAB/F,
LAB/F, NAB/F and FMI values are showed in TABLE II. Fig.3
depicts the proposed method result image, the other methods
result images and their local magnified version of “Disk”
source image. Thus in the non-referenced image, the proposed
method also has better results and quality.
For assessment of the proposed algorithm and comparing it
with the previous algorithms, evaluation performance metrics
731
THE MSE AND SSIM COMPARISION OF THE VARIOUS IMAGE
FUSION APPROACHES ON REFERENCE IMAGES.
Method
Average values for 12 pairs image
created from image shown in Fig.2
IV. CONCLUSION
Two new multi-focus image fusion methods in DCT
domain based on energy of Laplacian (EOL) and variance of
Laplacian (VOL) were introduced in this paper. We calculate
EOL and VOL directly in DCT domain using vector
processing. Considering EOL and VOL as a measurement of
image contrast in multi-focus image fusion process caused
better results compared with the previous works. Also due to
simple implementation of the proposed algorithms in DCT
domain, it is appropriate for using in real-time applications.
Accuracy of the proposed methods is assessed by applying the
proposed algorithms and the others on the several referenced
images and one non-referenced image by evaluating the results
with the various performance metrics. The results show
superiority of the output image quality for the proposed
algorithms comparison to some precious algorithms.
SSIM
MSE
DCT+Average [22]
0.8983
68.9065
DWT [17]
0.9542
20.5945
SIDWT [18]
0.9563
17.3855
DCT+Variance[19]
0.9683
19.2826
DCT+AC-Max [23]
0.9894
5.1054
DCT+SF [24]
0.9870
6.9794
DCT+EOL (proposed)
0.9944
2.2789
DCT+VOL (proposed)
0.9950
1.8950
DCT+Variance+CV [19]
0.9895
10.4003
REFERENCES
DCT+AC-Max+CV [23]
0.9965
1.7855
DCT+SF+CV [24]
0.9958
2.8608
DCT+ EOL +CV
(proposed)
0.9972
1.0205
DCT+ VOL +CV
(proposed)
0.9971
0.9297
[1] B.K.S. Kumar, M..N.S Swamy, and M. O. Ahmad, “Multiresolution
DCT decomposition for multifocus image fusion,” in 2013 26th Annual
IEEE Canadian Conference on Electrical and Computer Engineering
(CCECE), Regina, pp. 1-4, 2013.
[2] M. B. A. Haghighat, A. Aghagolzadeh, and H. Seyedarabi, “Real-time
fusion of multi-focus images for visual sensor networks,” in 2010 6th
Iranian Machine Vision and Image Processing (MVIP), Isfahan, pp. 1-6,
2010.
[3] F. Castanedo, J. García, M. Patricio, and J. M. Molina, “Analysis of
distributed fusion alternatives in coordinated vision agents,” 2008 11th
International Conference on Information Fusion, Cologne, pp. 1-6 ,
2008.
[4] S. Li, and B. Yang, “Multifocus Image Fusion Using Region
Segmentation and Spatial Frequency,” Image and Vision Computing,
vol. 26, no. 7, pp. 971 – 979, 2 July 2008.
[5] S. Mahajan and A. Singh, “A Comparative Analysis of Different Image
Fusion Techniques,” IPASJ International Journal of Computer Science
(IIJCS), vol. 2, no. 1, pp. 634-642, 2014.
[6] S. Pertuz, D. Puig, and M. A. Garcia, “Analysis of focus measure
operators for shape-from-focus,” Pattern Recognition, vol. 46, no. 5, pp.
1415-1432, 2013.
[7] W. Hongmei, N. Cong, L. Yanjun, and C. Lihua, “A Novel Fusion
Algorithm for Multi-focus Image,” In Applied Informatics and
Communication, vol. 227, pp. 641-647, 2011.
[8] P. Kaur and M. Kaur, “A Comparative Study of Various Digital Image
Fusion Techniques: A Review,” International Journal of Computer
Applications, vol.114, no. 4, 2015.
[9] W. Huang and Z. Jing, “Evaluation of focus measures in multi-focus
image fusion,” Pattern Recognition Letters, vol. 28, no. 4, pp. 493-500,
2007.
[10] S. Pertuz, D. Puig, and M. A. Garcia, “Analysis of focus measure
operators for shape-from-focus,” Pattern Recognition, vol. 46, no. 5, pp.
1415-1432, 2013.
[11] H. Zhao, Q. Li and H. Feng, “Multi-focus color image fusion in the
HSI space using the sum-modified-laplacian and a coarse edge map,”
Image and Vision Computing, vol. 26, no. 9, pp. 1285-1295, 2008.
[12] V. Kazemi, H. Seyedarabi, and A. Aghagolzadeh. “Multifocus image
fusion based on compressive sensing for visual sensor networks.” 22nd
Iranian Conference on Electrical Engineering (ICEE), 2014.
[13] Z. Zhang and R. S. Blum, “A categorization of multiscaledecomposition-based image fusion schemes with a performance study
for a digital camera application,” Proceedings of the IEEE, vol. 87, no.
8, pp. 1315-1326, 1999.
THE QAB/F, LAB/F, NAB/F AND FMI COMPARISION OF THE
VARIOUS IMAGE FUSION APPROACHES ON NON-REFERENCED IMAGES.
“DISK”
Method
QAB/F
LAB/F
NAB/F
FMI
DCT+Average [22]
0.5187
0.4782
0.0063
0.9013
DWT [17]
0.6302
0.2552
0.3362
0.9039
SIDWT [18]
0.6694
0.2764
0.1564
0.9049
DCT+Variance[19]
0.7165
0.2612
0.0478
0.9070
DCT+AC-Max [23]
0.6763
0.2910
0.0696
0.9057
DCT+SF [24]
0.7213
0.2600
0.0415
0.9086
DCT+EOL (proposed)
0.7271
0.2522
0.0444
0.9093
DCT+VOL (proposed)
0.7272
0.2521
0.0444
0.9094
DCT+Variance+CV [19]
0.7192
0.2734
0.0163
0.9100
DCT+AC-Max+CV [23]
0.6990
0.2922
0.0186
0.9102
DCT+SF+CV [24]
0.7269
0.2662
0.0153
0.9106
DCT+ EOL +CV
(proposed)
0.7291
0.2664
0.0096
0.9109
DCT+ VOL +CV
(proposed)
0.7288
0.2666
0.0097
0.9108
732
[14] H. Eltoukhy, S. Kavusi, “A computationally efficient algorithm for
multi-focus image reconstruction,” in Proceedings of SPIE Electronic
Imaging, vol. 5017, pp. 332–341, 2003.
[15] P. J. Burt, and E. H. Adelson. “The Laplacian pyramid as a compact
image code,” IEEE Transactions on Communications, vol. 31, no. 4,
pp. 532-540, 1983.
[16] V. Petrovic and C. Xydeas, “Gradient-Based Multiresolution Image
Fusion,” IEEE Transactions on Image Processing, vol. 13, no. 2, pp.
228-237, 2004.
[17] H. Li, B. Manjunath and S. Mitra, “Multisensor Image Fusion Using the
Wavelet Transform,” Graphical Models and Image Processing, vol. 57,
no. 3, pp. 235-245, 1995.
[18] O. Rockinger, “Image sequence fusion using a shift-invariant wavelet
transform,” in Proceedings of IEEE International Conference on Image
Processing, vol. 3, Santa Barbara, pp. 288-291, 1997.
[19] M. Haghighat, A. Aghagolzadeh and H. Seyedarabi, “Multi-focus image
fusion for visual sensor networks in DCT domain,” Computers &
Electrical Engineering, vol. 37, no. 5, pp. 789-797, 2011.
[20] M. A. Naji and A. Aghagolzadeh, “Multi-focus image fusion in DCT
domain based on correlation coefficient,” in 2015 2nd International
Conference on Knowledge-Based Engineering and Innovation (KBEI),
2015, pp. 632-639.
[21] M. A. Naji and A. Aghagolzadeh, “A new multi-focus image fusion
technique based on variance in DCT domain,” in 2015 2nd International
Conference on Knowledge-Based Engineering and Innovation (KBEI),
2015, pp. 478-484.
[22] J. Tang, “A contrast based image fusion technique in the DCT domain,”
Digital Signal Processing, vol. 14, no. 3, pp. 218-226, 2004.
[23] Y. Phamila and R. Amutha, “Discrete Cosine Transform based fusion of
multi-focus images for visual sensor networks,” Signal Processing, vol.
95, pp. 161-170, 2014.
[24] L. Cao, L. Jin, H. Tao, G. Li, Z. Zhuang and Y. Zhang, “Multi-Focus
Image Fusion Based on Spatial Frequency in Discrete Cosine Transform
Domain,” IEEE Signal Processing Letters, vol. 22, no. 2, pp. 220-224,
2015.
[25] [Online].Available:
[26]
[27]
[28]
[29]
[30]
[31]
http://www.imageprocessingplace.com/root_files_V3/image_databases.h
tm. [Accessed: Dec- 2015].
[Online]. Available: http://www.metapix.de/toolbox.htm. [Accessed:
Dec - 2015].
[Online].Available:
https://www.mathworks.com/matlabcentral/fileexchange/59499-multifocus-image-fusion-in-dct-domain. [Accessed: Dec - 2015].
Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image
quality assessment: from error visibility to structural similarity,” IEEE
Transactions on Image Processing, vol. 13, no.4, pp. 600-612, 2004.
Zhou Wang and A. Bovik, “Mean squared error: Love it or leave it? A
new look at Signal Fidelity Measures”, IEEE Signal Processing
Magazine, vol. 26, no. 1, pp. 98-117, 2009.
V. Petrović and C. Xydeas, “Objective image fusion performance
characterisation,” in 2009 Tenth IEEE International Conference on
Computer Vision (ICCV), Beijing, 2005, pp. 1866-1871.
M. Haghighat, A. Aghagolzadeh and H. Seyedarabi, “A non-reference
image fusion metric based on mutual information of image features,”
Computers & Electrical Engineering, vol. 37, no. 5, pp. 744-756, 2011.
Fig. 3. Source images ‘‘DISK” and the fusion results. (a) The first source image with focus on the right. (b) The second source image with focus on the
left. (c) DCT + Average result. (d) DWT result. (e) SIDWT result. (f) DCT + Variance result. (g) DCT + Ac-Max result. (h) DCT + Spatial frequency
result. (i) The result of DCT + EOL (Proposed). (j) )The result of DCT + VOL (Proposed). (k), (l), (m), (n), (o), (p), (q) and (r) are the local magnified
version of (c), (d), (e), (f), (g), (h), (i) and (j), respectively.
733
| 1cs.CV
|
FedCSIS 2017, IEEE, ISSN 2300-5963
doi 10.15439/978-83-946253-7
pp.537-541
Least Square Method Robustness of Computations
What is not usually considered and taught
Vaclav Skala
Department of Computer Science and Engineering
Faculty of Applied Sciences, University of West Bohemia
CZ 306 14 Plzen, Czech Republic
http://www.VaclavSkala.eu
Abstract — There are many practical applications based on the
Least Square Error (LSE) approximation. It is based on a square
error minimization “on a vertical” axis. The LSE method is simple
and easy also for analytical purposes. However, if data span is
large over several magnitudes or non-linear LSE is used, severe
numerical instability can be expected.
The presented contribution describes a simple method for
large span of data LSE computation. It is especially convenient if
large span of data are to be processed, when the “standard”
pseudoinverse matrix is ill conditioned. It is actually based on a
LSE solution using orthogonal basis vectors instead of
orthonormal basis vectors. The presented approach has been used
for a linear regression as well as for approximation using radial
basis functions.
Keywords—Least square error; approximation regression;
radial basis function; approximation; condition number; linear
algebra; geometric algebra; projective geometry.
from the given points in this space. This algorithm is quite
complex and solution can be found in [18].
It should be noted, that all methods above do have one
significant drawback as values are taken in a squared value. This
results to an artifact that small values do not have relevant
influence to the final entity as the high values. Some methods
are trying to overcome this by setting weights to each measured
data [3]. It should be noted that the TLSE was originally derived
by Pearson [16](1901). Deep comprehensive analysis can be
found in [8][13][21][22]. Differences between the LSE a TLSE
methods approaches are significant, see Fig.1.
y
y
x
x
I. INTRODUCTION
Wide range of applications is based on approximation of
acquired data and the LSE minimization is used, known also as
a linear or polynomial regression. The regression methods have
been heavily explored in signal processing and geometrical
problems or with statistically oriented problems. They are used
across many engineering fields dealing with acquired data
processing. Several studies have been published and they can be
classified as follows:
“standard” Least Square Error (LSE) methods fitting data to
a function 𝑦 = 𝑓(𝒙) , where 𝒙 is an independent variable
and 𝑦 is a measured or given value,
“orthogonal” Total Least Square Error (TLSE) fitting data to
a function 𝐹(𝒙) = 0 , i.e. fitting data to some 𝑑 − 1dimensional entity in this 𝑑-dimensional space, e.g. a line
in the 𝐸 2 space or a plane in the 𝐸 3 space [1][6][8][21][22],
“orthogonally Mapping” Total Least Square Error (MTLSE)
methods for fitting data to a given entity in a subspace of
the given space. However, this problem is much more
complicated. As an example, we can consider data given in
and we need to find an optimal line in 𝐸 𝑑 , i.e. one
dimensional entity, in this 𝑑-dimensional space fitting
optimally the given data. Typical problem: Find a line in
the 𝐸 𝑑 space that has the minimum orthogonal distance
Research was supported by the and National Science Foundations (GACR)
project No. 17-05534S.
Fig. 1.a: Least Square Error
Fig.1.b: Total Least Square Error
In the vast majority the Least Square Error (LSE) methods
measuring vertical distances are used. This approach is
acceptable
in
the
case
of
explicit
functional
dependences 𝑓(𝑥, 𝑦) = ℎ, resp. 𝑓(𝑥, 𝑦, 𝑧) = ℎ. However, it
should be noted that a user should keep in a mind, that smaller
differences than 1.0, will have significantly smaller weight than
higher differences than 1.0 as the differences are taken in a
square resulting to dependences in scaling of data approximated,
i.e. the result will depend on physical units used, etc. The main
advantage of the LSE method is that it is simple for fitting
polynomial curves and it is easy to implement. The standard
LSE method leads to over determined system of linear
equations. This approach is also known as polynomial
regression.
Let us consider a data set Ω = {〈𝑥𝑖 , 𝑦𝑖 , 𝑓𝑖 〉}𝑛𝑖=1 , i.e. data set
containing for 𝑥𝑖 ,𝑦𝑖 and measured functional value 𝑓𝑖 , and we
want to find parameters 𝒂 = [𝑎, 𝑏, 𝑐, 𝑑]𝑇 for optimal fitting
function, as an example:
𝑓(𝑥, 𝑦, 𝒂) = 𝑎 + 𝑏𝑥 + 𝑐𝑦 + 𝑑𝑥𝑦
Minimizing the vertical squared distance 𝐷, i.e.:
(1)
FedCSIS 2017, IEEE, ISSN 2300-5963
doi 10.15439/978-83-946253-7
𝑛
where 𝒃 = (𝑏1 , … , 𝑏𝑛 ), 𝝃 = (𝜉1 , … , 𝜉𝑚 ) and 𝑚 is a number of
parameters, 𝑚 < 𝑛.
2
𝐷 = min ∑(𝑓𝑖 − 𝑓(𝑥𝑖 , 𝑦𝑖 , 𝒂)) =
𝑎,𝑏,𝑐,𝑑
(2)
𝑖=1
𝑛
𝑚𝑖𝑛 ∑(𝑓𝑖 − (𝑎 + 𝑏𝑥𝑖 + 𝑐𝑦𝑖 + 𝑑𝑥𝑖 𝑦𝑖 ))
𝑎,𝑏,𝑐,𝑑
2
𝑖=1
Conditions for an extreme are given as:
𝜕𝑓(𝑥, 𝑦, 𝒂)
(3)
= [1, 𝑥, 𝑦, 𝑥𝑦]𝑇
𝜕𝒂
Applying this on the expression of 𝐷 we obtain
𝑛
𝜕𝐷
𝜕𝑓(𝑥, 𝑦, 𝒂)
∑(𝑓𝑖 − (𝑎 + 𝑏𝑥𝑖 + 𝑐𝑦𝑖 + 𝑑𝑥𝑖 𝑦𝑖 ))
= 0 (4)
𝜕𝒂
𝜕𝒂
𝑖=1
It leads to conditions for 𝒂 = (𝑎, 𝑏, 𝑐, 𝑑) parameteters in the
form of a linear system of equations 𝑨𝒙 = 𝒃:
𝑨=
𝑛
𝑛
𝑛
∑
𝑖=1
𝑛
∑
𝑖=1
𝑛
𝑛
𝑛
∑
𝑥𝑖
∑
𝑦𝑖
∑
𝑥𝑖 𝑦𝑖
𝑥𝑖
∑
𝑥𝑖2
∑
𝑥𝑖 𝑦𝑖
∑
𝑥𝑖2 𝑦𝑖
𝑦𝑖
∑
𝑥𝑖 𝑦𝑖
∑
𝑦𝑖2
∑
𝑥𝑖 𝑦𝑖2
∑
𝑥𝑖2 𝑦𝑖
∑
𝑥𝑖 𝑦𝑖2
∑
𝑥𝑖2 𝑦𝑖2
𝑖=1
𝑛
𝑖=1
𝑛
𝑖=1
𝑛
∑ 𝑥𝑦
[ 𝑖=1 𝑖 𝑖
𝑛
𝒃 = [∑
𝑖=1
𝑖=1
𝑛
𝑖=1
𝑛
𝑖=1
𝑛
𝑖=1
𝑖=1
𝑛
𝑛
𝑓𝑖 , ∑
𝑖=1
𝑛
𝑓𝑖 𝑥𝑖 , ∑
𝑖=1
𝑖=1
(5)
𝑖=1
𝑛
𝑖=1
𝑛
𝑓𝑖 𝑦𝑖 , ∑
𝑖=1
]
𝑻
𝑓𝑖 𝑥𝑖 𝑦𝑖 ]
The selection of bilinear form was used to show the LSE
method application to a non-linear case, if the case of a linear
function, i.e. 𝑓(𝑥, 𝑦, 𝒂) = 𝑎 + 𝑏𝑥 + 𝑐𝑦, the 4th row and column
are to be removed. Note that the matrix 𝑨 is symmetric and the
function 𝑓(𝒙) might be more complex, in general.
Several methods for LSE have been derived [4][5][10], however
those methods are sensitive to the vector 𝒂 orientation and not
robust in general as a value of ∑𝑛𝑖=1 𝑥𝑖2 𝑦𝑖2 might be too high in
comparison with the value 𝑛, which has an influence to
robustness of a numerical solution. In addition, the LSE methods
are sensitive to a rotation as they measure vertical distances. It
should be noted, that rotational and translation invariances are
fundamental requirements especially in geometrically oriented
applications.
The LSE method is usually used for a small size of data and
span of a domain is relatively small. However, in some
applications the domain span can easily be over several decades,
e.g. in the case of Radial Basis Functions (RBF) approximation
for GIS applications etc. In this case, the overdetermined system
can be difficult to solve.
II. NUMERICAL STABILITY
Let us explore a simple example, when many points 𝒙𝑖 ∈ 𝐸 2 ,
i.e. 𝒙𝑖 = (𝑥𝑖 , 𝑦𝑖 ) , are given with relevant associated values
𝑏𝑖 , 𝑖 = 1, … , 𝑛. Expected functional dependency can be
expressed (for a simplicity) as 𝑦 = 𝑎1 + 𝑎2 𝑥 + 𝑎3 𝑦. The LSE
leads to an overdetermined system of equations
𝑨𝑇 𝑨 𝝃 = 𝑨𝑇 𝒃
If the values 𝑥𝑖 , 𝑦𝑖 over a large span, e.g. 𝑥𝑖 , 𝑦𝑖 ∈ 〈100 , 105 〉,
the matrix 𝑨𝑇 𝑨 is extremely ill conditioned. This means that the
reliability of a solution depends on the distribution of points in
the domain. Situation gets worst when a non-linear polynomial
regression is to be used and dimensionality of the domain is
higher.
As an example, let us consider a simple case, when points
form regular orthogonal mesh and values are generated using 𝑅5
distribution scheme (equidistant in a logarithmic scale)
as (𝑥𝑖 , 𝑦𝑖 ) ∈ 〈10, 105 〉 × 〈10, 105 〉. It can be easily found using
MATLAB that conditional number 𝑐𝑜𝑛𝑑(𝑨𝑇 𝑨) ≅ 1011 .
In the following, we will show how the condition number
might be decreased significantly using orthogonal basis vectors
instead of the orthonormal ones.
III.
𝑖=1
𝑛
𝒙 = [𝑎, 𝑏, 𝑐 , 𝑑 ]𝑻
pp.537-541
(6)
PROJECTIVE NOTATION AND GEOMETRY ALGEBRA
The LSE approximation is based on a solution of a linear
system of equations, i.e. 𝑨𝒙 = 𝒃. Usually the Euclidean
representation is used. However if the projective space
representation is used [19] , it is transformed into homogeneous
linear system of equations, i.e. 𝑩𝜻 = 𝟎. Rewriting the Eq.(6), we
obtain
(7)
𝑩𝜻 = 𝟎
where
𝑩 = [−𝑨𝑇 𝒃|𝑨𝑇 𝑨]
(8)
𝜻 = (𝜁0 : 𝜁1 , … , 𝜁𝑚 )
𝜁
and 𝜉𝑖 = 𝑖⁄𝜁 , 𝑖 = 1, … , 𝑚; 𝜁0 is the homogeneous coordinate
0
in the projective representation, matrix 𝑩 size is 𝑚 × (𝑚 + 1).
Now, a system of homogeneous linear equations is to me solved.
It can be shown that a system of homogeneous linear
equations 𝑨𝒙 = 𝟎 is equivalent to the extended cross-product,
actually outer-product [19][20]. In general, solutions of the both
cases 𝑨𝒙 = 𝟎 and 𝑨𝒙 = 𝒃, i.e. homogeneous and nonhomogeneous system of linear equations, is the same and no
division operation is needed as the extended cross-product (outer
product) does not require any division operation at all. Applying
this we get:
(9)
𝜻 = (𝜁0 : 𝜁1 , … , 𝜁𝑚 ) = 𝜷1 ∧ 𝜷2 ∧ … ∧ 𝜷𝑚−1 ∧ 𝜷𝑚
where
(10)
𝜷𝑖 = [−𝑏𝑖0 : 𝑏𝑖1 , … , 𝑏𝑖𝑚 ]𝑇 𝑖 = 1, … , 𝑚
The extended cross-product can be rewritten using determinant
of (𝑚 + 1) × (𝑚 + 1) as
𝒆0
𝒆1
𝒆2 ⋯ 𝒆𝑚
−𝑏10 𝑏11 𝑏12 ⋯ 𝑏1𝑚
(11)
𝜻 = det [ ⋮
⋮
⋮
⋱
⋮ ]
−𝑏𝑚0 𝑏𝑚1 𝑏𝑚2 ⋯ 𝑏𝑚𝑚
where 𝒆0 are orthonormal basis vectors in the 𝑚-dimensional
space. As a determinant is a multilinear, we can multiply any 𝑗
column by a value 𝑞𝑗 ≠ 0
FedCSIS 2017, IEEE, ISSN 2300-5963
𝜻′ =
𝒆′0
−𝑏 ′
det [ 10
⋮
′
−𝑏𝑚0
where
𝒆1′
′
𝑏11
⋮
′
−𝑏𝑚1
𝒆′2
′
𝑏12
⋮
′
−𝑏𝑚2
𝒆𝑗
𝑞𝑗
⋯
⋯
⋱
⋯
doi 10.15439/978-83-946253-7
𝒆′𝑚
′
𝑏1𝑚
⋮
′
−𝑏𝑚𝑚
]
(12)
pp.537-541
Using the approach presented above, the conditional number
𝑇 𝑨) ≅ 2. 106 .
̅̅̅̅̅̅
was decreased significantly to 𝑐𝑜𝑛𝑑(𝑨
𝑏∗𝑗
(13)
𝑞𝑗
where 𝒆𝑗′ are orthogonal basis vectors in the 𝑚-dimensional
space.
𝒆𝑗′ =
′
𝑏∗𝑗
=
From the geometrical point of view, it is actually a
“temporary” scaling on each axis including the units. Of course,
a question remains – how to select the 𝑞𝑗 value. The 𝑞𝑗 is to be
selected as
𝑞𝑗 = max {|𝑏𝑖𝑗 |}
𝑖=1,…,𝑚
(14)
where 𝑗 = 1, … , 𝑚. Note that the matrix 𝑩 is indexed
as (0, … , 𝑚) × (0, … , 𝑚).
Fig.3: Conditionality of the modified matrix depending on number of data set
size, i.e. number of points
Comparing the condition numbers of the original and
modified matrices, we can see significant improvement of
matrix conditionality as
10
Applying this approach, we get a modified system
′ )
(15)
𝜻′ = (𝜁0′ : 𝜁1′ , … , 𝜁𝑚
= 𝜷1′ ∧ 𝜷′2 ∧ … ∧ 𝜷′𝑚−1 ∧ 𝜷′𝑚
where
′
′
′ 𝑇
]
(16)
𝜷′𝑖 = [−𝑏𝑖0
: 𝑏𝑖1
, … , 𝑏𝑖𝑚
′
𝑇 𝑨], i.e.
̅̅̅̅̅̅
̅ ´ = [−𝑨𝑇 𝒃|𝑨
where 𝜷𝑖 are coefficients of the matrix 𝑩
modified matrix 𝑩 as described above, for the orthogonal (not
orthonormal) vector basis.
𝑇
6.10
(19)
𝜐 = 𝑐𝑜𝑛𝑑(𝑨 𝑨)⁄
≅
= 3.104
𝑇
̅̅̅̅̅̅
𝑐𝑜𝑛𝑑(𝑨 𝑨) 2.106
In the case of a little bit more complex function defined by
Eq.(1), i.e. 𝑓(𝑥, 𝑦) = 𝑎 + 𝑏𝑥 + 𝑐𝑦 + 𝑑𝑥𝑦 we obtain
The approximated 𝑓(𝑥, 𝑦) value is computed as
𝑓(𝑥, 𝑦) = 𝑎𝑞1 + 𝑏𝑞2 𝑥 + 𝑐𝑞3 𝑦
in the case of 𝑓(𝑥, 𝑦) = 𝑎 + 𝑏𝑥 + 𝑐𝑦, or
(17)
(18)
𝑓(𝑥, 𝑦) = 𝑎𝑞1 + 𝑏𝑞2 𝑥 + 𝑐𝑞3 𝑦 + 𝑑𝑞4 𝑥𝑦
in the case 𝑓(𝑥, 𝑦) = 𝑎 + 𝑏𝑥 + 𝑐𝑦 + 𝑑𝑥𝑦 and similarly for the
general case of a regression function 𝑦 = 𝑓(𝒙, 𝒂).
The above presented modification is simple. However, what
is the influence of this operation?
Fig.4: Conditionality of the original matrix depending on number of data set
size, i.e. number of points
IV. MATRIX CONDITIONALITY
Let us consider a recent simple example again, when points
are generated from (𝑥𝑖 , 𝑦𝑖 ) ∈ 〈10, 105 〉 × 〈10, 105 〉. It can be
found that conditional number 𝑐𝑜𝑛𝑑(𝑨𝑇 𝑨) ≅ 6. 1010 using
MATLAB, Fig.2, if 𝑓(𝑥, 𝑦) = 𝑎 + 𝑏𝑥 + 𝑐𝑦 is used for the LSE.
Fig.5: Conditionality of the modified matrix depending on number of data set
size, i.e. number of points
In this case of the LSE defined by Eq.(1) the conditionality
improvement is even higher, as
20
Fig.2: Conditionality histogram of the original matrix depending on number of
data set size, i.e. number of points
𝑇
6.10
(20)
𝜐 = 𝑐𝑜𝑛𝑑(𝑨 𝑨)⁄
≅
= 109
𝑇
̅̅̅̅̅̅
𝑐𝑜𝑛𝑑(𝑨 𝑨) 6.1011
It means that better numerical stability is obtained by a simple
operation. All graphs clearly shows also dependency on a
number of points used for the experiments (horizontal axis).
FedCSIS 2017, IEEE, ISSN 2300-5963
doi 10.15439/978-83-946253-7
The geometric algebra brings also an interesting view on
problems with numerical solutions. Let us consider vectors ̂
𝜷𝑖
with coordinates of points, i.e.
̂ 𝑖 = [𝑏𝑖1 , … , 𝑏𝑖𝑚 ]𝑇 𝑖 = 1, … , 𝑚
(21)
𝜷
̂𝑖 ∧ 𝜷
̂𝑗 = 𝜸
̂𝑖𝑗 defines a bivector, which is an oriented
Then 𝜷
̂𝑖𝑗 ‖
surface, given by two vectors in 𝑚-dimensional space and ‖𝜸
gives the area represented by the bivector ̂𝜸𝑖𝑗 .
So, the proposed approach of introducing orthogonal basis
functions instead of the orthonormal ones, enable us to
“eliminate” influence of “small” bivectors in the original LSE
computation and increase precision of numerical computation.
Of course, if the regression is to be applied, the influence of
the 𝑞𝑗 values must be applied. By the presented approach we
actually got values 𝜁𝑖′ using the orthogonal basis vectors instead
of orthonormal. It means, that the estimated value by a
regression, using recent simple example, is
(22)
𝑓(𝑥, 𝑦) = 𝑞1 𝑎1 + 𝑞2 𝑎2 𝑥 + 𝑞3 𝑎3 𝑦
In the case of the least square approximation, we want to
minimize using a polynomial of degree 𝑛.
min ‖𝑓(𝑥) − 𝑃𝑛 (𝑥)‖
𝑘
(23)
𝑃𝑛 (𝑥) = ∑ 𝑎𝑖 𝑥 𝑖
𝑖=0
The 𝐿2 norme of a function 𝑓(𝑥) an an interval 〈𝑎, 𝑏〉 is defined
2
𝑏
‖𝑓(𝑥)‖ = √(∫ 𝑓(𝑥)𝑑𝑥 )
𝑏
𝑏
∑ 𝑎𝑖 ∫ 𝑥 𝑖+𝑘 𝑑𝑥 = ∫ 𝑥 𝑘 𝑓(𝑥)𝑑𝑥
𝑖=0
𝑎
(29)
𝑎
where 𝑘 = 1, … , 𝑛.
It means that the LSE problem is the polynomial (what has
been expected)
𝑘
𝑃𝑛 (𝑥) = ∑ 𝑎𝑖 𝑥 𝑖
(30)
𝑖=0
However, there is a direct connection with well known Hilbert’s
matrix. It can be shown that elements of the Hilbert’s matrix
(𝐻𝑛+1 (𝑎, 𝑏))𝑖,𝑘 of the size (𝑛 + 1) × (𝑛 + 1) are equivalent to
𝑏
1
(31)
(𝐻𝑛+1 (𝑎, 𝑏))𝑖,𝑘 = ∫ 𝑥 𝑖+𝑘 𝑑𝑥 =
1+𝑖+𝑘
𝑎
If interval 〈𝑎, 𝑏〉 = 〈0,1〉 is used, standard Hilbert’s
matrix 𝑯𝑛 (0,1) is obtained, which is extremely ill-conditioned.
VI. HILBERT’S MATRIX CONDITIONALITY
V. LEAST SQUARE METHOD WITH POLYNOMIALS
𝑃𝑛 (𝑥)
pp.537-541
𝑛
(24)
We should answer a question, how the conditional number
of the Hilbert’s matrix can be improved if orthogonal basis is
used instead of orthonormal one as an experimental test.
A simple experiment can prove that the proposed method
does not practically change the conditionality of the Hilbert’s
matrix 𝑯𝑛 (0,1). However, as the LSE approximation is to be
used for large span of data, it is reasonable to consider a general
case and explore conditionality of the 𝑯𝑛 (𝑎, 𝑏) matrix,
e.g. 𝑯5 (0, 𝑏), for demonstration.
𝑎
Minimizing square of the distance of a function of 𝑘 + 1
parameters 𝜑(𝒂) = 𝜑(𝑎0 , … , 𝑎𝑛 ) and using “per-partes” rule,
we obtain
𝑏
𝜑(𝒂) = ∫ [𝑓(𝑥) − 𝑃𝑛 (𝑥)]2 𝑑𝑥
𝑎
𝑏
=∫
[𝑓(𝑥)]2
𝑎
𝑛
𝑏
𝑑𝑥 − 2 ∑ 𝑎𝑖 ∫ 𝑥𝑖 𝑓(𝑥)𝑑𝑥
𝑖=0
𝑛 𝑛
𝑎
(25)
Fig.6: Conditionality of the 𝑯5 (0, 𝑏) for different values of 𝑏 using MATLAB
(numerical problems can be seen for 𝑏 > 650)
𝑏
+ ∑ ∑ 𝑎𝑖 𝑎𝑗 ∫ 𝑥 𝑖+𝑗 𝑑𝑥
𝑎
𝑖=0 𝑗=0
For a minimum a vector condition
𝜕𝜑(𝒂)
=𝟎
𝜕𝒂
must be valid. It leads to conditions
𝑛
𝑏
𝑏
𝜕𝜑(𝒂)
𝑘
= 0 − 2 ∫ 𝑥 𝑓(𝑥)𝑑𝑥 + ∑ 𝑎𝑖 ∫ 𝑥 𝑖+𝑘 𝑑𝑥
𝜕𝑎𝑘
𝑎
𝑎
𝑖=0
𝑛
𝑏
+ ∑ 𝑎𝑗 ∫ 𝑥
𝑗+𝑘
(26)
(27)
𝑑𝑥
Fig.7: Conditionality of the 𝑯5 (0, 𝑏) for different values of 𝑏
using logarithmic scaling for vertical axis
𝑎
𝑖=0
and by simple algebraic manipulations we obtain:
𝑏
𝑛
𝑘
𝑏
2 [− ∫ 𝑥 𝑓(𝑥)𝑑𝑥 + ∑ 𝑎𝑖 ∫ 𝑥
𝑎
and therefore
𝑖=0
𝑎
𝑖+𝑘
𝑑𝑥 ] = 0
(28)
It can be seen, that 𝑐𝑜𝑛𝑑(𝑯5 (0,800)) = 6. 1023 . If the
14
̅̅̅̅̅̅̅̅̅̅̅̅̅̅
proposed approach is applied 𝑐𝑜𝑛𝑑(𝑯
5 (0,800)) = 2,5. 10
for the modified matrix, Fig.8 - Fig.9.
FedCSIS 2017, IEEE, ISSN 2300-5963
doi 10.15439/978-83-946253-7
pp.537-541
VII. CONCLUSIONS
The proposed method of application orthogonal vector basis
instead of the orthonormal one decreases conditional number of
a matrix used in the least square method. This approach
increases robustness of a numerical solution especially when
domain data range is high. It can be used also for solving systems
of linear equations in general, e.g. if radial basis function
interpolation or approximation is used.
ACKNOWLEDGMENT
Fig.8: Conditionality of the modified 𝑯5 (0, 𝑏)
The author would like to thank to colleagues at the
University of West Bohemia in Plzen for fruitful discussions and
to anonymous reviewers for their comments and hints, which
helped to improve the manuscript significantly. Special thanks
belong to Zuzana Majdišová and Michal Šmolík for independent
experiments, images and generation in MATLAB.
REFERENCES
Fig.9: Conditionality of the modified 𝑯5 (0, 𝑏)
using logarithmic scaling for vertical axis
It means that the conditionality improvement
𝑐𝑜𝑛𝑑(𝑯5 (0,800))
6.1023
𝜐=
≅
≈ 109
(32)
̅̅̅̅̅̅̅̅̅̅̅̅̅̅
2,5.1014
𝑐𝑜𝑛𝑑(𝑯
5 (0,800))
This is a similar ratio as for the simple recent examples.
A change of the size of bivectors ‖𝜷𝑖 ∧ 𝜷𝑗 ‖ can be used as a
practical result using RBF approximation, which changes from
the interval 〈𝑒𝑝𝑠, 1010 〉 to 〈𝑒𝑝𝑠, 8. 102 〉, which significantly
increases robustness of the RBF approximation, Fig.10.
Fig.10: Bivector histogram sizes for original LSE matrix and modified one
The proposed approach has been used for St.Helen’s volcano
approximation by 10 000 points instead of 6 743 176 original
points, see Fig.11.
Fig.11: LSE approximation error with RBF approximation of St.Helen’s
(image generated in MATLAB by Michal Smolik)
[1] Abatzoglou,T., Mendel,J. 1987. Constrained total least squares, IEEE
Conf. Acoust., Speech, Signal Process. (ICASSP’87), Vol. 12, 1485–1488.
[2] Alciatore,D., Miranda,R. 1995. The best least-square line fit, Graphics
Gems V (Ed.Paeth,A.W.), 91-97, Academic Press.
[3] Amiri-Simkooei,A.R.,Jazaeri,S. 2012.Weighted total least squares
formulated by standard least squares theory,J.Geodetic Sci.,2(2):113-124.
[4] Charpa,S., Canale,R. 1988. Numerical methods for Engineers, McGrawHill.
[5] Chatfield,C. 1970. Statistics for technology, Penguin Book.
[6] de Groen,P. 1996 An introduction to total least squares, Nieuw Archief
voor Wiskunde,Vierde serie, deel 14, 237–253
[7] DeGroat,R.D., Dowling,E.M. 1993 The data least squares problem and
channel equalization. IEEE Trans. Signal Processing, Vol. 41(1), 407–411.
[8] Golub,G.H., Van Loan,C.F. 1980. An analysis of the total least squares
problem. SIAM J. on Numer. Anal., 17, 883–893.
[9] Jo,S., Kim,S.W. 2005. Consistent normalized least mean square filtering
with noisy data matrix. IEEE Trans. Signal Proc.,Vol. 53(6), 2112–2123.
[10] Kryszig,V. 1983. Advanced engineering mathematics, John Wiley & Sons.
[11] Lee,S.L. 1994. A note on the total least square fit to coplanar points,
Tech.Rep. ORNL-TM-12852, Oak Ridge National Laboratory.
[12] Levy,D., 2010. Introduction to Numerical Analysis, Univ.of Maryland
[13] Nievergelt,Y. 1994. Total least squares: State of the Art regression in
numerical mathematics, SIAM Review, Vol.36, 258-264
[14] Nixon,M.S., Aguado,A.S. 2012. Feature extraction & image processing for
computer vision, Academic Press.
[15] Markowsky,I., VanHueffel,S., 2007. Overview of total least square
methods, Signal Processing, 87 (10), 2283-2302.
[16] Pearson,K., 1901. On line and planes of closest fit to system of points in
space, Phil.Mag., Vol.2, 559-572
[17] Skala,V., 2016.Total Least Square Error Computation in E2: A New
Simple, Fast and Robust Algorithm, CGI 2016 Proc., ACM, pp.1-4, Greece
[18] Skala,V., 2016. A new formulation for total Least Square Error method in
d-dimensional space with mapping to a parametric line, ICNAAM 2015,
AIP Conf. Proc.1738, pp.480106-1 - 480106-4, Greece
[19] Skala,V., 2017. Least Square Error method approximation and extended
cross product using projective representation, ICNAAM 2016 conf., to
appear in ICNAAM 2016 proceedings, AIP Press
[20] Skala,V., 2008. Barycentric Coordinates Computation in Homogeneous
Coordinates, Computers & Graphics, Elsevier, , Vol.32, No.1, pp.120-127
[21] Van Huffel,S., Vandewalle,J 1991. The total least squares problems:
computational aspects and analysis. SIAM Publications, Philadelphia PA.
[22] van Huffel,S., Lemmerling,P. 2002. Total least squares and errors-invariables modeling: Analysis, algorithms and applications. Dordrecht, The
Netherlands: Kluwer Academic Publishers
[23] Perpendicular regression of a line (download 2015-12-20)
http://www.mathpages.com/home/kmath110.htm
[24] Skala,V., 2017. RBF Interpolation with CSRBF of Large Data Sets, ICCS
2017, Procedia Computer Science, Vol.108, pp. 2433-2437, Elsevier
[25] Skala,V., 2017. High Dimensional and Large Span Data Least Square
Error: Numerical Stability and Conditionality, accepted to ICPAM 2017
| 1cs.CV
|
Distance Labelings on Random Power Law Graphs
Huacheng Yu1 and Hongyang Zhang2
1
arXiv:1712.08709v1 [cs.SI] 23 Dec 2017
2
Harvard University
Department of Computer Science, Stanford University, CA, USA
Abstract
A Distance Labeling scheme is a data structure that can answer shortest path queries on a
graph. Experiment results from several recent studies (Akiba et al.’13, Delling et al.’14) found
very efficient and very accurate labeling schemes, which scale to social and information networks
with tens of millions of vertices and edges. Such a finding is not expected in the worst case,
since even for graphs with maximum degree 3, it is known that any distance labeling requires
Ω(n3/2 ) space (Gavoille et al.’03). On the other hand, social and information networks have a
heavy-tailed degree distribution and small average distance, which are not captured in the worst
case.
In this paper, we fill in the gap between empirical and worst case results. We consider
distance labeling schemes on random graph models with a power law degree distribution. For
such graphs, we show that simple breadth-first-search based algorithm can find near optimal
labeling schemes. The intuition behind our proof reveals that the distances between different
pairs of vertices are almost independent, even for polynomially many pairs.
1
Overview
Distance Labeling refers to a family of data structures for answering distance queries [28]. Each
vertex is assigned a “labeling”; To answer a query between the distance of two vertices x and
y, one is only allowed to use the labels of x and y to compute dist(x, y). Labeling schemes are
designed to speed up distance queries on large graphs, where computing shortest path from scratch
is expensive. In this paper, we study distance labeling schemes on random graph models for social
and information networks.
A simple yet commonly used scheme is landmark based labelings (also known as 2-hop covers [19], hub labeling [1]). The idea is to find central landmarks that lie on the shortest paths
of many sources and destinations. Meanwhile, every vertex stores a set of landmarks as well as
its distance to each landmark. To answer a distance query dist(x, y), we simply find a common
landmark z in the landmark sets of x and y to minimize the sum of distances dist(x, z) + dist(z, y).
By cleverly finding local and global landmarks, Akiba et al. [4] and Delling et al. [22] found
that only a few hundred landmarks per vertex suffices to recover all-pairs distances exactly, in a
collection of social, Web, and computer networks with tens of millions of edges. Such a finding
does not hold in worst case, since no distance labellings can always recover the exact distance while
using sub-quadratic amount of space. Even for graphs with maximum degree 3, it is known that
any distance labeling scheme requires Ω(n1.5 ) space [28].
Existing models for social and information networks build on random graphs with a fixed degree
distribution [24, 15, 42]. Informally, we assume that the degree sequence of our graph is given, and
then we draw a “uniform” sample from graphs that have the same or very similar degree sequences.
Random graphs capture the small world phenomenon [15], because the average distance grows
logarithmically in the number of vertices. They serve as a basic block to richer models with more
realistic features, e.g. community structures [31], shrinking diameters in temporal graphs [34]. In
this work, we fill in the gap between empirical and worst case results, by studying distance labelings
on random graphs:
Given a random graph from the Chung-Lu model with a power law degree distribution of
exponent β, how much storage does a distance labeling scheme require overall, in order
to answer distance queries with no distortion?
In the Chung-Lu model [14], each vertex x has a weight (expected degree) px . For every pair
of vertices x and y, there is an undirected and unweighted edge between them with probability
proportional to px · py , independent of other edges. Hence, Chung-Lu model generalizes ErdősRenyi graph to arbitrary degree distributions. Later on, we will discuss the implication of our
results for configuration model and directed Chung-Lu model as well. In the rest of the paper, we
use the term “random power law graph” to refer to a graph that is sampled from the Chung-Lu
model, where the weight of each vertex is independently drawn from a power law distribution with
mean value ν > 1 and exponent β. We are interested in the regime when β > 2 — this covers most
of the empirical power law degree distributions that people have observed on social and information
networks [16].
1.1
Results
When the degree distribution has finite variance (β > 3), we show that breadth first search produces
√
a 2-hop cover which only requires each vertex to store Õ( n) landmarks. As a complement, the
total length of any distance labeling schemes that answer distance queries exactly is almost surely
Ω(n1.5 ). The same conclusion also applies to Erdős-Renyi graphs G(n, nc ) when c > 1, or when
c = (1 + ε) log n.
Theorem 1. Let G n (p) be a random power law graph model with average degree ν > 1 and exponent
β > 3. For a random graph G = (V, E) drawn from G n (p), we have that:
p
• Almost surely there exists a 2-hop cover F such that |F (x)| ≤ O( n log3 n) for all x ∈ V .
• Almost surely any distance labeling scheme will output a labeling whose total length is Ω̃(n3/2 ).
We then present an algorithm such that on a random power law graph when 2 < β < 3
(infinite variance of the degree distribution), it generates at most Õ(n(β−2)/(β−1) ) landmarks per
vertex when β ≥ 2.5; and Õ(n(3−β)/(4−β) ) landmarks per vertex when 2 < β < 2.5 (See Figure 1 for
an illustration). We also show that when 2 < β < 3, any distance labeling scheme will generate
5−β
labels of total size n 2 −o(1) almost surely.
Theorem 2. Let G n (p) be a random power law graph model with average degree ν > 1 and exponent
2 < β ≤ 3. For a random graph G = (V, E) drawn from G n (p), we have that:
• Almost surely there exists a 2-hop cover F such that |F (x)| ∼ O(n
all x ∈ V .
1
1
, 4−β
)
1−min( β−1
· log3 n) for
• For any distance labeling scheme, almost surely it will output a labeling whose total length is
5−β
n 2 −o(1) .
Our algorithm starts from the following observation: do a breadth-first search from each vertex
√
x until either the entire connected component has been explored or n vertices has been traversed
— let F (x) denote the set of vertices we discovered. For any two vertices x and y in the same
component, F (x) and F (y) have a nonempty intersection, thus ensuring that there is a common
landmark on the shortest path between (x, y). This simple procedure generates a 2-hop cover where
√
each vertex stores Õ( n) landmarks.
√
When 2 < β < 3, instead of stopping when n vertices are explored, we stop before reaching the
layer containing the maximum weighted vertex. The vertices that we already discovered together
with the maximum weighted vertex is essentially a (+1)-stretch scheme with at most Õ(n(β−2)/(β−1) )
landmarks per vertex. However, the size of the boundary, which contains the maximum weighted
√
vertex, can be much larger than n. This is because the branching process grows doubly exponentially near the boundary [14]. Therefore, we preprocess a set of high degree vertices first, then
carefully add vertices on the boundary to resolve the (+1)-stretch. It is also not hard to obtain
1.5
Storage (ny )
1.4
1.3
1.2
1.1
1.0
2.00
2.25
2.50
β
2.75
3.00
Figure 1: An illustration of the results: The x-axis is the exponent of the power law degree
distribution and for each value in the y-axis it means that the amount of storage is Õ(ny ).
a (+2)-stretch scheme with Õ(nβ/2−1 ) landmarks per vertex — when β < 2.5, this improves the
3-approximate distance oracle of Chen et al. [13], which requires O(n(β−2)/(2β−3) space per vertex.
Now we describe the intuition behind our lower bound for the β > 3 case. Consider two
randomly chosen vertices x, y from V . We alread know that dist(x, y) is close to the average
distance (on the order of O(log n)). Let d be slightly smaller than the average distance. While
dist(x, y) will be at least 2d + 1 with high probability, it is crucial that Pr[dist(x, y) = 2d + 1] is
already non-negligible (e.g. Ω(1/ log n) if we set d appropriatly). Hence the information of whether
dist(x, y) is equal to 2d + 1 or not, is worth Ω(1/ log n) bits. If we can construct n1.5 pairs of such
√
(x, y), then we could obtain the desired lower bound. Towards this goal, we note that even for n
vertices, we could still maintain the neighborhood growth “almost” independently, up to distance
d. In Section 3, we implement this idea via a martingale argument and an entropy argument, which
yields a lower bound of Ω̃(n1.5 ).
To prove the lower bound when 2 < β < 3, we adopt the high level plan described above.
However, the neighborhood growth has very high variance, because there are lots of high degree
vertices. To overcome the barrier, we use a technique from Van Der Hofstad (Chapter 3 [42]). We
carefully construct a set of “good” path, so that with high probability, a vertex will follow our good
path during the neighborhood growth. Our lower bound is nearly tight when β is close to 2, and
has a small gap when β < 2.5. It would be interesting to close the gap when 2.5 < β < 3. We
believe that the right answer should be Ω(n1.5 ) when β is close to 3.
In Section 6, we test our algorithm on real world graphs. We found that our algorithm achieves
fairly accurate results — the 80%-percentitle multiplicative error is less than 0.25 in our experiment.
In addition, the algorithm is scalable to preprocess graphs with millions of edges in several minutes.
One limitation is that our algorithm is designed for graphs with small average distance and
a heavy tailed distribution. The second limitation is that we only derived results for Chung-Lu
model. However, our technical proofs only rely on upper and lower bounding the growth rate of
neighborhood growth. And it is well-known that configuration model have the same neighborhood
growth rate [42], hence we expect our high level intuition to hold for the configuration model [24].
Finally, our lower bound technique only applies to labeling schemes. We refer the interested reader
to Section 7 for more discussion about future work.
1.2
Related work
Landmark based Labelings The problem of computing the optimal landmark based labelings
can be formulated as an integer program, similar to the set cover problem [19]. It is NP-hard
to compute the optimal landmark labelings (or 2-hop cover), and a log n-approximation can be
obtained via a greedy algorithm [19]. See also the references [29, 23, 8, 7] for a line of followup
work. Another closely related line of work is approximate distance oracle. We refer the reader to
the excellent survey of Sommer [39] on this topic. as well as the classic work [6, 17, 41], and recent
developments [20, 36, 3, 37, 43, 38, 5, 12] for further reading. On the practical side, numerous studies
have demenstated the effectiveness of landmark based labelings on large graphs [21, 18, 4, 22, 11].
Notably, Bahmani and Goel [9] implemented the distance sketch of Das Sarma et al. [21] in a
distributed setting.
Power Law Graphs It has been empirically observed that many social and information networks
have a heavy-tailed degree distribution [16] — concretely, the number of vertices whose degree is
x, is proportional to x−β . For real world graphs, the coefficient β vary from 2 to 10, and are often
greater than 3 [25].
Previous work of Chen et al. [13] proved that the space complexity of the classic Thorup and
Zwick distance oracle [41] can be reduced on random graphs with a power law degree distribution,
while achieving stretch 3, when 2 < β < 3. They left open the question of finding distance oracles
with better stretch and less space. Gavoille et al. [27] reported a stretch 5 compact routing scheme
(distance oracles for distributed settings) that also requires much less space on random power law
graphs than in worst case. Enachescu et al. [26] presented a compact routing scheme that achieves
stretch 2 using space O(n1.75 ) on Erdős-Renyi graphs. Existing mathematical models on special
families of graphs related to distance queries include road networks [2], planar graphs [35] and
graphs with doubling dimension [30]. However none of them can capture the expansion properties
that have been observed on sub-networks of real-world social networks [34]. Apart from the ChungLu model and the configuration model that we have mentioned, the preferential attachment graph
is also well-understood [24]. It would be interesting to see if our results extend to preferential
attachment graphs as well. The Kronecker model [32] allows a richer set of features by extending
previous random graph models, however its mathematical properties are not as well-understood as
the other three models.
Organization: The rest of the paper is organized as follows. In Section 2 we define the random
graph model. In Section 3 we present our technical results along with the proof sketches. In Section
4 and 5 we fill in the missing proofs from Section 3. Section 6 describes our experiments. Finally,
we discuss future extensions of our results in Section 7.
2
Preliminaries
In this section, we introduce notations and define graph models. Consider an undirected graph
G = (V, E). Let n = |V | be the number of vertices.
For any vertex x ∈ V , Let dx denote the
P
degree of x. For a set of vertices S, let dS = x∈S dx denote the sum of their degrees. We use
the notation x ∼ y to indicate that (x, y) ∈ E. For two disjoint sets S and T , S ∼ T means there
exists an edge between S and T ; S ≁ T means there does not exist any edge between S and T . Let
distG (x, y) denote the distance of x and y in G (we drop the subscript G if there is no ambiguity).
For any integer 1 ≤ k ≤ n − 1, let Γk (x) = {y ∈ V : dist(x, y) = k} denote the set of vertices whose
distance from x is equal to i. And let Ni (x) = {y ∈ V : dist(x, y) ≤ i} denote the set of vertices
whose distance from x is at most i.
2.1
Distance Labelings
A distance labeling scheme consists of two algorithms [28, 39]:
• Preprocessing: given a graph G = (V, E), output a vertex-labeling L : V → {0, 1}∗ ;
• Query: given input L(x) and L(y), compute dist(x, y) without accessing any other information.
P
For a set of vertices S, the total label size of S is defined as x∈S |L(x)|. The total label size of L
is given by |L(V )|. The maximum size of L over all vertices is given by maxx∈V |L(x)|.
In a landmark based labeling scheme, the labeling of each vertex x consists of a subset of vertices
and their distances to x. A function F : V → 2V is said to be a 2-hop cover [19], if for every pair of
vertices x and y in the same connected component of G, there exists a vertex z ∈ F (x) ∩ F (y) such
that dist(x, y) = dist(x, z) + dist(z, y). When x and y are not reachable from each other, then the
intersection of F (x) and F (y) should be empty. Apart from F , a 2-hop cover also stores dist(x, y)
for every y ∈ F (x).
The query algorithm for dist(x, y) is given by
min
z∈F (x)∩F (y)
dist(x, z) + dist(z, y).
If no common landmark in F (x) and F (y) is found, we return disconnected. Each query takes no
more than O(|F (x)| + |F (y)|) time.
2.2
Chung-Lu Model
In Chung-Lu model, each vertex x ∈ V has a weight px > 0, which corresponds to x’s expected
degree. Given weight vector p over V P
, the Chung-Lu model defines a probability distribution
P over
the set of all graphs G n . Let vol(S) = x∈S px denote the volume of S. And let vol2 (S) := x∈S p2x
denote the second moment of S. Each edge (x, y) is chosen independently with probability
px · py
,1 .
Pr[x ∼ y] = min
vol(V )
Thus, px is approximately the expected degree of x, and vol(V ) is approximately the expected
number of edges. Let G n (p) denote such a probability distribution over G n . We use the notation
G ∈ G n (p) to refer to a graph drawn from the distribution defined by G n (p). We have the following
convenient Proposition for bounding the probability of whether two sets connect or not.
Proposition 3. Let G = (V, E) ∈ G n (p) be a random graph. For any two disjoint set of vertices
S and T ,
vol(S)vol(T )
, and
vol(V )
vol(S)vol(T )
Pr[S ≁ T ] ≤ exp −
.
vol(V )
Pr[S ∼ T ] ≤
Proof. For the first bound, we have:
Pr[S ∼ T ] = 1 −
YY
x∈S y∈T
≤ 1 − (1 −
≤ 1 − (1 −
(1 − min(
XX
px py
, 1))
vol(V )
min(
x∈S y∈T
px py
, 1))
vol(V )
X X px py
p(S)p(T )
)=
vol(V )
vol(V )
x∈S y∈T
and
Pr[S ≁ T ] =
YY
x∈S y∈T
(1 − min(
≤ exp −
≤ exp(−
XX
x∈S y∈T
px py
, 1))
vol(V )
px py
min(
, 1)
vol(V )
p(S)p(T )
).
vol(V )
2.3
Power Law Distribution
Let f : [xmin , ∞) → R denote the probability density function of a power law [16] distribution with
β−1
. The expectation of f (·) exists when
exponent β > 1, i.e. f (x) = Zx−β , where Z = (β − 1) · xmin
β > 2:
β−1
.
ν = xmin ·
β−2
The second moment is finite, only when β > 3, which is equal to
ω = x2min ·
β−1
.
β−3
In the random power law graph model, the weight of each vertex x is drawn independently from
a power law distribution (with the same mean ν and exponent β). Given the weight vector p, we
then sample a random graph according to the Chung-Lu model G n (p).
It is shown in Chung and Lu [15] that if ν > 1, then almost surely a graph G ∈ G n (p) has a
unique giant component. If ν < 1, almost surely all connected components have at most O(log n)
vertices. In this paper, we are interested in cases when the average degree ν is a constant greater
than 1. In this case, it is also known that the diameter of G is O(log n) with high probability.
3
Proof Overview
In this section, we state our results formally along with a sketch for the proof. First, we consider
the case when the degree distribution has finite variance.
Theorem 1 (restate). Let G n (p) be a random power law graph model with average degree ν > 1
and exponent β > 3. For a random graph G = (V, E) drawn from G n (p), we have that:
p
• Almost surely there exists a 2-hop cover F such that |F (x)| ≤ O( n log3 n) for all x ∈ V .
• Almost surely any distance labeling scheme will output a labeling whose total length is Ω̃(n3/2 ).
√
For the first part, we simply observe that for each vertex, if we add the closest n vertices to
the landmark set of every vertex, then the landmark sets of every pair of vertices will intersect with
high probability, i.e. we have obtained a 2-hop cover. The proof can be found in Appendix A.
For the lower bound, we divide the set of vertices whose weight is at most 2ν to groups of size
√
n. The following Proposition shows that for each group, with high probability the total label
√
√
size of Ω̃( n). This reduction will simplify the proof because the joint neighborhood growth of n
vertices is almost independent in early stages.
√
Proposition 4. Let S be a set of n vertices, where
every vertex has weight no more than 2ν. The
−O(1+1/γ)
total label size of S is at least Ω n · log
n with high probability, where γ = (β − 3)/2.
Let us first show that the above proposition directly implies the lower bound in Theorem 1.
Proof of Theorem 1. We know that at least n/2 vertices have weight at most 2ν, by an averaging
√
argument. Divide them into groups of size n, and apply Proposition 4 on each group. Clearly,
√
√
except for o( n) groups, most groups
there are at least Θ( n) disjoint
groups. In expectation,
√
will have label size at least Ω n · log−O(1+1/γ) n . Hence Markov’s inequality, at least Ω( n) of
the groups have total labeling size at least Ω n · log−O(1+1/γ) n with probability 1 − o(1), thus
they have a total labeling size of Ω̃ n1.5 . This proves the theorem.
To build up intuition towards the proof of Proposition 4, we consider Erdős-Renyi graph G =
G(n, p) where p = c/n, for a constant c > 0. Our lower bound is derived via an entropy argument.
We will carefully exploit the information given by the pairwise distances of S. Note that the average
n
distance of G is roughly 2log
log c (see e.g. [10]). We will consider distances slightly smaller than the
average distance. Let d =
log n
2 log c
− O(log log n). We observe the following facts.
• For every x ∈ S, with constant probability we have that |Γd (x)| = Θ(cd ).
2d
• For every pair of vertices x, y ∈ S, the probability that dist(x, y) ≤ 2d + 1 is at most O( cn ) =
1
).
O(
polylog(n)
• If |Γd (x)| and |Γd (y)| are both on the order of Θ(cd ) and dist(x, y) > 2d, the probability that
2d
1
dist(x, y) = 2d + 1 is Θ( cn ) = Θ(
).
polylog(n)
We know that the distance labeling of S determine the pairwise distances of S. In particular, they
determine whether each pair x, y has distance exactly 2d+ 1 or more, if dist(x, y) is not yet revealed
by Γd (x) and Γd (y). This is worth roughly 1/polylog(n) bit of information. In expectation, we
expect to have Θ(n) such pairs from S, which implies our lower bound on the labeling size of S.
To implement the above plan, we need to deal two more issues. First, we need to argue that
there are Θ(n) pairs with high probability. Secondly, for general degree distributions, there exists
high degree vertices which introduces high variance to neighborhood growth. To resolve the first
issue, we construct a martingale to grow the neighborthood of each vertex in S. To resolve the
second issue, we note the second moment of the degree distribution is finite. Duing the martingale
process, we carefully bound the second moment of degree sequence. We leave the full proof to
Section 4.
Now we consider the case when the degree distribution has infinite variance.
Theorem 2 (restate). Let G n (p) be a random power law graph model with average degree ν > 1
and exponent 2 < β ≤ 3. For a random graph G = (V, E) drawn from G n (p), we have that:
• Almost surely there exists a 2-hop cover F such that |F (x)| ∼ O(n
all x ∈ V .
1
1
, 4−β
)
1−min( β−1
· log3 n) for
• For any distance labeling scheme, almost surely it will output a labeling whose total length is
5−β
n 2 −o(1) .
Our upper bound uses the fact that G contains a heavy vertex whose weight is approximately
n
. We first add all high degreevertices to the landmark set of every vertex. Then we do
1
β−1
β−2
a breadth-first search, but stop right before the boundary size exceeding Θ̃(n β−1 ), and put all
vertices that we have explored in the landmark set. We claim that this gives up a (+1)-sketch
labeling.
To see this, for two vertices x, y, if their landmark sets intersect with each other, then we can
already compute their distances correctly. Otherwise, the bottom layer of x and y have distance
at most two (through the heavy vertex) with high probability. Therefore, we only have to check
whether they have distance one, i.e., there is an edge between them.
To resolve the (+1)-stretch, for each vertex on the boundary, we add all of its neighbors with a
higher degree to the landmark set. Clearly, this fixes the (+1)-stretch, if there is an edge connecting
Algorithm 1 AlgSkewDegree
Input: An undirected graph G = (V, E); Parameters δ and K.
1: Let H = {x ∈ V : dx ≥ K}
2: for x ∈ V do
β−2
β−2
3:
(F (x), l(x)) = AlgBfs(x, δn β−1 )
(add closest δn β−1 vertices)
4:
F (x) = F (x) ∪ H
(add all high degree vertices)
5:
for y ∈ Γl(x)−1 (x) do
6:
if dy ≤ K then
7:
F (x) = F (x) ∪ {z ∈ N (y) : dy ≤ dz ≤ K}
(add all neighbors of y with a higher degree)
8:
end if
9:
end for
10: end for
the two boundaries. To bound the landmark size, we refer the reader to Appendix B. A complete
description of the algorithm is shown below.
For each vertex x and 0 ≤ k ≤ n−1, αk (x) is the number of edges between Γk (x) and V \Nk−1 (x).
l(x) is the first non-negative integer that satisfies: αl(x)−1 (x) > δn(β−2)/(β−1) or Γl(x) (x) = ∅. It is
clear that l(x) always exists. The procedure AlgBfs(x, T ) expands from x until reaching a level
set whose volume is at least T ; And the output will be all the vertices visited so far — the precise
definition can be found in Appendix A. Set δ = 4ν log2 n and
(√
n
if 2.5 ≤ β ≤ 3
K=
1
n (4−β)(β−1) if 2 < β < 2.5.
Remark One can also obtain a (+2)-stretch labeling by setting K = Õ(nβ/2−1 ); The maximum
labeling size will be K. We omit the proof.
3−β
For the lower bound, we introduce a reduction to groups of size n 2 .
3−β
Proposition 5. Let S be a set of n 2 vertices with weights between [a, b] such that a, b =
nΘ(1/ log log n) and b > 2a. Then any labeling scheme must generate a total label size of at least
n3−β−O(1/ log log n) for S with high probability.
We first show how to reduce the lower bound part of Theorem 2 to the above proposition.
Proof of Theorem 2. Consider the set of vertices with weights between a and b, and divide them
3−β
β−1
into groups of size n 2 . The number of groups is n 2 −o(1) with high probability, because
there are n1−o(1) such vertices. For each group, by Proposition 5, the total label size is at least
n3−β−O(1/ log log n) except with o(1) probability. By Markov’s inequality, with probability 1 − o(1),
β−1
at least a constant fraction (i.e. n 2 −o(1) ) of the groups have label sizes at least n3−β−O(1/ log log n) .
5−β
5−β
Hence, we conclude that the total label size is at least n 2 −O(1/ log log n) = n 2 −o(1) with 1 − o(1)
probability.
To prove the proposition, we perform a breadth first search from S to distance d such that each
vertex is “expected” to have a neighborhood of volume roughly nβ/2−1 at distance d. We show:
3−β
1. There exists a subset of vertices S ′ ⊆ S, such that: i) |S ′ | = Θ(n 2 ); ii) for each vertex
β
x ∈ S ′ , vol(Γd (x)) ≈ Θ(n 2 −1 ); iii) for any two vertices x, y ∈ S ′ , their “neighborhoods” are
disjoint.
2. For a constant fraction of vertices x in S ′ , Γd (x) is connected to a vertex h(x) whose weight
√
is about O( n), and their pairwise distance is at least 2d + 4.
3. Distance between x and y being at least 2d + 4 implies that h(x) and h(y) are not connected,
which happens with roughly constant probability. However, there are n3−β such pairs, and
thus every label of S which implies such a distance function could only occur with exp(−n3−β )
probability. The total label size of S is at least n3−β with high probability.
The detailed proof can be found in Section 5.
4
Proof of Proposition 4
√
2 (V )
Let r = vol
vol(V ) and d = logr n − c, where c ≤ O(log log n) will be determined later. We describe
an iterative process to grow the neighborhood of S up to distance d. Let S = {x1 , x2 , . . .} be any
ordering of its vertices. Denote by G1 = (V1 , E1 ), where V1 = V and E1 = E. For any i ≥ 1, define
T (xi ) to be the set of of vertices in Gi whose distance is at most d from xi . Define L(xi ) to be the
set of vertices in Gi whose distance is equal to d from xi . More formally,
(
{y : dGi (xi , y) ≤ d}, if xi ∈ Vi
T (xi ) :=
∅,
otherwise.
L(xi ) := {y ∈ T (xi ) : dGi (xi , y) = d}
We then define Fi = Fi−1 ∪ T (xi ) (F0 := ∅ by default). Denote by Gi+1 to be the induced subgraph
of Gi on the remaining vertices Vi+1 = V \Fi .
We note that in the above iterative process, the neighborhood growth of xi only depends on
the degree sequence of Vi . In order to bound rate of growth, we introduce the following Lemma.
The result is standard (see e.g. [15]) – we leave the proof to Appendix C.1.
Lemma 6. Let G n (p) be a random graph model with weight sequence p satisfying the following
properties:
1. vol(V ) = (1 + o(1))ν · n for some constant ν;
2. vol2 (V ) = (1 + o(1))ω · n for some constant ω;
3. vol2+γ (V ) = τ · n for some positive constant γ < 1/2 and τ , where vol2+γ (S) :=
4. The growth rate r =
vol2 (V )
vol(V )
is bounded away from 1 (ν > ω).
P
x∈S
p2+γ
x ;
Then for any vertex x with a constant weight, the set of vertices Γk (x) at distance exactly k from
x has:
1. E[vol(Γk (x))] = O r k for every k ≤ logr n;
2. Pr[vol(Γk (x)) ≥ Ω r k ] ≥ Ω(1) for every k ≤ 21 logr n.
As a corollary, we have that Pr[dist(x, y) ≤ k + 1] ≤ O(r k /n) for every k ≤ 21 logr n, where y is any
vertex with constant weight.
√
We now claim that with high probability, at least c1 n vertices x ∈ S have that vol(L(x)) ≥
√
Ω( n · r −c ), for some absolute constant c1 and c = (3 + 1/γ) log r log n; The proof consists of two
parts:
√
2+ 1
a) For every 1 ≤ i ≤ n, conditional on xi ∈ Vi and vol(Fi−1 ) ≤ n/ log γ n, we have that
√
vol(L(xi )) ≥ Ω( n · r −c ) with constant probability.
√
b) By constructing a martingale and then apply Azuma-Hoeffding inequality, at least c1 n
√
vertices x ∈ S have that vol(L(x)) ≥ Ω( n · r −c ).
To prove Claim a), we note that the subgraph Gi on Vi is also a random graph sampled from
Chung-Lu model. The vertices at distance d from xi in this subgraph is exactly Li . As the total
weight of the subgraph is smaller than the vol(V ), the normalizer for the probability of an edge
changes if restricted to Vi . To adjust for this change, we set
vol(Fi−1 )
(i)
, ∀ y ∈ Vi .
py = py · 1 −
vol(V )
Then we have
(i)
(i)
py · pz
py · pz
.
Pr[y ∼ z] = P
=
(i)
vol(V
)
p
x
x∈Vi
Hence we see that Gi is equivalent to a random graph drawn from degree sequence pi . The growth
2+1/γ
2 (Vi )
n, by Hölder’s inequality,
rate is ri := vol
vol(V ) . When vol(Fi−1 ) ≤ n/ log
γ
1
vol2 (Fi−1 ) ≤ vol(Fi−1 ) 1+γ · O(n 1+γ ) ≤ o(n/ log n).
Thus, the growth rate ri > 1. By our growth Lemma 6, with constant probability,
√
vol(V )
≥ Ω(r d ) = Ω( n · r −c ).
vol(Li ) ≥ Ω rid ·
vol(Vi )
Hence we proved Claim a).
To prove Claim b), we consider
1
Xi :=
0
the following random variable, for any 1 ≤ i ≤
√
n.
if xi ∈
/ Vi , or vol(Fi−1 ) > n/ log2+1/γ ,
√
or vol(Li ) ≥ Ω( n · r −c )
otherwise.
We have Pr[Xi = 1 | X1 , . . . , Xi−1 ] ≥ Ω(1) by Claim a). Thus by Azuma-Hoeffding inequality,
P|S|
√
i=1 Xi ≥ Ω( n) with probability 1 − o(1). We shall prove below that the contributions to
P|S|
√
i=1 Xi from the first two predicates is o( n). Hence by taking union bound, we obtain Claim b).
√
• First, the number of xi such that xi ∈
/ Vi is o( n) with high probability. Note that xi ∈
/ Vi
implies that there exists some vertex xj ∈ Fi−1 such that dist(xi , xj ) ≤ d. On the other hand,
for any two vertices x, y ∈ S, Pr[dist(x, y) ≤ d] ≤ O(r d /n), by Lemma 6. Hence, the expected
√
number of vertex pairs in S whose distance is at most d, is O(r d /n) ≤ o(1/ n). By Markov’s
√
inequality, with probability 1 − o(1) , only o( n) vertex pairs have distance at most d in S.
√
Hence for at most o( n) i’s we have that xi ∈
/ Vi .
1
√
• Secondly for all 1 ≤ i ≤ n, vol(Fi ) ≤ n/ log2+ γ with high probability. This is because
the set of vertices Fi is a subset of Nd (xi ), the vertices within distance d to xi . Thus, by
Lemma 6, we have
d
E[vol(Ti )] ≤ E[vol(Nd (xi ))] ≤ O(r ).
Thus, the expected volume of Fi is at most
√
O(i · r d ) = O( n · r d ) = O(log−1 n) × n/ log2+1/γ ,
√
because d = logr n − c and c ≥ (3 + 1/γ) log r log n. Hence by Markov’s inequality, the
probability that vol(F|S| ) > n/ log2+1/γ n is at most O(log−1 n).
Now we are ready to prove Proposition 4. Given the labelings of S, we can recover the pairwise
distances for all vertex pairs in S. Let distS : S × S → N denote the distance function restricted
to all pairs in S. We show that with high probability, the total labeling size of S is at least
Ω(n · r −2c ) = Ω(n · log−O(1+1/γ) n) (recall that c = (3 + 1/γ) log r log n). Consider the following two
cases:
1. ∃c21 · n/4 pairs (xi , xj ) such that distS (xi , xj ) ≤ 2d + 1. By Lemma 6, we know that
Pr[dist(xi , xj ) ≤ 2d + 1] = O(r 2d /n), for any xi , xj ∈ S. Hence the expected number of
pairs with distance at most 2d + 1 in S, is at most O(r 2d /n). By Markov’s inequality, the
probability that a random graph induces any such distance function is O(r −2c ) = o(1).
2. The number of pairs such that distS (xi , xj ) ≤ 2d + 1 is at most c21 · n/4 in S. Let A =
√
{(x, y) ∈ S × S | dist(x, y) > 2d + 1, and vol(L(x)), vol(L(y)) > Ω( n · r −c )}. By Claim b),
the size of A is at least c1 n(c1 n − 1)/2 − c21 n/4 ≥ c21 n/5. For any (x, y) ∈ A, L(x) and L(y)
are clearly disjoint. Conditional on {T (x)} for all x ∈ S, the probability of the existences of
edges between Li and Lj are unaffected.
i
h
|S|
Pr distS (x, y) > 2d + 1, ∀(x, y) ∈ A | {Ti }i=1
Y
≤
Pr [L(x) 6∼ L(y) | L(x) ∩ L(y) = ∅,
(x,y)∈A
√
and vol(L(x)), vol(L(y)) ≥ Ω( n · r −c )
Y
vol(L(x))vol(L(y))
≤
exp −
vol(V )
(x,y)∈A
c2 n/5
≤ exp −Ω(r −2c ) 1
≤ exp(−Ω(n · r −2c )).
Now let c2 be a sufficiently small value (e.g. 1/ log log n suffices). The number of labelings of
−2c
size less than c2 n · r −2c is at most 2c2 n·r . The probability that the total label size of |S| is
−2c
at most c2 n · r −2c , is at most 2c2 n·r × exp(−Ω(n · r −2c )) = o(1) by union bound.
By taking a union bound on the two cases, we prove that with probability 1 − o(1), the total
label size of S has to be at least Ω(n · r −2c ) = Ω(n · log−O(1+1/γ) n).
5
Proof of Proposition 5
We divide the proof into three parts. Let d =
log log log n
.
1
log β−2
In part 1, we specify the set of good
path. We argue that the growth of S to the d-th level follows our good path with high probability.
√
In part 2, we connect to vertices with weight n in the (d + 1)-th level. In part 3, we use the
entropy argument to show that with high probability, the label size of S is at least n3−β−O(ε) , where
ε = 1/ log log n.
Part I: neighborhood growth For all 0 ≤ i ≤ d + 1, let
d−i
µi := n(β/2−1−ε)(β−2)
σi := w
1/(β−2)2i (3−β)
, and
,
where w = log log n. Denote by ai = µi /σi and bi = µi σi . µi can be thought of as the “expected”
volume at distance i from a vertex x ∈ S. If the volume at distance i from x always stays inside
[ai , bi ], then we think of x as a “good” vertex.
It is easy to verify that both a0 and b0 are nΘ(1/ log log n) and b0 > 2a0 . Let S = {x1 , x2 , . . . } be
3−β
an arbitrary vertex set of size n 2 such that all xi have weights between a0 and b0 . Clearly, for
any i 6= j, the neighborhood growth of i and j are correlated. However, one would expect that the
√
correlation is small, so long as the volume of the neighborhood has not reached O( n). We leverage
the observation by exploring the neighborhood of S one vertex at a time. Denote by G1 = (V1 , E1 )
where V1 = V and E1 = E. For 1 ≤ i ≤ |S|, we consider the following inductive process:
1. If xi ∈ Vi , let 1 ≤ λi ≤ d be the maximum k that still satisfy vol(Γk (xi )) ∈ [ak , bk ] in graph
Gi (Recall that Γk (xi ) is the set of vertices at distance exactly k from xi in Gi );
2. Denote by Ti the set of vertices within distance min{d, λi + 1} from xi in Gi ;
3. If λi = d, let L(xi ) = Γd (xi ); otherwise, let L(xi ) = ∅;
We then define Fi = Fi−1 ∪ Ti (F0 = ∅ by default). Let Gi+1 be the subgraph of Gi on remaining
vertices Vi+1 = Vi \Ti . In the above process, we keep expanding the neighborhood of xi until we
reach distance d or we find a distance λi + 1 such that the volume of vertices at distance λi + 1
from xi is not in [aλi +1 , bλi +1 ]. If we reach distance d, then L(xi ) is the set of vertices at distance
d from xi .
In order to bound the growth rate of xi on graph Gi . We introduce the following Lemma.
This result follows standard arguments (see e.g. [42]) – a proof can be found in Appendix C.2 for
completeness.
Lemma 7. Let c1 , c2 , c3 > 0 be absolute constants. Let S be a set of fixed values within [1, nβ/2−1 ]
whose size is at most 2d. Let G n (p) be a random graph with weight sequence p satisfying vol(V ) =
Θ(n), and for any t ∈ S,
X
py ≥ c1 × nt2−β
y:py ≥t
X
y:py ≤t
X
y:py ≥t
p2y ≤ c2 × nt3−β ,
py ≤ c3 × nt2−β .
Then we have the following facts regarding neighborhood growth:
a) Following a good path: Let x be a fixed vertex and 1 ≤ k ≤ d+1. Suppose that vol(Γi (x)) ∈
[ai , bi ] for any 1 ≤ i < k, then vol(Γk (x)) ∈ [ak , bk ], with probability at least 1 − O(1/wβ−2 ),
where w = log log n;
b) Average distance: let x, y be two vertices such that px , py ∈ [a0 , b0 ], then Pr[dist(x, y) ≤
2d + 3] = o(1).
We make the following crucial claim.
Claim 8. With probability 1 − o(1), at least Θ(n
3−β
2
) vertices xi in S have vol(L(xi )) ∈ [ad , bd ].
To prove the claim, let us consider the following random variables. Define
(
1 vol(L(xi )) ∈ [ad , bd ],
Xi =
0 otherwise.
We show that Xi = 1 with high probability for all 1 ≤ i ≤ |S|. We first verify that Pr[xi ∈
/ Vi ] =
Pr[xi ∈ Fi−1 ] ≤ o(1). Consider any vertex z ∈ V and 1 ≤ j ≤ i − 1, we have that
d−1
X
p z · bl
Pr[z ∈ Tj ] ≤
vol(Vj )
l=0
≤ pz · n(β/2−1−ε)(β−2)−1 · wO(log
2
log n)
.
Thus, by union bound from 1 ≤ j ≤ i − 1, we have
1
Pr[z ∈ Fi−1 ] ≤ pz · n 2 (β
2 −5β+5)−ε(β−2)+o(ε)
= pz · n λ ,
(1)
i.e. denote the exponent by λ above.
Next, we verify that the weight sequence of Gi satisfies the premises of Lemma 7 with high
probability. It suffices to verify the first premise – the second and the third hold because Vi is a
subset of V . It’s not hard to see that the initial weight sequence of G1 = G satisfies all the premises
by Chernoff bound (details omitted). It suffices to show that Fi−1 has small volume, i.e., we only
remove a small volume from G in total. By Equation (1), we have:
X
E[vol(Fi−1 )] ≤
pz · min{1, pz · nλ }
z∈V
=
X
z:pz ≥n−λ
≤ O(n
pz +
1−λ(2−β)
X
z:pz <n−λ
p2z · nλ
+ n1−λ(3−β)+λ )
= O(n1+λ(β−2) ).
The second inequality above is because of Lemma 7. It is not hard to verify that
1 + λ(β − 2)
1
≤ 1 − (β − 2)2 − Θ(ε).
2
Having bounded the expected volume of Fi−1 , we obtain that with high probability only a total
2
volume of o(n1−(β−2) /2 ) is from V in Vi . Thus, we obtain the first premise of Lemma 7, because
2
nt2−β = Ω(n1−(β−2) /2 ) for any t ≤ nβ/2−1 .
Now we can apply Lemma 7 to Gi to obtain that Pr[Xi = 0] = o(1). Finally, by Markov’s
P
3−β
3−β
inequality, i (1 − Xi ) ≥ 12 · n 2 with o(1) probability. Hence at least 0.99n 2 vertices in S have
vol(L(xi )) ∈ [ad , bd ] with high probability. Denote by S1 ⊆ S the set of such vertices.
Part II: connecting to heavy vertices In this part, we show how L(xi ) and certain set of high
degree vertices are connected. Let A = {x ∈ V : px ∈ [(ad /w)1/(β−2) , 2(ad /w)1/(β−2) ]} \ F|S| . We
claim that there is a constant fraction of the vertices x in S1 such that each L(x) is connected to
a different vertex in A. More formally, we make the following claim.
Claim 9. With 1 − o(1) probability, there exists a set S2 ⊆ S1 and a function h : S2 → A such that
3−β
|S2 | ≥ 13 n 2 , for every x ∈ S2 , h(x) connects to some vertex in L(x) and h is an injection.
We first show that A has a large volume. By Equation (1), any vertex with weight at most
1
2(ad /w)1/(β−2) < n1/2 belongs to F|S| with probability at most n 2 (β−2)(β−3) = o(1). Thus, the
volume of A is
vol(A) = Θ(nw/ad )
with 1 − o(1) probability. This implies that |A| ≥ Θ(nw/(ad (ad /w)1/(β−2) )) > ω(n(3−β)/2 ).
Now we construct the set S2 and the function h as follows:
1. Go over all vertices x in S1 , if L(x) has a neighbor y in A that is “unused”, add x to S2 and
set h(x) to y;
2. Mark y as “used”.
This procedure will generate a set S2 of size at least 13 n
3−β
2
with high probability, since
1. Only an o(1) fraction of the vertices in A are marked as “used”, as |A| ≫ n
3−β
2
;
2. vol(L(x)) ≥ ad by Claim 8, and thus,
Pr[∀y ∈ A, s.t. y “unused”, L(x) ≁ y] ≤ exp(−vol(L(x)) · vol(A)/vol(V ))
≤ exp(−ad · nw/ad n)
= o(1).
The claim follows by Markov’s inequality.
Part III: upper bounding the label size of S Consider the distance function on G restricted
to all vertex pairs in S, distS : S × S → N. Clearly, distS can be determined from the labels of S.
We consider three cases:
1. The random graph does not satisfy the bound on |S2 | in Claim 9. By the above argument,
such a graph can only be generated with o(1) probability.
2. If there exists 0.01 · n3−β vertex pairs from S whose distance is at most 2d + 3, we claim
that the probability that a random graph induces any such distance function is at most o(1).
By Lemma 7, the probability that two vertices have distance at most 2d + 3 is o(1), hence
the expected number of vertex pairs in S within distance 2d + 3 is o(n3−β ). The claim then
follows by Markov’s inequality.
3. If the number of vertex pairs from S within distance 2d + 3 is at most 0.01 · n3−β , then we
infer that there are 0.2n(3−β)/2 vertices in S2 whose pairwise distance is at least 2d + 4. Let
B ⊆ S2 × S2 be the set of all such vertex pairs. The size of B is least 0.01n3−β . Note that the
distance function distS determines the set B, and we have |B| ≥ Θ(n3−β ), for any (x, y) ∈ B,
h(x) and h(y) must not be connected by an edge. Hence, for any such set B,
i
h
3−β
Pr dist(x, y) ≥ 2d + 4, ∀(x, y) ∈ B | |S2 | = Θ(n 2 )
Y
≤
Pr[h(x) ≁ h(y)]
(x,y)∈B
β−3
≤ (1 − Θ((ad /w)2/(β−2) /n))Θ(n )
≤ exp −(ad /w)2/(β−2) · nβ−4
2
1
2
= exp −n−2ε/(β−2) · w− β−2 − 3−β log log n · nβ−3
= exp −nβ−3−O(ε) .
That is, each of such labels for S could only occur with probability exp −nβ−3−O(ε) . However,
there are only O(2s ) different labels of total size at most s. Then by union bound, the probability
that a random graph induces labels for S of size at most nβ−3−O(ε) is o(1). This finishes the proof.
6
Experiment
In this section, we evaluate our algorithms on a collection of large networks. We compare with the
algorithm of Akiba et al.’s [4] and the Thorup-Zwick distance oracle [41, 13]. The first algorithm
produces an exact landmark labeling via recursively pruning during breadth first search over all
vertices – we will refer to it as PrunedLabel later. The second algorithm adapts the 3-approximate
distance oracle of Thorup and Zwick [41], via picking high degree vertices as global landmarks –
we refer to it as BallGrow. In Table 1 we list the graphs used in our experiment. More details
are available at Stanford Large Network Dataset Collection [33].
graph
Twitter
Stanford
Google
BerkStan
# nodes
81,306
281,903
875,713
685,230
# edges
1,768,149
2,312,497
5,105,039
7,600,595
category
Social
Web
Web
Web
90% effective diameter
4.5
9.7
8.1
9.9
average distance
3.8
5.2
6.0
6.3
Table 1: Basic statistics of graphs in experiments.
Implementation We implemented all three algorithms in Scala. The graph library we used is
available at https://github.com/teapot-co/tempest. We run the experiments on Amazon EC2
m4.4xlarge instance, with 64GB of RAM and 16 Intel Xeon 2.3GHz CPUs. We used a variant of
AlgSkewDegree in the experiments 6. We hand tune the two parameters used in the algorithm.
For PrunedLabel, a vertex ordering is required: we simply sort all vertices by indegree plus
outdegree. For BallGrow, it is necessary to specify the number of global landmarks; we handtune
this parameter and choose the numer of high degree vertices as global landmarks accordingly.
We measure accuracy over 2000 randomly sampled pairs of source/destination vertices. We
look at the 80 and 90-percentile multiplicative error (|estimated-distance / true-distance −1 |).
Algorithm 2 A description of our algorithm in experiments.
Input: A directed graph G = (V, E); Parameters d and K.
1: σ = vertices ordered by (indegree + outdegree)
2: for i ≤ n do
3:
if i ≤ K then
4:
computeGlobalLm(σi )
5:
else
6:
computeLocalLm(σi )
7:
end if
8: end for
9: procedure computeGlobalLm(x)
10:
{(y, dist(x, y)), ∀ y ∈ V } = Run a forward BFS
11:
{(y, dist(y, x)), ∀ y ∈ V } = Run a backward BFS
12: end procedure
13: procedure computeLocalLm(x)
14:
Run a forward BFS from x up to distance d, prune any node from {σi }K
i=1 .
15: end procedure
Results Table 2 compares the landmark size and running time of the three tested algorithms.
Table 3 compares the accuracy. Looking at accuracy, we found that both our algorithm and
BallGrow are fairly accurate on the three Web graphs. However, our algorithm does slightly
worse for the first test cases. From the performance comparison, we found that both our algorithm
and BallGrow are more scalable compared to PrunedLabel. This is to be expected, since
PrunedLabel is designed to gaurantee exact distances. Our algorithm found smaller landmark
sets compared to BallGrow and PrunedLabel in three out of four tests, and runs faster than
PrunedLabel on the two largest instance.
Twitter
Stanford
Google
BerkStan
Landmark size per node
Ours PrunedLabel BallGrow
227
261
637
82
95
367
215
285
276
63
155
742
Ours
<1
<1
7.2
1.0
Running time (min)
PrunedLabel BallGrow
10.5
1.8
<1
2.0
84
8.2
46.9
8.6
Table 2: Comparison of performances over our algorithm, PrunedLabel and BallGrow. The
landmark size is equal to the total number of forward and backward landmarks stored, divided by
the total number of vertices.
Twitter
Stanford
Google
BerkStan
90% error
Ours BallGrow
0.5
0.0
0.07
0.08
0.0
0.0
0.125 0.1
80% error
Ours BallGrow
0.25
0.125
0.0
0.0
0.0
0.0
0.0
0.0
Table 3: Comparion of accuracy. The accuracy of PrunedLabel is not listed becaue it is gauranteed to output exact distances.
7
Discussions
In this work, we studied distance labeling schemes on random graphs. We showed that simple
breadth first search based algorithms are near optimal. Our experiments suggest that the algorithms
we developed are effective on real world graphs that have small average distance and power law
degree distribution. Apart from closing the gap between upper and lower bounds, we discuss about
future work below.
Extensions. The Chung-Lu model has a natural extension to directed graphs. Consider two power
law distributions f in (x) and f out(x) with mean value bigger than 1, representing the indegree
and outdegree distributions, respectively. Each node v is associated with two parameters
out
in
out
(·), respectively. For any two nodes u and v, there is a directed
pin
v ∼ f (·) and pv ∼ f
out in
edge from u to v with probability pu M·pv , where M is a normalization term. We sketch a
√
heuristic argument which shows that O( n) landmarks per node suffices to get a 2-hop cover:
√
If we do a breadth-first search forward from every node x to include Õ( n) landmarks for
√
x as well as a backward BFS from x to include Õ( n) landmarks, then for every pair of
nodes x and y, the forward frontier of x and the reverse frontier of y will intersect with high
probability. A second possible extension is to consider configuration models with a power law
degree distribution. We believe all of our proofs can be extended to configuration models,
since our technical tools only involve bounding the growth of branching processes from every
node; We leave the details to future work.
Beyond worst case analysis. It would be interesting to consider deterministic characterizations
that will ensure short distance labelings. Do constant expansion ensure the existence of subquadratic distance labeling schemes? The high level intuition behind our algorithmic result is
that as long as the breadth-first search process grows neither too fast nor too last, but rather
at a uniform rate, then it is possible to obtain a “short” labeling scheme. Here is a more
concrete instantiation: consider any graph G = (V, E) such that for all x ∈ V and 0 ≤ i ≤ n,
√
(x)|
∈ [l, c], where c > l > 1 are fixed values. Now for each x ∈ V ,
either |Ni (x)| ≤ n or |Γ|Γi+1
i (x)|
√
add all the vertices within distance log2c n plus the closest n vertices to the landmark set of
x. It is not hard to see that this gives a 2-hop cover. The total size of the 2-hop cover is
log c n
P logc n
O(l 2 + i=12 Hcii ), where Hi is the total number of length-i path in G. For random power
law graphs when β > 3, l and c are asymptotically equal to r = ω/ν, E[Hi ] = O(nr i ) and the
above formula becomes O(n3/2 log n).
√
Distance oracles. For the finite variance case, the query complexity is Θ( n) for labeling schemes.
However, if we want faster query schemes, the situation seems very mysterious. Can we even
√
obtain sub-quadratic data structures with o( n) query complexity? In another direction, we
suspect that random graphs might provide a candidate hard instance for the set intersection
conjecture [40].
References
[1] Ittai Abraham, Daniel Delling, Andrew V Goldberg, and Renato F Werneck. A hub-based
labeling algorithm for shortest paths in road networks. In International Symposium on Experimental Algorithms, pages 230–241. Springer, 2011.
[2] Ittai Abraham, Amos Fiat, Andrew V Goldberg, and Renato F Werneck. Highway dimension,
shortest paths, and provably efficient algorithms. In Proceedings of the twenty-first annual
ACM-SIAM symposium on Discrete Algorithms, pages 782–793. Society for Industrial and
Applied Mathematics, 2010.
[3] Ittai Abraham and Cyril Gavoille. On approximate distance labels and routing schemes with
affine stretch. In International Symposium on Distributed Computing, pages 404–415. Springer,
2011.
[4] Takuya Akiba, Yoichi Iwata, and Yuichi Yoshida. Fast exact shortest-path distance queries
on large networks by pruned landmark labeling. In Proceedings of the 2013 ACM SIGMOD
International Conference on Management of Data, pages 349–360. ACM, 2013.
[5] Stephen Alstrup, Søren Dahlgaard, Mathias Bæk Tejs Knudsen, and Ely Porat. Sublinear
distance labeling. arXiv preprint arXiv:1507.02618, 2015.
[6] Ingo Althöfer, Gautam Das, David Dobkin, Deborah Joseph, and José Soares. On sparse
spanners of weighted graphs. Discrete & Computational Geometry, 9(1):81–100, 1993.
[7] Haris Angelidakis, Yury Makarychev, and Vsevolod Oparin. Algorithmic and hardness results for the hub labeling problem. In Proceedings of the Twenty-Eighth Annual ACM-SIAM
Symposium on Discrete Algorithms, pages 1442–1461. Society for Industrial and Applied Mathematics, 2017.
[8] Maxim Babenko, Andrew V Goldberg, Haim Kaplan, Ruslan Savchenko, and Mathias Weller.
On the complexity of hub labeling. In International Symposium on Mathematical Foundations
of Computer Science, pages 62–74. Springer, 2015.
[9] Bahman Bahmani and Ashish Goel. Partitioned multi-indexing: bringing order to social search.
In Proceedings of the 21st international conference on World Wide Web, pages 399–408. ACM,
2012.
[10] Béla Bollobás. Random graphs. In Modern Graph Theory, pages 215–252. Springer, 1998.
[11] Michele Borassi, Pierluigi Crescenzi, and Luca Trevisan. An axiomatic and an average-case
analysis of algorithms and heuristics for metric properties of graphs. In Proceedings of the
Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 920–939. SIAM,
2017.
[12] Shiri Chechik. Approximate distance oracles with improved bounds. In Proceedings of the
Forty-Seventh Annual ACM on Symposium on Theory of Computing, pages 1–10. ACM, 2015.
[13] Wei Chen, Christian Sommer, Shang-Hua Teng, and Yajun Wang. Compact routing in powerlaw graphs. In International Symposium on Distributed Computing, pages 379–391. Springer,
2009.
[14] Fan Chung and Linyuan Lu. The average distances in random graphs with given expected
degrees. Proceedings of the National Academy of Sciences, 99(25):15879–15882, 2002.
[15] Fan RK Chung and Linyuan Lu. Complex graphs and networks, volume 107. American
mathematical society Providence, 2006.
[16] Aaron Clauset, Cosma Rohilla Shalizi, and Mark EJ Newman. Power-law distributions in
empirical data. SIAM review, 51(4):661–703, 2009.
[17] Edith Cohen. Size-estimation framework with applications to transitive closure and reachability. Journal of Computer and System Sciences, 55(3):441–453, 1997.
[18] Edith Cohen, Daniel Delling, Fabian Fuchs, Andrew V Goldberg, Moises Goldszmidt, and
Renato F Werneck. Scalable similarity estimation in social networks: Closeness, node labels,
and random edge lengths. In Proceedings of the first ACM conference on Online social networks,
pages 131–142. ACM, 2013.
[19] Edith Cohen, Eran Halperin, Haim Kaplan, and Uri Zwick. Reachability and distance queries
via 2-hop labels. SIAM Journal on Computing, 32(5):1338–1355, 2003.
[20] Hagai Cohen and Ely Porat. On the hardness of distance oracle for sparse graph. arXiv
preprint arXiv:1006.1117, 2010.
[21] Atish Das Sarma, Sreenivas Gollapudi, Marc Najork, and Rina Panigrahy. A sketch-based
distance oracle for web-scale graphs. In Proceedings of the third ACM international conference
on Web search and data mining, pages 401–410. ACM, 2010.
[22] Daniel Delling, Andrew V Goldberg, Thomas Pajor, and Renato F Werneck. Robust distance
queries on massive networks. In European Symposium on Algorithms, pages 321–333. Springer,
2014.
[23] Daniel Delling, Andrew V Goldberg, Ruslan Savchenko, and Renato F Werneck. Hub labels:
Theory and practice. In International Symposium on Experimental Algorithms, pages 259–270.
Springer, 2014.
[24] Richard Durrett. Random graph dynamics, volume 200. Citeseer, 2007.
[25] Nicole Eikmeier and David F Gleich. Revisiting power-law distributions in spectra of real world
networks. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, pages 817–826. ACM, 2017.
[26] Mihaela Enachescu, Mei Wang, and Ashish Goel. Reducing maximum stretch in compact
routing. In INFOCOM 2008. The 27th Conference on Computer Communications. IEEE.
IEEE, 2008.
[27] Cyril Gavoille, Christian Glacet, Nicolas Hanusse, and David Ilcinkas. Brief announcement:
Routing the internet with very few entries. In Proceedings of the 2015 ACM Symposium on
Principles of Distributed Computing, pages 33–35. ACM, 2015.
[28] Cyril Gavoille, David Peleg, Stéphane Pérennes, and Ran Raz. Distance labeling in graphs.
In Proceedings of the twelfth annual ACM-SIAM symposium on Discrete algorithms, pages
210–219. Society for Industrial and Applied Mathematics, 2001.
[29] Andrew V Goldberg, Ilya Razenshteyn, and Ruslan Savchenko. Separating hierarchical and
general hub labelings. In International Symposium on Mathematical Foundations of Computer
Science, pages 469–479. Springer, 2013.
[30] Ken-ichi Kawarabayashi, Philip N Klein, and Christian Sommer. Linear-space approximate
distance oracles for planar, bounded-genus and minor-free graphs. In International Colloquium
on Automata, Languages, and Programming, pages 135–146. Springer, 2011.
[31] Tamara G Kolda, Ali Pinar, Todd Plantenga, and Comandur Seshadhri. A scalable generative
graph model with community structure. SIAM Journal on Scientific Computing, 36(5):C424–
C452, 2014.
[32] Jure Leskovec, Deepayan Chakrabarti, Jon Kleinberg, Christos Faloutsos, and Zoubin Ghahramani. Kronecker graphs: An approach to modeling networks. Journal of Machine Learning
Research, 11(Feb):985–1042, 2010.
[33] Jure Leskovec and Andrej Krevl. SNAP Datasets: Stanford large network dataset collection.
http://snap.stanford.edu/data, June 2014.
[34] Jure Leskovec, Kevin J Lang, Anirban Dasgupta, and Michael W Mahoney. Statistical properties of community structure in large social and information networks. In Proceedings of the
17th international conference on World Wide Web, pages 695–704. ACM, 2008.
[35] Shay Mozes and Christian Sommer. Exact distance oracles for planar graphs. In Proceedings
of the twenty-third annual ACM-SIAM symposium on Discrete Algorithms, pages 209–222.
SIAM, 2012.
[36] Mihai Patrascu and Liam Roditty. Distance oracles beyond the thorup-zwick bound. In
Foundations of Computer Science (FOCS), 2010 51st Annual IEEE Symposium on, pages
815–823. IEEE, 2010.
[37] Mihai Patrascu, Liam Roditty, and Mikkel Thorup. A new infinity of distance oracles for
sparse graphs. In Foundations of Computer Science (focs), 2012 Ieee 53rd Annual Symposium
on, pages 738–747. IEEE, 2012.
[38] Ely Porat and Liam Roditty. Preprocess, set, query! Algorithmica, 67(4):516–528, 2013.
[39] Christian Sommer. Shortest-path queries in static networks.
(CSUR), 46(4):45, 2014.
ACM Computing Surveys
[40] Christian Sommer, Elad Verbin, and Wei Yu. Distance oracles for sparse graphs. In Foundations of Computer Science, 2009. FOCS’09. 50th Annual IEEE Symposium on, pages 703–712.
IEEE, 2009.
[41] Mikkel Thorup and Uri Zwick. Approximate distance oracles. Journal of the ACM (JACM),
52(1):1–24, 2005.
[42] Remco Van Der Hofstad. Random graphs and complex networks. Available on http://www.
win. tue. nl/rhofstad/NotesRGCN. pdf, page 11, 2009.
[43] Christian Wulff-Nilsen. Approximate distance oracles with improved preprocessing time. In
Proceedings of the twenty-third annual ACM-SIAM symposium on Discrete Algorithms, pages
202–208. SIAM, 2012.
A
Upper Bound of Theorem 1
In this section we consider the degree sequence p when it is drawn from a power law distribution
with finite second moment.
Proposition 10. Let f denote a power law distribution with mean value ν > 1 and exponent β > 3.
1
Let p denote n independent samples from f (·) and 0 < ε < 1/2 be a fixed value. Let d ∼ o(n β−1 )
be a threshold value. The following holds almost surely:
1
i) The maximum weight max p ≤ o(n β−1 ).
P
ii) The volume of V is vol(V ) = ni=1 px = νn ± O(n1/2+ε ).
iii) The second moment below d is
X
x∈V
px 2 1px ≤d = (ω −
Zd3−β
)n ± O(n(1+ε)/2 d).
β−3
iv) The second moment above d is
X
x∈V
px 2 1px ≥d =
Z
nd3−β (1 + o(1)).
β−3
1
+ε
v) The second moment vol2 (V ) = ωn ± O(n 2(β−2) ).
P
vi) The first moment below d is x∈V px 1px ≤d ≤ (ν −
vii) The first moment above d is
P
x∈V
px 1px ≥d ≤
Zd2−β
β−2 )n
Zd2−β
β−2 n(1 +
+ o(n).
o(1)).
The proof is via standard concentration inequality – we leave it to the reader.
We will assume that Proposition 10 holds for the degree sequence p, and condition on p being
fixed. Given a random graph G = (V, E), we present an algorithm (shown below) for finding a
2-hop cover of G.
Algorithm 3 AlgBoundedVar
Input: An undirected graph G = (V, E); A parameter δ.
1: for x ∈ V do
√
2:
(F (x), l(x)) = AlgBfs(x, δ n)
3: end for
4: procedure AlgBfs(x, t)
5:
S = {x}
6:
α0 (x) = dx ; k = 0
7:
while αk (x) ≤ t ∧ |Γk+1 (x)| > 0 do
8:
S = S ∪ Γk (x)
9:
Y = {(y, z) ∈ E : y ∈ Γk (x), z ∈ Γk+1 (x)}
10:
k = k + 1P
11:
αk (x) = y∈Γk (x) dy − |Y |
12:
end while
13:
return (S, k)
14: end procedure
(estimate the next boundary size)
√
√
Lemma 11. Algorithm 3 with parameter δ n where δ = 5 ν log n finds a 2-hop cover F of G
with high probability.
Proof. Consider a fixed vertex x ∈ V , we first show that unless F (x) contains the entire
√ connected
component of x, the Algorithm will stop with a boundary layer whose volume is Ω( n log n) with
√
probability at least 1 − n−2 . If l(x) = 1, then either x is an
isolated node, or dx ≥ δ n. Since
√
√
n
dx ≤ n, this happens with probability at most exp(− (δ−1)
) ∼ o(n−4 ) by Proposition 21.
2
√
When l(x) = k + 1 ≥ 2, we show that vol(Γk (x)) ≥ δ n/3 with high probability. The termination
√
condition implies that αk (x) ≥ δ n. Consider the process in which αk (x) is generated: we keep
branching out from x until we reach the k-th level from x, then we reveal the edges between Γk (x)
√
and V \Nk−1 (x). Conditional on a = vol(Γk (x)) ≤ δ n/3, αk (x) is the sum of independent 0-1-2
random variables (because the edges inside Γk (x) are counted twice), and
X
X
py pz
E[αk (x)] =
vol(V )
y∈Γk (x) z ∈N
/ k−1 (x)
√
X pz
δ n
=a≤
≤a
vol(V )
3
z∈V
Therefore by Chernoff bound,
√
√
Pr[αk (u) ≥ δ n | vol(Γk (u)) ≤ δ n/3]
√
δ n
)
≤ exp(−
18
∼ o(n−4 ).
√
Secondly, we show that if two level sets have volume Ω( n log n) but are disjoint, then the
probability that there is no edge between them is very small. Let x and y be any two vertices. Let 1 ≤ k1 ≤ n − 1 and 1 ≤ k2 ≤ n − 1. Let Ωk1 ,k2 denote the set of graphs satisfying
√
√
vol(Γk1 (x)) ≥ δ n/3, vol(Γk2 (y)) ≥ δ n/3 and Γk1 (x) ∩ Γk2 (y) = ∅. Then by Proposition 3
Pr[Γk1 (x) ≁ Γk2 (y) | Ωk1 ,k2 ] ≤ exp(−
δ2 n
)
9vol(V )
∼ o(n−6 ).
Now we are ready to bound the probability that F is not a 2-hop cover. Let ΩS denote the set
of graphs such that
√
1. Γl(x) (x) = ∅ or vol(Γl(x)−1 (x)) ≥ δ n/3, for all nodes x;
2. if Γl(x) (x) and Γl(y) (y) are both non-empty, then F (x) ∩ F (y) is non-empty, for all x, y ∈ V ,
where x 6= y.
It’s clear that if G ∈ ΩS , then the algorithm successfully finds a 2-hop cover. The probability that
Condition (1) or (2) does not hold is at most o(n2 ) by taking union bound.
Lemma 12. Let x be a fixed node. Let 0 ≤ k ≤ O(log n). Let Ωk denote the set of graphs such
that
√
vol(Γi (x)) ≤ 4δ n, for any 0 ≤ i ≤ k − 1,
and
√
Then Pr[αk (x) ≤ δ n | Ωk ] ≤ n−2 .
√
vol(Γk (x)) > 4δ n.
Proof. Let a = vol(Γk (x)) and b = vol(Nk−1 (x)). Then conditional on Ωk , a and b satisfies
√
√
a > 4δ n and b ≤ 4kδ n
And αk is the sum of independent 0-1-2 random variables. Let µ denote its expected value, then
µ=
X
X
y∈Γk (x) z ∈N
/ k−1 (x)
py pz
b
= a(1 −
)
vol(V )
vol(V )
Since vol(V ) ∼ Θ(n) by Proposition 10, b/vol(V ) ∼ o(1). By Chernoff bound,
√
√
δ n
Pr[αk ≤ δ n | Ωk ] ≤ exp(−
) ∼ o(n−2 ).
4
Lemma 13. Let x be a fixed node. Let 0 ≤ k ≤ O(log n). Denote by Ω∗k the set of graphs such that
√
αi ≤ δ n, for all 0 ≤ i ≤ k
√
Then Pr[vol(Γk (x)) > 4δ n, Ω∗k ] ≤ (k + 1)n−2 .
Proof. For k = 0, the claim follows by applying Lemma 12. For k > 0, we repeatedly apply Lemma
12 for all values of i smaller than k. For 0 ≤ i ≤ k, denote by Si ⊂ Ω∗k the set of graphs such that:
√
√
(1) vol(Γk (x)) > 4δ n; (2) vol(Γj (x)) ≤ 4δ n for any 0 ≤ j ≤ i − 1. We claim that
Pr[Si ] − Pr[Si+1 ] ≤ n−2 , for 0 ≤ i ≤ k − 1,
and Pr[Sk ] ≤ n−2 . These two combined together would imply this conclusion.
To see this,
√
Pr[Si ] − Pr[Si+1 ] = Pr[vol(Γi (x)) > 4δ n, Si ]
√
≤ Pr[αi ≤ δ n, Ωi ] ≤ n−2
The first inequality is because if a graph G is in Si (in particular condition (2)) and the volume of
√
√
Γi (x) is bigger than 4δ n, then G ∈ Ωi . And the condition on Si ⊂ Ω∗k implies αi ≤ δ n. The
second inequality is because of Lemma 12. One can similarly argue about Pr[Sk+1 ] and we omit
the details.
Lemma 14. The following holds almost surely:
p
• |F (x)| ∼ O( n log3 n) for all x ∈ V .
p
• The algorithm runs in time O( n log3 n) for all x ∈ V .
Proof. We first bound the number of landmarks added before reaching the boundary layer. Since
√
αi (x)
p ≤ δ3 n, for i = 0, . . . , l(x) − 2, and l(x) ≤ O(log n), we conclude that there are at most
O( n log n) landmarks before layer l(x) − 1. The rest of the proof shows that |Γl(x)−1 (x)| ∼
p
√
O( n log3 n). Set c = 8 log n + 8δω/ν. When l(x) = 1, clearly dx ≤ n by Proposition 10.
√
When l(x) = k + 1 ≥ 2, we have that G ∈ Ω∗k−1 and hence vol(Γk−1 (x)) ≤ 4δ n with probability
√
at least 1 − kn−2 , by Lemma 13. We now argue that vol(Γk (x)) ≤ c n with high probability.
√
Conditional on a = vol(Γk−1 (x)) ≤ 4δ n, vol(Γk (x)) is the sum of independent random variables
√
that are all bounded in [0, n], with expected value µ as
X
µ=
Pr[z ∼ Γk−1 (y)] × py
(by Proposition 3)
y ∈N
/ k−1 (x)
≤
≤
X
y ∈N
/ k−1 (x)
py 2 a
vol(V )
vol2 (V )
ω
× a ∼ a( + o(1)).
vol(V )
ν
The last line is because vol2 (V ) = ωn(1 + o(1)) and vol(V ) = νn(1 + o(1)), by Proposition 10. By
Chernoff bound,
√
Pr[vol(Γk (x)) > c n | vol(Γk−1 (x)) ≤ a]
√
c n−µ
) ≤ exp(−3 log n) ∼ o(n−2 ),
≤ exp(− √
2 n
√
because µ ≤ ων a(1 + o(1)) and a ≤ 4δ n.
√
√
For the last part, we argue that if vol(Γk (x)) ≤ c n, then αk (x) ≤ 3c n. This is because
√
√
conditional on a = vol(Γk (x)) ≤ c n, the expected value of αk (x) is at most c n (details omitted).
Since αk (x) is the sum of independent 0-1-2 random variables, by Chernoff bound
√
√
Pr[αk (x) > 3c n | vol(Γk )(x) ≤ c n]
√
c n
)
≤ exp(−
4
∼ o(n−2 ).
In summary,
√
Pr[l(x) = k + 1, αk (x) > 3c n] ∼ O(n−2 )
Taking a union bound over k ≤ O(log n) and x ∈ V , we obtain the desired conclusion.
B
Upper Bound of Theorem 2
In this section, we consider degree sequence p whose second moment is infinite.
Proposition 15. Let f denote the probability density function of a power law distribution with
mean value ν > 1 and exponent 2 < β ≤ 3. Let p denote n independent samples from f (·). Let
√
log n ≤ d ≤ 2 n be any fixed value and let ε(n) be a function that goes to 0 when n goes to infinity.
Then almost surely the following holds:
1
i) The maximum weight max p ≥ ε(n)n β−1 .
P
ii) The sum of weights beyond d is x∈V px 1px ≥d ∼ o(n).
iii) The volume of V is vol(V ) = νn ± o(n).
√
iv) Let log n < K ≤ 2 n be a fixed value. Set
c(K) =
Then
(
X
x∈V
5−2β
3Zxmin
2β−5
3Z
5−2β
5−2β K
if 2.5 ≤ β ≤ 3
if 2 < β < 2.5
px 4−β 1px ≤K ≤ c(K)n.
v) Let c > 1 denote a fixed constant value. For any vertex x ∈ V ,
X
y∈V
py 1 py ≤px ≤2√n ≤ 6 max(
c
√
cβ−2 Z
npu 2−β , n log n).
β−2
The proof is via standard concentration inequality (details omitted). In the following we assume
that p satisfies all properties in Proposition 15. In the following, we prove that Algorithm 1 is correct
in Lemma 16 and bound its output size in Lemma 19.
Lemma 16. Algorithm 1 with parameter δ and K finds a 2-hop cover F with high probability.
Proof. Consider the random variable l(x) that is computed in the Algorithm for each node x. Let
ΩS denote the set of graphs that satisfies
Γl(x) (x) = ∅ or dist(v ∗ , x) ≤ l(x), ∀x ∈ V
where v ∗ is the node with the maximum weight. We argue that Algorithm 1 finds a 2-hop cover
for any G ∈ ΩS , and
1 − Pr[ΩS ] ≤ 2/n
This would imply that Algorithm 1 succeeds with probability at least 1 − 2/n.
We first argue that Algorithm 1 is correct if G ∈ ΩS . Let x and y be two different vertices in
V . If x and y are not reachable from each other, then clearly F (x) ∩ F (y) = ∅. If x and y are
reachable from each other, consider their distance dist(x, y). Note that when Γl(x) (x) (or Γl(y) (y))
is empty, then F (x) (or F (y)) includes the entire connected component that contains x (or y).
Therefore, y ∈ F (x), vice versa. When none of them are empty, we know that dist(x, v ∗ ) ≤ l(x)
and dist(y, v ∗ ) ≤ l(y) since G ∈ ΩS . We consider three cases:
• If dist(x, y) ≤ l(x) + l(y) − 2, then there exists a node z such that dist(x, z) ≤ l(x) − 1 and
dist(y, z) ≤ l(y) − 1. By our construction, z is in F (x) and F (y).
• If dist(x, y) = l(x) + l(y) − 1, then consider the two nodes z and z ′ on one of the shortest path
from x to y, with dist(x, z) = l(x) − 1 and dist(y, z) = l(y). If either dz or dz ′ is at least K,
then they have been added as a landmark to every node in V . Otherwise, assume without
loss of generality that dz ≥ dz ′ . Then our construction adds z into F (y) and clearly z is also
in F (x), hence z is a common landmark for x and y.
• If dist(x, y) = l(x) + l(y), then clearly v ∗ is a common landmark for x and y.
We now bound 1 − Pr[ΩS ]. Clearly,
1 − Pr[ΩS ]
X
≤
Pr[Γl(x) (x) 6= ∅, dist(v ∗ , x) > l(x)]
x∈V
=
X
X n−1
x∈V k=0
Pr[l(x) = k + 1, Γk+1 (x) 6= ∅, dist(v ∗ , x) > k + 1]
Note that l(x) = k + 1 and Γk+1 (x) 6= ∅ is the same as the event that:
1
• αi (x) ≤ δn1− β−1 , for i = 0, . . . , k − 1;
• αk (x) > δn
1
1− β−1
.
Hence,
Pr[l(x) = k + 1, Γk+1 (x) 6= ∅, dist(v ∗ , x) > k + 1]
≤ Pr[αk (x) > δn
≤ Pr[αk (x) > δn
1
1− β−1
1
1− β−1
+ Pr[vol(Γk (x)) >
δn
, dist(v ∗ , x) > k + 1]
, vol(Γk (x)) ≤
1
1− β−1
3
δn
1
1− β−1
3
]
, dist(v ∗ , u) > k + 1]
(2)
(3)
For Equation (2), consider how αk (x) is discovered when we do the level set expansion from node
1
x. Conditioned on a = vol(Γk (x)) ≤ δn1− β−1 /3, αk (x) is the sum of 0-1-2 independent random
1− 1
variables, with expected value less than δn β−1 /3. Hence by Chernoff bound, Equation (2) is at
1− 1
1− 1
most exp(−δn β−1 /6) ∼ o(n−3 ). For Equation (3), conditioned on vol(Γk (x)) ≥ δn β−1 /2 and
v∗ ∈
/ Nk (x),
1−
1
δn β−1 pv∗
) ∼ o(n−3 )
Pr[v ≁ Γk (x)] ≤ exp(−
2vol(V )
∗
The first inequality is because of Proposition 3. The second inequality is because vol(V ) ∼ νn±o(n)
by Proposition 15. In summary, 1 − Pr[ΩS ] ≤ 2/n.
We now consider the size of our landmark scheme. There are three parts in each landmark set:
(1) the heavy nodes whose degree is at least K; (2) all the level sets before the last layer; (3) the
last layer that we carefully constructed. It’s not hard to bound the first part, since the degree of a
node is concentrated near its weight, and the number of nodes whose weight is Ω(K) is O(nK 1−β ).
The second part can be bounded by the maximum number of layers, hence the diameter of G,
which is O(log n). For the third part, the idea is that before adding all the nodes on the boundary
layer, we already have a (+1)-stretch scheme. Therefore, for a given vertex x, it is enough if we
only add neighbors whose degree is bigger than dx — this reduces the amount of vertices from dx
to O(d3−β
x ).
1
We first show that the volume of all the level sets is at most O(δn1− β−1 ) before the boundary
layer. For the rest of the section, let αk = αk (x) for any 0 ≤ k ≤ n−1, unless there is any ambiguity
on the vertex we are considering. Recall that αk denotes the number of edges between Γk (x) and
V \Nk−1 (x).
Lemma 17. Let x be a fixed node. Let k be an integer less than ≤ O(log n). Let Ωk denote the set
of graphs such that
1− 1
vol(Γi (x)) < 4δn β−1 , for any 0 ≤ i ≤ k − 1,
and
vol(Γk (x)) > 4δn
Then Pr[αk ≤ δn
1
1− β−1
1
1− β−1
.
| Ωk ] ≤ n−2 .
Proof. Let a = vol(Γk (x)) and b = vol(Nk−1 (x)). Conditioned on Ωk ,
1
1
a > 4δn1− β−1 and b ≤ 4kδn1− β−1 .
Clearly, the random variable αk is the sum of independent 0-1 random variables. Let µ denote its
√
expected value. For each y ∈ Γk (x), we know that py ≤ a = O( n). Let µy denote the expected
number of edges between y and V \Nk−1 (x), then
X
µy =
min(
z:z6=y∧z ∈N
/ k−1 (x)
≥ py (1 −
because of Proposition 15. And µ =
Chernoff bound,
b+
P
P
py pz
, 1)1pz ≤√n
vol(V )
pz 1pz ≥√n
) = py (1 − κ(n))
vol(V )
z∈V
y∈Γk (x) µy
= (1 − o(1))a. Let c =
µ
δn
1−
1
β−1
≥ 2 − o(1). By
1
Pr[αk ≤ δn
1
1− β−1
(c − 1)2 δn1− β−1
) ∼ o(n−2 )
| Ωk ] ≤ exp(−
4
Lemma 18. Let x be a fixed vertex. Let 0 ≤ k ≤ O(log n). Denote by Ω∗k the set of graphs such
that
1− 1
αi ≤ δn β−1 , for any 0 ≤ i ≤ k
1
Then Pr[vol(Γk (x)) > 4δn1− β−1 , Ω∗k ] ≤ (k + 1)n−2
Proof. When k = 0, the claim is proved by Lemma 17. When k ≥ 1, we will repeatedly apply
Lemma 17 to prove the statement. For any values of i smaller than or equal to k, let Si ⊂ Ω∗k
denote the set of graphs that also satisfy: (1) vol(Γj (x)) ≤ 4δn
1
1− β−1
. We show that Pr[Si ] − Pr[Si+1 ] ≤
vol(Γk (x)) > 4δn
Our Lemma follows from the two claims.
For the first part,
n−2
1
1− β−1
, for any 0 ≤ j ≤ i − 1; (2)
if 0 ≤ i ≤ k − 1, and Pr[Sk ] ≤ n−2 .
1
Pr[Si ] − Pr[Si+1 ] = Pr[vol(Γi (x)) > 4δn1− β−1 , Si ]
1
≤ Pr[Ωi , αi ≤ δn1− β−1 ] ≤ n−2
1
The first inequality is because if G ∈ Si and G satisfies vol(Γi (x)) > 4δn1− β−1 , then G ∈ Ωi . Also
1
αi ≤ δn1− β−1 since G ∈ Si ⊂ Ω∗k . The second inequality is because of Lemma 17. The other part
can be proved similarly and we omit the details.
Now we are ready to bound the size of our landmark scheme.
Lemma 19. The following holds almost surely
• |F (x)| ∼ O(n
1
1
, 4−β
)
1−min( β−1
· log3 n) for all x ∈ V ;
• The algorithm terminates in time O(n
1
1
2−min( β−1
, 4−β
)
· log3 n).
Remark To implement Line 7, one can first sort N (x) for each x ∈ V , in descending order on
their degrees, and then create a separate list that truncates the nodes whose degree is at least K.
Given this list, one can find the set of neighbors of x whose degree is between [dx , K]. The amount
of time it takes to sort N (x) is O(dx log dx ) ∼ O(dx log n). Hence the total amount of time it takes
to sort all the adjacency lists is O(|E| log n) = O(n log n).
We will use the following lemma for technical reasons — the proof is deferred to the end of the
section.
Lemma 20. Let x be a fixed node with weight px ≤ 2K. Denote by
Sx = {y ∈ N (x) : dx ≤ dy and dy ≤ K}
and let dˆx = |Sx |. Then
where c1 =
192Z
ν(β−2)
Pr[dˆx ≥ max(c1 px 3−β , c2 log n)] ≤ n−3
and c2 = 130.
Proof of Lemma 19: We first bound the number of nodes in H. By Proposition 22, with probability
1 − n−1
1
1
1−min( β−1
, 4−β
)
|H| = O(nK 1−β ) = O(n
)
Secondly, we bound the number of landmarks added before reaching the boundary layer. For any
1
vertex x, with i = 0, . . . , l(x) − 2, |Γi (x)| ≤ αi (x) = O(δn1− β−1 ). Since l(x) ≤ O(log n), the total
1
landmarks for these layers are at most O(n1− β−1 log3 n). The rest of the proof will bound the
number of landmarks on the boundary layer with depth l(x) − 1.
Denote by
X
dˆy 1dy ≤K for x ∈ V, 0 ≤ k ≤ n − 1
πk (x) =
y∈Γk (x)
Hence πl(x)−1 (x) gives the number of landmarks added on the boundary layer.
1−min(
1
,
1
)
1−
1
3Z
β−1 4−β , and ∆ = max(c ψ, c δn
β−1 log n),
Set c3 = |2β−5|
max(x5−2β
1
2
min , 1), ψ = 12c3 δn
where c1 and c2 are defined in Lemma 20. We show that πl(x)−1 (x) ≤ ∆ with probability 1 − n−2
for the rest of the proof — our conclusion follows by taking union bound over x ∈ V and 1 ≤ l(x) ≤
O(log n).
When l(x) = 1, π0 (x) = dx ≤ K ≤ ∆. When l(x) = k + 1 ≥ 2, we know that G ∈ Ω∗k−1 . Hence
1
by Lemma 18, vol(Γk−1 (x)) ≤ 4δn1− β−1 with high probability. More concretly,
Pr[l(x) = k + 1, πk (x) ≥ ∆]
1
≤ (k + 1)n−2 + Pr[l(x) = k + 1, vol(Γk−1 (x)) ≤ 4δn1− β−1 , πk (x) ≥ ∆]
Denote by
wk =
X
y∈Γk (x)
py 3−β 1py ≤2K .
(4)
1
1−
Conditional on a = vol(Γk−1 (x)) ≤ 4δn β−1 , we show that wk ≤ ψ with high probability. Denote
1− 1
by Ωw the set of graphs satisfying a ≤ 4δn β−1 . Conditioned on Ωw , wk is the sum of independent
random variables that are all bounded in [0, (2K)3−β ]. Hence
X
E[wk ] =
Pr[y ∼ Nk−1 (x)]py 3−β 1py ≤2K
y ∈N
/ k−1 (x)
≤
a
(
vol(V )
≤
X
a
(
py 4−β 1py ≤2K )
vol(V )
X
y ∈N
/ k−1 (x)
py 4−β 1py ≤2K )
(by Proposition 3)
y∈V
aφ(K)n
≤
vol(V )
aφ(K)
∼
ν
ψ
≤ .
3
(by Proposition 15)
(vol(V ) = νn ± o(n) by Proposition 15)
The last line follows by a ≤ 4δn
Chernoff bound on wk ,
1
1− β−1
and φ(K)n
Pr[wk > ψ | Ωw ] ≤ exp(−
1
1− β−1
≤ c2 n
1
1
1−min( β−1
, 4−β
)
. Now we apply
ψ
) ∼ o(n−2 )
4(2K)3−β
because when 2.5 ≤ β ≤ 3,
ψ
K 3−β
And when 2 < β < 2.5,
= Θ(n
1
− 3−β
1− β−1
2
ψ
) = Θ(n
(β−1)2 −2
2(β−1)
)
(3−β)(β−2)
K 3−β
= Θ(n (4−β)(β−1) )
Hence the second part in Equation (4) is bounded by o(n−2 ) plus
Pr[l(x) = k + 1, vol(Γk−1 (x)) ≤ 4δn
≤ Pr[wk ≤ ψ, αk−1 ≤ δn
1
1− β−1
≤ Pr[wk ≤ ψ, |Γk (x)| ≤ δn
1
1− β−1
, wk ≤ ψ, πk (x) ≥ ∆]
, πk (x) ≥ ∆]
1
1− β−1
, πk (x) ≥ ∆]
In the reminder of the proof we show the above Equation is at most n−2 . Denote by
X
dˆy 1py ≤2K
πk′ (x) =
y∈Γk (x)
By Proposition 21, Pr[dy ≤ K | py > 2K] ≤ exp(−K/8) ∼ o(n−3 ) for any y ∈ V . Hence
πk′ (x) = πk (x) with probability at least 1 − o(n−2 ). Lastly, we have
Pr[wk ≤ ψ, |Γk (x)| ≤ δn
1
1− β−1
, πk′ (x) ≥ ∆] ≤ n−2
Otherwise, there exists a vertex y ∈ Γk (x) such that py ≤ 2K and dˆy ≥ max(c1 py 3−β , c2 log n),
1−
1
because ∆ ≥ max(c1 ψ, c2 δn β−1 log n). This happens with probability at most n−2 , by taking
union bound over every vertice with Lemma 20.
Proof of Lemma 20: When px ≤ c2 log n/2,
Pr[dˆx ≥ c2 log n] ≤ Pr[dx ≥ c2 log n] ≤ o(n−3 )
Now suppose that px > c2 log n/2. Consider any vertex y whose weight is at most px /8. Then
Pr[dy ≥ dx ] ≤ Pr[dy ≥ px /4] + Pr[dx < px /4] ∼ o(n−4 )
The second inequality is because of Proposition 21. Hence y is not in Sx .
Now if py ≥ 2K, then Pr[dy ≤ K] ≤∼ o(n−4 ). Hence y is also not in Sx . Lastly, let X denote the
set of vertices whose weight is between [ p8x , 2K] and who is connected to x. We have
X
E[X] =
y∈V \{x}:px /8≤py ≤2K
≤
px py
vol(V )
√
4px
8Z
max(
npx 2−β , n log n)
vol(V )
β−1
≤ max(c1 px 3−β , c2 log n)/3
The first inequality is because of Proposition 15. The second inequality is because vol(V ) =
√
νn + o(n) by Proposition 15, and px ≤ 2K ≤ 2 n. From here it is not hard to obtain that
Pr[|X| ≤ max(c1 px 3−β , c2 log n)] ∼ o(n−3 ).
C
Random Graph Toolbox
The following Lemma characterizes the probability that a vertex’s actual degree deviates from its
weight.
Proposition 21. Let G = (V, E) ∈ G n (p) be a random graph. Let x be a fixed vertex with weight
px and degree dx in G. Then
1. If c ≥ 3, then
Pr[dx ≥ cpx ] ≤ exp(−
2. If 0 < c < 1, then
Pr[dx ≤ cpx ] ≤ exp(−
(c − 1)px
)
2
(1 − c)2 px
)
8
Proof. Let µ = E[dx ]. First,
µ=
X
y∈V \{x}
≤
min(
px py
, 1)
vol(V )
X px py
= px
vol(V )
y∈V
By Chernoff bound, for any c ≥ 3,
cpx − µ
)
2
(c − 1)px
≤ exp(−
)
2
Pr[dx ≥ cpx ] ≤ exp(−
since cpx − µ ≥ 2µ.
1− 1
ν
n β−1 , then for any y ∈ V where py ≤ t, we know that px py ≤
On the other hand, let t = 2ε(n)
vol(V ) by Proposition 15. Hence
P
px
y∈V py 1py ≥t
−
)
µ ≥ px (1 −
vol(V )
vol(V )
By Proposition 15,
X
y∈V
py 1py ≥t ∼ o(n)
Since px ∼ o(n) and vol(V ) = νn + o(n), we conclude that µ = px (1 − o(1)). By Chernoff bound,
for any 0 < c < 1,
Pr[dx ≤ cpx ] ≤ exp(−
px (1 − c)2
(cpx − µ)2
) ≤ exp(−
)
4µ
8
for large enough n.
The following Proposition characterizes the number of nodes whose degree is at least K.
√
Proposition 22. Let G = (V, E) ∈ G n (p) be a random graph. Let 8 log n ≤ K ≤ n denote a fixed
β−1
1−β , log n).
value and S = {x ∈ V : dx ≥ K}. With probability at least 1−n−1 , |S| ≤ 3 max( Z3
β−1 nK
K
Proof. Let Y1 = {x ∈ V : px ≥ K
3 } and Y2 = {x ∈ V : px < 3 and K ≤ dx }. Clearly, S ⊂ Y1 ∪ Y2 .
We first show that Y2 is empty with probability at least 1 − n−1 . Consider a fixed node x ∈ V with
weight px ≤ K/3. By Proposition 21,
Pr[dx ≥ K] ≤ exp(−K) ∼ o(n−2 )
Hence Pr[Y2 6= ∅] = o(n−1 ) by union bound.
β−1
1−β . Then by Chernoff
We then bound the size of Y1 . The expected value of Y1 is Z3
β−1 nK
bound, it’s not hard to obtain the desired conclusion (details omitted).
C.1
Proof of Growth Lemma 6: β > 3
We first show an upper bound for the expected volume for each level Γk (x).
The expected volume of Γk (x) is O(r k ) Let us first fix Γk (x), and consider the set Γk+1 (x).
For a vertex y, the probability that it is in Γk+1 (x) is at most
py ·
vol(Γk (x))
vol(V )
by Proposition 3. Thus, the expected volume of Γk+1 (x) conditioned on vol(Γk (x)) is at most
E[vol(Γk+1 (x)) | vol(Γk (x))]
X
vol(Γk (x))
≤
p2y ·
vol(V )
y6∈Nk (x)
vol2 (V )
vol(V )
= vol(Γk (x)) · r.
≤ vol(Γk (x)) ·
On the other hand, vol(Γ0 (x)) = O(1). Thus, we have E[vol(Γk (x))] = O(r k ). This proves Item 1.
Two fixed constant-weight vertices are close with very low probability Fix two vertices
x, y ∈ S. By Item 1,
k
E[vol(Nk (x))] ≤ O(r ).
However, for each i, the probability that y is at distance i from x conditioned on Ni−1 (x) is at
most
vol(Γi−1 (x))
Pr[y ∈ Γi (x) | Ni−1 (x)] ≤ py ·
.
vol(V )
The probability y is within distance k + 1 from x is at most
k+1
X
Pr[y ∈ Nk+1 (x)] ≤
i=1
Pr[y ∈ Γi (x)]
k+1
X
py
·
≤
E[vol(Γi−1 (x))]
vol(V )
i=1
Nk (x)
= py ·
vol(V )
≤ O(r k ).
\Nk (x))
With large probability, Γk+1 (x) has volume not much smaller than Γk (x) · vol2 (V
vol(V )
Conditioned on Γk (x), the probability that a vertex y ∈
/ Nk (x) is in Γk+1 (x) is at least
−py ·
1−e
vol(Γk (x))
vol(V )
by Proposition 3.
For any T > 0, we have
X
y:py >T
We also have
X
y:py ≤T
p2y ≤ T −γ ·
p3y ≤ T 1−γ ·
X
y:py >T
X
y:py ≤T
p2+γ
≤ τ · n · T −γ .
y
p2+γ
≤ τ · n · T 1−γ .
y
Let us focus on all y’s with weight at most T . By that fact that 1 − e−x ≥ x − x2 /2 when x ≥ 0,
the expected volume of Γk+1 (x) ∩ {y : py ≤ T } conditioned on Nk (x) is at least
−py ·
py 1 − e
X
p2y ·
y6∈Nk (x):py ≤T
≥
X
y6∈Nk (x):py ≤T
vol(Γk (x))
vol(V )
vol(Γk (x)) 1
−
vol(V )
2
X
y6∈Nk (x):py ≤T
p3y ·
vol(Γk (x))
vol(V )
2
X
vol(Γk (x)) 2
1 X 3
vol(Γk (x))
2 vol(Γk (x))
py ·
−
py ·
−
≥ (vol2 (V \ Nk (x))) ·
vol(V )
vol(V )
2
vol(V )
y:py ≤T
y:py ≥T
vol2 (V \ Nk (x))
vol(Γk (x)) 2
−γ vol(Γk (x))
1−γ
≥ vol(Γk (x)) ·
−τ ·n·T ·
−τ ·n·T
·
vol(V )
vol(V )
vol(V )
vol(Γ
(x))
vol2 (V \ Nk (x))
k
− τ · T −γ + T 1−γ ·
.
≥ vol(Γk (x)) ·
vol(V )
vol(V )
Note that “y ∈ Γk+1 (x)” are independent events conditioned on Nk (x) for different y ∈
/ Nk (x).
Now we apply Chernoff Bound to lower bound the probability that the volume of Γk+1 (x) is too
small. The above inequality holds for every T > 0. In the following, we set T = vol(Γk (x))1/2 .
k (x))
When vol(Γk (x)) ≤ vol(V )2/3 , T −γ ≥ T 1−γ · vol(Γ
vol(V ) , the expected volume of Γk+1 (x) conditioned on Nk (x) is at least:
E[vol(Γk+1 (x) ∩ {y : py ≤ T }) | Nk (x)]
vol2 (V \ Nk (x))
− 2τ · vol(Γk (x))−γ/2 .
≥ vol(Γk (x)) ·
vol(V )
Since each py ≤ T = vol(Γk (x))1/2 , by Chernoff bound, we have
vol2 (V \ Nk (x))
− vol(Γk (x))−γ/3 Nk (x)
Pr vol(Γk+1 (x)) ≤ vol(Γk (x)) ·
vol(V )
1/2−2γ/3
),
≤ 2−Θ(vol(Γk (x))
(5)
as long as vol(Γk (x)) = O(n2/3 ) and vol(Γk (x)) sufficiently large.
With constant probability, Γk (x) has volume at least Ω(r k ) Fix a sufficiently large constant
C, denote by E0 the event that x has a neighborhood of volume at least C. Then it is not hard to
verify that for any constant C, the probability of E0 is at least a constant:
Pr[vol(Γ1 (x)) ≥ C] ≥ ΩC (1).
Moreover, for i ≥ 1, denote by Ei the event that either
vol2 (V \ Ni (x))
vol(Γi+1 (x)) > vol(Γi (x)) ·
− vol(Γi (x))−γ/3
vol(V )
or
vol(Γi (x)) ≥ n2/3 .
By the argument above,
1/2−2γ/3 )
Pr[Ei | Ni (x)] ≤ 2−Θ(vol(Γi (x))
We claim that these events have the following properties.
.
Claim 23. When Ei occurs for all 0 ≤ i < k, we must have either vol(Γk (x)) ≥ Ω(r k ) or
vol(Nk (x)) ≥ n2/3 for sufficiently large C.
Claim 24. All events Ei ’s (0 ≤ i < k) occur simultaneously with constant probability.
Before proving the two claims, let us first show that they together imply that Pr[vol(Γk (x)) ≥
Ω(r k )] ≥ Ω(1).
By Markov’s inequality, the first inequality in the lemma statement and k ≤ 12 logr n, we have
√
Pr[vol(Nk (x)) ≥ n2/3 ] ≤ O( n/n2/3 ) = o(1).
Therefore, we have the lower bound
Pr[vol(Γk (x)) ≥ Ω(r k )]
≥ Pr[E0 , . . . , Ek−1 ] · Pr[vol(Γk (x)) ≥ Ω(r k ) | E0 , . . . , Ek−1 ]
≥ Pr[E0 , . . . , Ek−1 ] · (1 − Pr[vol(Nk (x)) ≥ n2/3 | E0 , . . . , Ek−1 ])
≥ Pr[E0 , . . . , Ek−1 ] · (1 − Pr[vol(Nk (x)) ≥ n2/3 ]/ Pr[E0 , . . . , Ek−1 ])
= Ω(1).
This proves the lemma.
Proof of Claim 23: Assume Ei occurs for all 0 ≤ i < k and vol(Nk (x)) < n2/3 . The goal is to show
that in this case, we must have vol(Γk (x)) ≥ Ω(r k ).
In particular, vol(Nk (x)) < n2/3 implies that vol(Ni (x)) < n2/3 and vol(Γi (x)) ≤ n2/3 for every
i ≤ k. By Hölder’s inequality, we also have
1
γ
vol2 (Ni (x)) ≤ vol2+γ (Ni (x)) 1+γ · vol(Ni (x)) 1+γ
1
≤ τ 1+γ · n
γ
1− 3(1+γ)
.
Thus, the event Ei (i > 0) implies
1
− γ
vol(Γi+1 (x)) > vol(Γi (x)) · r − τ 1+γ · n 3(1+γ) − vol(Γi (x))−γ/3 .
1
−
γ
Let r̂ = r − τ 1+γ · n 3(1+γ) − C −γ/3 . For sufficiently large C, r̂ >
vol(Γi (x)) ≥ C · r̂ i−1 ≥ C · r (i−1)/2 for i ≥ 1 inductively:
√
r > 1. First we can show
• By the definition of E0 , vol(Γ1 (x)) ≥ C;
• If vol(Γi (x)) ≥ C · r̂ i−1 , then we have
1
− γ
vol(Γi+1 (x)) ≥ vol(Γi (x)) · r − τ 1+γ · n 3(1+γ) − vol(Γi (x))−γ/3
1
− γ
≥ vol(Γi (x)) · r − τ 1+γ · n 3(1+γ) − C −γ/3
≥ vol(Γi (x)) · r̂.
(6)
Thus, we have vol(Γi (x)) ≥ Ω(r̂ i ) ≥ Ω(r i/2 ). By Equation (6) again, we have
1
vol(Γk (x)) > vol(Γk−1 (x)) · (r − τ 1+γ · n
≥ vol(Γ1 (x)) ·
≥ r k−1 · C ·
k−1
Y
i=1
k−1
Y
i=1
k
= r · αk ,
γ
− 3(1+γ)
1
(r − τ 1+γ · n
(1 − τ
1
1+γ
·n
− r −(k−1)γ/6 )
!
γ
− 3(1+γ)
γ
− 3(1+γ)
− r −iγ/6 )
!
· r −1 − r −iγ/6−1 )
where αk is decreasing as k increases. α 1 logr n is lower bounded by a constant α. Thus, vol(Γk (x)) ≥
2
α·r k ≥ Ω(r k ). This proves the claim. Proof of Claim 24: By Lemma 23, conditioned on E0 , . . . Ei−1 ,
we have either
vol(Γi (x)) ≥ α · r i
or
Thus, by Equation (5), we have
vol(Ni (x)) ≥ n2/3 .
Pr[Ei | E0 , . . . , Ei−1 ] ≤ 2−Θ(r
Ω(i) )
.
Since Pr[E0 ] = Ω(1), we may lower bound the probability that all events happen simultaneously
!
k−1
Y
−Θ(r Ω(i) )
1−2
Pr[E0 , . . . , Ek−1 ] ≥ Ω
i=0
≥ Ω(1).
C.2
Proof of Growth Lemma 7: 2 < β < 3
Let us first upper bound the volume of k-neighborhood of any vertex x.
Upper bounding vol(Γk (x)) We will upper bound vol(Γk (x)) in terms of vol(Γk−1 (x)). First,
we have vol(Γ0 (x)) = vol(x) by definition. For any vertex y, the probability that it is connected to
some vertex in Γk−1 (x) is at most
X
z∈Γk−1 (x)
py ·
vol(Γk−1 (x))
pz
= py ·
.
vol(V )
vol(V )
Thus, the probability that Γk (x) contains any vertex with very high weight is low:
h
i
Pr ∃y, py ≥ (vol(Γk−1 (x)))1/(β−2) · w, y ∈ Γk (x) | vol(Γk−1 (x))
X
vol(Γk−1 (x))
≤
py ·
vol(V )
y:py ≥(vol(Γk−1 (x)))1/(β−2) ·w
vol(Γk−1 (x))
1
·
≤ O n·
vol(Γk−1 (x)) · wβ−2
vol(V )
= O(1/wβ−2 ).
That is, the highest weight in Γk (x) is at most
w · (vol(Γk−1 (x)))1/(β−2)
with probability at least 1 − O(1/wβ−2 ). Denote this event by Ek . We have
E[vol(Γk (x)) | Ek , vol(Γk−1 (x))] ≤
X
y:py ≤(vol(Γk−1 (x)))1/(β−2) ·w
By Markov’s inequality, we have
p2y ·
vol(Γk−1 (x))
vol(V )
(3−β)/(β−2)
≤ O n · vol(Γk−1 (x))
·w
≤ O vol(Γk−1 (x))1/(β−2) · w3−β .
3−β
vol(Γk−1 (x))
·
vol(V )
Pr[vol(Γk (x)) ≥ vol(Γk−1 (x))1/(β−2) · w | Ek ] ≤ O(1/wβ−2 ).
Since Ek occurs with high probability, by union bound, we have
Pr[vol(Γk (x)) ≥ vol(Γk−1 (x))1/(β−2) · w] ≤ O(1/wβ−2 ).
On the other hand, we have
1/(β−2)
vol(Γk−1 (x))1/(β−2) · w ≤ bk−1
·w
2(k−1) (3−β)
= (ck−1 · w1/(β−2)
2k−1 (3−β)+1
= ck · w1/(β−2)
)1/(β−2) · w
2k (3−β))·(β−2+(3−β)(β−2)2k )
= ck · w(1/(β−2)
≤ bk .
Thus, we have Pr[vol(Γk (x)) > bk ] ≤ O(1/wβ−2 ).
Lower Bounding vol(Γk (x)) For any vertex y, if py · pz ≥ vol(V ) for some z ∈ Γk−1 (x), then y
must be connected to Γk−1 (x), otherwise the probability that y does not connect to Γk−1 (x) is at
most
P
py ·pz
Y
py · pz
− z∈Γ
k−1 (x) vol(V )
1−
≤e
vol(V )
z∈Γk−1 (x)
−py ·
=e
vol(Γk−1 (x))
vol(V )
.
−p ·
vol(Γk−1 (x))
That is, in either case, if y ∈
/ Nk−1 (x), then Pr[y ∈
/ Γk (x)] ≤ e y vol(V ) . When n is large
enough, we have bk−1 < ak , and Nk−1 (x) does not contain high weight vertex by the premises of
the lemma. Therefore, the probability that Γk (x) contains no high weight vertex is low:
Pr[∀y, s.t. py ≥ ak , y ∈
/ Γk (x)] ≤
Y
−py ·
e
vol(Γk−1 (x))
vol(V )
y:py ≥ak
P
vol(Γk−1 (x))
− y:py ≥a py ·
vol(V )
=e
k
−Ω(a2−β
·ak−1 )
k
≤e
= e−Ω(w
= e−Ω(w
1/(β−2)2k−1 (3−β)−1/(β−2)2k−2 (3−β) )
1/(β−2)2k−1 )
≤ e−w = 1/ log n.
In particular, it implies that the probability that vol(Γk (x)) < ak is at most 1/ log n.
Finally, by union bound, the probability that vol(Γk (x)) ∈ [ak , bk ] is at least 1 − O(1/wβ−2 ) as
the lemma states.
Lower Bounding dist(x, y)
By the above argument, we have
Pr[dist(x, y) ≤ 2d + 3]
≤ Pr[∃0 ≤ i, j ≤ d + 1, ∃u ∈ Γi (x), v ∈ Γj (y), u ∼ v]
≤
=
≤
d+1
X
Pr[∃u ∈ Γi (x), v ∈ Γj (y), u ∼ v]
i,j=0
Γi (x),Γj (y)
i,j=0
d+1
X
d+1
X
E
[Pr[∃u ∈ Γi (x), v ∈ Γj (y), u ∼ v | Γi (x), Γj (y)]]
(Pr[∃u ∈ Γi (x), v ∈ Γj (y), u ∼ v | Γi (x) ∈ [ai , bi ], Γj (y) ∈ [aj , bj ]]
i,j=0
+ Pr[Γi (x) ∈
/ [ai , bi ]] + Pr[Γj (y) ∈
/ [aj , bj ]])
d+1
X
bi · bj /vol(V ) + O(1/wβ−2 )
≤
i,j=0
≤ O(b2d+1 /n + d2 /wβ−2 )
= O(n−ε/(β−2) + log2 log log n/ logβ−2 log n) = o(1).
This proves the lemma.
| 8cs.DS
|
Modular Responsive Web Design using Element Queries
Lucas Wiener
Tomas Ekholm
Philipp Haller
EVRY
Stockholm, Sweden
KTH Royal Institute of
Technology
Stockholm, Sweden
KTH Royal Institute of
Technology
Stockholm, Sweden
arXiv:1511.01223v1 [cs.SE] 4 Nov 2015
[email protected]
[email protected]
[email protected]
ABSTRACT
1. INTRODUCTION
Responsive Web Design (RWD) enables web applications
to adapt to the characteristics of different devices such as
screen size which is important for mobile browsing. Today,
the only W3C standard to support this adaptability is CSS
media queries. However, using media queries it is impossible
to create applications in a modular way, because responsive
elements then always depend on the global context. Hence,
responsive elements can only be reused if the global context
is exactly the same, severely limiting their reusability. This
makes it extremely challenging to develop large responsive
applications, because the lack of true modularity makes certain requirement changes either impossible or expensive to
realize.
Responsive Web Design (RWD) is an approach to make an
application respond to the viewport size and device characteristics. This is currently achieved by using CSS media
queries that are designed to conditionally design content by
the media, such as using serif fonts when printed and sansserif when viewed on a screen [30].
In this paper we extend RWD to also include responsive
modules, i.e., modules that adapt their design based on
their local context independently of the global context. We
present the ELQ project which implements our approach.
ELQ is a novel implementation of so-called element queries
which generalize media queries. Importantly, our design
conforms to existing web specifications, enabling adoption
on a large scale. ELQ is designed to be heavily extensible using plugins. Experimental results show speed-ups of
the core algorithms of up to 37x compared to previous approaches.
In this paper we focus on the presentation layer of web applications. As it stands, using media queries to make the presentation layer responsive precludes modularity. The problem is that there is no way to make a module responsive
without making it context-aware, due to the fact that media
queries can only target the viewport; this means that responsive elements can only respond to changes of the (global)
viewport. Thus, a responsive module using media queries
is layout dependent and has both reduced functionality and
limited reusability [33]. As a result, media queries can only
be used for RWD of non-modular static applications. In a
world where no better solution than media queries exists
for RWD, changing the layout of a responsive application
becomes a cumbersome task.
Categories and Subject Descriptors
D.2.13 [Software Engineering]: Reusable Software—reusable
libraries; I.7.2 [Document and Text Processing]: Document Preparation—hypertext/hypermedia, markup languages
Keywords
Responsive web design, Element queries, CSS, Modularity,
Web
In order to reduce complexity and enable reusability, applications are typically composed of modules, i.e., interchangeable and independent parts that have a single and
well-defined responsibility [20]. In order for a module to be
reusable it must not assume in which context it is being
used.
1.1 The Problem Exemplified
Imagine an application that displays the current weather of
various cities as widgets, by using a weather widget module.
The module should be responsive so that more information,
such as a temperature graph over time, is displayed when
the widget is big. When the widget is small it should only
display the current temperature. Users should also be able
to add, remove and resize widgets.
Such an application cannot be built using media queries,
since the widgets can have varying sizes independent of the
viewport (e.g., the width of one widget is 30% while another
is 40%). To overcome this problem we must change the application, so that widgets always have the same sizes. This
implies that the size of the module and the media query
breakpoints are coupled/intertwined, i.e. they are proportional to each other. The problem now is that we have removed the reusability of the weather module, since it re-
quires the specific width that is correctly proportional to
the media query breakpoints.
compatible with Android version 4), Safari version 5, and
Opera version 12.
Imagine a company working on a big application that uses
media queries for responsiveness (i.e., each responsive module assumes to have a specific percentage of the viewport
size). The ability to change is desired by both developers
and stakeholders, but is limited by this responsive approach.
The requirement of changing a menu from being a horizontal
menu at the top to being a vertical menu on the side implies that all responsive modules break, since the assumed
proportionality of each module is changed. Even worse, if
the menu is also supposed to hide on user input, the responsiveness of the module breaks, since the layout changes
dynamically. The latter requirement is impossible to satisfy
in a modular way without element queries.
One could argue that a solution does not need to be executed
on the client side, but instead generate media queries on the
server side for all modules with respect to the current application layout. However, this approach is insufficient, since it
limits modules to applications with static layouts [33]. Also,
the generated media queries would not be able to respond
to the user changing properties of elements such as layout
and font size.
Additionally, it is popular to define breakpoints relative to
the font size so that conditional designs respect the size of
the content [10]. Media queries can only target the font size
of the document root, limiting their functionality drastically.
With element queries breakpoints may be defined relative to
the font size of the targeted element.
As we can see, even with the exemplified limited requirements there are still significant restrictions when using media queries for responsive modules.
1.2 Requirements
The desired behavior of a responsive module is having its
inner design respond to the size of its container instead of
the viewport. Only then is a responsive module independent
of its layout context. Realizing responsive modules requires
CSS rules that are conditional upon elements, instead of the
global viewport. We have identified the following requirements of a solution:
• It must provide the possibility for an element to automatically respond to changes of its parent’s properties.
• It must conform to the syntax of HTML, CSS, and
JavaScript to retain the compatibility of tools, libraries
and existing projects.
• It must have adequate performance for large applications that make heavy use of responsive modules.
• It must enable developers to write encapsulated style
rules, so that responsive modules may be arbitrarily
composed without any conflicting style rules.
1.3 Approach
In this paper we extend the concept of RWD to also include
responsive modules. The W3C has discussed such a feature under the name of element queries given its analogy to
media queries [32]. This paper presents a novel implementation of element queries in JavaScript named ELQ that
enables new possibilities of RWD. Our approach satisfies all
requirements given in Section 1.2. We have released ELQ
as an open-source library under the MIT license.1 The implementation supports all major browsers, including Internet Explorer version 8, Chrome version 42 (the last version
1
https://github.com/elqteam/elq
1.4 Contributions
This paper makes the following contributions:
• A new design for element queries that enables responsive modules while conforming to the syntax of HTML,
CSS, and JavaScript.
• Our approach is the first to enable nested elements
that are responsive in a modular way, i.e., modules
fully encapsulate any styling required for RWD. As a
side effect, responsive modules may also be arbitrarily
styled with CSS independent of their context.
• An extensible architecture that enables plugins to significantly extend the behavior of ELQ, our library implementation. This makes it possible to create plugins
in order to enable new features and to ease integration
of ELQ into existing projects.
• A new implementation that offers substantially higher
performance than previous approaches. The implementation batch-processes DOM operations in order
to avoid layout thrashing (i.e., forcing the layout engine to perform multiple independent layouts).
• A run-time cycle detection system that detects and
breaks cycles stemming from cyclic rules due to unrestricted usage of element queries [33].
The rest of the paper is organized as follows. Section 2 introduces ELQ and its API from a user’s perspective. In
Section 3 we introduce ELQ’s plugin architecture. Section 4
provides an overview of the main components of ELQ’s implementation. In Section 5 we evaluate the performance of
ELQ and report on case studies. Section 6 discusses limitations of ELQ and related libraries, as well as the current
state of standardization of element queries. Section 7 relates
ELQ to prior work, and Section 8 concludes.
2. OVERVIEW OF ELQ
Media queries and element queries are similar in the sense
that they both enable developers to define conditional designs that are applied by specified criteria. The main difference is the type of criteria that can be used. With media
queries critera of the device, document, and media are used,
while element criteria are used with element queries. It can
somewhat simplified be described as that media queries target the document root and up such as viewport, browser,
device, and input mechanisms. Element queries target the
document root and down, i.e., elements of the document.
ELQ is designed to be plugin-based for increased flexibility
and extensibility. By providing a good library foundation
and plugins it is up to developers to choose the right plugins
for each project. In addition, by letting the plugins satisfy
the requirements it is easy to extend the library with new
plugins when new requirements arise.
An element breakpoint is defined as a point of an element
property range which can be used to define conditional behavior, similar to breakpoints of media queries. For example, an element breakpoint of 500 pixels in width enables
conditional styling depending on if the element is narrower
or wider than 500 pixels. An element may have multiple
breakpoints. An element breakpoint state is defined as the
state of the element breakpoint relative to the current element property value. For example, if an element that is 300
pixels wide has two width breakpoints of 200 and 400 pixels
the element breakpoint states are “wider than 200 pixels”
and “narrower than 400 pixels”.
When the breakpoint states of an element changes, ELQ performs cycle detection in order to detect and handle possible
style cycles. If a cycle is detected, the new element breakpoint states are not applied in order to avoid an infinite loop
of layouts. The cycle detection system is implemented as an
conservative algorithm, and may in some cases detect false
positives.
2.1 The API
In this section, we use the elq-breakpoints API (that is
bundled with ELQ as default) that let use define element
breakpoints. The main idea is to define element breakpoints
of interest so that children can be conditionally styled in
CSS by targeting the different element breakpoint states.
As CSS3 does not support custom at-rules/selectors [31],
responsive elements are annotated in HTML by element attributes. ELQ then observes the annotated elements in order to automatically update breakpoint state classes. Although not written in the examples, the API also supports
attributes defined with the data- prefix to conform to the
HTML standard [29].
The following example shows the HTML of an element that
has two annotated width breakpoints at 300 and 500 pixels:
<d i v c l a s s =”f o o ” e l q e l q −b r e a k p o i n t s
e l q −b r e a k p o i n t s−w i d t h s =”300 500”>
<p>When i n doubt , mumble. </p>
</d i v>
When ELQ has processed the element it will have two classes
that reflect each breakpoint state. For instance, if the element is 400 pixels wide, the element has the two classes elqmin-width-300px and elq-max-width-500px. Similarly, if
the element is 200 pixels wide the classes are instead elqmax-width-300px and elq-max-width-500px. So for each
breakpoint only the min/max part changes. It may seem
alien that the classes describe that the width of the element
is both maximum 300 and 500 pixels. This is because we
have taken a user-centric approach so that the CSS usage of
the classes is similar to the API of media queries. However,
developers are free to change this API by creating plugins.
Now that we have defined the breakpoints of the element,
we can conditionally style it in CSS by using the classes as
shown in listing 1.
. f o o . e l q −max−width −300px {
background−c o l o r : b l u e ;
}
. f o o . e l q −min−width −300px . e l q −max−width −500px {
background−c o l o r : g r e e n ;
}
. f o o . e l q −min−width −500px p {
c o l o r : white ;
}
Listing 1: Example usage of the breakpoint state
classes in CSS.
In order for the conditional styles to be applied, the elements that have breakpoints must be activated by the ELQ
JavaScript runtime. ELQ can either be required as a module
(by the CommonJS or AMD syntax), or it can be included
in a HTML script tag which then will expose a global constructor Elq. The following is an example of how to create
an ELQ instance and activating elements:
// C r e a t e an i n s t a n c e
v a r e l q = Elq ( ) ;
// P l u g i n s may be r e g i s t e r e d
e l q . u s e ( myPlugin ) ;
e l q . u s e ( myOtherPlugin ) ;
// A c t i v a t e e l e m e n t s .
v a r e l e m e n t s = document . q u e r y S e l e c t o r A l l ( ”[ e l q ] ” ) ;
e l q . a c t i v a t e ( elements ) ;
In this example we create an ELQ instance and register two
plugins with it. Then we query the document for all elements with an elq attribute (as annotated in the previous
example) and then pass them as an argument to the activation method of ELQ. It should be noted that it is up to
developers how to activate elements; annotating elements
with elq is used for simplicity in the example. The only
requirement is that conditionally-styled elements are processed by the activate method at some point. This can,
for example, also be achieved with a plugin that listens to
DOM mutations to perform the activation automatically, or
a plugin that parses CSS and activates all elements that have
conditional styles defined.
2.1.1 Nested modules
The elq-breakpoints API is sufficient for applications that
do not need nested breakpoint elements, and similar features
are provided by related libraries such as [22, 34]. However,
using such an API in responsive modules still limits composability, since modules then may not exist in an outer
responsive context.
The reason this API is not sufficient for nested modules is
that there is no way to limit the CSS matching search of
the selectors. The last style rule of the example given in
listing 1 specifies that all paragraph elements should have
white text if any ancestor breakpoints element is wider than
500 pixels. Since the ancestor selector may match elements
outside of the module, such selectors are dangerous to use
in the context of responsive modules. The problem may be
somewhat reduced by more specific selectors and such, but
it cannot be fully solved for arbitrary styling [33].
To solve this problem, we provide a plugin that let us define
elements to “mirror” the breakpoints classes of the nearest
ancestor breakpoints element (the target of the mirror element). This means that the mirror element always reflects
the element breakpoint states of the target. Then, the conditional style of the mirror element may be written as a
combinatory selector that is relative to the nearest ancestor
breakpoints element. The following is an example usage of
the mirror plugin to enable nested modules:
<d i v c l a s s =”f o o ” e l q e l q −b r e a k p o i n t s
e l q −b r e a k p o i n t s−w i d t h s =”300 500”>
<d i v c l a s s =”f o o ” e l q e l q −b r e a k p o i n t s
e l q −b r e a k p o i n t s−w i d t h s =”300 500”>
<p e l q e l q −m i r r o r > . . . < / p>
</d i v>
<p e l q e l q −m i r r o r > . . . < / p>
</d i v>
In this example, the paragraph elements always have the
same element breakpoint classes as the parent elq-breakpoints
elements. This enables us to write CSS that does not traverse the ancestor tree:
. foo {
/∗ So t h a t t h e n e s t e d modules
have d i f f e r e n t s i z e ∗/
wi d t h : 50%;
}
. f o o p . e l q −min−width −500px {
c o l o r : white ;
}
In the examples we have given so far we have annotated element breakpoints manually; however, this does not properly
show the power and flexibility of ELQ’s API. Therefore, the
next section presents an API that combines JavaScript and
generated CSS in order to create a flexible grid API.
2.1.2 A grid API
In this section we present a plugin that defines an API
that enables developers to use responsive grids consisting of
twelve columns and utility classes very similar to the ones
defined by the CSS Bootstrap framework [19]. The goal of
the API is to provide an abstraction of element queries, so
that developers may focus on responsivity using classes instead of the syntax presented in previous sections.
The following is an example grid:
<d i v c l a s s =”c o n t a i n e r ”>
<d i v c l a s s =”row”>
<d i v c l a s s =”c o l −500−4 c o l −700−6”>
...
</d i v>
<d i v c l a s s =”c o l −500−4” c o l −700−6>
...
</d i v>
<d i v c l a s s =”c o l −500−4 hidden −700−up”>
...
</d i v>
</d i v>
</d i v>
The example grid is defined to be single columned when
the width of the grid is below 500 pixels, triple columned
when the width is between 500 and 700 pixels, and double
columned for when the width is above 700 pixels. The last
column is hidden when the width is above 700 pixels.
The column classes define the behaviour of the grid, and
have the syntax col-[breakpoint]-[size]. The [breakpoint] part of a column class is relative to the parent row
and can be any positive number including an optional unit.
Currently, the supported units are px, em, rem. If the unit
is omitted, px is assumed. Grids may also be nested.
The plugin traverses the grid structure to initialize all columns
and possible nested grids. It also generates and applies the
CSS needed for each column breakpoint automatically to
the document. This enables developers to easily create responsive grids in nestable modules.
3. EXTENSIONS VIA PLUGINS
For example, if annotating HTML is undesired it is possible
to create a plugin that instead generates element breakpoints
by parsing CSS. Likewise, if adding breakpoint state classes
to elements is undesired it is possible to create a plugin that
does something else when an element breakpoint state has
changed.
A plugin is defined by a plugin definition object and has the
structure shown in listing 2.
var m y P l u g i n D e f i ni ti o n = {
getName : f u n c t i o n ( ) {
r e t u r n ”my−p l u g i n ”;
},
getVersion : function () {
r e t u r n ”0 . 0 . 0 ”;
},
isCompatible : f u n c t i o n ( e l q ) {
return true ;
},
make : f u n c t i o n ( e l q , o p t i o n s ) {
return {
// Implement p l u g i n i n s t a n c e methods .
...
};
}
};
Listing 2: The structure of plugin definition objects.
All of the methods are invoked when registered to an ELQ
instance. The getName and getVersion methods tells the
name and version of the plugin. The isCompatible tells if
the plugin is compatible with the ELQ instance that it is
registered to. In the make method the plugin may initialize
itself to the ELQ instance and return an object that defines
the plugin API accessible by ELQ and other plugins.
ELQ invokes certain methods of the plugin API, if implemented, to let plugins decide the behavior of the system.
Those methods are the following:
• activate(element) Called when an element is requested
to be activated, in order for plugins to initialize listeners and element properties.
• getElements(element) Called in order to let plugins
reveal extra elements to be activated in addition to the
given element.
• getBreakpoints(element) Called to retrieve the current breakpoints of an element.
• applyBreakpointStates(element, breakpointStates)
Called to apply the given element breakpoint states of
an element.
Plugins may also use an extended API of ELQ that offers
access to subsystems such as the plugin handler, cycle detector, batch processor, etc. The extended API is exposed
to plugins as an argument to the make method of the plugin definition object. In addition, plugins may set behavior
properties of an element by the element.elq property. It is
also possible for plugins to define own behavior properties
for inter-plugin collaboration, or for storing plugin-specifc
element state. Examples of behavior properties of the ELQ
core are:
In addition, plugins may also listen to the following ELQ
events:
• resizeDetection Indicates if resize detection should
be performed.
• resize(element) Emitted when an ELQ element has
changed size.
• cycleDetection Indicates if cycle detection should be
performed.
• breakpointStatesChanged(element, breakpointStates)
Emitted when an element has changed element breakpoint states (e.g., when the width of an element changed
from being narrower to being wider than a breakpoint).
• updateBreakpoints Indicates if the element should be
passed through the update flow.
• applyBreakpointStates Indicates that plugins may
apply breakpoint states of the element (for some elements it is only necessary to emit element breakpoint
state changes, without applying them to the actual element [33]).
There are two main flows of the ELQ system; activating an
element and updating an element. When ELQ is requested
to activate an element, the following flow occurs:
1. Initialize the element by installing properties and a
system that handles listeners.
2. Call the getElements method of all plugins to retrieve
any additional elements to activate. Perform an activation flow for all additonal elements.
3. Call the activate method of all plugins, so that pluginspecific initialization may occur.
4. If any plugin has requested ELQ to detect resize events
of the element, install a resize detector to the element.
5. Pass the element through the update flow.
3.1 Example Plugin Implementation
The elq-breakpoints API that enables developers to annotate breakpoints in HTML, as described in Section 2.1,
is implemented as two plugins. This shows that even the
core functionality of ELQ is implemented in terms of plugins. The first plugin parses the breakpoints of the element
attributes. The second plugin applies the breakpoint states
as classes.
The following is a simplified implementation of the make
method (see listing 2) of the parsing plugin:
f unc tion a c t i v a t e ( element) {
i f ( ! e l e m e n t . h a s A t t r i b u t e ( ” e l q −b r e a k p o i n t s ”) ) {
return ;
}
element .
element .
element .
element .
The update flow is as follows:
1. Call the getBreakpoints method of all plugins to retrieve the breakpoints of the element.
2. Calculate the breakpoint states of the element.
3. If any state has changed since the previous update:
(a) Perform cycle detection. If a cycle is detected,
then abort the flow and emit a warning.
(b) Call the applyBreakpointStates method of all
plugins in order for plugins to apply the new element breakpoint states.
(c) Emit an breakpointStatesChanged event.
Of course, there are options to disable some of the steps
such as cycle detection and applying breakpoint states. In
additon to being triggered by the activation flow and plugins,
the update flow is also triggered by element resize events.
elq
elq
elq
elq
.
.
.
.
resizeDetection
updateBreakpoints
applyBreakpointStates
cycleDetection
=
=
=
=
true
true
true
true
;
;
;
;
}
f unc tion getBreakpoints ( element) {
// P a r s e t h e ” e l q −b r e a k p o i n t s −∗” a t t r i b u t e s
// and r e t r i e v e t h e i r b r e a k p o i n t s .
return . . . ;
}
// Return t h e p l u g i n API
return {
activate : activate ,
getBreakpoints : getBreakpoints
};
In the activate method the plugin registers that resize detection is needed for the element and that it should be passed
through the update flow. It also enables the application of
breakpoint states and run-time cycle detection. Although
not shown in the simplifed implementation, applyBreakpointStates and cycleDetection are in some cases disabled.
The plugin that applies the element breakpoint states simply implements the applyBreakpointStates method to alter the className property of the element using the given
element breakpoint states.
4. IMPLEMENTATION
4.1 Batch Processing
Batch processing is the foundation of the performance gains
of our approach, and is therefore used by several subsystems. ELQ uses a leveled batch processor, which is implemented as a stand-alone project.2 It serves two purposes:
to process batches in different levels to avoid layout thrashing, and to automatically process batches asynchronously
to enable multiple synchronous calls being grouped into a
pending batch.
Being able to process a batch in levels is important when
different types of operations, that are to be processed in a
specific order (usually to avoid layout thrashing), needs to
be grouped together in a batch. For example, a function
that doubles an element’s width and reads the new calculated height benefits by being batch processed in three levels: reading the width, mutating the width, and reading the
height. The following is an example implementation of such
function that uses the leveled batch processor:
var b a t c h P r o c e s s o r = . . . ;
f u n c t i o n doubleWidth ( e l e m e n t , c a l l b a c k ) {
// F i r s t l e v e l : r e a d i n g t h e wi d t h .
v a r wi d t h = e l e m e n t . o f f s e t W i d t h ;
v a r newWidth = ( wi d t h ∗ 2 ) + ”px ”;
// Second l e v e l : m u t a t i n g t h e wi d t h .
// Th i s i s e x e c u t e d i n l e v e l 0 o f t h e b a t c h .
b a t c h P r o c e s s o r . add ( 0 , f u n c t i o n mutateWidth ( ) {
e l e m e n t . s t y l e . wi d t h = newWidth ;
});
// Th i r d l e v e l : r e a d i n g t h e h e i g h t .
// Th i s i s e x e c u t e d i n l e v e l 1 o f t h e batch ,
// a f t e r l e v e l 0 . Changing t h e l e v e l number
// from ”1 ” t o ”0 ” r e s u l t s i n l a y o u t t h r a s h i n g .
b a t c h P r o c e s s o r . add ( 1 , f u n c t i o n r e a d H e i g h t ( ) {
var h e i g h t = element . o f f s e t H e i g h t ;
callback ( height ) ;
});
}
It should be noted that the first level is executed asynchronously by the function, and not handled by the actual
batch processor. Since each batch is delayed to execute asynchronously, all synchronous calls of the method is grouped
into a pending batch. If the batch would not automatically
be delayed, layout thrashing would occur when the method
is called multiple times. This results in a 45-fold speedup,
when applied to 1000 elements, of the function compared to
not processing the batch in levels. It also results in a simple
API that allows multiple synchronous calls without causing
layout thrashing, like so:
The activate method of ELQ is implemented similarly so
it may also be called multiple times synchronously, without
performance penalties, like the following example:
var elements = [ . . . ] ;
elements . forEach ( e l q . a c t i v a t e ) ;
4.2 Element Resize Detection
Unfortunately, there is no standardized resize event for arbitrary elements [28]. A naive approach to detecting element
resize events is to have a script continously check elements
if they have resized given some interval (polling). This approach is appealing because it does not mutate the DOM,
supports arbitrary elements, and it provides excellent compatibility. However, in order to prevent the responsive elements lagging behind the size changes of the user interface,
polling needs to be performed quite frequently. The problem is that each poll forces the layout queue to be flushed
since the computed style of elements needs to be retrieved
in order to know if elements have resized or not [33]. Since
the polling is performed all the time the overall page performance is decreased even if the page is idle, which is undesired
especially for mobile devices running on battery.
It is desired to instead have an event-based approach that
only performs additional computations when an actual element resize has happened. This is achieved by the resize
detection subsystem of ELQ by using two independent injecting approaches, both originally presented by [6]. These
appraoches are limited to non-void elements, i.e., elements
that may have content. It is a reasonable limitation since
void elements can easily be wrapped with non-void elements
without affecting the page visually. Like the batch processor, the resize detection subsystem is also implemented as a
stand-alone project.3
Object-based resize detection. Only documents emit resize events in modern browsers and therefore such events
can only be observed for frame elements (since a frame element has its own document). This approach injects object
elements into the target element, which can be listened to
resize events since object elements are frames. The object
is styled so that it always matches the size of the target element and so that it does not affect the page visually. This
approach has good browser compatability and excellent resize detection performance, but imposes severe performance
impacts during injection since object elements use a significant amount of memory as shown in Section 5.1.
Scroll-based resize detection. This approach injects an
var elements = [ . . . ] ;
elements . forEach ( f u n c t i on ( element ) {
doubleWidth ( e l e m e n t , f u n c t i o n ( h e i g h t ) {
...
});
});
element that contains multiple overflowing elements that listen to scroll events. The overflowing elements are styled so
that scroll events are emitted when the target element is
resized. For detecting when the target element shrinks, two
elements are needed; one for handling the scrollbars and
one for causing them to scroll. Similarly, for detecting when
the target element expands, two elements are needed in the
same way. As this solution only injects div elements, it offers
greater opportunities for optimizations. The main algorithm
2
3
https://github.com/wnr/batch-processor
https://github.com/wnr/element-resize-detector
that is performed when an element e is to be observed for
resize events is the following:
1. Get the computed style of e.
2. If the element is positioned (i.e., position is not static)
the next step is 4.
3. Set the position of e to be relative. Here additional
checks can be performed to warn the developer about
unwanted side effects of doing this.
4. Create the four elements needed (two for detecting
when e shrinks, and two for detecting when e expands)
and attach event handlers for the scroll event of the
elements. When the elements have been styled and
configured properly, they are added as children to an
additional container element that is injected into e.
penalties of the library. Also, it is hard to compare performance results of related libraries since the functionality is
different. Fortunately, element resize detection is the common denominator of all automatic libraries and the results
of this system can be compared faithfully. Measurements
and graphs show evaluations performed in Chrome version
42 unless stated otherwise.
The object-based approach (as presented by [6]) performs
well when detecting resize events, which it does with a delay
of 30 ms for 100 elements. However, the injection performance is not great as presented in figure 1. As shown by
the graph, the injection can be performed with adequate
performance as long as the number of elements is low. The
approach does not scale well as the number of elements increases. This is probably due to the fact that the heap memory usage grows roughly by 0.55 MB per element.
5. The current size of e is stored and the scrollbars of the
injected elements are positioned correctly.
Injection time [s]
6
6. The algorithm waits for the scroll event handlers to
be called asynchronously by the layout engine (they are
called since the previous step repositioned the scrollbars). When the handlers have been called, the injection is finished and observers can be notified on resize
events of e when scroll events occur.
4
2
0
0
Layout thrashing can be avoided by using the leveled batch
processor described in Section 4.1, which results in a significant performance improvement as shown in Section 5.1. The
algorithm steps are batch processed in the following levels:
1. The read level: Step 1 is performed to obtain all necessary information about e. The information is stored
in a shared state so that all other steps can obtain the
information without reading the DOM.
2. The mutation level: Steps 2, 3 and 4 are performed,
which mutate the DOM. All mutations performed in
this level can be queued by layout engines.
3. The forced layout level: Step 5 is performed, which
forces the some layout engines to perform a layout.
Since repositioning a scrollbar in some layout engines forces
a layout, such operations need to be performed after that all
other queueable operations have been executed. Therefore,
step 5 is performed in level 3 as the last step. Even though
some layout engines are unable to queue the repositing of
scrollbars, it is still beneficial to batch process the algorithm
since only pure layouts need to be performed (instead of
having to recompute styles and synchronize the DOM and
render trees before each layout). As step 6 is performed by
the layout engine asynchronously and does not interact with
the DOM, it does not need to be batch processed.
5. EMPIRICAL EVALUATION
5.1 Performance
Only the performance of the element resize detection system has been performed. This due to the fact that detecting element resize events entails the significant performance
200
400
600
Number of elements
Object-based approach
Figure 1: The injection performance of the objectbased approach.
As the scroll-based approach (as presented by [6]) does not
inject object elements the memory footprint is reduced significantly, which improves the injection performance. The
amount of used memory is too low for reliable measurements.
See figure 2 for graphs that show how the ELQ scroll-based
approach performs compared to the other two approaches.
As evident in the figure, the optimized ELQ approach has
significantly reduced injection times. It achieves a 37-fold
speedup compared to the object-based approach and a 17fold speedup compared to the scroll-based approach when
preparing 700 elements for resize detection. It also performs
well when detecting resize events, which it does with a delay
of 25 ms for 100 elements.
Browsers
Chrome v. 42
Firefox v. 40
Safari v. 9
Internet Explorer v. 11
iOS Safari v. 9
Android v. 5 Chrome v. 39
Injection
scroll
object
30 ms
550 ms
150 ms
1000 ms
100 ms
400 ms
350 ms
6700 ms
350 ms
1600 ms
40 ms
1000 ms
Resize detection
scroll
object
25 ms
20 ms
70 ms
30 ms
30 ms
20 ms
100 ms
80 ms
150 ms
60 ms
20 ms
10 ms
Table 1: Performance of ELQ’s two resize detection
strategies, operating on 100 elements.
ELQ uses the object-based approach as a fallback for legacy
browsers. Therefore the performance of the ELQ resize detection system is at minimum as performant as related approaches. See table 1 for the performance of ELQ’s two resize detection strategies in different browsers. As shown in
the table, the scroll-based approach is suitable for reduced
page load, but may in some cases be preferred for better
resize detection performance.
5.2 Case studies
In this section we aim to provide answers to the following
questions:
• How can ELQ be used to modularize existing responsive code bases?
• How much effort is this modularization?
In order to answer the questions, we have adapted the popular Bootstrap framework4 to use element queries instead of
media queries. According to its website, “Bootstrap is the
most popular HTML, CSS, and JS framework for developing
responsive, mobile first projects on the web.” [19]
In order to modularize Bootstrap, we redefine the behavior
of its responsive elements so that they no longer respond
to the viewport but to enclosing container elements. The
following observation guides our modularization: all responsive elements should respond to their closest enclosing container or container-fluid element. Both classes are used
in Bootstrap to define new parts of a page (e.g., a grid is required to have a container ancestor). We also enable them to
be nestable, which is important to satisfy the requirement
of composable modules. The breakpoints of the container
elements are defined using the elq-breakpoints API (see
Section 2.1). Since the Bootstrap API uses a predefined set
of breakpoints, they are all added to the container elements
with JavaScript.
According to this design, we convert all responsive elements
of Bootstrap to elq-mirror elements, since they need to mirror the breakpoints of the nearest ancestor elq-breakpoints
element. Since container elements may be nested, they have
both the elq-breakpoints and elq-mirror behavior.
The breakpoints of Bootstrap are defined as the following
constants:5
@ sc r e e n−sm−min : 480 px ;
@ sc r e e n−md−min : 992 px ;
@ sc r e e n−l g −min : 1200 px ;
The following example shows how Bootstrap’s style definitions are changed from using media queries to using ELQ’s
element queries:
/∗ F i l e ” l e s s / g r i d . l e s s ” o f B o o t s t r a p . ∗/
// O r i g i n a l B o o t s t r a p u s i n g media q u e r i e s .
. container {
@media ( min−wi d t h : @ sc r e e n−sm−min ) {
wi d t h : @ c o n t a i n e r −sm ;
}
...
}
. container {
&. e l q −min−width−@{ s c r e e n −sm−min} {
wi d t h : @ c o n t a i n e r −sm ;
}
...
}
By using the power of preprocessors, ELQ element queries
become as pleasant to work with as media queries. In fact,
only about 0.6% of the style code (LESS syntax) need to be
altered. Most changes are similar to the one shown above,
which replaces the media query syntax with the ELQ element queries syntax. This is especially advantageous when
keeping a forked project up to date with the original project,
as fewer diverged lines implies a lowered risk of merge conflicts.
In summary we have shown that it is easy to adapt existing responsive code to use ELQ’s element queries instead of
media queries. With only a small number of changes, the
widely used Bootstrap framework can be modularized.
Industrial use of ELQ. In addition to the Bootstrap case
study, we have been gathering experience with the application of ELQ in large financial applications at EVRY. Our
practical experience shows that complex applications require
a variety of features to be supported by element queries.
Such features can be provided effectively by ELQ plugins.
6. DISCUSSION
6.1 Limitations
Inherent to all current implementations of element queries
is that the conditional style is applied “one layout behind”.
Since a layout pass needs to have been performed in order
for an element to change size, the conditional styles defined
by the element queries cannot be applied until next layout.
Therefore, the element will display invalid design until another layout has been performed. The flash of invalid design
is usually so short that users do not notice it, but in some
cases developers need to work around this issue to avoid
more apparent results.
Another caveat is presented by the element resize detection
approaches, as they mutate the DOM. Developers need to
be aware of this as CSS selectors and JavaScript may also
match the injected elements. This is easily avoided by good
practices.
It should be noted that all limitations described only affects the elements that uses the element queries functionality. ELQ does not impose potential problems to other parts
of the DOM other than where applied explicitly.
Currently ELQ only supports breakpoints for the width and
height element properties, as it has been identified as the
general use case [33]. In the future, we aim to support plugins to define custom breakpoint properties.
// ELQ B o o t s t r a p u s i n g e l e m e n t q u e r i e s .
4
6.2 Standardization
5
It is stated on the W3C’s www-style mailing list [32] by
Zbarsky of Mozilla, Atkins of Google and Sprehn of Google
This evaluation uses Bootstrap version 3.3.2.
The Bootstrap CSS is generated using the LESS preprocessor [25] whose syntax we use.
Implementation
MagicHTML [9]
EQCSS [13]
Element Media Queries [21]
Localised CSS [3]
Grid Style Sheets 2.0 [8]
Class Query [27]
breakpoints.js [26]
MediaClass [17]
ElementQuery [16]
Responsive Elements [15]
SickleS [18]
Responsive Elements [34]
breaks2000 [12]
eq.js [22]
Element Queries [7]
CSS Element Queries [24]
Selector queries and responsive containers [14]
ELQ
Syntax
Custom
Custom
Custom
Custom
Custom
-
CSS
CSS
CSS
CSS
CSS
Resize detection
Viewport only
Non-void elements
Arbitrary elements
Arbitrary elements
Viewport only
Viewport only
Viewport only
Viewport only
Viewport only
Viewport only
Viewport only
Viewport only
Non-void elements
Non-void elements
Arbitrary elements
Page dynamism
Static
Dynamic
Dynamic
Dynamic
Dynamic
Static
Dynamic
Dynamic
Dynamic
Dynamic
Dynamic
Dynamic
Dynamic
Dynamic
Dynamic
Dynamic
Dynamic
Composability
Full support
Partial support
-
Cycle detection
-
Non-void elements
Dynamic
Full support
Yes
Table 2: Classification of related approaches to modular RWD.
that element queries are infeasible to implement without restricting them. By limiting element queries to specially separated container elements that can only be queried by child
elements, many of the problems are resolved [2, 1]. Therefore, the Responsive Issues Community Group (RICG) is
currently investigating the possibility of standardizing container queries.
Unfortunately, even such limited container queries requires
significant effort to implement due to the complex changes
to browsers required [1]. Atkins argues that a full implementation that avoids the double layout issue is unlikely to
be implemented, and therefore it might be wiser to pursue
sub-standards that aids third-party solutions instead.
In the future, we hope that ELQ may use the aiding substandards pushed by RICG, to achieve greater flexibility and
performance. A standardized resize event would enable us
to avoid injecting elements, and to reduce the code base
of ELQ significantly. Support for custom at-rules/selectors
would also enable us to define a more natural API in CSS.
Finally, being able to tell elements to ignore children while
computing their size would decrease the need for cycle detection.
7.
RELATED WORK
Table 2 attempts to classify all existing approaches, of modular RWD, known to us. We discuss these approaches according to two different aspects: (a) syntax extensions and
(b) resize detection.
Syntax extensions. The libraries [9, 13, 21, 3, 8] have in
common that they require developers to write custom CSS,
unlike ELQ. Since they do not conform to the CSS standard,
new features are supported through custom CSS parsed using JavaScript. As shown by [13, 8] quite advanced features
can be implemented this way. Additionally, adding new CSS
features implies that it is possible to implement a solution
to element queries that does not require any changes to the
HTML, which may be preferable since all styling then can
be written in CSS. However, there are numerous drawbacks
with libraries that require custom CSS. Extending the CSS
syntax violates the requirement of compatability and also
introduces a compilation step which decreases the performance [33].
Resize detection. The libraries [13, 26, 17, 16, 15, 18, 34,
12, 22] simply observe the viewport resize event, which may
be enough for static pages, but not enough to satisfy the
requirements of reusable responsive modules [33]. Approach
[27] does not detect resize events at all. Like ELQ, [3, 14,
21, 8, 7, 24] observe elements for resize events. The libraries
[3, 14] use polling while ELQ and [21, 8, 7, 24] use different
injection approaches, as described in Section 4.2. As shown
in Section 5.1, the injection approaches used by related libraries have significanly less performance than the element
resizing detection system used in ELQ.
Constraint-based CSS. CCSS [4] proposes a more general
and flexible alternative to CSS. As the name suggests, the
idea of CCSS is to layout documents based on constraints.
According to its authors, the constraint-based approach provides extended features and reduced complexity compared
to CSS. To solve the constraints CCSS uses the Cassowary
constraint solving algorithm [5].
The Grid Style Sheets library [8] builds upon the ideas of
CCSS and uses a JavaScript port [23] of Cassowary to solve
the constraints at runtime. While not directly offering element queries, the library enables the possibility to conditionally style elements by element criteria and thus makes it a
good candidate to solve the problem of responsive modules.
However, the library has two major issues: performance and
browser compatibility [11]. One approach to resolve both issues is to precompute the layout in a compilation step at the
server. However, precompiling styles implies static layouts.
The authors discuss other approaches [11] that would increase the performance while limiting the dynamism of page
layout. In contrast, ELQ only considers element queries, but
without these limitations and with higher performance.
8. CONCLUSION
Responsive Web Design (RWD) enables web applications
to adapt to the characteristics of different devices, which is
achieved using CSS media queries. However, using media
queries it is impossible to create responsive applications in
a modular way, because responsive elements then always
depend on the global context.
This paper extends RWD to also include responsive modules
through element queries. We present ELQ, an open-source
implementation of our approach, that conforms to the current standards of HTML, CSS and JavaScript. It enables
developers to create responsive modules that are independent of their context, and a way to encapsulate their conditional style rules. The element resize detection of ELQ, used
to automatically evaluate element queries on changes of responsive elements, performs up to 37x better than previous
algorithms.
Using a case study based on the popular Bootstrap framework we show that large code bases using media queries can
be converted to using ELQ’s element queries with little effort. Changing only about 0.6% of the LOC of style related
code was sufficient to enable the use of Bootstrap in responsive modules. We also report on first commercial usage of
ELQ.
We believe ELQ is an important contribution to realizing a
modular form of element queries, in particular since standardization bodies like the RICG do not intend to standardize a complete solution. In the future we intend to improve ELQ by using forthcoming standards developed by the
RICG to avoid some current limitations.
9. ACKNOWLEDGMENTS
The authors would like to thank EVRY for sponsoring the
ELQ project including the supporting projects for element
resize detection and batch processing.
10. REFERENCES
[1] CSS containment draft. Retrieved April 29, 2015 from
https://github.com/ResponsiveImagesCG/cq-usecases/issues/7.
[2] RICG IRC log. Retrieved April 29, 2015 from
http://ircbot.responsiveimages.org/bot/log/respimg/2015-03[3] C. Ashton. Localised CSS. Retrieved April 29, 2015
from
https://github.com/ChrisBAshton/localised-css.
[4] G. J. Badros, A. Borning, K. Marriott, and
P. Stuckey. Constraint cascading style sheets for the
web. In Proceedings of the 12th annual ACM
symposium on User interface software and technology,
pages 73–82. ACM, 1999.
[5] G. J. Badros, A. Borning, and P. J. Stuckey. The
cassowary linear arithmetic constraint solving
algorithm. ACM Trans. Comput.-Hum. Interact,
8(4):267–306, 2001.
[6] D. Buchner. Backalleycoder. Retrieved March 23, 2015
from http://www.backalleycoder.com/.
[7] D. Buchner. Element Queries. Retrieved April 29,
2015 from
https://github.com/csuwildcat/element-queries.
[8] e. a. Dan Tocchini. Grid Style Sheets 2.0. Retrieved
April 29, 2015 from http://gridstylesheets.org/.
[9] G. Felipe. MagicHTML. Retrieved April 29, 2015 from
https://github.com/gabriel-felipe/MagicHTML.
[10] L. Gardner. The EMs have it: Proportional media
queries FTW! Retrieved April 28, 2015 from
http://blog.cloudfour.com/the-ems-have-it-proportional-medi
[11] Grid Style Sheets. Element queries with
precompilation. Retrieved June 8, 2015 from
https://github.com/gss/engine/issues/178.
[12] D. Hägglund. breaks2000. Retrieved April 29, 2015
from
https://github.com/judas-christ/breaks2000.
[13] T. Hodgins and M. Euzière. EQCSS. Retrieved April
29, 2015 from http://elementqueries.com/.
[14] A. Hume. Selector queries and responsive containers.
Retrieved April 29, 2015 from
https://github.com/ahume/selector-queries/.
[15] K. Hunaid. Responsive Elements. Retrieved April 29,
2015 from
https://github.com/kumailht/responsive-elements.
[16] T. Matanich. ElementQuery. Retrieved April 29, 2015
from
https://github.com/tysonmatanich/elementQuery.
[17] J. Neal. MediaClass. Retrieved April 29, 2015 from
https://github.com/jonathantneal/MediaClass.
[18] T. Nguyen. SickleS. Retrieved April 29, 2015 from
http://singggum3b.github.io/SickleS/.
[19] M. Otto and J. Thornton. Bootstrap. Retrieved
October 15, 2015 from http://getbootstrap.com/.
[20] D. L. Parnas. On the criteria to be used in
decomposing systems into modules. Communications
of the ACM, 15(12):1053–1058, 1972.
[21] F. Remy. Element Media Queries. Retrieved April 29,
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
2015 from
https://github.com/FremyCompany/prollyfill-min-width/.
S. Richard. eq.js. Retrieved April 29, 2015 from
https://github.com/Snugug/eq.js.
A. Russell. Cassowary/JS. Retrieved April 28, 2015
from
https://github.com/slightlyoff/cassowary.js.
M. J. Schmidt. CSS Element Queries. Retrieved April
29, 2015 from
https://github.com/marcj/css-element-queries.
A. Sellier. LESS. Retrieved October 10, 2015 from
http://lesscss.org/.
J. Stoutenburg. breakpoints.js. Retrieved April 29,
2015 from
https://github.com/reusables/breakpoints.js.
M. Stow. Class Query. Retrieved April 29, 2015 from
https://github.com/stowball/Class-Query.
W3C. Document object model events. Retrieved
March 14, 2015 from
http://www.w3.org/TR/DOM-Level-2/events.html.
W3C. HTML 5.1. Retrieved October 10, 2015 from
http://www.w3.org/html/wg/drafts/html/master/dom.html.
W3C. Media queries. Retrieved April 19, 2015 from
http://www.w3.org/TR/css3-mediaqueries/.
W3C. Selectors level 3. Retrieved March 19, 2015 from
http://www.w3.org/TR/css3-selectors/.
W3C. W3C public mail archive: The
:min-width/:max-width pseudo-classes. Retrieved
April 28, 2015 from
https://lists.w3.org/Archives/Public/www-style/2013Mar/0368.html.
L. Wiener. ELQ: Extensible Element Queries for
Modular Responsive Web Components. Master’s
thesis, KTH Royal Institute of Technology, Stockholm,
Sweden, 2015.
C. Worrell. Responsive Elements. Retrieved April 29,
2015 from
https://github.com/coreyworrell/responsive-elements.
20
0.15
Injection time [s]
Injection time [s]
15
0.1
10
5
0.05
0
0
200
400
Number of elements
ELQ scroll-based solution
600
0
500
1,000
1,500
Number of elements
Object-based solution
Scroll-based solution
ELQ scroll-based solution
Figure 2: The left graph shows the injection time of the ELQ scroll-based approach. The right graph shows
all three approaches, including graph predictions by polynomial regression.
| 6cs.PL
|
The NOESIS Network-Oriented Exploration, Simulation, and Induction
System
Vı́ctor Martı́nez,∗ Fernando Berzal,† and Juan-Carlos Cubero‡
arXiv:1611.04810v2 [cs.SI] 23 Jun 2017
Department of Computer Science and Artificial Intelligence, University of Granada, Spain
Network data mining has become an important area of study due to the large number of problems
it can be applied to. This paper presents NOESIS, an open source framework for network data
mining that provides a large collection of network analysis techniques, including the analysis of
network structural properties, community detection methods, link scoring, and link prediction,
as well as network visualization algorithms. It also features a complete stand–alone graphical
user interface that facilitates the use of all these techniques. The NOESIS framework has been
designed using solid object–oriented design principles and structured parallel programming. As a
lightweight library with minimal external dependencies and a permissive software license, NOESIS
can be incorporated into other software projects. Released under a BSD license, it is available
from http://noesis.ikor.org.
Contents
I. Introduction
II. The design of the NOESIS framework
A. System architecture
B. Core classes
C. Supported data formats
D. Graphical user interface
III. Network analysis tools
A. Network models
∗ Electronic address: [email protected]
† Electronic address: [email protected]
B. Structural properties of networks
C. Network visualization techniques
4
5
IV. Network data mining techniques
A. Community detection
B. Link scoring and prediction
6
6
8
1
2
2
2
2
3
3
3
I. INTRODUCTION
Data mining techniques are intended to extract information from large volumes of data (Tan et al., 2006).
Data mining includes tasks such as classification, regression, clustering, or anomaly detection, among others.
Traditional data mining techniques are typically applied
to tabulated data. Novel techniques have also been devised for semi-structured or structured data, since exploiting the relationships among instances from a dataset
leads to new research and development opportunities
(Getoor and Diehl, 2005).
For example, network data mining has been used
to predict previously unknown protein interactions in
protein-protein interaction networks (Martı́nez et al.,
2014). It has also been used to study and predict future
author collaborations and tendencies in co-authorship
networks (Pavlov and Ichise, 2007). Different network
mining techniques are used by popular internet search
engines to rank the most relevant websites (Page et al.,
1999). These are only some examples of the large number
of applications of network data mining.
There are many software tools that facilitate the analysis of networked data. Some tools provide closed solutions for end users who need to work with their own
network data sets, whereas other tools cater to software
V. Conclusion
11
Acknowledgments
11
References
address: [email protected]
11
‡ Electronic
developers as software libraries that collect network analysis algorithms. Most tools allow the analysis of network topology and the computation of different structural properties of networks having thousands or even
millions of nodes. Some of them also include implementations of specific network data mining techniques, such
as community detection algorithms or predictive models, including link prediction (Lü and Zhou, 2011) and
epidemic models (Keeling and Eames, 2005).
For instance, Pajek (Batagelj and Mrvar, 1998),
NodeXL (Smith et al., 2009), Gephi (Bastian et al.,
2009), and UCINET (Borgatti et al., 2002) are widely
used for social network analysis (SNA). Graphviz (Ellson et al., 2002) and Cytoscape (Shannon et al., 2003)
are two well–known alternatives for network visualization. Finally, igraph (Csardi and Nepusz, 2006) and
NetworkX (Schult and Swart, 2008) are two popular software libraries of network algorithms. A more comprehensive and up–to–date list of available software tools can
be found at Wikipedia: https://en.wikipedia.org/
wiki/Social_network_analysis_software.
NOESIS, whose name stands for Network–Oriented
Exploration, Simulation, and Induction System, is a software framework for network analysis and mining. It tries
to combine the best features of closed social network analysis tools and extensible algorithm libraries, while provid-
2
The NOESIS Network-Oriented Exploration, Simulation, and Induction System
GraphicalDUserDInterface
3rdDparty
applications
Application
generator
NOESISDAPI
DAL:
DataDAccessDLayer
Data
sources
NOESIS
ReflectiveDKernel
DDHAL:DHardwareDAbstractionDLayer
FIG. 1 The NOESIS framework architecture and its core subsystems.
ing support for parallel execution, something that most
listed tools lack. NOESIS is completely written in Java
and its source code has been released under a permissive
BSD open source license.
Our paper is structured as follows. In Section 2, the
NOESIS architectural design principles are briefly described. Section 3 covers the network analysis techniques
included in NOESIS. Network data mining techniques are
surveyed in Section 4. Finally, Section 5 describes the
NOESIS project current status and roadmap.
II. THE DESIGN OF THE NOESIS FRAMEWORK
NOESIS has been designed to be an easily–extensible
framework whose architecture provides the basis for the
implementation of network data mining techniques. In
order to achieve this, NOESIS is designed around abstract interfaces and a set of core classes that provide
essential functions, which allows the implementation of
different features in independent components with strong
cohesion and loose coupling. NOESIS components are
designed to be maintainable and reusable.
A. System architecture
The NOESIS framework architecture and its core subsystems are displayed in Figure 1. These subsystems are
described below.
The lowest-level component is the hardware abstraction layer (HAL), which provides support for the execution of algorithms in a parallel environment and hides implementation details and much of the underlying technical complexity. This component provides different building blocks for implementing well-studied parallel programming design patterns, such as MapReduce (Dean
and Ghemawat, 2008). For example, we would just
write result = (double) Parallel.reduce( index ->x[index]
* y[index], ADD, 0, SIZE-1) to compute the dot product
of two vectors in parallel. The HAL does not only implement structured parallel programming design patterns,
but it is also responsible for task scheduling and parallel
execution. It allows the adjustment of parallel execution
parameters, including the task scheduling algorithm.
The reflective kernel is at the core of NOESIS and
provides its main features. The reflective kernel provides the base models (data structures) and tasks (algorithms) needed to perform network data mining, as
well as the corresponding meta-objects and meta-models,
which can be manipulated at run time. It is the underlying layer that supports a large collection of network
analysis algorithms and data mining techniques, which
are described in the following section. Different types
of networks are dealt with using an unified interface, allowing us to choose the particular implementation that
is the most adequate for the spatial and computational
requirements of each application. Algorithms provided
by this subsystem are built on top of the HAL building
blocks, allowing the parallelized execution of algorithms
whenever possible.
The data access layer (DAL) provides an unified interface to access external data sources. This subsystem
allows reading and writing networks in different file formats, providing implementations for some of the most
important standardized network file formats. This module also enables the development of data access components for other kinds of data sources, such as network
streaming.
Finally, an application generator is used to build
a complete graphical user interface following a model
driven software development (MDSD) approach. This
component provides a friendly user interface that allows users without programming skills to use most of the
NOESIS framework features.
B. Core classes
The core classes and interfaces shown in Figure 2 provide the foundation for the implementation of different
types of networks with specific spatial and computational requirements. Basic network operations include
adding and removing nodes, adding and removing links,
or querying a node neighborhood. More complex operations are provided through specialized components.
NOESIS supports networks with attributes both in
their nodes and their links. These attributes are defined
according to predefined data models, including categorical and numerical values, among others.
C. Supported data formats
Different file formats have been proposed for network
datasets. Some data formats are more space efficient,
II
The design of the NOESIS framework
NetworkWriter
NetworkReader
3
Network
Attribute
AttributeNetwork
LinkAttribute
NetworkRenderer
FIG. 2 Some of the NOESIS core classes and interfaces represented as an UML class diagram.
III. NETWORK ANALYSIS TOOLS
NOESIS is designed to ease the implementation of network analysis tools. It also includes reusable implementations of a large collection of popular network–related
techniques, from graph visualization (Tamassia, 2013)
and common graph algorithms, to network structural
properties (Newman, 2010) and network formation models (Jackson, 2008). The network analysis tools included
in NOESIS and the modules that implement them are
introduced in this section.
whereas others are more easily parseable.
NOESIS supports reading and writing network data
sets using the most common data formats. For example,
the GDF file format is a CSV-like format used by some
software tools such as GUESS and Gephi. It supports
attributes in both nodes and links. Another supported
file format is GML, which stands for Graph Modeling
Language. GML is a hierarchical ASCII-based file format. GraphML is another hierarchical file format based
on XML, the ubiquitous eXtensible Markup Language
developed by the W3C.
Other file formats are supported by NOESIS, such as
the Pajek file format, which is similar to GDF, or the
file format of the datasets from the Stanford Network
Analysis Platform (SNAP) (Leskovec and Krevl, 2014).
D. Graphical user interface
In order to allow users without programming knowledge to use most of the NOESIS features, a lightweight
easy–to–use graphical user interface is included with the
standard NOESIS framework distribution. The NOESIS
GUI allows non–technical end users loading, visualizing,
and analyzing their own network data sets by applying
all the techniques provided with NOESIS.
Some screenshots of this GUI are shown in Figure 3.
A canvas is used to display the network in every moment. The network can be manipulated by clicking or
dragging nodes. At the top of the window, a menu gives
access to different options and data mining algorithms.
The Network menu allows loading a network from an
external source and exporting the results using different
file formats, as well as creating images of the current
network visualization both as raster and vector graphics
image. The View menu allows the customization of the
network appearance by setting specific layout algorithms
and custom visualization styles. In addition, this menu
allows binding the visual properties of nodes and links to
their attributes. The Data menu allows the exploration
of attributes for each node and link. Finally, the Analysis
menu gives access to most of the techniques that will be
described in the following sections.
A. Network models
NOESIS implements a number of popular random network generation models, which are described by probability distributions or random processes. Such models have
been found to be useful in the study and understanding
of certain properties or behaviors observed in real-world
networks.
Among the models included in NOESIS, the ErdösRényi model (Erdős and Rényi, 1959) is one of the simplest ones. The Gilbert model (Gilbert, 1959) is similar
but a probability of existence is given for links instead.
The anchored network model is also similar to the two
previous models, with the advantage of reducing the occurrence of isolated nodes, but at the cost of being less
than perfectly random. Finally, the connected random
model is a variation of the anchored model that avoids
isolated nodes.
Other models included in NOESIS exhibit specific
properties often found in real-world networks. For example, the Watts–Strogatz model (Watts and Strogatz,
1998) generates networks with small-world properties,
that is, low diameter and high clustering. This model
starts by creating a ring lattice with a given number of
nodes and a given mean degree, where each node is connected to its nearest neighbors on both sides. In the
following steps, each link is rewired to a new target node
with a given probability, avoiding self-loops and link duplication.
Despite the small-world properties exhibited by networks generated by the Watts–Strogatz model are closer
to real world networks than those generated by models
based on the Erdös-Rényi approach, they still lack some
important properties observed in real networks. The
Barabási–Albert model (Albert and Barabási, 2002) is
another well-known model that generates networks whose
node degree distribution follows a power law, which leads
to scale-free networks. This model is driven by a preferential attachment process, where new nodes are added
and connected to existing nodes with a probability proportional to their current degree. Another model with
very similar properties to Barabási–Albert’s model is the
Price’s citation model (Newman, 2003).
In addition to random network models, a number of
regular network models are included in NOESIS. These
4
The NOESIS Network-Oriented Exploration, Simulation, and Induction System
FIG. 3 Different screenshots of the NOESIS graphical user interface.
FIG. 4 Random networks generated using the Erdös-Rényi model (left), the Watts–Strogatz model (center), and the Barabási–
Albert model (right).
models generate synthetic networks that are useful in the
process of testing new algorithms. The networks regular
models include complete networks, where all nodes are
interconnected; star networks, where all nodes are connected to a single hub node; ring networks, where each
node is connected to its closest two neighbors along a
ring; tandem networks, like ring model but without closing the loop; mesh network, where nodes are arranged in
rows and columns, and connected only to their adjacent
nodes; toruses, meshes where nodes in the extremes of
the mesh are connected; hypercubes; binary trees; and
isolates, a network without links.
B. Structural properties of networks
Network structural properties allow the quantification
of features or behaviors present in the network. They
can be used, for instance, to measure network robustness
or reveal important nodes and links. NOESIS considers three types of structural properties: node properties,
node pair properties (for pairs both with and without
links among them), and global properties.
NOESIS provides a large number of techniques for analyzing network structural properties. Many structural
properties can be computed for nodes. For example, indegree and out-degree, indicate the number of incoming
and outgoing links, respectively. Related to node degree,
two techniques to measure node degree assortativity have
III
Network analysis tools
been included: biased (Piraveenan et al., 2008) and unbiased (Piraveenan et al., 2010) node degree assortativity.
Node assortativity is a score between −1 and 1 that measures the degree correlation between pairs of connected
nodes. The clustering coefficient can also be computed
for nodes. The clustering coefficient of a node is the fraction of its neighbors that are also connected among them.
Reachability scores are centrality measures that allow
the analysis of how easy it is to reach a node from other
nodes. The eccentricity of a node is defined as the maximum distance to any other node (Hage and Harary,
1995). The closeness, however, is the inverse of the sum
of the distance from a given node to all others (Bavelas,
1950). An adjusted closeness value that normalizes the
closeness according to the number of reachable nodes can
also be used. Inversely to closeness, average path length
is defined as the mean distance of all shortest paths to
any other node. Decay is yet another reachability score,
computed as the summation of a delta factor powered by
the path length to any other node (Jackson, 2008). It is
interesting to note that with a delta factor close to 0, the
measure becomes the degree of the node, whereas with a
delta close to 1, the measure becomes the component size
of the component the node is located at. A normalized
decay score is also available.
Betweenness, as reachability, is another way to measure node centrality. Betweenness, also known as Freeman’s betweenness, is a score computed as the count of
shortest paths the node is involved in (Freeman, 1977).
Since this score ranges from 2n − 1 to n2 − (n − 1) for
n the number of nodes in strongly-connected networks, a
normalized variant is typically used.
Finally influence algorithms provide a different perspective on node centrality. These techniques measure
the power of each node to affect others. The most popular influence algorithm is PageRank (Page et al., 1999),
since it is used by the Google search engine. PageRank computes a probability distribution based on the
likelihood of reaching a node starting from any other
node. The algorithm works by iteratively updating node
probability based on direct neighbors probabilities, which
leads to convergence if the network satisfies certain properties. A similar algorithm is HITS (Kleinberg, 1999),
which stands for hyperlink-induced topic search. It follows an iterative approach, as PageRank, but computes
two scores per node: the hub, which is a score related
to how many nodes a particular node links, and the authority, which is a score related to how many hubs link
a particular node. Both scores are connected by an iterative updating process: authority is updated according to the hub scores of nodes connected by incoming
links and hub is updated according to authority scores of
nodes connected by outgoing links. Eigenvector centrality is another iterative method closely related to PageRank, where nodes are assigned a centrality score based
on the summation of the centrality of their neighbors
nodes. Katz centrality considers all possible paths, but
penalizes long ones using a given damping factor (Katz,
5
1953). Finally, diffusion centrality (Kang et al., 2012)
is another influence algorithm based on Katz centrality.
The main difference is that, while Katz considers infinite
length paths, diffusion centrality considers only paths of
a given limited length.
In the following example, we show how to load a network from a data file and compute its structural properties using NOESIS, its PageRank scores in particular:
FileReader fileReader =
new FileReader("karate.gml");
NetworkReader reader =
new GMLNetworkReader(fileReader);
Network network = reader.read();
PageRank task = new PageRank(network);
NodeScore score = task.call();
Different structural properties for links can also be
computed by NOESIS. For example, link betweenness,
which is the count of shortest paths the link is involved
in, or link rays, which is the number of possible paths
between two nodes that cross a given link. Some of these
properties are used by different network data mining algorithms.
C. Network visualization techniques
Humans are still better than machines at the recognition of certain patterns when analyzing data in a visual
way. Network visualization is a complex task since networks tend to be huge, with thousands nodes and links.
NOESIS enables the visualization of networks by providing the functionality needed to render the network and
export the resulting visualization using different image
file formats.
NOESIS provides different automatic graph layout techniques, such as the well–known Fruchterman–
Reingold (Fruchterman and Reingold, 1991) and
Kamada–Kawai (Kamada and Kawai, 1989) force–based
layout algorithms. Force–based layout algorithms assign
forces among pairs of nodes and solve the system to reach
an equilibrium point, which usually leads to an aesthetic
visualization.
Hierarchical layouts (Tamassia, 2013), which arrange
nodes in layers trying to minimize edge crossing, are also
included. Different radial layout algorithms are included
as well (Wills, 1999). These layouts are similar to the hierarchical ones, but arrange nodes in concentric circles.
Finally, several regular layouts are included. These layouts are common for visualizing regular networks, such
as meshes or stars.
NOESIS allows tuning the network visualization look
and feel. The visual properties of nodes and links can
be customized, including color, size, borders, and so on.
In addition, visual properties can be bound to static or
dynamic properties of the network. For example, node
sizes can be bound to a specific centrality score, allowing
the visual display of quantitative information.
6
The NOESIS Network-Oriented Exploration, Simulation, and Induction System
FIG. 5 A dolphin social network (Lusseau et al., 2003) represented using different network visualization algorithms: random
layout (top left), Kamada–Kawai layout (top right), Fruchterman–Reingold layout (bottom left), and circular layout using
average path length (bottom right).
IV. NETWORK DATA MINING TECHNIQUES
A. Community detection
Network data mining techniques exist for both unsupervised and supervised settings. NOESIS includes
a wide array of community detection methods (Lancichinetti and Fortunato, 2009) and link prediction techniques (Liben-Nowell and Kleinberg, 2007). These algorithms are briefly described below.
Community detection can be defined as the task of
finding groups of densely connected nodes. A wide range
of community detection algorithms have been proposed,
exhibiting different pros and cons. NOESIS features different families of community detection techniques and
implements more than ten popular community detection
algorithms. The included algorithms, their time complexity, and their bibliographic references are shown in
Table I.
NOESIS provides hierarchical clustering algorithms.
Agglomerative hierarchical clustering treats each node
IV
Network data mining techniques
7
FIG. 6 Community detection methods applied to Zachary’s karate club network (Zachary, 1977): Fast greedy partitioning
(top left), Kernighan-Lin bi-partitioning (top right), average-link hierarchical partitioning (bottom left), and complete-link
hierarchical partitioning (bottom right).
as a cluster, and then iteratively merges clusters until all
nodes are in the same cluster (Fortunato, 2010). Different
strategies for the selection of clusters to merge have been
implemented, including single-link (Sibson, 1973), which
selects the two clusters with the smallest minimum pairwise distance; complete-link (Defays, 1977), which selects
the two clusters with the smallest maximum pairwise distance; and average-link (Liu, 2011), which selects the two
clusters with the smallest average pairwise distance.
Modularity-based techniques are also available in our
framework. Modularity is a score that measures the
strength of particular division into modules of a given
network. Modularity–based techniques search for com-
munities by attempting to maximize their modularity
score (Newman and Girvan, 2004). Different greedy
strategies, including fast greedy (Newman, 2004) and
multi-step greedy (Schuetz and Caflisch, 2008), are available. These greedy algorithms merge pairs of clusters
that maximize the resulting modularity, until all possible merges would reduce the network modularity.
Partitional clustering is another common approach.
Partitioning clustering decomposes the network and performs an iterative relocation of nodes between clusters.
For example, Kernighan-Lin bi-partitioning (Kernighan
and Lin, 1970) starts with an arbitrary partition in two
clusters. Then, iteratively exchanges nodes between both
8
The NOESIS Network-Oriented Exploration, Simulation, and Induction System
Type
Name
Time complexity
Reference
Single-link (SLINK)
O(v 2 )
(Sibson, 1973)
Hierarchical Complete-link (CLINK)
O(v 2 log v)
(Defays, 1977)
Average-link (UMPGA)
O(v 2 log v)
(Liu, 2011)
Fast greedy
O(kvd log v)
(Newman, 2004)
Modularity
Multi-step greedy
O(kvd log v)
(Schuetz and Caflisch, 2008)
2
Kernighan-Lin bi-partitioning
O(v log v)
(Kernighan and Lin, 1970)
Partitional
K-means
O(kvd)
(MacQueen et al., 1967)
Ratio cut algorithm (EIG1)
O(v 3 )
(Hagen and Kahng, 1992)
Spectral
Jordan and Weiss NG algorithm (KNSC1)
O(v 3 )
(Ng et al., 2002)
Spectral k-means
O(v 3 )
(Shi and Malik, 2000)
Overlapping BigClam
O(v 2 )
(Yang and Leskovec, 2013)
TABLE I
Computational time complexity and bibliographic references for the community detection techniques provided by NOESIS. In
the time complexity analysis, v is the number of nodes in the network, d is the maximum node degree, and k is the desired
number of clusters.
clusters to minimize the number of links between them.
This approach can be applied multiple times to subdivide the obtained clusters. K-means community detection (MacQueen et al., 1967) is an application of the traditional k-means clustering algorithm to networks and
another prominent example of partitioning community
detection.
Spectral community detection (Fortunato, 2010) is another family of community detection techniques included
in NOESIS. These techniques use the Laplacian representation of the network, which is a network representation
computed by subtracting the adjacency matrix of the network to a diagonal matrix where each diagonal element
is equal to the degree of the corresponding node. Then,
the eigenvectors of the Laplacian representation of the
network are computed. NOESIS includes the ratio cut
algorithm (EIG1) (Hagen and Kahng, 1992), the Jordan
and Weiss NG algorithm (KNSC1) (Ng et al., 2002), and
spectral k-means (Shi and Malik, 2000).
Finally, the BigClam overlapping community detector
is also available in NOESIS (Yang and Leskovec, 2013).
In this algorithm, each node has a profile, which consists
in a score between 0 and 1 for each cluster that is proportional to the likelihood of the node belonging to that
cluster. Also, a score between pairs of nodes is defined
yielding values proportional to their clustering assignment overlap. The algorithm iteratively optimizes each
node profile to maximize the value between connected
nodes and minimize the value among unconnected nodes.
In the following example, we show how to load a network from a data file and detect communities with the
KNSC1 algorithm using NOESIS:
FileReader fileReader =
new FileReader("mynetwork.net");
NetworkReader reader =
new PajekNetworkReader(fileReader);
Network network = reader.read();
CommunityDetector task =
new NJWCommunityDetector(network);
Matrix results = task.call();
B. Link scoring and prediction
Link scoring and link prediction are two closely related
tasks. On the one hand, link scoring aims to compute a
value or weight for a link according to a specific criteria.
Most link scoring techniques obtain this value by considering the overlap or relationship between the neighborhood of the nodes at both ends of the link. On the other
hand, link prediction computes a value, weight, or probability proportional to the likelihood of the existence of a
certain link according to a given model of link formation.
The NOESIS framework provides a large collection of
methods for link scoring and link prediction, from local
methods, which only consider the direct neighborhood of
nodes, to global methods, which consider the whole network topology. As the amount of information considered
is increased, the computational and spatial complexity of
the techniques also increases. The link scoring and prediction methods available in NOESIS are shown in Table
II.
Among local methods, the most basic technique is the
common neighbors score (Newman, 2001), which is equal
to the number of shared neighbors between a pair of
nodes. Most techniques are variations of the common
neighbors score. For example, the Adamic–Adar score
(Adamic and Adar, 2003) is the sum of one divided by
the logarithm of the degree of each shared node. The
resource–allocation index (Zhou et al., 2009) follows the
same expression, but directly considers the degree instead of the logarithm of the degree. The adaptive degree penalization score (Martı́nez et al., 2016) also follows
the same approach, yet automatically determines an adequate degree penalization by considering properties of
the network topology. Other local measures consider the
IV
Network data mining techniques
9
Type Name
Time complexity
Reference
Common Neighbors count
O(vd3 )
(Newman, 2001)
Adamic–Adar score
O(vd3 )
(Adamic and Adar, 2003)
Resource–allocation index
O(vd3 )
(Zhou et al., 2009)
3
Adaptive
degree
penalization
score
O(vd
)
(Martı́nez
et al., 2016)
Local
Jaccard score
O(vd3 )
(Jaccard, 1901)
Leicht-Holme-Newman score
O(vd3 )
(Leicht et al., 2006)
3
Salton score
O(vd )
(Salton and McGill, 1986)
Sorensen score
O(vd3 )
(Sørensen, 1948)
Hub promoted index
O(vd3 )
(Ravasz et al., 2002)
hub depressed index
O(vd3 )
(Ravasz et al., 2002)
Preferential attachment score
O(vd2 )
(Barabási and Albert, 1999)
Katz index
O(v 3 )
(Katz, 1953)
Leicht-Holme-Newman score
O(cv 2 d)
(Leicht et al., 2006)
Random walk
O(cv 2 d)
(Pearson, 1905)
Global
Random walk with restart
O(cv 2 d)
(Tong et al., 2006)
Flow propagation
O(cv 2 d)
(Vanunu and Sharan, 2008)
3
Pseudoinverse Laplacian score
O(v )
(Fouss et al., 2007)
Average commute time score
O(v 3 )
(Fouss et al., 2007)
Random forest kernel index
O(v 3 )
(Chebotarev and Shamis, 2006)
TABLE II
Computational time complexity and bibliographic references for the link scoring and prediction methods provided by
NOESIS. In the time complexity analysis, v is the number of nodes in the network, d is the maximum node degree, and c
refers to the number of iterations required by iterative global link prediction methods.
number of shared neighbors, but normalize their value
according to certain criteria. For example, the Jaccard
score (Jaccard, 1901) normalizes the number of shared
neighbors by the total number of neighbors. The local
Leicht-Holme-Newman score (Leicht et al., 2006) normalizes the count of shared neighbors by the product of both
neighborhoods sizes. Similarly, the Salton score (Salton
and McGill, 1986) also normalizes, this time using the
square root of the product of both node degrees. The
Sorensen score (Sørensen, 1948) considers the double of
the count of shared neighbors normalized by the sum
of both neighbors size. The hub promoted and hub depressed scores (Ravasz et al., 2002) normalize the count
of shared neighbors by the minimum and the maximum
of both nodes degree respectively. Finally, the preferential attachment score (Barabási and Albert, 1999) only
considers the product of both node degrees.
Global link scoring and prediction methods are more
complex than local methods. For example, the Katz score
(Katz, 1953) sums the influence of all possible paths between two nodes, incrementally penalizing paths by their
length according to a given damping factor. The global
Leicht-Holme-Newman score (Leicht et al., 2006) is quite
similar to the Katz score, but resorts to the dominant
eigenvalue to compute the final result.
Random walk techniques simulate a Markov chain of
randomly-selected nodes (Pearson, 1905). The idea is
that, starting from a seed node and randomly moving
through links, we can obtain a probability vector where
each element corresponds to the probability of reaching
each node. The classical random walk iteratively multiplies the probability vector by the transition matrix,
which is the row-normalized version of the adjacency matrix, until convergence. An interesting variant is the random walk with restart (Tong et al., 2006), which models the possibility of returning to the seed node with a
given probability. Flow propagation is another variant
of random walk (Vanunu and Sharan, 2008), where the
transition matrix is computed by performing both row
and column normalization of the adjacency matrix.
Finally, some spectral techniques are also available in
NOESIS. Spectral techniques, as we mentioned when discussing community detection methods, are based on the
Laplacian matrix. The pseudoinverse Laplacian score
(Fouss et al., 2007) is the inner product of the rows of
the corresponding pair of nodes from the Laplacian matrix. Other spectral technique is the average commute
time (Fouss et al., 2007), which is defined as the average
number of steps that a random walker starting from a
particular node takes to reach another node for the first
time and go back to the initial node. Despite it models
a random walk process, it is considered to be a spectral
technique because it is usually computed in terms of the
Laplacian matrix. Given the Laplacian matrix, it can be
computed as the diagonal element of the starting node
plus the diagonal element of the ending node, minus two
times the element located in the row of the first node and
the column of the second one.
10
The NOESIS Network-Oriented Exploration, Simulation, and Induction System
FIG. 7 Different link scoring techniques applied to Les Miserables coappearance network (Knuth, 2009): common neighbors
(top left), preferential attachment score (top right), Sorensen score (bottom left), and Katz index (bottom right). Link width
in the figure is proportional to the link score.
Finally, the random forest kernel score (Chebotarev
and Shamis, 2006) is a global technique based on the
concept of spanning tree, i.e. a connected undirected
sub-network with no cycles that includes all the nodes
and some or all the links of the network. The matrix-tree
theorem states that the number of spanning trees in the
network is equal to any cofactor, which is a determinant
obtained by removing the row and column of the given
node, of an entry of its Laplacian representation. As a
result of this, the inverse of the sum of the identity matrix
and the Laplacian matrix gives us a matrix that can be
interpreted as a measure of accessibility between pairs of
nodes.
Using network data mining algorithms in NOESIS is
simple. In the following code snippet, we show how to
generate a Barabsi-Albert preferential attachment network with 100 nodes and 1000 links, and then compute
the Resource Allocation score for each pair of nodes using
NOESIS:
Network network =
new BarabasiAlbertNetwork(100, 1000);
LinkPredictionScore method =
new ResourceAllocationScore(network);
Matrix result = method.call();
V
Conclusion
V. CONCLUSION
Currently, the NOESIS project comprises more than
thirty five thousand lines of code organized in hundreds
of classes and dozens of packages. NOESIS relies on a library of reusable components that, with more than forty
thousand lines of Java code, provide a customizable collection framework, support for the execution of parallel
algorithms, mathematical routines, and the model-driven
application generator used to build the NOESIS graphical user interface.
NOESIS can ease the development of applications that
involve the analysis of any kind of data susceptible of
being represented as a network. NOESIS provides an
efficient, scalable, and developer–friendly framework for
network data mining, released under a permissive Berkeley Software Distribution free software license. Our
framework can be downloaded from its official website
at http://noesis.ikor.org.
Acknowledgments
The NOESIS project is partially supported by the
Spanish Ministry of Economy and the European Regional
Development Fund (FEDER), under grant TIN2012–
36951, and the Spanish Ministry of Education under the
program “Ayudas para contratos predoctorales para la
formación de doctores 2013” (grant BES–2013–064699).
We are grateful to Aarón Rosas, Francisco–Javier Gijón,
and Julio–Omar Palacio for their contributions to the implementation of community detection methods in NOESIS.
References
Adamic, L. A. and Adar, E. (2003). Friends and neighbors
on the web. Social Networks, 25(3):211–230.
Albert, R. and Barabási, A.-L. (2002). Statistical mechanics
of complex networks. Reviews of Modern Physics, 74(1):47.
Barabási, A.-L. and Albert, R. (1999). Emergence of scaling
in random networks. Science, 286(5439):509–512.
Bastian, M., Heymann, S., Jacomy, M., et al. (2009). Gephi:
an open source software for exploring and manipulating
networks. International AAAI Conference on Weblogs and
Social Media, 8:361–362.
Batagelj, V. and Mrvar, A. (1998). Pajek-program for large
network analysis. Connections, 21(2):47–57.
Bavelas, A. (1950). Communication patterns in task-oriented
groups. The Journal of the Acoustical Society of America,
22(6):725–730.
Borgatti, S. P., Everett, M. G., and Freeman, L. C. (2002).
UCINET for Windows: Software for social network analysis. Technical report, Analytic Technologies.
Chebotarev, P. and Shamis, E. (2006). Matrix-forest theorems. arXiv preprint math/0602575.
Csardi, G. and Nepusz, T. (2006). The igraph software package for complex network research. International Journal of
Complex Systems, 1695(5):1–9.
11
Dean, J. and Ghemawat, S. (2008). MapReduce: simplified
data processing on large clusters. Communications of the
ACM, 51(1):107–113.
Defays, D. (1977). An efficient algorithm for a complete link
method. The Computer Journal, 20(4):364–366.
Ellson, J., Gansner, E., Koutsofios, L., North, S. C., and
Woodhull, G. (2002). Graphviz - open source graph drawing tools. In Graph Drawing, pages 483–484. Springer.
Erdős, P. and Rényi, A. (1959). On random graphs. Publicationes Mathematicae Debrecen, 6:290–297.
Fortunato, S. (2010). Community detection in graphs. Physics
Reports, 486(3):75–174.
Fouss, F., Pirotte, A., Renders, J.-M., and Saerens, M. (2007).
Random-walk computation of similarities between nodes of
a graph with application to collaborative recommendation.
IEEE Transactions on Knowledge and Data Engineering,
19(3):355–369.
Freeman, L. C. (1977). A set of measures of centrality based
on betweenness. Sociometry, pages 35–41.
Fruchterman, T. M. and Reingold, E. M. (1991). Graph drawing by force-directed placement. Software: Practice and
Experience, 21(11):1129–1164.
Getoor, L. and Diehl, C. P. (2005). Link mining: a survey.
ACM SIGKDD Explorations Newsletter, 7(2):3–12.
Gilbert, E. N. (1959). Random graphs. The Annals of Mathematical Statistics, 30(4):1141–1144.
Hage, P. and Harary, F. (1995). Eccentricity and centrality
in networks. Social Networks, 17(1):57–63.
Hagen, L. and Kahng, A. B. (1992). New spectral methods for
ratio cut partitioning and clustering. IEEE Transactions on
Computer-Aided Design of Integrated Circuits and Systems,
11(9):1074–1085.
Jaccard, P. (1901). Étude comparative de la distribution florale dans une portion des alpes et des jura. Bulletin de la
Société Vaudoise des Sciences Naturelles, 37:547–579.
Jackson, M. O. (2008). Social and Economic Networks.
Princeton University Press, Princeton, NJ, USA.
Kamada, T. and Kawai, S. (1989). An algorithm for drawing
general undirected graphs. Information Processing Letters,
31(1):7–15.
Kang, C., Molinaro, C., Kraus, S., Shavitt, Y., and Subrahmanian, V. (2012). Diffusion centrality in social networks.
In Proceedings of the 2012 International Conference on Advances in Social Networks Analysis and Mining (ASONAM
2012), pages 558–564. IEEE Computer Society.
Katz, L. (1953). A new status index derived from sociometric
analysis. Psychometrika, 18(1):39–43.
Keeling, M. J. and Eames, K. T. (2005). Networks and
epidemic models. Journal of the Royal Society Interface,
2(4):295–307.
Kernighan, B. W. and Lin, S. (1970). An efficient heuristic
procedure for partitioning graphs. Bell System Technical
Journal, 49(2):291–307.
Kleinberg, J. M. (1999). Authoritative sources in a hyperlinked environment. Journal of the ACM (JACM),
46(5):604–632.
Knuth, D. E. (2009). The Stanford GraphBase: A Platform for Combinatorial Computing. Addison-Wesley Professional, 1st edition.
Lancichinetti, A. and Fortunato, S. (2009). Community detection algorithms: a comparative analysis. Physical Review
E, 80(5):056117.
Leicht, E. A., Holme, P., and Newman, M. E. (2006). Vertex
12
The NOESIS Network-Oriented Exploration, Simulation, and Induction System
similarity in networks. Physical Review E, 73(2):026120.
Leskovec, J. and Krevl, A. (2014). SNAP Datasets: Stanford
large network dataset collection. http://snap.stanford.
edu/data.
Liben-Nowell, D. and Kleinberg, J. (2007).
The linkprediction problem for social networks. Journal of the
American Society for Information Science and Technology,
58(7):1019–1031.
Liu, B. (2011). Web Data Mining, 2 ed. Berlin, Germany:
Springer Berlin Heidelberg.
Lü, L. and Zhou, T. (2011). Link prediction in complex networks: A survey. Physica A: Statistical Mechanics and its
Applications, 390(6):1150–1170.
Lusseau, D., Schneider, K., Boisseau, O. J., Haase, P.,
Slooten, E., and Dawson, S. M. (2003). The bottlenose
dolphin community of doubtful sound features a large proportion of long-lasting associations. Behavioral Ecology and
Sociobiology, 54(4):396–405.
MacQueen, J. et al. (1967). Some methods for classification
and analysis of multivariate observations. In Proceedings of
the Fifth Berkeley Symposium on Mathematical Statistics
and Probability, number 14 in 1, pages 281–297. Oakland,
CA, USA.
Martı́nez, V., Berzal, F., and Cubero, J.-C. (2016). Adaptive
degree penalization for link prediction. Journal of Computational Science, 13:1–9.
Martı́nez, V., Cano, C., and Blanco, A. (2014). Prophnet: A
generic prioritization method through propagation of information. BMC Bioinformatics, 15(Suppl 1):S5.
Newman, M. (2010). Networks: An Introduction. Oxford
University Press.
Newman, M. E. (2001). Clustering and preferential attachment in growing networks.
Physical Review E,
64(2):025102.
Newman, M. E. (2003). The structure and function of complex networks. Society for Industrial and Applied Mathematics Review, 45(2):167–256.
Newman, M. E. (2004).
Fast algorithm for detecting
community structure in networks. Physical Review E,
69(6):066133.
Newman, M. E. and Girvan, M. (2004). Finding and evaluating community structure in networks. Physical Review E,
69(2):026113.
Ng, A. Y., Jordan, M. I., Weiss, Y., et al. (2002). On spectral
clustering: Analysis and an algorithm. Advances in Neural
Information Processing Systems, 2:849–856.
Page, L., Brin, S., Motwani, R., and Winograd, T. (1999).
The pagerank citation ranking: Bringing order to the web.
Technical report, Stanford InfoLab.
Pavlov, M. and Ichise, R. (2007). Finding experts by link
prediction in co-authorship networks. Finding Experts on
the Web with Semantics Workshop, 290:42–55.
Pearson, K. (1905). The problem of the random walk. Nature,
72:342.
Piraveenan, M., Prokopenko, M., and Zomaya, A. (2008). Local assortativeness in scale-free networks. Europhysics Letters, 84(2):28002.
Piraveenan, M., Prokopenko, M., and Zomaya, A. Y. (2010).
Classifying complex networks using unbiased local assortativity. In ALIFE, pages 329–336.
Ravasz, E., Somera, A. L., Mongru, D. A., Oltvai, Z. N., and
Barabási, A.-L. (2002). Hierarchical organization of modularity in metabolic networks. Science, 297(5586):1551–
1555.
Salton, G. and McGill, M. J. (1986). Introduction to Modern
Information Retrieval. McGraw-Hill, Inc.
Schuetz, P. and Caflisch, A. (2008). Efficient modularity optimization by multistep greedy algorithm and vertex mover
refinement. Physical Review E, 77(4):046112.
Schult, D. A. and Swart, P. (2008). Exploring network structure, dynamics, and function using NetworkX. In Proceedings of the 7th Python in Science Conferences (SciPy 2008),
volume 2008, pages 11–16.
Shannon, P., Markiel, A., Ozier, O., Baliga, N. S., Wang,
J. T., Ramage, D., Amin, N., Schwikowski, B., and
Ideker, T. (2003). Cytoscape: a software environment
for integrated models of biomolecular interaction networks.
Genome Research, 13(11):2498–2504.
Shi, J. and Malik, J. (2000). Normalized cuts and image segmentation. Transactions on Pattern Analysis and Machine
Intelligence, 22(8):888–905.
Sibson, R. (1973). Slink: an optimally efficient algorithm
for the single-link cluster method. The Computer Journal,
16(1):30–34.
Smith, M. A., Shneiderman, B., Milic-Frayling, N.,
Mendes Rodrigues, E., Barash, V., Dunne, C., Capone, T.,
Perer, A., and Gleave, E. (2009). Analyzing (social media)
networks with NodeXL. In Proceedings of the Fourth International Conference on Communities and Technologies,
pages 255–264. ACM.
Sørensen, T. (1948). A method of establishing groups of equal
amplitude in plant sociology based on similarity of species
and its application to analyses of the vegetation on danish
commons. Biologiske Skrifter, 5:1–34.
Tamassia, R. (2013). Handbook of Graph Drawing and Visualization. CRC Press.
Tan, P.-N., Steinbach, M., Kumar, V., et al. (2006). Introduction to data mining. Pearson, Addison Wesley, Boston.
Tong, H., Faloutsos, C., and Pan, J.-Y. (2006). Fast random
walk with restart and its applications. In Proceedings of
the Sixth International Conference on Data Mining, ICDM
’06, pages 613–622. IEEE Computer Society.
Vanunu, O. and Sharan, R. (2008). A propagation-based algorithm for inferring gene-disease assocations. In German
Conference on Bioinformatics, pages 54–52.
Watts, D. J. and Strogatz, S. H. (1998). Collective dynamics
of ‘small-world’ networks. Nature, 393(6684):440–442.
Wills, G. J. (1999). Nicheworksinteractive visualization of
very large graphs. Journal of Computational and Graphical
Statistics, 8(2):190–212.
Yang, J. and Leskovec, J. (2013). Overlapping community
detection at scale: a nonnegative matrix factorization approach. In Proceedings of the Sixth ACM International
Conference on Web Search and Data Mining, pages 587–
596. ACM.
Zachary, W. W. (1977). An information flow model for conflict and fission in small groups. Journal of Anthropological
Research, pages 452–473.
Zhou, T., Lü, L., and Zhang, Y.-C. (2009). Predicting missing
links via local information. The European Physical Journal
B, 71(4):623–630.
| 2cs.AI
|
arXiv:1610.04807v3 [cs.DS] 10 Apr 2017
Local Max-Cut in Smoothed Polynomial Time
Omer Angel
Sébastien Bubeck∗
University of British Columbia
Department of Mathematics
Canada
[email protected]
Microsoft Research;
USA
[email protected]
Yuval Peres
Fan Wei
Microsoft Research;
USA
[email protected]
Stanford University
Department of Mathematics
USA
[email protected]
ABSTRACT
1 INTRODUCTION
In 1988, Johnson, Papadimitriou and Yannakakis wrote that “Practically all the empirical evidence would lead us to conclude that
finding locally optimal solutions is much easier than solving NPhard problems". Since then the empirical evidence has continued to
amass, but formal proofs of this phenomenon have remained elusive. A canonical (and indeed complete) example is the local maxcut problem, for which no polynomial time method is known. In a
breakthrough paper, Etscheid and Röglin proved that the smoothed
complexity of local max-cut is quasi-polynomial, i.e., if arbitrary
bounded weights are randomly perturbed, a local maximum can
be found in ϕn O (log n) steps where ϕ is an upper bound on the random edge weight density. In this paper we prove smoothed polynomial complexity for local max-cut, thus confirming that finding
local optima for max-cut is much easier than solving it.
Let G = (V , E) be a connected graph with n vertices and w : E →
[−1, 1] be an edge weight function. The local max-cut problem asks
to find a partition of the vertices σ : V → {−1, 1} whose total cut
weight
1 Õ
w(uv) 1 − σ (u)σ (v) ,
(1)
2
CCS CONCEPTS
• Theory of computation → Graph algorithms analysis;
KEYWORDS
Smoothed analysis, Max-cut, Polynomial running time, Hopfield
network, Nash equilibrium, Potential game, Sherrington-Kirkpatrick
model
ACM Reference format:
Omer Angel, Sébastien Bubeck, Yuval Peres, and Fan Wei. 2017. Local MaxCut in Smoothed Polynomial Time. In Proceedings of 49th Annual ACM
SIGACT Symposium on the Theory of Computing, Montreal, Canada, June
2017 (STOC’17), 9 pages.
DOI: 10.1145/3055399.3055402
∗ The
corresponding author
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected].
STOC’17, Montreal, Canada
© 2017 ACM. 978-1-4503-4528-6/17/06. . . $15.00
DOI: 10.1145/3055399.3055402
uv ∈E
is locally maximal, in the sense that one cannot increase the cut
weight by changing the value of σ at a single vertex (recall that finding the global maximum of (1) is NP-hard). This problem comes up
naturally in a variety of contexts. For example (Schäffer and Yannakakis
1991) showed that local max-cut is complete for the complexity
class Polynomial-Time Local Search (PLS). It also appears in the
party affiliation game, (Fabrikant et al. 2004): this is an n-player
game where each player v ∈ V selects an action σ (v) ∈ {−1, 1} and
Í
the resulting payoff for player v is sign uv ∈E w(uv) 1 − σ (u)σ (v) .
It is easy to see that a local maximum of (1) exactly corresponds
to a Nash equilibrium for the party affiliation game. Yet another
appearance of this problem is in the context of Hopfield networks,
(Hopfield 1982): this is a collection of neurons with weighted connections between them, where each neuron is in one of two states
(either firing or not firing) and with state update at random times
by thresholding the sum of incoming weights from firing neurons.
It is again easy to see that such dynamics make the state configuration converge (for undirected weights) to a local maximum of (1)
(with σ (u) representing the state of neuron u and w(uv) the weight
of the connection between neurons u and v).
There is a natural algorithm to find a local maximum of (1),
sometimes referred to as the FLIP algorithm: Start from some initial partition σ , and until reaching a local maximum, repeatedly
find a vertex for which flipping the sign of σ would increase the
cut weight - and carry out this flip. (To be precise, this is a family of algorithms corresponding to different ways of selecting the
improving change when there are multiple possibilities.) This algorithm also corresponds to a natural dynamics for the party affiliation game, and a specific implementation (random selection of an
improving vertex) exactly corresponds to the asynchronous Hopfield network dynamics described above. However, it is easy to see
that there exists weight functions such that FLIP takes an exponential number of steps before reaching a local maximum. As noted
in (Johnson et al. 1988) (who introduced the PLS class), this seems
STOC’17, June 2017, Montreal, Canada
at odd with empirical evidence suggesting that algorithms such
as FLIP usually reach a local maximum in a reasonable time. This
conflicting situation naturally motivates the study of the smoothed
complexity of local max-cut: is it true that after adding a small
amount of noise to the edge weights, the FLIP algorithm terminates in polynomial time with high probability? In this paper we
answer this question affirmatively, provided that a small amount
of noise is added to all vertex pairs (i.e., even to non-edges); in
other words, we assume that G is a complete graph. We note that
a similar subtlety arises in the smoothed analysis of the simplex
algorithm by (Spielman and Teng 2004) where noise is added to
every entry of the constraint matrix (in particular, the null entries
are also smoothed).
We now introduce the problem formally, discuss existing results,
and state our main contributions. Let X = (X e )e ∈E ∈ [−1, 1]E be
a random vector with independent entries. One should think of
X e as the original edge weight w(e) plus some independent small
noise. We assume that X e has a density fe with respect to the
Lebesgue measure, and we denote ϕ = maxe ∈E k fe k∞ . In this paper the phrase with high probability means with probability at
least 1−on (1) with respect to X . We consider the space of spin configurations {−1, 1}V , and for a spin configuration σ ∈ {−1, 1}V we
denote by σ (v) the value of σ at vertex v. We are interested in the
random map H : {−1, 1}V → R (usually called the Hamiltonian)
defined by:
1 Õ
H(σ ) = −
Xuv σ (u)σ (v).
(2)
2
uv ∈E
Our objective is to find a local maximum of H with respect to the
Hamming distance d(σ, σ ′ ) = #{v : σ (v) , σ ′ (v)}. Equivalently,
we are looking for a locally optimal cut in the weighted graph
(G, X ) (since (1) and (2) differ by the half of the total weight of
all edges).
We say that σ ′ is an improving move from σ if d(σ ′, σ ) = 1
and H(σ ′ ) > H(σ ). We will sometimes refer to a sequence of improving moves as an improving sequence. The FLIP algorithm iteratively performs improving moves until reaching a configuration with no improving move. An implementation of FLIP specifies
how to choose the initial configuration and how to choose among
the improving moves available at each step. (Etscheid and Röglin
2014) show that for any graph with smoothed weights, with high
probability, any implementation of FLIP will terminate in at most
nC log(n) steps, for some universal constant C > 0.
Our main result is that FLIP terminates in a polynomial number
of steps for the complete graph. Since our results are asymptotic
in n, in the rest of the paper we assume n ≥ n 0 for some universal
constant n 0 .
Theorem 1.1. Let G be the complete graph on n vertices, and assume the edge weights X = (X e )e ∈E are independent random variables with |X | ≤ 1 and density bounded above by ϕ. For any η > 0,
with high probability any implementation of FLIP terminates in at
most O(ϕ 5n 15+η ) steps, with implicit constant depending only on η.
Corollary 1.2. Under the assumptions of Theorem 1.1, the expected number of steps of any implementation of FLIP is O(n 15), with
implicit constant depending only on ϕ.
Omer Angel, Sébastien Bubeck, Yuval Peres, and Fan Wei
Note that any implementation is a very broad category. It includes an implementation where an adversary with unbounded
computational power chooses each improving step. Theorem 1.1
implies that even in this case the number of steps is polynomial
with high probability.
Remark 1.3. The edge weights are assumed to be bounded only for
simplicity. Our methods can be used to give the same bound as long
as the edge weights have finite variance, and a polynomial bound as
long as the weights have a polynomial tail. Indeed, if P(|X | > t) ≪
t −δ , then with high probability all edge weights are at most n 3/δ ;
rescaling the edge weights by n 3/δ increases ϕ by a corresponding
factor, giving a bound of order n 15+15/δ +η .
Remark 1.4. In the the classical Sherrington-Kirkpatrick model
(Sherrington and Kirkpatrick 1975), a mean field model for a spin
glass, the Hamiltonian is exactly a scaled version of the random map
defined in (2) and when X i j are i.i.d. Gaussian random variables for
all pairs i, j. Therefore our Theorem 1.1 implies that in the in the
Sherrington-Kirkpatrick model, the maximal length of a monotone
path (along which the energy is decreasing) in the random energy
landscape is O(n 15+η ).
Theorem 1.1 can be equivalently stated as follows.
Theorem 1.5. Let G be the complete graph. Assume the edge weights
X = (X e )e ∈E are independent random variables with |X | ≤ 1 and
density bounded above by ϕ. The probability that there is an improving sequence of length Ω(ϕ 5 n 15+η ) is o(1).
We say that a sequence L is ϵ-slowly improving from an initial
state σ0 if each step of L increases H by at most ϵ (and more than
0). Our main task will be to prove the following proposition:
Proposition 1.6. Fix η > 0 and let ϵ = n −(12+η)ϕ 5 . Then with
high probability, there is no ϵ-slowly improving sequence of length
2n from any σ0 .
Proposition 1.6 implies Theorem 1.5 as follows. Since X e ∈ [−1, 1],
the maximum total improvement for H is at most n 2 . If there exists
an improving sequence of length at least Ω(n 15+η ϕ 5 ) then there
must exist an improving sequence of length 2n with total improvement less than O(n −(12+η)ϕ −5 ). Apart from Section 5 the rest of the
paper is dedicated to proving Proposition 1.6.
We believe that the exponent 15 in Theorem 1.1 is far from tight.
In fact we make the conjecture that local max-cut is in smoothed
quasi-linear time:
Conjecture 1.7. Let G be the complete graph on n vertices, and
assume the edge weights X = (X e )e ∈E are independent random variables with |X | ≤ 1 and density bounded above by ϕ. With high probability any implementation of FLIP terminates in at most n(ϕ log n)c
steps where c > 0 is a universal constant.
This quasi-linear time behavior could quite possibly extend to
an arbitrary graph G; however, the first step should be to show
smoothed polynomial complexity in this setting (that is, to generalize Theorem 1.1 to an arbitrary graph). Some graphs are easier
than others. E.g., (Elsässer and Tscheuschner 2011) observed that
for graphs with maximum degree O(log(n)) endowed with Gaussian edge weights, with high probability, any implementation of
Local Max-Cut in Smoothed Polynomial Time
FLIP terminates in a polynomial number of steps. (Since, with high
probability, each improving move increases H significantly.) In the
final section of the paper we show that a natural approach to generalize our result to arbitrary graphs cannot work; the proof relies on
a new result on combinatorics of words, which is of independent
interest.
2
PRELIMINARIES
In this section we provide a high-level overview of the proof of
Proposition 1.6. We also state and prove some lemmas which will
be useful in our analysis.
Recall that we work in the state space {−1, 1}V and that a move
flips the sign of a single vertex. Each move can be viewed as a linear
operator, which we define now. For any σ ∈ {−1, 1}V and v ∈ V ,
we denote by σ −v the state equal to σ except for the coordinate
corresponding to v which is flipped. For such σ, v there exists a
vector α = α(σ, v) ∈ {−1, 0, 1} E such that H(σ −v ) = H(σ ) + hα, X i.
More specifically α = (αuw )uw ∈E is defined by
αuv = σ (v)σ (u) ∀u , v
(3)
αuw = 0
if v < {u,w }
Crucially, note that α does not depend on X . We say that v is an
improving move from a configuration σ if hα, X i > 0. It will be
convenient to identify a move with the corresponding vector α.
Thus we may talk of improving vectors (meaning that hα, X i > 0).
Similarly, we say that certain moves are linearly independent if the
corresponding vectors are.
2.1
Basic idea for the analysis
We first observe that for non-zero α ∈ ZE , the random variable
hα, X i also has density bounded by ϕ. Thus for a fixed move from
σ to σ −v , one has P(hα, X i ∈ (0, ϵ]) ≤ ϕϵ. Naively (ignoring correlations), one could expect that for a fixed sequence of moves with
corresponding vectors α 1 , . . . , α ℓ ,
P ∀i ∈ [ℓ], hαi , X i ∈ (0, ϵ] ≤ (ϕϵ)ℓ .
(4)
A rigorous and more general statement in this direction is given in
the following lemma.
STOC’17, June 2017, Montreal, Canada
is to fix the above calculation when the length of the sequence is
replaced by the linear rank of the sequence of improving moves. A
particularly important task for us will be to show that given any
sequence of length Ω(n) of potentially improving moves, one can
always find many αi ’s which are linearly independent. (Some sequences of moves cannot possibly be improving, e.g., if the same
coordinate is flipped twice in a row.)
Given an initial configuration σ0 and a sequence of moves L
of length ℓ, let the corresponding move operators be α 1 , . . . , α ℓ .
Consider the |E| × ℓ matrix A L = [αi ]iℓ=1 whose ith column is the
vector αi ∈ {−1, 0, 1} E (thus each row is indexed by an edge e ∈ E).
Note that the vectors αt , and thus also the matrix A L depends
(implicitly) on the initial spin state σ0 . The maximum number of
linearly independent moves in L is the rank of the matrix A L ; and
thus we may apply Lemma 2.1 with k being this rank.
This turns out not to be sufficient for our needs. However, if a
sequence of moves L is an improving sequence from some initial
state, then every contiguous segment of L is also improving from
some (different) state. We use the term block to refer to a contiguous segment of some sequence of moves under consideration (we
will formally define it in Section 3). Thus to bound the probability that L is improving we can instead consider only a segment of
our choice of L. Note that there are two competing effects in the
choice of a segment: on the one hand the probability that a block is
ϵ-slowly improving is generally much larger than the probability
that the full sequence is ϵ-slowly improving; on the other hand any
given block appears in many different sequences, which yields an
improvement in the union bound.
Our proof will proceed in two key steps: (i) find a block of L
with relatively high rank (this is done in Section 3), and (ii) apply
the union bound we alluded to above in a more efficient way so as
to replace the term 2n (counting possible initial configurations) by
a smaller term (Section 4). To this end, we will want to find a block
in L which has a high rank and in which the number of distinct
symbols is as small as possible.
2.2 Preliminary linear algebra
We now provide some preliminary results which prepare us to find
Lemma 2.1 (Lemma A.1 (Etscheid and Röglin 2014)). Let α 1 , . . . , αk a lower bound for the rank of the matrix A L corresponding to a
sequence L = (v 1 , . . . ,v ℓ ). (Here vt ∈ V denotes the vertex which
be k linearly independent vectors in ZE . Then the joint density of
moves at step t.) Denote by σt the spin configuration after step t.
k
(hαi , X i)i ≤k is bounded by ϕ . In particular, if sets Ji ⊂ R have
The following statement is a direct consequence of equation (3).
measure at most ϵ each, then
Lemma 2.2. The vector αt is supported precisely on the edges inP ∀i ∈ [k], hαi , X i ∈ Ji ≤ (ϕϵ)k .
cident to vt . The entry in αt corresponding to the edge {vt , u} is
−σt (vt )σt (u), which is also equal to σt −1 (vt )σt −1 (u).
This lemma is stated slightly differently from Lemma A.1 of
(Etscheid and Röglin 2014) but the same proof applies. If in a seWe now make the following simple observation.
quence of moves all moves are linearly independent, then (4) holds.
Lemma 2.3. The rank of A L does not depend on the initial configUnder this assumption, a union bound implies that the probability
uration σ0 .
there exists an initial configuration and a sequence of ℓ improving moves which improves by at most ϵ (since each step improves
Proof. Let A L be obtained from some initial configuration σ0
by at most ϵ), is smaller than 2n n ℓ (ϕϵ)ℓ , since there are 2n iniand let A L′ be obtained from another initial configuration σ0′ . Both
matrices are derived from the same sequence L. For any vertex u
tial configurations and at most n ℓ sequences of length ℓ. In other
and time t we have that σt (u)σt′(u) = σ0 (u)σ0′(u). Thus the row corwords, with high probability, any sequence of length Ω(n) would
responding to an edge {u,v } in A L is σ0 (u)σ0 (v)σ0′ (u)σ0′(v) times
improve the cut value by at least Ω(1/poly(n)). Since H is bounded
the corresponding row in A L′ , and thus these two matrices have
by poly(n), as a consequence of this, the FLIP algorithm should
the same rank.
reach a local maximum after at most poly(n) steps. The challenge
STOC’17, June 2017, Montreal, Canada
Rather than working with the matrix A L directly, we will consider the matrix A = AL whose t-th column is −σt (vt )αt (for
t ∈ [ℓ]). Obviously A has the same rank as A; in light of this
and of Lemma 2.3, we define the rank of a sequence of moves
by rank(L) = rank(AL ). For future reference, we give the following alternative definition of the matrix A (the two definitions are
equivalent by Lemma 2.2).
Definition 2.4. For a given sequence L = (v 1 , . . . ,v ℓ ), let A = AL
be the |E| × ℓ matrix with rows indexed by edges. For an edge e =
{u,v } and time t such that u , vt , the entry A[e, t] = 1vt =v σt (u).
Thus the t-th entry of the row corresponding to an edge e =
{u,v } is non-zero, if and only if vt ∈ {u,v }. If vt = v, then the
t-th entry of the row A[{u,v }] is the spin of u (the other endpoint
of the edge) at time t, i.e., σt (u) (which also equals σt −1 (u) since
u , v = vt ).
3
BOUNDING THE RANK OF L
The goal of this section is to prove Lemma 3.1 which gives a lower
bound on the rank of L in terms of simple combinatorial properties
of L. First we introduce some notation.
For any sequence of moves L, a vertex that appears only once in
L is called a singleton; vertices that appear at least twice are called
repeated vertices. Let ℓ(L) be the length of L; Let s 1 (L) be the
number of singletons in L, and let s 2 (L) be the number of repeated
vertices in L. Denote by s(L) = s 1 (L) + s 2 (L) the total number of
distinct vertices that appear in L. When the sequence of moves
L is clear from the context, we shall use ℓ, s, s 1 and s 2 to denote
ℓ(L), s(L),s 1 (L) and s 2 (L), respectively.
A block of a sequence L = (v 1 ,v 2 , . . . ,v ℓ ) is a contiguous segment from the sequence, i.e. (vi , . . . , v j ) of length j −i + 1 for some
i ≤ j. We denote this block by L[i, j]. A maximal block (w.r.t. inclusion) of L which consists of only singletons is called a singleton block. A maximal block of L which consists of only repeated
vertices is called a transition block. Thus L is naturally partitioned into alternating singleton and transition blocks. Note that
a repeated vertex might appear only once in a specific transition
block, in which case it must appear also in at least one other transition block. For every v in L, let b(v) be the number of transition
blocks containing v. Let T1 , . . . ,Tk denote the transition blocks,
and x + = max(x, 0). Throughout the proof, we use u,v, w etc. to
denote vertices in V ; sometimes for the purpose of enumeration,
we might also use integers 1, 2, . . . to denote vertices in V which
should cause no confusion.
The next lemma is the main result of this section.
Lemma 3.1. For any sequence of moves L one has
(i) rank(L) ≥ min(s(L), n − 1).
Furthermore, if s(L) < n and L does not visit any state more than
once, then
(ii) rank(L) ≥ s(L) + s 2 (L)/2.
Í
Í
(iii) rank(L) ≥ s 1 (L) + i s(Ti ) = s(L) + v (b(v) − 1)+ , where
the sum is over the transition blocks of L.
Note that L visits a state more than once if σi = σ j for some
i < j, or equivalently the block L[i + 1, j] contains every vertex an
even number of times. (This clearly is a property of L, independent
of σ0 ). If a sequence is improving, then it cannot revisit any state.
Omer Angel, Sébastien Bubeck, Yuval Peres, and Fan Wei
We can safely disregard any sequence which fails this condition in
later analysis.
Proof. (i) Without loss of generality, suppose 1, 2, . . . , s are the
only vertices appearing in L, and suppose that s < n. Let ti be some
time at which vertex i appears in L; Consider the s × s sub-matrix
of A restricted to the columns ti ’s and the rows corresponding to
edges {i, n} for i = 1, 2, . . . , s. By our choice of ti , the column ti has
a non-zero entry at the row corresponding to {i, n}, and no others,
and thus has full rank s. If s = n apply the above reasoning to the
set of times {t 1 , . . . , tn−1 }.
(ii) We first make the following simple observation. Given a sequence L which does not revisit any state, if vertex v is moved at
least twice, then the block between any two consecutive moves of
v contains some vertex u an odd number of times in this block. This
is clear, since any block in L contains some vertex an odd number
of times by an earlier argument.
We create an auxiliary directed graph H as follows. The vertices of H are the n vertices of G. For each repeated vertex v, there
must be a vertex u that appears an odd number of times between
the first two times v appears. We pick one such u arbitrarily, and
add to H a directed edge from v to u. Note that H might contain
both an edge and its reverse (e.g. for the sequence L = 1, 2, 1, 3, 2).
Each repeated vertex has one out-going edge in H , and so H has
exactly s 2 directed edges. Moreover, directed cycles (including cycles of length 2) in H are vertex-disjoint, and their total length is
at most s 2 . Let us define a sub-graph of H by removing one edge
from each directed cycle of H . Since the cycles are vertex-disjoint
(since the out-degree for each vertex is at most 1), we remove at
most s 2 /2 edges, and obtain an acyclic sub-graph of H with at least
s 2 /2 edges.
Since not all vertices appear in L, suppose without loss of generality that vertex n does not appear in L. Part (ii) of the lemma now
follows from the following.
Claim 3.2. For any acyclic sub-graph H ′ of H , the following edges
correspond to linearly independent rows in A: All edges of H ′, together with {v, n} for vertices v ∈ L.
We prove this by induction on the number of edges in H ′. If
H ′ is the empty subgraph, these are precisely the rows used to
prove part (i). Now suppose H ′ is not empty. Since H ′ is acyclic,
there must be a vertex v with in-degree 0 and unique outgoing edge
Í
e = {v, u}. Suppose we have a linear combination i λi A[{i, n}] +
Í
e ∈H ′ µ e A[e] = 0, where A[e] is the row corresponding to e and
the sum is over the edges of the claim. Let t 1 , t 2 be the first two
times that v moves. By the definition of A (see Definition 2.4), the
t 1 -th and t 2 -th entry of A[{v, n}] are both σt1 (n) = σt2 (n) (since n
does not move). Furthermore, since u appears an odd number of
times between the first two appearance of v we have that the t 1 -th
entry and t 2 -th entry of A[e] are of opposite signs. Furthermore,
since v has out-degree 1 in H ′ and in-degree 0, among the rows
we have picked, only the rows A[{v, n}] and A[e] have non-zero
entries in positions t 1 , t 2 . We thus have λv ± µ e = 0, implying
λv = µ e = 0. Thus the linear combination involves only edges of
H ′ \ e and edges to n. Applying the inductive hypothesis to H ′ \ e
gives that the linear combination is trivial.
Local Max-Cut in Smoothed Polynomial Time
(iii) Suppose without loss of generality that 1, . . . , s 2 are the
repeated vertices in L. By the definition of b(i), there exist times
t 1 (i), t 2 (i), . . . , tb (i ) (i) in different transition blocks at which i moves,
and for any 2 ≤ j ≤ b(vi ), there is a singleton vertex wi, j that appears in the block L[t j−1 (i), t j (i)].
We claim that the following rows are linearly independent. For
each v in L the edge {v, n}, and for each repeated vertex i, the rows
ei, j = {i, wi, j } for j = 2, . . . , b(i).
For any repeated vertex vi , among the rows we have picked, the
ones which have non-zero entries at times t 1 (i), . . . , tb (i ) (i) correspond to the rows of {i, n}, and ei, j for j = 2, . . . , b(vi ). At those
columns, by Lemma 2.3, we can assume the row A[{i, n}] has all
ones. The row A[ei, j ] has entries 1 before the (unique) appearance
of wi, j and −1 after the appearance. Thus the minor for these rows
and the sequence of times {t 1 (i), . . . , tb (i ) (i)} has the form
1
1
1
1
.
..
1
−1
1
1
..
.
1
−1
−1
1
..
.
1
−1
−1
−1
..
.
· · ·
· · ·
· · · .
· · ·
. .
.
This clearly has full rank b(i). For singleton vertices v appearing
at time t = tv , the only selected row with no-zero t-th entry corresponds to edge {v, n}. Thus if we group together columns for
the repeated vertices, the selected rows of A have a block structure, with blocks of the form above along the diagonal and zeros
elsewhere. It follows that
Õ
Õ
rank(A) ≥ s 1 +
b(i) = s +
(b(i) − 1)+ .
i ≤s2
4
i ≤s2
PROOF OF PROPOSITION 1.6
STOC’17, June 2017, Montreal, Canada
this contradicts criticality of B.
Lemma 4.2. Suppose s(B) < n. For a critical block B as in Lemma
4.1, we have
β
s 1 (B).
rank(B) ≥ s(B) +
1+β
Proof. We apply Lemma 3.1(iii) to B. Let T1 , . . . ,Tk be the transition blocks of B. If the whole of B is a transition block, i.e. s 1 (B) =
0, then s(T1 ) = s(B) and rank(B) = s(B) by Lemma 3.1(iii) yields
the claim. Otherwise, each Ti is a proper sub-block of B, and by
criticality of B we find ℓ(Ti ) < (1 + β)s(Ti ) for each Ti . Thus
rank(B) ≥ s 1 +
Õ
vertices i in B
b(i) = s 1 (B) +
≥ s 1 (B) +
k
Õ
i =1
s(Ti )
k
1 Õ
ℓ(Ti )
1 + β i =1
ℓ(B) − s 1 (B)
1+β
β
ℓ(B)
+
s 1 (B),
≥
1+β 1+β
Í
where we have used that ℓ(B) = s 1 (B)+ ki=1 ℓ(Ti ), since each letter
is either a singleton or part of one of the Ti .
By Lemma 4.1, ℓ(B) = ⌈(1 + β)⌉s(B), and the claim follows.
≥ s 1 (B) +
Corollary 4.3. For a critical block B with s(B) < n, we have
β
1
s 1 (B), s 2 (B) .
rank(B) ≥ s(B) + max
1+β
2
In particular,
In this section we prove Proposition 1.6, and thus conclude the
proof of our main result (Theorem 1.1). We first show in Subsection
4.1 that any improving sequence contains a certain special block
which we can use to obtain high rank. Then we conclude the proof
of Proposition 1.6 in Section 4.3 with an “improved” union bound
argument.
Proof. The two bounds come from Lemmas 3.1 and 4.2. Since
s 1 (B) + s 2 (B) = s(B), the last bound is obtained by a convex combination of the two preceding bounds.
4.1
4.2 A better bound on improving sequences
Finding a critical block with large rank
We start with a simple combinatorial lemma. Fix some β > 0. We
say that a block B is critical if ℓ(B) ≥ (1 + β)s(B), and every block
B ′ strictly contained in B has ℓ(B ′) < (1 + β)s(B ′).
Lemma 4.1. Fix any positive integer n ≥ 2 and a constant β > 0.
Given a sequence L consisting of s(L) < n letters and with length
ℓ(L) ≥ (1+ β)s, there exists a critical block B in L. Moreover, a critical
block satisfies ℓ(B) = ⌈(1 + β)s(B)⌉.
Proof. A block satisfying ℓ(B) ≥ (1 + β)s(B) exists, since the
whole sequence L satisfies this. A minimal (w.r.t. inclusion) block
that satisfies this will by definition be a critical block.
We now show that B satisfies ℓ(B) = ⌈(1 + β)s(B)⌉. If ℓ(B) ≥
⌈(1 + β)s(B)⌉ + 1, remove the last vertex from B, thus obtaining B ′.
Then ℓ(B ′) = ℓ(B) − 1, while s(B) ≥ s(B ′) ≥ s(B) − 1. For the block
B ′ we thus have
ℓ(B ′) = ℓ(B) − 1 ≥ ⌈(1 + β)s(B)⌉ ≥ ⌈(1 + β)s(B ′)⌉ ;
rank(B) ≥
1 + 4β
s(B).
1 + 3β
Lemma 2.1 implies that the probability that a sequence L is ϵ-slowly
improving from any given σ0 is at most (ϕϵ)rank(L) , and therefore
the probability that L is ϵ-slowly improving from some σ0 is at most
2n (ϕϵ)rank(L) . For sequences with large rank this is sufficiently small
for our needs. However, for sequences with small rank and small
s a better bound is needed. The next novel ingredient of our proof
is an improvement of this bound that reduces the factor of 2n , provided s(L) is small.
Lemma 4.4. Suppose the random weights X e a.s. have |X e | ≤ 1.
Then
s
4n
P(L is ϵ-slowly improving from some σ ) ≤ 2
(8ϕϵ)rank(L) .
ϵ
The key idea is that instead of taking a union over the initial
state σ0 for the non-moving vertices, we only consider the influence of the non-moving vertices on the moving vertices.
STOC’17, June 2017, Montreal, Canada
Proof. Without loss of generality, we may assume that the vertices that appear in L = L[1, ℓ] are 1, . . . , s, and that s + 1, . . . , n do
not appear. We separate H(σ ) = H0 (σ ) + H1 (σ ) + H2 (σ ), where Hj
is the sum over edges with j endpoints that appear in L ∪ {s + 1},
for j ∈ {0, 1, 2}. The reason for including s + 1 will become clear
later.
With a given initial state σ0 , let σt be the state after flipping the
state of vt . For u > s (so u does not appear in L), we have that
σt (u) is constant over t ≤ ℓ and thus H0 (σt ) = H0 (σ0 ) for all t ≤ ℓ.
Moreover, as in (3), we get
n
Õ
H1 (σt ) − H1 (σt −1 ) = −σt (vt )
Xvt ,u σ0 (u) = σt (vt )Q(vt ),
u=s+2
Ín
where Q(v) = − u=s+2
Xvt ,u σ0 (u). One may think of Q as a constant external field acting on the s moving vertices. Finally, the increments of H2 are linear functionals of the weights on edges with
both endpoints in {1, . . . , s, s + 1}. We denote these functionals by
ᾱt , so that
Omer Angel, Sébastien Bubeck, Yuval Peres, and Fan Wei
by Lemma 3.1(i). By Lemma 2.1, each term in (5) is bounded by
(ϕϵ)n−1 , and so
provided
H(σt ) − H(σt −1 ) = hᾱt , X i + σt (vt )d(vt ) + δ t ,
where |δ t | ≤ ϵ. If the sequence is ϵ-slowly increasing, then
hᾱt , X i + σt (vt )d(vt ) ≤ 2ϵ,
and thus hᾱt , X i lies in the union of two intervals of length 4ϵ centered at ±d(vt ). Note that rank(ᾱt ) = rank(L), since we included
in ᾱ the contributions from the stationary vertex s + 1. (This holds
also if s = n.) By Lemma 2.1, the probability of this event is at
most (8ϵϕ)rank(L) . Crucially, if we know (σ0 (i))i ≤s+1 and d(v) for
v = 1, . . . , s, then the event under consideration is the same for all
2n−(s+1) possible configurations σ0 .
The claim now follows by a union bound over the possible values of (σ0 (i))i ≤s+1 and d(v).
4.3
Proof of Proposition 1.6
Proof of Proposition 1.6. Fix β = 1. Let R be the event that
there exists an initial configuration σ0 and a sequence L of length
2n which is ϵ-slowly improving. Our goal is to show P(R) = o(1).
We consider two cases: either the sequence L has s(L) = n or
else s(L) < n. Call these events R 0 and R 1 . We bound P(R 0 ) by a
union bound over sequences:
Õ Õ
P(L is ϵ-slowly improving from σ0 ). (5)
P(R 0 ) ≤
σ0 L:s(L)=n
The summation is over all initial configurations σ0 and all possible
sequences of improving moves L from σ0 with n moving vertices.
There are 2n initial configurations and at most n 2n sequences of
length 2n. Since s = n, each such sequence has rank(L) ≥ n − 1
(6)
is small.
We turn to the event R 1 , that there exists an initial configuration
σ0 and an ϵ-slowly improving sequence L of length 2n such that
s(L) < n. By Lemma 4.1, on the event R 1 for some s < n there
exists a critical block using precisely s vertices and some initial
configuration such that the block is ϵ-slowly improving from that
configuration. Thus
Õ
P(R 1 ) ≤
P(B is ϵ-slowly improving from some σ ). (7)
critical B
By definition, a critical block has ℓ(B) = 2s(B). By Corollary 4.3,
it has rank(B) ≥ 5s(B)/4. Thus by Lemma 4.4, for any critical block
we have
P(B is ϵ-slowly improving from some σ )
s(B)
s(B)
4n
≤2
(8ϕϵ)5s(B)/4 ≤ 2 64ϕ 5/4nϵ 1/4
.
ϵ
H(σt ) − H(σt −1 ) = σt (vt )Q(vt ) + hᾱt , X i.
Note that ᾱt is simply the restriction of αt to edges with both endpoints in {1, . . . , s + 1}. Observe that ᾱt depends on the first s + 1
coordinates of σ0 , but not on the other coordinates.
Since X e is assumed to be bounded, we have |Q(v)| ≤ n. Consider the set D = 2ϵZ ∩ [−n, n], of size at most n/ϵ + 1 ≤ 2n/ϵ. We
have that Q(v) is within ϵ of some element of d(v) ∈ D. Instead of
a union bound on σ0 , we now use a union bound over (σ0 (i))i ≤s+1
and the vector (d(v))v ≤s . From the above definitions it follows that
2n 2ϕϵ
P(R 0 ) ≤ 2n n 2n (ϕϵ)n−1 = o(1),
The number of critical blocks using s letters is at most n 2s , (which
is the number of sequences of length ℓ = 2s). Thus
s
Õ
P(R 1 ) ≤ 2
n 2s 64ϕ 5/4nϵ 1/4 .
(8)
s <n
This sum tends to 0 as n → ∞ when ϵ = n −(12+η)ϕ −5 with η >
0.
Remark 4.5. The proof above shows that for ϵ = αϕ −5 n −12 , we
have P(R 1 ) ≤ O(α 3/4 ) as α → 0 (since a critical block with β = 1
has s ≥ 3) and hence that the run time of the FLIP algorithm, divided
by n 15 is tight. The number of critical blocks with a given s can be
bounded by ns s 2s ≤ (ens)s which is less than n 2s for s ≤ n/e. Using
this gives
s
Õ
P(R 1 ) ≤ 2
Cϕ 5/4n 2sϵ 1/4 ,
(9)
s <n
and so P(R 1 ) decays super-polynomially in α.
Corollary 1.2 follows easily from the proof of Proposition 1.6:
Proof of Corollary 1.2. Suppose an increasing sequence of
length L ≥ 2n exists. Since the total weight of any cut is in [−n 2/4, n 2 /4,
there must be a block of size 2n in L such that the total improven2
≤ 2n 3 /L. Let R(n, L) be
ment along the block is at most ϵ = 2[L/2n]
the probability there is such a block using all n letters (R 0 above),
and R(s, L) the probability there is a critical block of length 2s using
s letters.
Let T be the number of steps before FLIP terminates. Then we
have
Õ
P(R(s, L)).
P(T ≥ L) ≤
s ≤n
and so
E(T ) =
∞
Õ
L=1
P(T ≥ L) ≤ n 15 +
Õ Õ
L>n 15 s ≤n
P(R(s, L)),
Local Max-Cut in Smoothed Polynomial Time
and we need to show that the last sum is O(n 15). For s = n, by (6),
P(R(n, L)) ≤
2n n 2n (ϕϵ)n−1
=
2n n 2n (ϕ2n 3 /L)n−1 ,
and the sum over L > n 15 is o(n 15 ).
For s > 4, by (9),
s
P(R(s, L)) ≤ 2 Cϕ 5/4n 2s(2n 3 /L)1/4 ,
and so
Õ
C
P(R(s, L)) ≤ (Cs)s n 11s/4 (n 15 )1−s/4 ≤ (Cs)s−1n 15−s .
s
L>n 15
Í
Í
Thus 4<s <n L>n 15 P(R(s, L)) = O(n 15 ).
For small s the bound above is not sufficient, and we need a
better rank bound. There are no critical blocks with s = 1 or s = 2.
It is easy to check that critical blocks with s = 3 all have rank 6. A
short exhaustive search yields that critical blocks with s = 4 have
rank 7 or 8. Since the number of sequences with s = 3 or s = 4 is
O(n s ), for s = 3, 4 we get
s
4n
P(R(s, L)) ≤ O n s
(8ϕϵ)s+3 = O(n 2s ϵ 3 ).
ϵ
Í
Í
Thus s=3, 4 L>n 15 P(R(s, L)) = o(1), which completes the proof.
5
A WORD THAT IS SPARSE AT EVERY SCALE
The quasi-polynomial proof in (Etscheid and Röglin 2014) (which
applies to any graph) relied crucially on the following lemma: for
any word of length ℓ = Ω(n) over an alphabet of size n, there
must exist a subword of some length ℓ ′ such that the number of
distinct letters which appear more than once in this subword is
Ω(ℓ ′ /log(n)) (see Lemma 5.1 below for a precise statement). In
some sense this says that “a word cannot be too sparse at every
scale” (a word is viewed as sparse if it is mostly made of letters
that appear only once). We provide here a simple new proof of this
statement. A natural approach to prove smoothed polynomial complexity for any graph (that is generalize Theorem 1.1 to arbitrary
graphs) would be to remove the log(n) term in this combinatorics
of words lemma (see paragraph after Lemma 5.1 for more details).
Our main contribution is this section is to show that such an improvement is not possible: we show by a probabilistic construction
that Lemma 5.1 is tight, that is there exist words which are sparse
at every scale to the extent allowed by the lemma. More specifically
we construct a word of length Ω(n) such that for any subword of
length ℓ ′ the number of repeating letters is O(ℓ ′/log(n)) (in fact
we prove a stronger version of this statement where ℓ ′ is replaced
by the number of distinct letters in the subword), see Theorem 5.2
below.
Lemma 5.1. Suppose a > 1, and that L is a sequence of length an
in an alphabet of n letters. Then there exists a block B in L such that
s 2 (B) s 2 (B)
a−1
≥
≥
.
s(B)
ℓ(B)
a log2 (n)
Proof. The first inequality holds trivially for every block B. Define the surplus of a sequence L to be ℓ(L) − s(L), i.e. the difference
between the number of elements and the number of distinct elements in the sequence. If a block B is a concatenation of B 1 and B 2
then its surplus is at most the total surplus of B 1 and B 2 plus s 2 (B).
STOC’17, June 2017, Montreal, Canada
Let m(ℓ) be the maximum surplus in any block of length ℓ in L.
Assume that for some ϵ, for every block B from L we have s 2 (B) ≤
ϵℓ(B). Then one has
m(2ℓ) ≤ 2m(ℓ) + ϵ · 2ℓ.
By recursing this inequality, with m(2ℓ − 1) ≤ m(2ℓ) and m(1) = 0
we get
m(an) ≤ ϵan log2 (n).
Since m(an) = an − s(L) ≥ (a − 1)n, this shows that ϵ has to be
a−1
which concludes the proof.
greater than a log
(an)
2
It is easy to check that the proof of the rank lower bound given
in Lemma 3.1(ii) (and (i)) applies to arbitrary graphs. By using
Lemma 5.1 above together with the union bound argument from
Section 4.3 one obtains an alternative proof to the quasi-polynomial
complexity result of (Etscheid and Röglin 2014). A tempting approach to prove a polynomial complexity result for any graph would
be to “simply” replace the log(n) term in Lemma 5.1 by some constant. The main result of this section is to show that this cannot
be done, and that the log(n) in Lemma 5.1 is tight up to possibly
constant factors. As noted above, this can be interpreted as saying
that there exist words which are sparse at every scale. In fact, we
prove something stronger, as stated in the following theorem.
Theorem 5.2. For every a > 1 there exists a C so that for every n
there is a sequence of length ℓ = [an] in n letters so that every block
B of L has s 2 (B)/s(B) ≤ C/log n. Moreover, for n > n 0 (a) one may
take C = 9a log(a).
This is stronger in that we have a bound on s 2 (B)/s(B) ≥ s 2 (B)/ℓ(B).
We remark that decreasing a makes the problem easier (just take
the first [a ′n] letters). We can assume all letters are used in the
sequence, otherwise we can replace some repetitions by unused
letters.
5.1 The probabilistic construction
The construction proving Theorem 5.2 is probabilistic, and implies
that there are many sequences with these properties. We do not
optimize the constant C here in order to keep the proof simple and
clean. A more careful analysis will improve C.
We create a sequence as follows. In stage one of the construction we write down the (potentially) repeated letters. Each repeated
letter is written in some random set of locations, possibly overwriting previous letters. Afterwards, in stage two, all positions where
no repeated letters have been written are filled in with new and
unique letters. Note that it is possible that a potentially repeated
letter is overwritten, and consequently appears only once or even
not at all in the final sequence.
The construction is defined in terms of integers b 0 , b 1 and γ
which we will specify later in the proof. The potentially repeated
letters are denoted by i and i ′ for i ∈ {b 0, . . . , b 1 −1}. Thus the total
number of potentially repeated letters is 2(b 1 −b 0 ). To simplify the
description, we construct an infinite sequence and truncate afterwards to the first ℓ letters. For each i ∈ [b 0 , b 1 ), split N to blocks of
size γi. In each block [kγi, (k + 1)γi) where k ∈ N, we choose uniformly one position; In that position write the letter i if k is even,
and i ′ if k is odd. All these choices are independent. (Creating an
infinite sequence at this stage avoids having shorter blocks at the
STOC’17, June 2017, Montreal, Canada
Omer Angel, Sébastien Bubeck, Yuval Peres, and Fan Wei
end.) A position that is left empty at the end of stage one is filled
in stage two.
5.2
Negative correlations
For t ≤ ℓ, let Ut be the event that position t is empty at the end
of stage one. We will prove that any block contains many unique
letters. If the Ut were independent this would follow from standard large deviation bounds for Binomial random variables. While
the Ut are not independent, they have a weaker property which is
sufficient for our needs. A collection of events {Ut } is called negatively correlated if for every subset S of indices and every t < S
we have
P(Ut |Us ∀s ∈ S) ≤ P(Ut ),
P(Utc |Usc
∀s ∈ S) ≤
P(Utc ).
(10)
(11)
Negative correlation of the (Ut ) will follow from the following
more general statement.
Proposition 5.3. Let A1 , . . . , Am be some finite sets, and pick
a uniform element from each set independently. Let Ux be the event
that element x is never picked. Then the Ux are negatively correlated.
This applies to our model, by taking the sets to be the intervals
[kγi, (k + 1)γi) for b 0 ≤ i < b 1 and all k.
Proof. The effect of conditioning on Us ∀s ∈ S is simple: The
element from Ai is chosen uniformly from Ai \ S. Clearly this can
only decrease the probability that an element t is not selected from
any Ai . Since selections are independent, this gives (10).
Now we prove (11). The claim is equivalent to proving
P(Ut |Usc ∀s ∈ S) ≥ P(Ut ),
which in turn is equivalent to
P(Usc ∀s ∈ S |Ut ) ≥ P(Usc ∀s ∈ S).
Let ai be the element picked from Ai . To obtain the law of (ai )
conditioned on Ut , start with the unconditioned selections, and resample each ai if ai = t, until another element is chosen. If initially
(in the unconditioned vector), every element of S is selected from
some Ai , then this is also true after the resampling, and so the
probability of such full occupation is increased.
We use the following generalized Chernoff bounds for negatively correlated events.
Î 1 −1 i −1/γ +x
Proof. Let f (x) = bi =b
. Then f is increasing in x,
i +x
0
and d = f (0). We have that
d γ ≤ f (1/γ ) · f (2/γ ) · · · f (1) =
b0
,
b1
as this is a telescoping product. Similarly,
d γ ≥ f (0) · f (−1/γ ) · · · f ((1 − γ )/γ ) =
b0 − 1
.
b1 − 1
Proof of Theorem 5.2. With a > 1 and n given, we apply the
probabilistic construction above with parameters
√
log n
b 0 = [log n]
b1 = n
γ =
.
2 log(2a)
1 as n →
Note that b 0 /b 1 = n −1/2+o(1) , and therefore d tends to 2a
∞.
We first claim that with good probability the resulting sequence
√
uses at most n letters. Stage one uses at most 2b 1 = 2 n letters. The
expected number of letters used in stage two is dℓ ≤ n/2 + o(n).
By Markov’s inequality, the whole sequence use at most n letters
with asymptotic probability at least 1/2.
Next we consider repetitions within (possibly smaller) blocks.
Since occurences of the letter i are at least γi apart, and similarly for
the letter i ′ , not all letters can appear multiple times in short blocks.
In particular, each block B ∈ L is certain to have s 2 (B) ≤ 2ℓ(B)/γ .
Moreover, blocks B with ℓ(B) ≤ γb 0 have no repeated letters by
our construction, so that s 2 (B) = 0 for such blocks.
To estimate s(B), we note that the number of letters in B is at
least the number of letters added to B in stage two:
s(B) ≥ u(B) :=
j
Õ
1Ut .
t =i
We have Eu(B) = dℓ(B). By the Chernoff bound Theorem 5.4 with
δ = 1/2 we have
p
P u(B) ≤ 12 dℓ(B) ≤ ( 2/e)d ℓ(B) .
2
For blocks of length at least γb 0 this is e −c log n = o(n −2). By a
union bound, with high probability every block of length at least
γb 0 has
dℓ(B)
s 2 (B) ≤ 2ℓ(B)/γ
and
s(B) ≥
,
Theorem 5.4 ((Panconesi and Srinivasan 1997)). Suppose U1 , . . . , Uk
2
Ík
are negatively correlated events, and let Y = i =1 1Ui be the number
4 . (Shorter blocks have s (B) = 0.)
2 (B)
and so ss(B)
≤ dγ
of bad events occur. Then for any constant δ ∈ (0, 1),
2
E[Y ]
8a log(2a)+o(1)
As n → ∞, this decays as
, implying the claim for
log(n)
P(Y ≤ (1 − δ )E[Y ]) ≤ (1 − δ )−(1−δ )e −δ
.
n large enough. By changing C we can get the claim also for all
smaller n.
5.3 Analysis of the construction
We first estimate the probability that a letter of the sequence
is
Îb 1 −1
1
filled in stage two. This probability is P(Ut ) = i =b 1 − γ i ,
0
which we denote by d.
Lemma 5.5.
1/γ
b0
b 0 − 1 1/γ
.
≤d ≤
b1 − 1
b1
Remark 5.6. The above construction can be used to show that for
any a > 0 and η > 0 there exist infinitely many graphs G (with
number of vertices tending to infinity), paired with some initial configurations σ0 and sequence of moves L, such that ℓ(L) ≥ a|V (G)|,
and for each block B ∈ L, rank(B) ≤ (1 + η)s(B). These graphs are
a significant obstacle to generalizing our main result (Theorem 1.1)
beyond the complete graph via rank arguments.
Local Max-Cut in Smoothed Polynomial Time
ACKNOWLEDGMENTS
We are grateful to Constantinos Daskalakis for bringing this problem to our attention, and for helpful discussions at an early stage of
this project. We thank the Bellairs Institute, where this work was
initiated. Most of this work was done at Microsoft Research Redmond during the first author’s visit and the last author’s internship.
O. Angel is supported in part by NSERC.
REFERENCES
R. Elsässer and T. Tscheuschner. 2011. Settling the complexity of local max-cut (almost) completely. In Proceedings of the 38th international colloquim conference on
Automata, languages and programming - Volume Part I (ICALP’11). 171–182.
M. Etscheid and H. Röglin. 2014. Smoothed analysis of local search for the maximumcut problem. In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on
Discrete Algorithms (SODA ’14). 882–889.
STOC’17, June 2017, Montreal, Canada
A. Fabrikant, C. Papadimitriou, and K. Talwar. 2004. The complexity of pure Nash
equilibria. In Proceedings of the thirty-sixth annual ACM symposium on Theory of
computing (STOC ’04). 604–612.
J. J. Hopfield. 1982. Neural networks and physical systems with emergent collective
computational abilities. Proceedings of the National Academy of Sciences 79, 8 (1982),
2554–2558.
D. S. Johnson, C. H. Papadimtriou, and M. Yannakakis. 1988. How easy is local search?
J. Comput. Syst. Sci. 37, 1 (1988), 79–100.
A. Panconesi and A. Srinivasan. 1997. Randomized Distributed Edge Coloring via an
Extension of the Chernoff–Hoeffding Bounds. SIAM J. Comput (1997), 350–368.
A. A. Schäffer and M. Yannakakis. 1991. Simple local search problems that are hard
to solve. SIAM J. Comput. (1991), 56–87.
David Sherrington and Scott Kirkpatrick. 1975.
Solvable Model of a SpinGlass.
Phys. Rev. Lett. 35 (Dec 1975), 1792–1796.
Issue 26. DOI:
https://doi.org/10.1103/PhysRevLett.35.1792
D. A. Spielman and S-H. Teng. 2004. Smoothed analysis of algorithms: Why the simplex algorithm usually takes polynomial time. J. ACM (2004), 385–463.
| 8cs.DS
|
COCHARACTER-CLOSURE AND THE RATIONAL
HILBERT-MUMFORD THEOREM
arXiv:1411.7849v5 [math.AG] 3 Oct 2016
MICHAEL BATE, SEBASTIAN HERPEL, BENJAMIN MARTIN, AND GERHARD RÖHRLE
Abstract. For a field k, let G be a reductive k-group and V an affine k-variety on which
G acts. Using the notion of cocharacter-closed G(k)-orbits in V , we prove a rational version
of the celebrated Hilbert-Mumford Theorem from geometric invariant theory. We initiate
a study of applications stemming from this rationality tool. A number of examples are
discussed to illustrate the concept of cocharacter-closure and to highlight how it differs from
the usual Zariski-closure.
Contents
1. Introduction
2. Preliminaries
3. Cocharacter-closure
4. The rational Hilbert-Mumford Theorem
5. Ascent and descent
6. Reduction to k-points
7. Geometric and rational conjugacy
8. Reduction to GLn
9. G-complete reducibility
10. The action of GL(W ) on End(W )
11. Further examples
References
1
5
10
11
14
17
19
23
25
27
30
33
1. Introduction
Let G be a (possibly non-connected) reductive algebraic group over an algebraically closed
field k, and let V be an affine variety over k on which G acts. A central problem in geometric
invariant theory is to understand the structure of the set of orbits of G on V . The closed
orbits are particularly important, because they can be identified with the points of the
quotient variety V //G. Given v ∈ V , it is well known that the closure G · v of the orbit
G · v contains a unique closed orbit O. The Hilbert-Mumford Theorem says there exists a
cocharacter λ of G such that the limit lima→0 λ(a) · v exists and lies in O—in fact, we can
replace O with any closed G-stable subset of V that meets G · v [29, Thm. 1.4]. This gives a
characterization of the closed orbits in terms of cocharacters: if the orbit G · v is not closed,
then there is a cocharacter λ of G such that lima→0 λ(a) · v exists and lies outside G · v.
2010 Mathematics Subject Classification. 20G15 (14L24).
Key words and phrases. Affine G-variety, cocharacter-closed orbit, rationality.
1
Conversely, if G · v is closed then lima→0 λ(a) · v lies in G · v for all λ such that the limit
exists (cf. Section 2.4).
A strengthening of the Hilbert-Mumford Theorem due to Hesselink [27], Kempf [29] and
Rousseau [48] shows that if G · v is not closed, then there is a class of so-called “optimal”
cocharacters λopt such that the limit lima→0 λopt (a) · v exists in V but lies outside G · v.
Each cocharacter λopt enjoys some nice properties: for instance, if G is connected, then
the parabolic subgroup Pλopt associated to λopt contains the stabiliser Gv . Moreover, if
G, V and the G-action are defined over a perfect subfield k0 of k, k/k0 is algebraic and
v ∈ V (k0 ), then λopt is Gal(k/k0 )-fixed and hence is defined over k0 . The (strengthened)
Hilbert-Mumford Theorem has become an indispensable tool in algebraic group theory and
has numerous applications in geometric invariant theory and beyond [41]: e.g., geometric
complexity theory [39], [40], nilpotent and unipotent elements of reductive groups [18], [44],
[43], [2], moduli spaces of bundles [24], good quotients in geometric invariant theory [25],
Hilbert schemes [42], moduli spaces of sheaves [28], the structure of the Horn cone [45],
Kähler geometry [56], filtrations for representations of quivers [58], symplectic quotients [41,
App. 2C], degenerations of modules [59] and G-complete reducibility [4], [9].
Now suppose k is an arbitrary field, not necessarily algebraically closed. The orbit G · v
is a union of G(k)-orbits. The structure of this set of G(k)-orbits can be very intricate. For
instance, if w ∈ G·v and v, w are k-points then one can ask whether w is G(k)-conjugate to v.
The answer is no in general; if k is perfect then this is controlled by the Galois 1-cohomology
of Gv (ks ) (see Remark 7.3(iv) or [10]). Things only get more complicated when one considers
the G(k)-orbits that are contained in G · v. Orbits of actions of reductive groups over nonalgebraically closed fields have come under increasing attention, particularly from number
theorists. For instance, suppose k is a global function field, let v ∈ V (k) and let C be the
set of all w ∈ V (k) such that w is G(kν )-conjugate to v for every completion kν of k; then
Conrad showed that C is a finite union of G(k)-orbits [19, Thm. 1.3.3]. Bremigan studied the
strong topology of the orbits when k is a local field [17] (see also Remark 7.3(v) below). Now
let V be the Lie algebra g with the adjoint action of G. When k is perfect, J. Levy proved
that if x ∈ g, then any two elements of the form lima→0 λ(a) · x that are semisimple are G(k)conjugate [32, 33]. We extend Levy’s result below, see Remark 4.4(ii). W. Hoffmann asked
a question about limits lima→0 λ(a) · x when k is a global field of arbitrary characteristic, see
Remark 7.7(ii). Both Levy and Hoffmann were motivated by constructions involving orbital
integrals and the Selberg trace formula.
In this paper we develop a theory of G(k)-orbits for k an arbitrary field, building on earlier
work of three of us with Tange [9]. To make our results as general as possible, we do not
endow the field k with any extra structure. Moreover, we need not always assume that v is a
k-point. The most difficult problems arise when k is non-perfect. However, even for groups
over perfect fields our methods apply; see Theorem 1.6(ii). Observe that for perfect fields
the hypotheses on the stabilizers needed for our main results hold automatically (e.g., see
Proposition 7.4).
As noted above, the concept of a closed orbit is fundamental in geometric invariant theory
over algebraically closed fields. A first problem is to devise a suitable analogue of this idea
for G(k)-orbits. One can define the notion of a k-orbit over G [15, 10.2, Def. 4], or study the
Zariski closure of a G(k)-orbit, but such constructions do not appear to be helpful here (cf.
the discussion in [9, Rem. 3.9]). Instead we adopt an approach involving cocharacters. Let
2
Yk (G) denote the set of k-defined cocharacters of G. In [9, Def. 3.8], we made the following
definition.
Definition 1.1. The orbit G(k) · v is cocharacter-closed over k provided for all λ ∈ Yk (G),
if v ′ := lima→0 λ(a) · v exists, then v ′ ∈ G(k) · v.
We now extend this definition to cover arbitrary subsets of V , and introduce the cocharacterclosure of a subset of V :
Definition 1.2. (a) Given a subset X of V , we say that X is cocharacter-closed (over k) if
for every v ∈ X and every λ ∈ Yk (G) such that v ′ := lima→0 λ(a) · v exists, v ′ ∈ X. Note
that this definition coincides with the one above if X = G(k) · v for some v ∈ V .
c
(b) Given a subset X of V , we define the cocharacter-closure of X (over k), denoted X , to be
c
c
the smallest subset of V such that X ⊆ X and X is cocharacter-closed over k. (This makes
sense because the intersection of cocharacter-closed subsets is clearly cocharacter-closed.)
It follows from the Hilbert-Mumford Theorem that G · v is cocharacter-closed over k if
c
and only if G · v is closed. It is obvious that G · v is contained in G · v. Note, however, that
this containment can be proper: e.g., see Example 11.1.
Our first main result is a rational version of the Hilbert-Mumford Theorem.
Theorem 1.3. Let v ∈ V . Then there is a unique cocharacter-closed G(k)-orbit O inside
c
G(k) · v . Moreover, there exists λ ∈ Yk (G) such that lima→0 λ(a) · v exists and lies in O.
By a standard fact, the closure of a geometric G-orbit is again a union of G-orbits, [13,
I 1.8 Prop.]. Thanks to Lemma 3.3(i), the rational counterpart holds for the cocharacterclosure of a G(k)-orbit in V . Therefore, we can mimic the usual “degeneration” partial order
on the G-orbits in V in this rational setting:
c
Definition 1.4. Given v, v ′ ∈ V , we write G(k) · v ′ ≺ G(k) · v if v ′ ∈ G(k) · v .
Then it is clear that ≺ is reflexive and transitive, so ≺ gives a preorder on the set of
G(k)-orbits in V . In general, the behavior of the G(k)-orbits can be quite pathological; e.g.,
see Example 3.4, Remark 7.3(i), Example 7.6(ii) and Remark 7.7(ii).
Our second main result holds under some mild hypotheses on the stabilizer Gv of v in G.
Theorem 1.5. Let v ∈ V and suppose that Gv is k-defined. Then the following hold:
(i) If G · v is Zariski-closed, then G(k) · v is cocharacter-closed over k.
(ii) Let k ′ /k be an algebraic field extension and suppose that G(k ′ )·v is cocharacter-closed
over k ′ . Then G(k) · v is cocharacter-closed over k. Moreover, the converse holds
provided that v ∈ V (k) and k ′ /k is separable.
(iii) Let S be a k-defined torus of Gv and set L = CG (S). Then G(k) · v is cocharacterclosed over k if and only if L(k) · v is cocharacter-closed over k.
(iv) Let w ∈ V and suppose that both G(k) · w ≺ G(k) · v and G(k) · v ≺ G(k) · w. Then
G(k) · v = G(k) · w.
In fact, we prove stronger versions of the results of Theorem 1.5 in Proposition 5.5, Theorems 5.7 and 7.1 and Corollary 7.2 below. Note that the rationality condition on the
centralizer Gv in Theorem 1.5 is satisfied in many instances, e.g., if v ∈ V (k) and k is perfect
(see Proposition 7.4).
3
Recall that G is k-anisotropic provided Yk (G) = {0}. Part (i) of the next theorem gives
a characterization of k-anisotropic reductive groups over an arbitrary field k in terms of
cocharacter-closed orbits. In the special case when k is perfect, we recover in part (ii) a
result of Kempf [29, Thm. 4.2].
Theorem 1.6.
(i) G is k-anisotropic if and only if for every k-defined affine G-variety
W and every w ∈ W (k), the orbit G(k) · w is cocharacter-closed over k.
(ii) Suppose k is perfect. Then G is k-anisotropic if and only if for every k-defined affine
G-variety W and every w ∈ W (k), the orbit G · w is closed in W .
Part (ii) of Theorem 1.6 follows from part (i), Theorem 1.5(ii) and the Hilbert-Mumford
Theorem. Observe that it suffices to consider the case when W is a k-defined rational Gmodule, cf. Remark 2.3. Characterizing anisotropy over perfect fields in terms of closed orbits
was a question of Borel, [12, Rem. 8.8 (d)]. As noted above, this question was answered by
Kempf. Birkes [11] proved Theorem 1.6(ii) over the reals and number fields, and Bremigan
[17] proved it for p-adic fields. Note that Theorem 1.6(ii) fails for non-perfect fields; see
Remark 5.9(ii).
The notion of a cocharacter-closed G(k)-orbit has already proved very useful in the context
of Serre’s notion of G-complete reducibility over k, see Section 9. In [9, Thm. 5.9] we gave
a geometric characterisation of the latter using the former. In Theorem 9.3, we strengthen
this result by removing the connectedness assumption on G from [9, Thm. 5.9].
In Corollary 9.5, we prove a general Galois descent result for G-complete reducibility
for arbitrary algebraic field extensions k ′ /k under very mild assumptions on char(k) which
guarantee smoothness of centralizers of subgroups, [26, Thm. 1.1]. General results of this
nature were previously only known when both fields are algebraically closed or perfect, [4,
Thms. 5.3, 5.8], or else when the extension k ′ /k is separable, [9, Thm. 5.11].
The paper is organized as follows. We spend most of Section 2 recalling some results
from [9]. In Section 3, we discuss the concept of cocharacter-closure in detail, and introduce
the notion of accessibility of G(k)-orbits and its relation to the cocharacter-closure of a
G(k)-orbit, see Lemma 3.3.
Theorem 1.3 is proved in Section 4 (Theorem 4.3). This is followed in Section 5 by a
discussion of various ascent/descent results for field extensions on the one hand and for Levi
subgroups on the other: see Theorems 5.4 and 5.7, which prove Theorem 1.5(ii) (second
assertion) and (iii). A technical result needed in the proof of Theorem 5.7 is postponed to
Theorem 6.1 in Section 6.
Section 7 addresses questions of geometric G-conjugacy versus rational G-conjugacy. In
particular, here we prove Theorem 1.5(i), (ii) (first assertion) and (iv) (see Corollary 7.2).
We close the section with a discussion of when the preorder ≺ from Definition 1.4 is a partial
order. It turns out that this is closely related to a descent result for geometric conjugacy,
see Corollary 7.15. In Section 8, we give an application of our results from Sections 5 and 7:
we show that Galois ascent and certain conjugacy results hold for general G provided they
hold for GLn .
In Section 9, we discuss applications to Serre’s notion of G-complete reducibility over
k. Under some mild rationality assumptions we obtain corresponding Galois and Levi ascent/descent results, see Corollary 9.7.
4
In Section 10, we discuss the notion of cocharacter-closure in the classical context of conjugacy of endomorphisms under the general linear group in great detail. Here the preorder
from Definition 1.4 is automatically antisymmetric (Corollary 10.3). We then consider modules over finitely generated k-algebras. Our methods give a geometric way to explore the
notion of degenerations of modules over a non-algebraically closed field.
We finish (Section 11) with some examples of unipotent classes in G2 and a representation
of SL2 , demonstrating that the notions of Zariski-closure and cocharacter-closure already
differ for an algebraically closed field.
2. Preliminaries
2.1. Basic notation. Let k be a field, let k denote a fixed algebraic closure, and let ks ⊆ k
denote the separable closure and ki ⊆ k the purely inseparable closure of k. Note that
ks = k if k is perfect. We denote the Galois group Gal(ks /k) = Gal(k/k) by Γ. We use the
notion of a k-scheme from [13, AG.11]: a k-scheme is a k-scheme together with a k-structure.
So k-schemes are assumed to be of finite type and reduced separated k-schemes are called
k-varieties. Furthermore, a subscheme of a scheme V over k or over k is always a subscheme
of V as a scheme over k and points of V are always closed points of V as a scheme over k.
By “variety” we mean “variety over k”.
Now let V be a k-variety. If k1 /k is an algebraic extension, then we write V (k1 ) for the
set of k1 -points of V . If W is a subvariety of V , then we set W (k1) = W (k) ∩ V (k1 ). Here
we do not assume that W is k-defined, so W (k1 ) can be empty even when k1 = ks . The
Galois group Γ acts on V (ks ); see, e.g., [53, 11.2]. Recall the Galois criterion for a closed
subvariety W of V to be k-defined: W is k-defined if and only if W (k) contains a Γ-stable
subset of V (ks ) which is dense in W (see [13, Thm. AG.14.4]).
2.2. Algebraic groups. All linear algebraic groups are assumed to be smooth. Let H be a
k-defined linear algebraic group. By a subgroup of H we mean a smooth subgroup. We let
Z(H) denote the centre of H and H 0 the connected component of H that contains 1. Recall
that H has a k-defined maximal torus [13, 18.2(i) Thm.]. For K a subgroup of H, we denote
the centralizer of K in H by CH (K).
For the set of cocharacters (one-parameter subgroups) of H we write Y (H); the elements
∗
of Y (H) are the homomorphisms from the multiplicative group k to H. We denote the
set of k-defined cocharacters by Yk (H). There is a left action of H on Y (H) given by
∗
(h · λ)(a) = hλ(a)h−1 for λ ∈ Y (H), h ∈ H and a ∈ k . The subset Yk (H) is stabilized by
H(k).
The unipotent radical of H is denoted Ru (H); it is the maximal connected normal unipotent subgroup of H. The algebraic group H is called reductive if Ru (H) = {1}. Note that
we allow a reductive group to be non-connected.
2.3. Reductive groups. Throughout the paper, G denotes a k-defined reductive algebraic
group, possibly disconnected. The crucial idea which allows us to deal with non-connected
groups is the introduction of so-called Richardson parabolic subgroups (R-parabolic subgroups)
of a reductive group G. We briefly recall the main definitions and results; for more details
and further results, the reader is referred to [4, Sec. 6] and [9, Sec. 2.2].
5
Definition 2.1. For each cocharacter λ ∈ Y (G), let Pλ = {g ∈ G | lim λ(a)gλ(a)−1 exists}
a→0
(see Section 2.4 for the definition of limit). Recall that a subgroup P of G is parabolic if G/P
is a complete variety. The subgroup Pλ is parabolic in this sense, but the converse is not
true in general. If we define Lλ = {g ∈ G | lim λ(a)gλ(a)−1 = g}, then Pλ = Lλ ⋉ Ru (Pλ ),
a→0
and we also have Ru (Pλ ) = {g ∈ G | lim λ(a)gλ(a)−1 = 1}. The subgroups Pλ for λ ∈ Y (G)
a→0
are called the Richardson parabolic (or R-parabolic) subgroups of G. Given an R-parabolic
subgroup P , a Richardson Levi (or R-Levi ) subgroup of P is any subgroup Lλ such that
λ ∈ Y (G) and P = Pλ .
If G is connected, then the R-parabolic subgroups (resp. R-Levi subgroups of R-parabolic
subgroups) of G are exactly the parabolic subgroups (resp. Levi subgroups of parabolic
subgroups) of G; indeed, most of the theory of parabolic subgroups and Levi subgroups
of connected reductive groups—including rationality properties—carries over to R-parabolic
and R-Levi subgroups of arbitrary reductive groups. In particular, Ru (P )(k) acts simply
transitively on the set of k-defined R-Levi subgroups of a k-defined R-parabolic subgroup
P . If P, Q are R-parabolic subgroups of G and P 0 = Q0 , then Ru (P ) = Ru (Q). Given
any maximal torus T of an R-parabolic subgroup P , there is a unique R-Levi subgroup L
of P such that T ⊆ L, and if P is k-defined then L is k-defined if and only if T is. If λ is
k-defined then Pλ and Lλ are k-defined. Conversely, if G is connected and P is a k-defined Rparabolic subgroup of G with a k-defined Levi subgroup L then there exists λ ∈ Yk (G) such
that P = Pλ and L = Lλ . (See [9, Rem. 2.4] for a counter-example with G non-connected.)
If H is a subgroup of G, then there is an obvious inclusion Y (H) ⊆ Y (G) of the sets of
cocharacters. When H is reductive and λ ∈ Y (H), there is then an R-parabolic subgroup
of H associated to λ, as well as an R-parabolic subgroup of G. In order to distinguish
between R-parabolic subgroups associated to different subgroups of G, we use the notation
Pλ (H), Lλ (H), etc., where necessary, but we write Pλ for Pλ (G) and Lλ for Lλ (G). Note
that Pλ (H) = Pλ ∩ H, Lλ (H) = Lλ ∩ H and Ru (Pλ (H)) = Ru (Pλ ) ∩ H.
More generally, for H ⊆ G a closed subgroup which is not necessarily reductive and λ ∈
Y (G) a cocharacter normalizing H, we define Pλ (H) = H ∩Pλ and Ru (Pλ (H)) = H ∩Ru (Pλ ).
We recall a smoothness result from [20, Prop. 2.1.8(3) and Rem. 2.1.11] which holds for these
intersections.
Proposition 2.2. Let H ⊆ G be a closed subgroup, and let λ ∈ Y (G) be a cocharacter that
normalizes H. Then the scheme-theoretic intersections H ∩ Pλ , H ∩ Ru (Pλ ) are smooth and
coincide with Pλ (H), Ru (Pλ (H)) respectively.
2.4. G-varieties and limits. Throughout the paper, V denotes an affine k-defined Gvariety. This means that we assume both V and the action of G on V are k-defined. By
a rational G-module, we mean a finite-dimensional vector space over k with a linear Gaction. If both the G-module and the action are k-defined, then we say the rational Gmodule is k-defined. For a subgroup H of G, we denote the set of H-fixed points in V by
V H = {v ∈ V | h · v = v for all h ∈ H}.
For v ∈ V , let Gv denote the (set-theoretic) stabilizer of v, and let G0v = (Gv )0 denote its
identity component. Let G(k) · v denote the orbit of v under G(k); we call this the rational
orbit. We write G · v for G(k) · v and call this the geometric orbit. We say that G · v is
6
separable provided that the orbit map G → G · v is a separable morphism. Equivalently,
G · v is separable if and only if the scheme-theoretic stabilizer Gv of v in G is smooth.
∗
For each cocharacter λ ∈ Y (G), we define a morphism of varieties φv,λ : k → V via the
formula φv,λ (a) = λ(a) · v. If this morphism extends to a morphism φbv,λ : k → V , then we
say that lim λ(a) · v exists, and set this limit equal to φbv,λ (0); note that such an extension, if
a→0
it exists, is necessarily unique. We sometimes call v ′ the limit of v along λ. If X is a closed
G-stable subset of V and v ∈ X then lim λ(a) · v belongs to X.
a→0
For λ ∈ Yk (G) we say that λ destabilizes v over k (for G) provided lim λ(a) · v exists, and
a→0
if lim λ(a) · v exists and does not belong to G(k) · v, then we say λ properly destabilizes v
a→0
over k (for G).
Remark 2.3. Sometimes we want to reduce the case of a general (k-defined) affine G-variety
V to the case of a (k-defined) rational G-module V0 . Such a reduction is possible, thanks to
[29, Lem. 1.1(a)], for example: given V , there is a k-defined G-equivariant embedding of V
inside some V0 . We set up some standard notation which is in force throughout the paper.
Let V be a rational G-module. Given λ ∈ Y (G) and n ∈ Z, we define
∗
Vλ,n := {v ∈ V | λ(a) · v = an v for all a ∈ k },
X
X
X
Vλ,≥0 :=
Vλ,n , Vλ,>0 :=
Vλ,n and Vλ,<0 :=
Vλ,n .
n≥0
n>0
n<0
Then Vλ,≥0 consists of the vectors v ∈ V such that lim λ(a) · v exists, Vλ,>0 is the subset of
a→0
vectors v ∈ V such that lim λ(a) · v = 0, and Vλ,0 is the subset of vectors v ∈ V such that
a→0
lim λ(a) · v = v. Furthermore, the limit map v 7→ lim λ(a) · v is nothing but the projection of
a→0
a→0
Vλ,≥0 with kernel Vλ,>0 and image Vλ,0 . If the G-module V is k-defined, then each Vλ,n and
Vλ,>0 , etc., is k-defined (cf. [13, II.5.2]).
It is sometimes possible to pass from a geometric point in V to an element of V n (k), using
the following technical lemma.
Lemma 2.4. Suppose V is a k-defined rational G-module. Let v ∈ V and let k1 /k be a
finiteP
field extension such that v ∈ V (k1 ). Let α1 , . . . , αn ∈ k1 be a basis for k1 over k. Write
v = i αi vi for certain (unique) vi ∈ V (k), and set v = (v1 , . . . , vn ) ∈ V n (k). Let G act
diagonally on V n . Then the following assertions hold:
(i) For λ ∈ Yk (G), the limit lima→0 λ(a) · v exists if and only if the limit lima→0 λ(a) · v
exists.
(ii) Let λ ∈ Yk (G) and suppose that the limits v ′ = lima→0 λ(a) · v and v′ = lima→0 λ(a) · v
exist. Then for any g ∈ G(k), we have v ′ = g · v if and only if v′ = g · v.
(iii) We have Gv ⊆ Gv and Gv (k) = Gv (k).
Proof. Parts (i) and (ii) are [9, Lem. 2.16], while part (iii) follows directly from the definitions.
The following result asserts that when considering fixed points one may approximate ksplit tori by k-defined cocharacters.
7
Lemma 2.5. Given a k-split torus S of G, there exists µ ∈ Yk (S) such that CG (S) = Lµ
and V S = V Im(µ) .
Proof. Let W be a k-defined finite-dimensional rational G-module. Since S is k-split, we
can write W as a direct sum of S-weight spaces W = W0 ⊕ · · · ⊕ Wr for some r, with
the corresponding weights α0 , . . . , αr of S, where α0 is the trivial weight. Given i > 0, set
Ci := {σ ∈ Yk (S) ⊗ Q | hσ, αi i = 0} (where we extend the pairing in the obvious way). Since
αi is non-trivial for each i > 0, the subspace Ci is proper for each i. Hence the complement
in Yk (S)⊗Q of C1 ∪...∪Cr is non-empty. Taking a point in this complement and multiplying
by a suitably large integer, we can find a cocharacter µ ∈ Yk (S) such that hµ, αi i =
6 0 for all
S
Im(µ)
i > 0. Then W = W
.
Now consider the affine G-variety X := G × V , where G acts on the first factor by
conjugation and on the second factor by the given action of G on V . By embedding X Gequivariantly in a rational G-module W and applying the argument in the first paragraph,
we can find a cocharacter µ ∈ Yk (S) such that X S = X Im(µ) . But X S = CG (S) × V S and
X Im(µ) = Lµ × V Im(µ) , which gives the result.
Next we recall a transitivity result for limits along commuting cocharacters.
Lemma 2.6. Let v ∈ V . Suppose λ, µ ∈ Y (G) are commuting cocharacters and that v ′ :=
lima→0 λ(a) · v and v ′′ := lima→0 µ(a) · v ′ both exist. Then lima→0 (nλ + µ)(a) · v = v ′′ for all
sufficiently large n.
Proof. It suffices to prove this when V is a vector space, and then the result is contained in
[9, Lem. 2.15].
The next result is similar to Lemma 2.6, but the assumption is that the limits of v exist
along both λ and µ.
Lemma 2.7. Let v ∈ V . Suppose λ, µ ∈ Y (G) are commuting cocharacters and suppose
v ∈ V is such that v ′ := lima→0 λ(a) · v and v ′′ := lima→0 µ(a) · v both exist. Let m, n ∈ N.
Then lima→0 λ(a) · v ′′ , lima→0 µ(a) · v ′ and lima→0 (mλ + nµ)(a) · v all exist and these limits
are equal.
Proof. It suffices to prove this when V is a vector space. Then, using the notation from
Remark 2.3, we have v ∈ Vλ,≥0 , the set of all points for which λ acts with a non-negative
weight. This set is closed in V and, since λ and µ commute, it is µ-stable. Therefore
v ′′ := lima→0 µ(a) · v ∈ Vλ,≥0 , and hence the limit along λ for v ′′ exists. Similarly, v ′ ∈ Vµ,≥0 ,
and hence the limit along µ for v ′ exists. Clearly v ∈ U := Vλ,≥0 ∩ Vµ,≥0 , so the limit along
mµ + nµ for v also exists.
Restricting attention to the subspace U, taking the limit along λ is the same as projecting
onto Uλ,0 := U ∩ Vλ,0 , and taking the limit along µ is the same as projecting onto Uµ,0 :=
U ∩ Vµ,0 . The combination of taking the limit along λ and then along µ is the same as
taking the limit along µ and then along λ, and the result is simply the projection of U onto
Uλ,0 ∩ Uµ,0 . It is clear that the same is true for taking the limit along mµ + nµ for v.
Lemma 2.8. Let v ∈ V and let S be a maximal k-defined torus of Gv . Suppose λ ∈
Yk (CG (S)) destabilizes v and does not fix v. Then λ properly destabilizes v over k for G.
8
Proof. Let v ′ := lima→0 λ(a) · v. Then S and Im(λ) generate a k-defined torus S ′ of Gv′ .
Now Im(λ) is not contained in S because λ does not fix v, so dim S ′ > dim S. But S is a
maximal k-defined torus of Gv , so v and v ′ cannot be G(k)-conjugate.
2.5. Results about Ru (Pλ )-conjugacy. First we recall two of the main results from [9]:
Theorem 2.9 ([9, Thm. 3.1]). Let v ∈ V such that Gv (ks ) is Γ-stable and let λ ∈ Yk (G)
such that v ′ := lima→0 λ(a) · v exists and is Ru (Pλ )(ks )-conjugate to v. Then v ′ is Ru (Pλ )(k)conjugate to v.
Theorem 2.10 ([9, Thm. 3.3]). Let v ∈ V and suppose k is perfect. Let λ ∈ Yk (G) such
that v ′ := lim λ(a) · v exists and is G(k)-conjugate to v. Then v ′ is Ru (Pλ )(k)-conjugate to
a→0
v.
Theorem 2.10 was first proved by H. Kraft and J. Kuttler for k algebraically closed of
characteristic zero in case V = G/H is an affine homogeneous space, cf. [49, Prop. 2.4] or
[24, Prop. 2.1.2].
The following result is an analogue of Theorem 2.9 involving descent to Levi subgroups;
the proof involves very similar ideas.
Proposition 2.11. Let v ∈ V . Suppose S is a k-defined torus of Gv , L = CG (S) and
λ ∈ Yk (L) is such that v ′ := lima→0 λ(a) · v exists and is Ru (Pλ )(k)-conjugate to v. Then v ′
is Ru (Pλ (L))(k)-conjugate to v.
Proof. Set P = Pλ and let u ∈ Ru (P )(k) be such that v ′ = u · v. Then u−1 · λ fixes v. Let
H be the subgroup of Gv generated by the image of u−1 · λ and S; then H is a k-defined
subgroup of Gv , and hence contains a k-defined maximal torus S ′ with S ⊆ S ′ . Moreover,
since λ commutes with S and Pu−1 ·λ = P , we have H ⊆ P ∩ Gv . Since u−1 · λ is a k-defined
cocharacter of H, we can find h ∈ H(k) such that µ := hu−1 · λ ∈ Yk (S ′ ). Since S ⊆ S ′ , we
have µ ∈ Yk (L). Now λ, µ ∈ Yk (L) give rise to the same parabolic subgroup P = Pλ = Pµ
of G, and hence Pλ (L) = Pµ (L). The R-Levi subgroups Lλ (L) and Lµ (L) are conjugate by
an element u0 ∈ Ru (Pλ (L))(k), i.e., u0 Lµ (L)u0 −1 = Lλ (L). We claim that u0 · µ = λ. To see
this, note that since u0 ∈ P , u0 Lµ u0 −1 is an R-Levi subgroup of P , and this Levi subgroup
contains the image of λ. Since Ru (P ) acts simply transitively on the set of R-Levi subgroups
of P , there is only one R-Levi subgroup of P containing the image of λ, namely Lλ itself,
and hence u0 Lµ u0 −1 = Lu0 ·µ = Lλ . But now, since u0 , h and u are all elements of P , we
have p := u0 hu−1 ∈ P and u0 · µ = p · λ. Writing p = u1 l, with u1 ∈ Ru (P ) and l ∈ Lλ , we
have p · λ = u1 · λ and
Lλ = Lp·λ = Lu1 ·λ = u1 Lλ u1 −1 .
Appealing to the simple transitivity of the action of Ru (P ) on the R-Levi subgroups of P
again, we see that u1 = 1, so u0 · µ = p · λ = λ, as required. Finally, we have found
u0 ∈ Ru (Pλ (L))(k) such that v ′ := lima→0 λ(a) · v exists and µ = u0 −1 · λ fixes v, so by [9,
Lem. 2.12], we have v ′ = u0 · v, which completes the proof.
2.6. k-rank. Let H be a subgroup of G, not necessarily k-defined. It makes sense to speak
of k-defined (or k-split) subgroups of H: a subgroup of H is k-defined (k-split) if it is kdefined (k-split) as a subgroup of the k-defined group G. Likewise we can speak of k-defined
cocharacters of H. The notion of maximal k-defined (or k-split) torus of H has the usual
meaning.
9
Below we will need the following result for H of the form Gv for some v ∈ V . Note that
even when v ∈ V (k), Gv need not be k-defined.
Lemma 2.12. Let H be a subgroup of G. Then any two maximal k-split tori of H are
H(k)-conjugate.
T
e be the closure of
e
Proof. Let H
γ∈Γ γ · H(ks ). Then H is k-defined by the Galois criterion.
e so S is a subgroup of H
e as S(ks ) is dense
If S is a k-defined torus of H then S(ks ) ⊆ H,
in S. Standard results for k-defined groups [53, Thm. 15.2.6] imply that any two maximal
e are H(k)-conjugate.
e
k-split tori of H
The assertion of the lemma now follows.
Definition 2.13. We define the k-rank of H to be the dimension of a maximal k-split torus
of H.
3. Cocharacter-closure
Recall the definitions of a cocharacter-closed G(k)-orbit and of a cocharacter-closed subset of V from the Introduction, Definitions 1.1 and 1.2. We begin with a few elementary
observations.
Remarks 3.1. (i). Note that the notions in Definition 1.2 depend on the given action of G
on V , because they are made with reference to the limits along cocharacters of G.
(ii). A closed G-stable set is cocharacter-closed.
(iii). The cocharacter-closed subsets of V form the closed sets of a topology on V (it is clear
that arbitrary intersections and unions of cocharacter-closed sets are cocharacter-closed, and
that the empty set and the whole space V are cocharacter-closed).
(iv). Clearly, if a set is cocharacter-closed over k, it is cocharacter-closed over k0 where k0
is any subfield of k (such that V is a k0 -defined G-variety).
c
(v) Let X ⊆ V . If X is not G(k)-stable then X need not be G(k)-stable. For example,
suppose G acts freely on V and let X = {v} for some v ∈ V . Then v is not destabilized by
any non-zero λ ∈ Y (G): for if v ′ = lima→0 λ(a) · v then λ fixes v ′ , which forces λ = 0 by the
c
freeness of the action. Hence X = X = {v}.
Recall the definition of the preorder ≺ from Definition 1.4. Examples 11.1 below compares
this preorder with the usual degeneration order on orbits in the case k = k. See also Example
3.4. We have the following related notion.
Definition 3.2. Suppose v, v ′ ∈ V . We say that the orbit G(k) · v ′ is 1-accessible from
G(k) · v if there exists λ ∈ Yk (G) such that lima→0 λ(a) · v exists and lies in G(k) · v ′ .
Similarly, for n ≥ 1, we say that G(k) · v ′ is n-accessible from G(k) · v provided there exists
a finite sequence of G(k)-orbits G(k) · v = G(k) · v1, G(k) · v2, . . . , G(k) · vn+1 = G(k) · v ′ with
G(k) · vi+1 1-accessible from G(k) · vi for each 1 ≤ i ≤ n. We say G(k) · v ′ is accessible from
G(k) · v if it is n-accessible for some n ≥ 1.
Note that this definition does not depend on the chosen representative of G(k) · v since if
lima→0 λ(a)·v = v ′′ ∈ G(k)·v ′ and g ∈ G(k) then lima→0 (g ·λ)(a)·(g ·v) = g ·v ′′ ∈ G(k)·v ′ . It
is clear from the definitions that if G(k)·v ′ is 1-accessible from G(k)·v then G(k)·v ′ ≺ G(k)·v.
Example 11.2 below shows that the converse to this is not true, but we do have the following:
Lemma 3.3. Suppose v ∈ V .
10
S
c
(i) We have G(k) · v = G(k) · v ′ , where the union is taken over all v ′ ∈ V such that
G(k) · v ′ is accessible from G(k) · v.
(ii) The preorder ≺ coincides with the accessibility relation; that is, given v, v ′ ∈ V ,
G(k) · v ′ ≺ G(k) · v if and only if there exists a finite sequence from G(k) · v to
G(k) · v ′ as in Definition 3.2.
Proof. (i). Let X denote the union defined above. Then given v ′ ∈ X, G(k) · v ′ is accessible
from G(k) · v. But now if lima→0 λ(a) · v ′ = v ′′ exists for some λ ∈ Yk (G), then G(k) · v ′′ is
1-accessible from G(k) · v ′ , and hence G(k) · v ′′ is accessible from G(k) · v, so v ′′ ∈ X. Hence
c
X is cocharacter-closed over k. Since G(k) · v ⊆ X, we have G(k) · v ⊆ X. But the reverse
inclusion is clear: by definition, the cocharacter-closure of G(k) · v must contain all orbits
1-accessible from G(k) · v, and all orbits 1-accessible from those orbits, and so on.
(ii). This follows from (i).
The following elementary example illustrates some of the complexities that can arise, even
over a field of characteristic 0.
Example 3.4. Let k = R and consider the group G = Gm acting on V = A1 by
a · z := a2 z.
The group G(k) = Gm (k) is just the multiplicative group of the field R, and there are three
orbits of G(k) on k-points of V : G(k) · (−1) = {x ∈ R | x < 0}, G(k) · 0 = {0} and
c
c
G(k) · 1 = {x ∈ R | x > 0}. We have G(k) · (−1) = G(k) · (−1) ∪ {0} and G(k) · 1 =
G(k) · 1 ∪ {0}. On the other hand, since the non-zero G(k)-orbits G(k) · 1 and G(k) · (−1)
are both infinite subsets of V , their Zariski closures are the whole of A1 . We also have
G · 1 = G · (−1) = {z ∈ A1 | z 6= 0}. This gives an example of how the cocharacter-closure
isn’t the same as the closure (or the closure intersected with the set of k-points) and how
different parts of the same G-orbit may be inaccessible from each other when viewed as
G(k)-orbits.
For more examples, see Sections 10 and 11.
4. The rational Hilbert-Mumford Theorem
In this section we prove Theorem 1.3. We start with a key technical result.
Proposition 4.1. Let v ∈ V . Suppose λ ∈ Yk (G) is such that v ′ := lima→0 λ(a) · v exists
but is not Ru (Pλ )(k)-conjugate to v. Let S be a k-split torus of Gv . Then there exists
τ ∈ Yk (CG (S)) such that ṽ := lima→0 τ (a) · v exists and lies outside G(k) · v. Moreover, there
exists a k-split torus S̃ in Gṽ such that dim S̃ > dim S.
Proof. It does no harm to assume that S is a maximal k-split torus of Gv : else S ⊂ S ′ for a
maximal k-split torus S ′ of Gv and then Yk (CG (S ′ )) ⊆ Yk (CG (S)), so the result for S follows
from the result for S ′ . Using Lemma 2.5, we can find µ ∈ Yk (S) such that V S = V Im(µ) and
L := CG (S) = Lµ . Now let T be a maximally split k-defined maximal torus of Pλ ∩Pµ . There
exists u ∈ Ru (Pλ )(k) such that u · λ ∈ Yk (T ). Moreover, lima→0 (u · λ)(a) · v = u · v ′ exists
and cannot be Ru (Pλ )(k)-conjugate to v, since then v ′ would also be Ru (Pλ )(k)-conjugate
to v, contradicting our hypothesis. Hence we may replace λ with u · λ and v ′ with u · v ′ if
we like and assume that λ ∈ Yk (T ).
11
Now there exists u1 ∈ Ru (Pµ )(k) such that ν := u1 · µ belongs to Yk (T ), and since µ
fixes v, lima→0 ν(a) · v = u1 · v exists. Since λ and ν belong to Yk (T ), they commute. By
Lemma 2.7, v ′′ := lima→0 (λ + nν)(a) · v exists for all positive integers n, and this limit does
not depend on n. Choosing n sufficiently large, we may assume that Pλ+nν ⊆ Pν = Pµ and
Ru (Pµ ) ⊆ Ru (Pλ+nν ) [9, Lem. 2.15]. In particular, u1 ∈ Ru (Pλ+nν )(k). Let σ = λ + nν for
such a choice of n. Note that, since v ′′ can be obtained as a limit along λ and as a limit along
ν (Lemma 2.7), λ and ν both fix v ′′ . Now ν = u1 · µ; so u1 Su1−1 ⊆ Gv′′ , since V S = V Im(µ) ,
and Im(λ) commutes with u1 Su1 −1 , since Lµ = CG (S). Hence Im(λ) and u1 Su1 −1 generate
a k-defined torus S ′′ of Gv′′ of dimension at least as large as the dimension of S. Note that
S ′′ is k-split, by [13, Thm. 15.4 (i)].
If S ′′ = u1 Su1 −1 , then Im(λ) ⊆ u1 Su1 −1 , and therefore λ fixes u1 · v. Since λ evaluates
in Pµ , u1 ∈ Ru (Pµ )(k) and Ru (Pµ ) is a closed connected normal subgroup of Pµ , we can
write u1 = xu2 with x ∈ P−λ (k), u2 ∈ Ru (Pλ )(k) by [53, Thm. 13.4.2, Cor. 13.4.4]. But
now [9, Lem. 2.13] implies that v ′ = u2 · v ∈ Ru (Pλ )(k) · v, which contradicts the hypothesis
on λ. Hence we must conclude that dim S ′′ > dim S. In particular, v and v ′′ are not
G(k)-conjugate, since S is a maximal k-split torus of Gv .
Finally, recall that u1 ∈ Ru (Pσ )(k), so ṽ := lima→0 (u1 −1 · σ)(a) · v = u1 −1 · v ′′ exists and is
also not G(k)-conjugate to v. But u1 −1 · σ = u1 −1 · λ + nµ ∈ Yk (Lµ ) = Yk (CG (S)) so, setting
τ = u1 −1 · σ and S̃ = u1 −1 S ′′ u1 , we are done.
We note here a consequence of Proposition 4.1 which we use at the end of Section 5.
Corollary 4.2. Let v ∈ V . If Gv contains a maximal k-split torus of G, then G(k) · v is
cocharacter-closed over k.
Proof. Let S be a maximal k-split torus of G contained in Gv . The result now follows from
Proposition 4.1 applied to S.
We can now give our generalisation of the Hilbert-Mumford Theorem. Recall in particular
the notions of cocharacter-closed and cocharacter-closure, Definitions 1.1 and 1.2, the preorder ≺ on the G(k)-orbits of V , Definition 1.4, and the relationship of these notions with
the notion of accessibility, Lemma 3.3. The key result, which proves Theorem 1.3, is the
following.
Theorem 4.3 (The rational Hilbert-Mumford Theorem). Suppose v ∈ V .
(i) There is a unique cocharacter-closed (over k) G(k)-orbit 1-accessible from G(k) · v.
c
(ii) The orbit from part (i) is the unique cocharacter-closed G(k)-orbit in G(k) · v .
c
(iii) If G(k) · v ′ ≺ G(k) · v, then the unique cocharacter-closed G(k)-orbits in G(k) · v and
c
G(k) · v ′ are equal.
Proof. (i). We first show existence. If G(k)·v is already cocharacter-closed over k, then there
is nothing to do. Otherwise, there exists λ ∈ Yk (G) such that v ′ = lima→0 λ(a) · v exists and
is not G(k)-conjugate to v. By Proposition 4.1, we can even assume that λ ∈ Yk (CG (S)),
where S is a maximal k-split torus of Gv , and that the k-rank of Gv′ is strictly greater than
the k-rank of Gv . If G(k) · v ′ is cocharacter-closed over k then we are done. If not, repeat
the process to find µ ∈ Yk (CG (S1 )) such that v ′′ = lima→0 µ(a) · v ′ exists and Gv′′ has strictly
greater k-rank than Gv′ . Note that, since µ ∈ Yk (CG (S1 )) and Im(λ) ⊆ S1 , λ and µ commute,
so by Lemma 2.6 G(k) · v ′′ is 1-accessible from G(k) · v. Again, if G(k) · v ′′ is cocharacterclosed over k then we are done. Otherwise, repeat the process again. Since the k-rank of the
12
stabilizer increases at each step, the process must terminate at a cocharacter-closed orbit
which is 1-accessible from G(k) · v.
For uniqueness, suppose G(k)·v1 and G(k)·v2 are two cocharacter-closed orbits 1-accessible
from G(k) · v. Choose λ1 , λ2 ∈ Yk (G) such that lima→0 λi (a) · v ∈ G(k) · vi for i = 1, 2.
Without loss we can assume that lima→0 λi (a) · v = vi for i = 1, 2. Then Pλ1 ∩ Pλ2 contains
a maximally split k-defined maximal torus of G and we can conjugate λ1 (resp. λ2 ) by an
element of Ru (Pλ1 )(k) (resp. an element of Ru (Pλ2 )(k)) so that λ1 and λ2 commute. Let
v ′ = lima→0 λ1 (a) · v2 = lima→0 λ2 (a) · v1 (these limits exist and are equal by Lemma 2.7).
Since G(k) · vi is cocharacter-closed over k for i = 1, 2, v ′ is G(k)-conjugate to v1 and v2 , and
hence G(k) · v1 = G(k) · v ′ = G(k) · v2 , as required. This completes the proof of (i).
(ii). Suppose G(k) · v ′ is the unique cocharacter-closed orbit 1-accessible from G(k) · v, as
c
provided by part (i), and suppose G(k) · v ′′ is another cocharacter-closed orbit in G(k) · v .
By Lemma 3.3(i), there is a finite sequence G(k) · v = G(k) · v1 , G(k) · v2 , . . . , G(k) · vn =
G(k) · v ′′ of orbits with G(k) · vi+1 1-accessible from G(k) · vi for each 1 ≤ i ≤ n − 1.
By choosing our representatives suitably, we have cocharacters λ and µ ∈ Yk (G) such that
lima→0 λ(a) · vn−2 = vn−1 and lima→0 µ(a) · vn−1 = vn = v ′′ . Now, by the usual argument, we
can find u1 ∈ Ru (Pλ )(k) and u2 ∈ Ru (Pµ )(k) such that σ := u1 · λ and τ := u2 · µ commute.
Moreover, lima→0 σ(a) · vn−1 = u1 · vn−1 and lima→0 τ (a) · vn−1 = u2 · v ′′ . Hence by Lemma 2.7,
lim σ(a) · (u2 · v ′′ ) exists and equals lim τ (a) · (u1 · vn−1 ).
a→0
a→0
′′
Call this common limit w. Since the orbit G(k) · v is cocharacter-closed over k, we have that
G(k) · w = G(k) · v ′′ . But, since u1 · vn−1 = lima→0 σ(a) · vn−2 and σ and τ commute, we can
apply Lemma 2.6 to conclude that there exists n ∈ N such that lima→0 (nσ + τ )(a) · vn−2 =
w ∈ G(k) · v ′′ ; in particular, G(k) · v ′′ is 1-accessible from G(k) · vn−2 . Continuing in this
way, we conclude that G(k) · v ′′ is 1-accessible from G(k) · v and hence G(k) · v ′′ = G(k) · v ′
by the uniqueness in part (i). This completes the proof of (ii).
c
c
c
(iii). If G(k) · v ′ ≺ G(k) · v, then v ′ ∈ G(k) · v and hence G(k) · v ′ ⊆ G(k) · v . This
c
c
means that any cocharacter-closed orbit in G(k) · v ′ is also contained in G(k) · v , and the
uniqueness from part (ii) gives the result.
Remarks 4.4. (i). See Example 10.7 for a linear algebra example illustrating the uniqueness
of the cocharacter-closed orbit in the cocharacter-closure.
(ii). In [33, Thm. 3.4], Levy considers the case where k is perfect and proves the following
result: Let v ∈ V (k) and assume that two limits s1 = lima→0 λ(a) · v, s2 = lima→0 µ(a) · v
along k-defined cocharacters λ and µ exist. Further assume that both G · s1 and G · s2 are
Zariski-closed. Then s1 and s2 are G(k)-conjugate. Since k is assumed to be perfect, it is
known that the orbits through s1 and s2 are Zariski-closed if and only if they are cocharacterclosed over k, compare Theorem 5.7 and Proposition 7.4(i) below. Hence Levy’s conjugacy
result follows from the uniqueness assertion of Theorem 4.3, and moreover holds for arbitrary
(possibly non-rational) points v ∈ V .
The following result is a rational version of the form of the Hilbert-Mumford Theorem
given in [29, Thm. 1.4]. In our setting, it is a direct consequence of Theorem 4.3.
Corollary 4.5. Suppose X is a G(k)-stable cocharacter-closed subset of V which meets
c
G(k) · v . Then there exists λ ∈ Yk (G) such that lima→0 λ(a) · v exists and lies in X.
13
c
Proof. Since X ∩ G(k) · v is G(k)-stable, it contains a G(k)-orbit. Since the intersection
is cocharacter-closed, it contains the cocharacter-closure of this orbit, and hence it contains
the corresponding unique cocharacter-closed orbit provided by Theorem 4.3(ii). However,
this cocharacter-closed orbit must also be the unique cocharacter-closed orbit contained in
c
G(k) · v , by Theorem 4.3(iii). Since this orbit is 1-accessible from G(k) · v by Theorem
4.3(i), we get the result required.
Remark 4.6. We do not know whether the strengthened version of the Hilbert-Mumford
Theorem discussed in the Introduction can be extended to arbitrary fields. For more on this,
see [9, Sec. 1].
5. Ascent and descent
Given a field extension k ′ /k, we use the terminology “Galois descent” to refer to a result which guarantees that a given G(k)-orbit is cocharacter-closed over k provided the
corresponding G(k ′ )-orbit is cocharacter-closed over k ′ , and vice versa for “Galois ascent”.
Likewise, we refer to “Levi descent” if the result at hand implies that a given L(k)-orbit
(for L a k-defined Levi subgroup) is cocharacter-closed over k provided the corresponding
G(k)-orbit is cocharacter-closed over k, and vice versa for “Levi ascent”.
We begin this section with a further consequence of Proposition 4.1.
Corollary 5.1. Let v ∈ V . Suppose G(k) · v is cocharacter-closed over k. Then for all
λ ∈ Yk (G) such that v ′ := lima→0 λ(a) · v exists, v ′ is Ru (Pλ )(k)-conjugate to v.
Proof. Suppose, for contradiction, that we have λ ∈ Yk (G) such that v ′ exists but is not
Ru (Pλ )(k)-conjugate to v. Then, by Proposition 4.1, there exists a point v ′′ which can be
obtained as the limit of v along a k-defined cocharacter of G but is not G(k)-conjugate to
v. But this contradicts the assumption that G(k) · v is cocharacter-closed over k.
Remark 5.2. Corollary 5.1 generalizes [9, Thm. 3.10] by removing the connectedness assumption on G.
As a direct consequence of Corollary 5.1, we see that cocharacter-closedness only depends
on the identity component of G.
Corollary 5.3. Let v ∈ V . The orbit G(k) · v is cocharacter-closed over k if and only if
G0 (k) · v is cocharacter-closed over k.
Proof. It is immediate from the definition that if G0 (k) · v is cocharacter-closed over k, then
G(k) · v is cocharacter-closed over k. Conversely, suppose G0 (k) · v is not cocharacter-closed
over k. Then there exists λ ∈ Yk (G0 ) = Yk (G) such that λ destabilizes v and lima→0 λ(a) · v
does not belong to Ru (Pλ (G0 ))(k) · v = Ru (Pλ (G))(k) · v. It follows from Corollary 5.1 that
G(k) · v is not cocharacter-closed over k, as required.
Here is our main result on Levi descent and split Levi ascent; it proves Theorem 1.5(iii).
Theorem 5.4. Let v ∈ V . Suppose S is a k-defined torus of Gv and set L = CG (S).
(i) If G(k) · v is cocharacter-closed over k, then L(k) · v is cocharacter-closed over k.
(ii) If S is k-split, then G(k) · v is cocharacter-closed over k if and only if L(k) · v is
cocharacter-closed over k.
14
Proof. (i). Suppose G(k) · v is cocharacter-closed over k, and let λ ∈ Yk (L) be such that
v ′ := lima→0 λ(a)·v exists. Then, by Corollary 5.1, v ′ is Ru (Pλ )(k)-conjugate to v. Now apply
Proposition 2.11 to get that v ′ is Ru (Pλ (L))(k)-conjugate—and hence L(k)-conjugate—to v.
(ii). The forward implication follows from part (i). Now suppose that G(k) · v is not
cocharacter-closed over k. Then, by Proposition 4.1, there exists τ ∈ Yk (L) such that
lima→0 τ (a) · v exists but is not G(k)-conjugate to v. But then this limit cannot be L(k)conjugate to v, and hence L(k) · v is not cocharacter-closed over k.
Our next result—which proves the second assertion of Theorem 1.5(ii)—is Galois descent
for separable algebraic extensions.
Proposition 5.5. Let v ∈ V such that Gv (ks ) is Γ-stable and let k ′ /k be a separable algebraic
extension. If G(k ′ ) · v is cocharacter-closed over k ′ then G(k) · v is cocharacter-closed over
k.
Proof. Suppose G(k ′ ) is cocharacter-closed over k ′ . Let λ ∈ Yk (G) such that v ′ := lima→0 λ(a)·
v exists. By Corollary 5.1, v ′ is Ru (Pλ )(k ′ )-conjugate to v. In particular, v ′ is Ru (Pλ )(ks )conjugate to v. The result now follows from Theorem 2.9.
It turns out to be more subtle to prove converse statements to Proposition 5.5—i.e.,
Galois ascent results—under appropriate hypotheses on v. Such a result would follow for
v ∈ V (k) if we could prove a rational version of the strengthened Hilbert-Mumford Theorem
(cf. Remark 4.6): for then the optimal ks -defined cocharacter λopt would be Γ-fixed and
hence would be k-defined. As a first lemma, we prove that G(k) · v is not cocharacter-closed
over k, provided the limit along some ks -defined cocharacter lies outside the geometric orbit.
Lemma 5.6. Let v ∈ V (k). Suppose there exists λ ∈ Yks (G) such that λ properly destabilizes
v over k̄. Then G(k) · v is not cocharacter-closed over k.
Proof. Let v ′ = lima→0 λ(a) · v ∈ VS
(ks ) and let Y ⊆ V (ks ) be the set of Galois-conjugates of
′
v ; note that Y is finite. Set S = y∈Y G · y. Since Y is finite, S is closed. This closed set
is also G-stable, Γ-stable and ks -defined, hence k-defined. Note also that since λ properly
destabilises v to v ′ , the dimension of the G-orbit of v ′ is strictly smaller than that of v,
and hence all G-orbits in S have dimension strictly smaller than that of G · v. This implies
that S ∩ G · v = ∅; in particular, S ∩ G(k) · v = ∅. Now, using the terminology of [9, §4],
v is uniformly S-unstable over ks , so we may thus apply [9, Cor. 4.9] to conclude that v
is uniformly S-unstable over k (note that {v} is k-closed). In other words, there exists a
k-defined cocharacter µ with lima→0 µ(a) · v ∈ S. The claim follows.
We get stronger ascent and descent results under some assumptions of k-definability on
Gv . In general, the stabilizer Gv is only ki -defined but not k-defined. If Gv is k-defined,
then we can say more. In particular, for v ∈ V (k), the stabilizer Gv is k-defined if G · v is
separable or if k is perfect, see Proposition 7.4. The following result proves Theorem 1.5(ii)
(second assertion) and (iii).
Theorem 5.7. Let v ∈ V . Suppose Gv has a maximal torus that is k-defined. Then the
following hold.
(i) Suppose v ∈ V (k). For any separable algebraic extension k ′ /k, G(k ′ )·v is cocharacterclosed over k ′ if and only if G(k) · v is cocharacter-closed over k.
15
(ii) Let S be a k-defined torus of Gv and let L = CG (S). Then L(k) · v is cocharacterclosed over k if and only if G(k) · v is cocharacter-closed over k.
Proof. The forward implication of (i) follows from Proposition 5.5. We prove the reverse
implication of (i) using induction on dim G. By Remark 2.3, we may assume that V is a kdefined rational G-module. The result holds trivially if dim G = 0. Assume the result holds
for any G′ such that dim G′ < dim G. Let k ′ /k be a separable algebraic extension. Suppose
G(k ′ )·v is not cocharacter-closed over k ′ . We prove that G(k)·v is not cocharacter-closed over
k. By Proposition 5.5, we can assume that k ′ = ks . So some λ ∈ Yks (G) properly destabilizes
v over ks . If λ properly destabilizes v over k̄, then G(k) · v is not cocharacter-closed over k,
by Lemma 5.6.
So suppose λ does not properly destabilize v over k̄. Then u · λ centralizes v for some
u ∈ Ru (Pλ ), by Theorem 2.10 for k̄. By hypothesis, there exists a k-defined maximal torus S ′
of Gv . Set G′ = CG (S ′ ), a reductive k-defined subgroup of G. Since λ properly destabilizes
v over ks , u · λ 6= λ, which implies that Im(u · λ) is not contained in Z(G0 ). It follows
that S ′ 6⊆ Z(G0 ), so dim G′ < dim G. By Theorem 5.4(ii) applied to the ks -split torus S ′ ,
G′ (ks ) · v is not cocharacter-closed over ks . By the induction hypothesis, G′ (k) · v is not
cocharacter-closed over k. So choose µ ∈ Yk (G′ ) such that v ′ := lima→0 µ(a) · v exists and is
not G′ (k)-conjugate to v. Then in particular, µ does not fix v. Hence µ properly destabilizes
v over k for G, by Lemma 2.8, and we are done.
Finally, we prove part (ii). We want to apply (i) to both G and L, so we first check that
Lv has a maximal torus that is k-defined. By assumption, Gv has a maximal torus T that is
e and T is
e be as in the proof of Lemma 2.12; then S ⊆ H
k-defined. Let H = Gv and let H
e
e
a maximal torus of H. As S and H are k-defined, we can choose a k-defined maximal torus
e containing S. Then dim T ′ = dim T , so T ′ is a maximal torus of both Gv and Lv ,
T ′ of H
as required.
If v ∈ V (k) then the result follows from Theorem 5.4. More generally, if k is infinite then
we can first apply Lemma 2.4 and replace an arbitrary v with some v ∈ V (k)n (note that
S ⊆ Gv as S(k) is dense in S). For arbitrary k and v we need the following argument,
which relies on constructions from Section 6. We use Theorem 6.1 below to replace V with
another k-defined G-variety W and v with some w ∈ W (k). By Theorem 6.1(i), L(k) · v is
cocharacter-closed if and only if L(k) · w is so, and likewise for G(k) · v. Moreover, Theorem
6.1(ii) assures that S ⊆ Gw and that Gw has a maximal torus that is k-defined. We may
thus apply Theorem 5.7(i) and assume k = ks . But then S is k-split, so the result follows
from Theorem 5.4.
Corollary 5.8. Let v ∈ V . If Gv contains a maximal k-torus of G, then G(k) · v is
cocharacter-closed over k.
Proof. Suppose Gv contains a maximal k-torus S of G. Then S is a k-defined maximal torus
of G, so S is a k-defined maximal torus of Gv . By Theorem 5.7(i), it is enough to show that
G(ks ) · v is cocharacter-closed over ks . But this follows from Corollary 4.2, as S splits over
ks .
We turn now to the proof of Theorem 1.6(i).
Proof of Theorem 1.6(i). Suppose G is k-anisotropic. Let W be an affine k-defined G-variety
and let w ∈ W . Since Yk (G) = {0}, the G(k)-orbit of w in W is cocharacter-closed over k.
16
(Note that we do not require that w ∈ W (k) here.) The converse follows from the argument
in the proof of [11, Lem. 10.1].
Remarks 5.9. (i). We recover the result (cf. [14, Cor. 3.8]) that if k is perfect and G is
k-anisotropic, then every element in G(k) is semisimple. For if g ∈ G(k) is not semisimple,
then its G(k)-conjugacy class in G is not closed, contradicting Theorem 1.6(ii).
(ii). Theorem 1.6(ii) fails for non-perfect fields. In [23, p. 488], Gille and QuéguinerMathieu give an example of a k-anisotropic semisimple group G of the form G = PGL1 (A),
where A is a simple central division algebra over k containing a field K such that K/k
is purely inseparable. Moreover, G contains a smooth unipotent subgroup admitting nontrivial k-points. If 1 6= g ∈ G(k) is unipotent, then its G(k)-conjugacy class is not closed.
So the conclusion of Theorem 1.6(ii) does not hold in this instance.
We finish with the following result, which is an easy consequence of Corollary 4.2 and
Theorem 1.5(ii); it shows that [11, Property (B)] holds for any perfect field (cf. [11, Thm.
7.2, Thm. 8.2]).
Proposition 5.10. Suppose k is perfect. Let v ∈ V . If Gv contains a maximal k-split torus
of G, then G · v is closed.
Remark 5.11. Proposition 5.10 fails for non-perfect fields: Remark 5.9(ii) gives an example
of this with G anisotropic. The assertion of Proposition 5.10 does hold for any k, however,
if we assume that G is split: for if Gv contains a maximal k-split torus S of G then S is a
maximal torus of G, so G · v is closed by Theorem 5.7(ii) applied to the field k.
6. Reduction to k-points
In this section we prove the following result which makes it possible to pass from a geometric point in V to a rational point in another affine variety W .
Theorem 6.1. Let v ∈ V . Then there exists an affine G-variety W and w ∈ W (k) with the
following properties:
(i) For any reductive k-defined subgroup M of G, M(k) · v is cocharacter-closed over k
if and only if M(k) · w is cocharacter-closed over k.
(ii) Let H ⊆ G be a connected k-defined subgroup. Then H ⊆ Gv if and only if H ⊆ Gw .
To prove Theorem 6.1 we require a series of lemmas. We first reduce from geometric points
to ks -points.
Lemma 6.2. To prove Theorem 6.1, we may assume v ∈ V (ks ).
Proof. By Remark 2.3, we may assume that V is a k-defined rational G-module. Let k1 /ks
be a finite field extension such that v ∈ V (k1 ). We now apply Lemma 2.4 to produce
v ∈ V n (ks ) for suitable n. It follows from Lemma 2.4(i) and (ii) (for the field ks ) that
M(k) · v is cocharacter-closed if and only if M(k) · v is so.
Suppose H is a connected k-defined subgroup of G. If H is contained in Gv , then by Lemma
2.4(iii) we have H ⊆ Gv ⊆ Gv . Conversely, suppose that H ⊆ Gv . Since H(ks ) is dense in
H, it is enough to show H(ks ) ⊆ Gv , which follows from the equality Gv (ks ) = Gv (ks ). We
conclude that we may replace V by V n and v by v to prove Theorem 6.1.
17
The next step is to make v ∈ V (ks ) become stable under the action of the Galois group.
We may achieve this by passing from V to the quotient by a finite group, using the following
result.
Lemma 6.3. Let F be a finite abstract group, regarded as a k-defined algebraic group endowed
with the usual k-structure (so that F = F (k)). Suppose F × G has a k-defined action on
V . Then the action of G on V descends to give a k-defined action of G on V /F . Let
π : V → V /F be the canonical G-equivariant projection. Then the following hold for v ∈ V :
(i) Let λ ∈ Y (G). Then λ destabilizes v if and only if λ destabilizes π(v).
(ii) For any reductive k-defined subgroup M ⊆ G, M(k) · v is cocharacter-closed over k
if and only if M(k) · π(v) is cocharacter-closed over k.
(iii) Let H ⊆ G be a connected subgroup. Then H ⊆ Gv if and only if H ⊆ Gπ(v) .
Proof. We first note that V /F is defined over k by [1, 2.2], so the statement of the proposition
makes sense.
(i). The forward direction is obvious. So suppose λ destabilizes π(v). By Remark 2.3,
there is a G-equivariant closed embedding of V in an F × G-module W . Let πW : W → W/F
be the canonical projection. We have an induced G-equivariant map φ from V /F to W/F ,
and λ destabilizes φ(π(v)) = πW (v). Hence it is enough to prove the result when V = W .
Clearly, we can take G to be Im(λ). Let V1 = Vλ,≥0 and let V2 = Vλ,<0 . Then V = V1 ⊕ V2
and V1 and V2 are both F -stable, since F commutes with G. The projection V → V2 is
G-equivariant and gives rise to a G-equivariant map V /F → V2 /F . Let π2 : V2 → V2 /F be
the canonical projection.
The group G acts on the dual space V2∗ . Let X1 , . . . , Xm ∈ V2∗ be a basis such that each
Xi is a weight vector for λ. Since the weights of λ on V2 are all negative, these weights
are all positive. We can regard the Xi as regular functions on V2 . Choose a generating set
f1 , . . . , fr for the ring of invariants k[V2 ]F . We can assume that each fj is a polynomial in
the Xi with no constant term and, since k[V2 ]F is G-stable, we can assume also that each fj
is a weight vector for λ. Clearly each of these weights is positive.
Write v = (v1 , v2 ). Since π(v) is destabilized by λ, so is π2 (v2 ). If f ∈ k[V2 ]F has weight
m > 0 with respect to λ, then f (λ(a) · π2 (v2 )) = (λ(a−1 ) · f )(π2 (v2 )) = a−m f (v2 ), and
so lima→0 λ(a) · π2 (v2 ) can only exist if f (v2 ) = 0. It follows that fj (v2 ) = 0 = fj (0) for
all j. This implies that π2 (v2 ) = π2 (0). But F is finite, so we must have v2 = 0. Hence
v = (v1 , 0) ∈ V1 and so λ destabilizes v, as required.
(ii). Suppose M(k) · v is cocharacter-closed over k. Let λ ∈ Yk (M) ⊆ Yk (G) such that
λ destabilizes π(v). Then λ destabilizes v by part (i), so there exists g ∈ M(k) such that
lima→0 λ(a) · v = g · v. Now g · π(v) = π(g · v) = π (lima→0 λ(a) · v) = lima→0 λ(a) · π(v). It
follows that M(k) · π(v) is cocharacter-closed over k.
Conversely, suppose M(k) · v is not cocharacter-closed over k. Then (F × M)(k) · v is not
cocharacter-closed over k, by Corollary 5.3. Hence there exists λ ∈ Yk (M) such that v ′ :=
lima→0 λ(a)·v exists and is not (F ×M)(k)-conjugate to v. Now π(v ′ ) = lima→0 λ(a)·π(v). If
π(v ′ ) is M(k)-conjugate to π(v), say g ·π(v) = π(v ′ ), then π(g ·v) = π(v ′ ), so g ·v ∈ F ·v ′ . But
then v ′ is (F × M)(k)-conjugate to v, a contradiction. Hence M(k) · π(v) is not cocharacterclosed over k.
(iii). We have Gv ⊆ Gπ(v) , hence one assertion of (iii) is clear. Conversely, suppose that
H ⊆ Gπ(v) is a connected subgroup. Then the orbit map H → H · v has image in the finite
set F · v. As H is connected, this implies that H ⊆ Gv , as required.
18
We are now in a position to prove Theorem 6.1.
Proof of Theorem 6.1. By Lemma 6.2, we may assume v ∈ V (ks ). Let 1 = γ1 , . . . , γr ∈ Γ be
chosen such that {v, γ2(v), . . . , γr (v)} is the Γ-orbit through v. Set v = (v, γ2 (v), . . . , γr (v)) ∈
V (ks )r . We first show that we may replace v by v, where we consider the diagonal action
of M on V r . Let λ ∈ Yk (M). As γ(λ) = λ for all γ ∈ Γ, we have that λ destabilizes v
if and only if it destabilizes the tuple v. If λ destabilizes v and lima→0 λ(a) · v = v ′ , then
lima→0 λ(a) · v = (γ1 (v ′ ), . . . , γr (v ′ )) =: v′ . For g ∈ M(k), the equation γ(g) = g for all γ ∈ Γ
implies that g · v = v ′ if and only if g · v = v′ . In particular, M(k) · v is cocharacter-closed
over k if and only if M(k) · v is so.
Now let H ⊆ G be a k-defined connected subgroup. If H ⊆ Gv , then clearly H ⊆ Gv .
Conversely, suppose H ⊆ Gv and let h ∈ H(ks ). As γ(h) ∈ H(ks ) for all γ ∈ Γ, we get
h ∈ Gv . Since H(ks ) is dense in H, this implies that H ⊆ Gv . It thus suffices to construct
W and w and prove the assertions in (i) and (ii) for v.
Let F = Sr be the symmetric group on r letters, and let F act on V r by permuting the
coordinates. Then the action of F commutes with the diagonal action of G, hence F ×G acts
on V r . As the assumptions of Lemma 6.3 are satisfied, we may replace V r by W = V r /Sr
and v by w = π(v) ∈ W (ks ). Now by construction, γ(w) = w for all γ ∈ Γ, hence w is a
k-point. This finishes the proof.
7. Geometric and rational conjugacy
Our next theorem allows us to relate geometric G-conjugacy to rational Ru (Pλ )-conjugacy,
provided the orbit has a k-defined stabilizer. The result relies on Theorem 2.10, but here we
do not require k to be perfect.
Theorem 7.1. Let v ∈ V and suppose that G0v is k-defined. Let λ ∈ Yk (G). Suppose that
v ′ = lima→0 λ(a) · v exists and is G-conjugate to v. Then v ′ is Ru (Pλ )(k)-conjugate to v.
Proof. Let g ∈ G satisfy v ′ = g · v. By Theorem 2.10, we can find u ∈ Ru (Pλ ) such that
v ′ = u−1 · v. Then u · λ is a cocharacter of G0v .
We first claim that Pu·λ(G0v ) is k-defined. Indeed, by Proposition 2.2 this group coincides
with the scheme-theoretic intersections G0v ∩Pu·λ = G0v ∩Pλ , which are therefore smooth. According to our assumptions both G0v and Pλ are k-defined, and thus the (smooth) intersection
is k-defined by [53, Prop. 12.1.5].
By [13, Thm. 18.2], Pu·λ (G0v ) contains a k-defined maximal torus S, which automatically
is a maximal torus of G0v . Then S is contained in some k-defined maximal torus T of Pu·λ .
Since T is contained in a k-defined R-Levi subgroup of Pu·λ and all such subgroups are
Ru (Pλ )(k)-conjugate, we can find x ∈ Ru (Pλ )(k) such that S ⊆ Lx·λ . After replacing λ with
x · λ and v ′ with x · v ′ , we may assume without loss of generality that λ centralises S. This
forces S to be contained in Gv′ , since v ′ = lima→0 λ(a) · v. By assumption, Gv′ is G-conjugate
to Gv and thus has the same rank. In particular, S is a maximal torus of Gv′ . But as λ is a
cocharacter of Gv′ that commutes with S, we deduce that Im(λ) ⊆ S ⊆ Gv . Hence v ′ = v,
which finishes the proof.
We record a number of consequences of Theorem 7.1 which include Theorem 1.5(i), (ii)
(first assertion) and (iv).
Corollary 7.2. Let v ∈ V and suppose that G0v is k-defined.
19
(i) Let λ ∈ Yk (G). Suppose that v ′ = lima→0 λ(a) · v exists and is G-conjugate to v. Then
v ′ is G(k)-conjugate to v.
(ii) Suppose G · v is Zariski-closed. Then G(k) · v is cocharacter-closed over k.
(iii) Let k ′ /k be an algebraic field extension and suppose that G(k ′ )·v is cocharacter-closed
over k ′ . Then G(k) · v is cocharacter-closed over k.
(iv) Let S be a k-defined torus of Gv and set L := CG (S). Let λ ∈ Yk (L) such that
v ′ := lima→0 λ(a) · v exists and v ′ ∈ G · v. Then v ′ ∈ Ru (Pλ (L))(k) · v.
(v) Let w ∈ V and suppose that both G(k) · w ≺ G(k) · v and G(k) · v ≺ G(k) · w. Then
G(k) · v = G(k) · w.
Proof. Parts (i) and (ii) are immediate consequences of Theorem 7.1. Part (iii) follows from
(i) and the fact that G(k ′ )-conjugacy implies G-conjugacy.
Part (iv) follows from Theorem 7.1 and Proposition 2.11.
For (v), by assumption, there exist cocharacters λ1 , . . . , λn ∈ Yk (G) and elements v1 , . . . ,
vn+1 ∈ V such that v1 = v, vn+1 = g · v for some g ∈ G(k), vj = g ′ · w for some j and some
g ′ ∈ G(k), and lima→0 λi (a) · vi = vi+1 for 1 ≤ i ≤ n. As v1 and vn+1 are G-conjugate, all of
the vi are G-conjugate (for if vi 6∈ G · vi−1 , then, as vi lies in the closure of G · vi−1 , the orbit
G · vi has strictly smaller dimension than G · vi−1 , which forces all of the subsequent orbits
G · vi , . . . , G · vn+1 to have smaller dimension than G · v). By (i), v2 is G(k)-conjugate to v1 .
In particular, G0v2 is again k-defined. Repeating the argument for v3 and so on, we find that
w is G(k)-conjugate to v, as required.
Remarks 7.3. (i). Note that Theorem 7.1 fails without the assumption on G0v . E.g., see [9,
Rem. 5.10] for the failure of Corollary 7.2(ii) without this assumption.
(ii). Suppose that G0v is k-defined and that G(k) · v is not cocharacter-closed over k. By
Theorem 7.1, there exists a k-defined cocharacter λ such that v ′ = lima→0 λ(a) · v exists and
does not belong to G · v. In the terminology of [9, §4], v is uniformly S-unstable over k for
S := G · v ′ , and v does not belong to S (cf. the proof of Lemma 5.6). In particular, there
exist non-trivial cocharacters which belong to the optimal class for v with respect to S over
k, see [9, Thm. 4.5].
(iii). Corollary 7.2 refines the descent assertions in Theorem 5.4 and Proposition 5.5 (Levi
and Galois descent), albeit under the stronger hypothesis that G0v is k-defined.
(iv). Corollary 7.2(i) is an instance where G-conjugacy descends to G(k)-conjugacy. This
property may be studied in cohomological terms as follows: Suppose that v, v ′ ∈ V (k) are
k-points which are G-conjugate. This means that v ′ ∈ (G · v)(k). Thus G(k)-conjugacy
would follow provided that the map G(k) → (G · v)(k) is surjective. By [21, III, §4, Cor.
4.7] this is automatic if all Gv -torsors over k are trivial, where Gv is the scheme-theoretic
stabilizer of v in G. If we further assume that either k is perfect or that the orbit map
G → G · v is separable, then the isomorphism classes of Gv -torsors over k are parametrized
by the Galois cohomology H 1 (Γ, Gv ), where Γ = Gal(ks /k) (see [21, III, §5, Cor. 3.5,3.6]).
But even in the separable case, H 1 (Γ, Gv ) does not always vanish. Our assumption that v ′
arises as the limit along a k-defined cocharacter λ allows us to circumvent all these technical
difficulties and to avoid additional assumptions.
(v). Bremigan considers the case where the field k has characteristic 0 and is complete
under a non-trivial real absolute value, [17]. In [17, Prop. 5.3], he proves that if G · v is
Zariski-closed, then G(k) · v is Hausdorff-closed (i.e., closed in the topology induced by the
extra structure on k). It follows that G(k) · v is cocharacter-closed over k, as the limit
20
along a k-defined cocharacter belongs to the Hausdorff-closure of G(k) · v. Hence, we obtain
Corollary 7.2(ii) in this case.
(vi). Corollary 7.2(v) does not imply that the relation ≺ on all G(k)-orbits is antisymmetric. It does, however, assert that antisymmetry holds if one of the two orbits has a k-defined
stabilizer.
The following result shows that the assumption in Theorem 7.1 that G0v is k-defined is
automatic in many naturally-occurring cases (e.g., if char(k) = 0).
Proposition 7.4. Let v ∈ V and assume that one of the following conditions is satisfied:
(i) v ∈ V (k) and k is perfect;
(ii) v ∈ V (ks ), Gv (ks ) is Γ-stable and G · v is separable.
Then G0v is k-defined. In particular, the assertions of Theorem 7.1 and Corollary 7.2 hold
for v.
Proof. Case (i) is [53, Prop. 12.1.2]. For (ii), suppose that v ∈ V (ks ), that Gv (ks ) is Γ-stable
and that G · v is separable. Then G0v is k-defined. Indeed, separability implies that Gv is
ks -defined (see [53, loc. cit.]). Due to the assumption on Gv (ks ), the Galois criterion for
k-definedness implies that Gv is k-defined, hence so is G0v .
Remark 7.5. There are situations where G0v is k-defined for all k-points v ∈ V (k). For
instance, Proposition 7.4 above implies that this is the case whenever k is perfect or all
G(k)-orbits are separable. In this case ≺ is antisymmetric on the set {G(k) · v | v ∈ V (k)},
hence is a partial order on this set. We do not know whether antisymmetry holds in general.
See also Proposition 7.11 below.
We give two simple examples. The first example illustrates that having Gv k-defined does
not imply separability in general (cf. Proposition 7.4(ii)). The second example shows that
even for separable orbits, the converse of Corollary 7.2(ii) is false in general.
Examples 7.6. (i).
2, and
consider the
adjoint action of G on
Let
G = SL2 , char(k) =
0 1
1 a
∈ V (k). Then Gv =
a ∈ k is k-defined, but G · v is
V = sl2 . Let v =
0 0
0 1
not separable.
(ii). Let G = PGL2 , char(k) = 2, acting on V = pgl2 by conjugation. Let A denote the
image in PGL2 of A ∈ GL2 , and likewise for images of elements of gl2 . Define v ∈ V (k) by
0 1
1 0
2
−1
v :=
, where x ∈ k but x 6∈ k. Then v = gṽg , where g :=
and ṽ :=
x2 0
x 1
x 1
0 1
a 0
. Then lima→0 λ(a) · ṽ = 0,
=
. Define λ ∈ Yk (G) by λ(a) =
0 x
0 0
0 a−1
so G · ṽ is not closed and hence G · v is not closed. Moreover, G · v is separable. However, it
is easily checked that G(k) · v is cocharacter-closed over k (the unique Borel subgroup whose
Lie algebra contains v is not k-defined). Note that 0 lies in G · v \ G · v, so G · v \ G · v does
indeed have a k-point.
Remarks 7.7. (i). The non-separability of the orbit map G → G · v in Examples 7.6(i) is
because z(g) is non-zero for G = SL2 in characteristic 2, whereas the group-theoretic centre
Z(G) vanishes. However, as z(g) consists of semisimple elements, this does not affect the
separability of the Ru (Pλ )-orbit through v.
21
In general, let v ∈ V (ks ) such that Gv (ks ) is Γ-stable, and suppose λ ∈ Yk (G) is such that
v ′ := lima→0 λ(a) · v exists and lies in G · v. Further suppose that the orbit Ru (Pλ ) · v is
separable. Then we may still conclude that v ′ ∈ Ru (Pλ )(k) · v, i.e., the assertion of Theorem
7.1 holds in this case. Indeed, according to Theorem 2.10, v ′ ∈ Ru (Pλ ) · v, and v ′ is a ks point. Clearly, H 1 (Gal(ks /ks ), Ru (Pλ )v ) vanishes. Due to the separability assumption, this
means that the map Ru (Pλ )(ks ) → (Ru (Pλ ) · v)(ks ) is surjective (see Remark 7.3(iv)), hence
v ′ ∈ Ru (Pλ )(ks ) · v. By Theorem 2.9, v ′ ∈ Ru (Pλ )(k) · v.
(ii). Hoffmann asked the following question: let v ∈ V (k) and suppose that G · v \ G · v has
a k-point. Does there exist λ ∈ Yk (G) such that v ′ := lima→0 λ(a) · v exists and v ′ 6∈ G · v?
Clearly, the answer is no if G(k) · v is cocharacter-closed over k. Example 7.6(ii) shows that
this is possible under the present hypotheses, even if we assume G · v is separable. If we
assume that G(k) · v is not cocharacter-closed over k and that G0v is k-defined, the answer is
yes, by Corollary 7.2(i).
If v ′ = lima→0 λ(a) · v then we have seen there is a complicated relationship between the
property that v ′ lies in G(k) · v and the property that v ′ lies in Ru (Pλ )(k) · v. If the latter
holds then we can, for instance, apply Theorem 2.9 and Proposition 2.11. The following open
question seems technical but resolving it is crucial to gaining a more complete understanding
of the behavior of the G(k)-orbits.
Question 7.8. Let v ∈ V and λ ∈ Yk (G) such that v ′ := lima→0 λ(a) · v exists. Suppose that
v ′ is G(k)-conjugate to v. Is v ′ then Ru (Pλ )(k)-conjugate to v?
Remarks 7.9. (i). Theorem 2.10 implies that Question 7.8 has a positive answer whenever k
is perfect.
(ii). By Theorem 7.1, Question 7.8 has a positive answer whenever G0v is k-defined. In
particular, see Proposition 7.4 for separability assumptions that are sufficient to imply this.
(iii). Corollary 5.1 shows that Question 7.8 has an affirmative answer for cocharacterclosed orbits.
(iv). In the context of G-complete reducibility, the fact that Question 7.8 has a positive
answer for algebraically closed fields has been used for instance by Stewart ([54, Cor. 3.6.2])
and Uchiyama ([57, Prop. 3.6]). For another application, see [9, Cor. 3.5].
In order to study the question of whether ≺ is antisymmetric, we also state the following
weaker version of Question 7.8.
Question 7.10. Let v ∈ V and λ ∈ Yk (G) such that v ′ := lima→0 λ(a) · v exists. Suppose that
v ′ is G(k)-conjugate to v. Is v ′ then G0 (k)-conjugate to v?
A positive answer to this question is related to antisymmetry as follows.
Proposition 7.11. Suppose that the assertion of Question 7.10 holds for any reductive group
acting on an affine variety. Then the preorder ≺ on G(k)-orbits is antisymmetric, hence is
a partial order.
Proof. Suppose we have v, v ′ ∈ V with G(k) · v ′ ≺ G(k) · v ≺ G(k) · v ′ . By Lemma 3.3 we can
find v1 , . . . , vn ∈ V , λ1 , . . . , λn ∈ Yk (G) such that all limits lima→0 λi (a) · vi (1 ≤ i ≤ n) exist,
v = v1 , vi = lima→0 λi−1 (a) · vi−1 for 2 ≤ i ≤ n, v ′ = g ′ · vj for some 1 ≤ j ≤ n, g ′ ∈ G(k) and
lima→0 λn (vn ) = g ·v for some g ∈ G(k). Now let v = (v1 , . . . , vn ) ∈ V n and let G = Cn ⋉Gn .
22
Here Cn = hσi denotes the cyclic group of order n acting on Gn by permuting the indices
cyclically (σ(i) = i + 1 mod n). Let λ = (λ1 , . . . , λn ) ∈ Yk (Gn ) = Yk (G).
Then G acts naturally on V n , and
lim λ(a) · v = (lim λ1 (a) · v1 , . . . , lim λn (a) · vn )
a→0
a→0
a→0
= (v2 , . . . , vn , g · v) = ((1, . . . , 1, g)σ) · v ∈ G(k) · v.
Since we assume that Question 7.10 has an affirmative answer for G, we can find u =
(u1 , . . . , un ) ∈ G0 (k) = G(k)n such that u · v = (v2 , . . . , vn , g · v). Hence v ′ = g ′ · vj =
(g ′ uj−1 · · · u1 ) · v ∈ G(k) · v. We conclude that G(k) · v = G(k) · v ′ .
Remark 7.12. Note that the proof of Proposition 7.11 requires that Question 7.10 has an
affirmative answer for all groups of the form Cn ⋉ Gn acting on V n , n ∈ N. In particular,
even though the assertion of Question 7.10 holds trivially for connected G, this does not
imply the antisymmetry of ≺ for connected G.
In order to obtain a converse of Proposition 7.11, we need the following lemma.
Lemma 7.13. Let v ∈ V , λ ∈ Yk (G). Suppose that v ′ = lima→0 λ(a) · v exists and v ′ = g · v
for some g ∈ G(k). Then G0 (k) · v is accessible from G0 (k) · v ′ .
Proof. As G/G0 is finite, there exists some n ∈ N such that g n ∈ G0 (k). Taking the limit
along g · λ maps v ′ = g · v to g · v ′ = g 2 · v, and then taking the limit along g 2 · λ maps g 2 · v to
g 2 · v ′ = g 3 · v. Continuing in this way, we see that the orbit through g n · v is accessible from
v ′ . As g n ∈ G0 (k), this implies that G0 (k) · v is accessible from G0 (k) · v ′ , as claimed.
Proposition 7.14. Suppose that the preorder ≺ is antisymmetric on the set of G0 (k)-orbits
in V . Then the assertion of Question 7.10 holds for G.
Proof. Let v ∈ V , λ ∈ Yk (G) and suppose that v ′ = lima→0 λ(a) · v exists and v ′ = g · v with
g ∈ G(k). By Lemma 7.13, G0 (k) · v is accessible from G0 (k) · v ′ . Conversely, it is immediate
that G0 (k) · v ′ is accessible from G0 (k) · v. By the antisymmetry of ≺, we conclude that
G0 (k) · v = G0 (k) · v ′ . Hence v ′ ∈ G0 (k) · v, as required.
Combining Propositions 7.11 and 7.14 we obtain the following result.
Corollary 7.15. The following are equivalent (for fixed k):
(i) The assertion of Question 7.10 holds for any reductive group acting on any affine
variety;
(ii) the preorder ≺ is antisymmetric for any reductive group acting on any affine variety;
(iii) the preorder ≺ is antisymmetric for any connected reductive group acting on any
affine variety.
8. Reduction to GLn
In this section we use Theorem 7.1 to prove a result which reduces certain questions
involving G(k)-orbits to the special case when G = GLn for some n. We start by recalling a
standard construction from GIT, cf. [1, 5.3]. Let G, M be reductive k-groups with G ⊆ M
and let V be an affine G-variety over k. We define an action of G on M × V by g ·
(m, v) = (mg −1 , g · v). Write M ×G V for the quotient of M × V by this G-action and let
πG : M × V → M ×G V be the canonical projection. Then πG is k-defined [1, 2.2]. The
23
group M acts on M × V by left multiplication on the first factor and trivially on the second
factor; this descends to give an action of M on M ×G V . If G = M then there is an obvious
k-defined G-equivariant isomorphism from G ×G V to V . Projection onto the first factor
gives a k-defined M-equivariant morphism η : M ×G V → M/G, where M acts on the coset
space M/G by left multiplication.
Theorem 8.1. Let G, M be reductive k-groups with G ⊆ M and let V be an affine Gvariety over k. Let k ′ /k be an extension of fields. Set W = M ×G V and define φ : V → W
by φ(v) = πG (1, v) (note that φ is a k-defined G-equivariant closed embedding: see [1], for
example). Then the following hold:
(i) for all v ∈ V and all λ ∈ Yk′ (M), if λ destabilizes φ(v) then u · λ ∈ Yk′ (G) for some
u ∈ Ru (Pλ (M))(k ′ );
(ii) for all v ∈ V and all λ ∈ Yk′ (G), if v ′ := lima→0 λ(a) · v exists and m · φ(v) = φ(v ′ )
for some m ∈ M(k ′ ) then m ∈ G(k ′ );
(iii) Mφ(v) = Gφ(v) = Gv ;
(iv) for all v ∈ V and all λ ∈ Yk′ (G), λ destabilizes v if and only if λ destabilizes φ(v);
(v) for all v ∈ V and all λ ∈ Yk′ (G), λ properly destabilizes v over k ′ for G if and only
if λ properly destabilizes φ(v) over k ′ for G if and only if λ properly destabilizes φ(v)
over k ′ for M;
(vi) for all v ∈ V , G(k ′ ) · v is cocharacter-closed over k ′ if and only if G(k ′ ) · φ(v) is
cocharacter-closed over k ′ if and only if M(k ′ ) · φ(v) is cocharacter-closed over k ′ .
Proof. By extending scalars, we can regard G and M as k ′ -groups and V as a k ′ -defined
G-variety. Hence we can assume without loss that k ′ = k.
Now let v ∈ V , let λ ∈ Yk (M) and suppose λ destabilizes φ(v). Let x be the coset
G ∈ M/G. Since the map η : M ×G V → M/G induced by projection onto the first factor
is M-equivariant, λ destabilizes η(φ(v)) = x. Set x′ = lima→0 λ(a) · x. The orbit M · x is
closed—it is the whole of M/G—and Mx = G is k-defined, so M(k) · x is cocharacter-closed
over k, by Corollary 7.2(ii). Hence by Corollary 5.1, there exists u ∈ Ru (Pλ (M))(k) such
that x = u · x′ . Then u · λ fixes x, so u · λ is a cocharacter of Mx = G. This proves (i). If we
assume in addition that λ is already a cocharacter of G then x′ = x; hence if m ∈ M(k) and
m · φ(v) = φ(v ′) then m · x = x by the M-equivariance of η, so m ∈ G(k) and (ii) follows.
This also proves part (iii) (take λ = 0 in (ii)).
Parts (iv)–(vi) now follow immediately from (i)–(ii).
We now see that to prove Galois ascent and to answer Question 7.8, it suffices to consider
the special case G = GLn .
Corollary 8.2.
(i) If the answer to Question 7.8 is yes when G = GLn then the answer
is yes for all G.
(ii) If Galois ascent holds when G = GLn then it holds for all G.
Proof. We can embed G as a k-defined subgroup of some GLn . Define φ as in Theorem 8.1.
Part (i) now follows from Theorem 8.1. For part (ii), suppose G(ks ) · v is not cocharacterclosed over ks . We want to show that G(k) · v is not cocharacter-closed over k. Now
GLn (ks ) · φ(v) is not cocharacter-closed over ks , by Theorem 8.1(vii) (with k ′ = ks ). By
assumption, Galois ascent holds for GLn , so GLn (k) · φ(v) is not cocharacter-closed over k.
The result now follows from Theorem 8.1(vii) (with k ′ = k).
24
9. G-complete reducibility
In this section we apply our previous results to the theory of G-completely reducible
subgroups over k. This was the original motivation for much of our work, and is an important
source of examples. The notion of a G-completely reducible subgroup was introduced by
Serre [51]; there are applications to the subgroup structure of reductive groups [34], [35],
spherical buildings [50] and the Langlands correspondence [31, Sec. 13]. For more details,
see [4] and [51].
First we recall the relevant definitions, then we explain the link with geometric invariant
theory.
Definition 9.1. A subgroup H of G is said to be G-completely reducible (G-cr) if whenever
H is contained in an R-parabolic subgroup P of G, there exists an R-Levi subgroup of P
containing H. Similarly, a subgroup H of G is said to be G-completely reducible over k if
whenever H is contained in a k-defined R-parabolic subgroup P of G, there exists a k-defined
R-Levi subgroup of P containing H.
In [4, Cor. 3.7], we show that G-complete reducibility has a geometric interpretation in
terms of the action of G on Gn , the n-fold Cartesian product of G with itself, by simultaneous
conjugation.
Let h ∈ Gn and let H be the algebraic subgroup of G generated by h, i.e., by the
components of the n-tuple h. Then G · h is closed in Gn if and only if H is G-cr [4, Cor. 3.7].
To generalize this to subgroups that are not topologically finitely generated, we employed
the following concept [9, Def. 5.4].
Definition 9.2. Let H be a subgroup of G and let G ֒→ GLm be an embedding of algebraic
groups. Then h ∈ H n is called a generic tuple of H for the embedding G ֒→ GLm if h
generates the associative subalgebra of Matm spanned by H. We call h ∈ H n a generic tuple
of H if it is a generic tuple of H for some embedding G ֒→ GLm .
Note that generic tuples exist for any embedding G ֒→ GLm provided n is sufficiently
large. We give a characterization of G-complete reducibility over k in terms of geometric
invariant theory.
Theorem 9.3. Let H be a subgroup of G and let h ∈ H n be a generic tuple of H. Then H
is G-completely reducible over k if and only if G(k) · h is cocharacter-closed over k.
Proof. This follows from the proof of [9, Thm. 5.9] (which is the special case when G is
connected), replacing [9, Thm. 3.10] with Corollary 5.3.
Remark 9.4. Suppose the subgroup H of G is k-defined. We can pick a generic tuple h ∈
H(ks )n of H for some n, and without loss we can assume that Γ permutes the entries of
the tuple. Let W = H n /Sn , where Sn acts by permuting the entries of tuples, and let
πW : H n → W be the canonical projection. Then w := πW (h) is a k-point, and it follows
from Lemma 6.3 and Theorem 9.3 that H is G-cr over k if and only if G(k) · w is cocharacterclosed over k.
With this characterization in hand, we may apply our earlier rationality results. Recall
(cf. [26, Def. 2.11]) that a prime p is called pretty good for G provided it is a good prime
with the extra property that both the root lattice in the character group and the coroot
25
lattice in the cocharacter group have no p-torsion. A prime p which is very good for G is
automatically pretty good, but the converse fails in general.
Corollary 9.5. Suppose p = char(k) is a pretty good prime for G and let H be a k-defined
subgroup of G. Let k ′ /k be an algebraic field extension. If H is G-completely reducible over
k ′ , then H is G-completely reducible over k.
Proof. Let h be a generic tuple of H. By [8, Cor. 3.4], the orbit G · h is separable, provided
that p is a very good prime for G. It follows from [26, Thm. 1.1] that the condition may be
relaxed to require only that p is pretty good for G. Moreover, since H is k-defined, H(ks )
is dense in H, so we can choose h ∈ H(ks )n such that h is a generic tuple of H. Then
Gh (ks ) = CG (H)(ks) is Γ-stable. Hence the conditions of Proposition 7.4(ii) are satisfied,
and the result follows from Theorem 9.3 together with Corollary 7.2(iii).
Remark 9.6. The assertion of Corollary 9.5 is false for inseparable field extensions without
the assumption on p = char(k): see [8, Ex. 7.22] for an example in G2 with p = 2 and [57,
Thm. 1.10] for an example in E7 with p = 2. Corollary 9.5 shows that this kind of pathology
can occur only in small characteristic.
Theorem 5.7 and Remark 9.4 immediately yield the following results on Galois ascent/
descent and Levi ascent/descent for G-complete reducibility. We note that we are able to
obtain results of this nature under slightly different hypotheses using a building-theoretic
approach: see [6], [3].
Corollary 9.7. Let H be a k-defined subgroup of G such that CG (H) is k-defined. Then the
following hold.
(i) For any separable algebraic extension k ′ /k, H is G-completely reducible over k ′ if and
only if H is G-completely reducible over k.
(ii) For a k-defined torus S of CG (H) let L = CG (S). Then H is G-completely reducible
over k if and only if H is L-completely reducible over k.
The Levi ascent/descent statement of Corollary 9.7(ii) provides a different proof—valid
for non-connected G—of Serre’s result [50, Prop. 3.2].
The next result has been obtained by McNinch in [38, Prop. 4.1.1]. We give a different
proof, which relies on Theorem 7.1.
Corollary 9.8. Let H be a k-defined linearly reductive subgroup of G. Then H is Gcompletely reducible over k.
Proof. Since H is k-defined, as in the proof of Corollary 9.5 we may choose a generic tuple
h ∈ H(ks )n of H such that Gh (ks ) is Γ-stable. By Theorem 9.3, it suffices to show that
G(k) · h ⊆ Gn is cocharacter-closed over k. By [5, Lem. 2.4] and [9, Thm. 5.8(iii)], the orbit
G · h is closed.
We have cg (h) = cg (H) = Lie(CG (H)) = Lie(CG (h)), where the second equation follows
from [46, Lem. 4.1], and the other equations from [9, Lem. 5.5] and its proof. Thus the orbit
G · h is separable. Therefore, it follows from Corollary 7.2(ii) and Proposition 7.4(ii) that
G(k) · h is cocharacter-closed over k, which finishes the proof.
Remark 9.9. Observe that Corollary 9.8 is false if H is not k-defined: e.g., let G be k-split
and let S be a maximal k-defined torus. Suppose λ ∈ Yk (S) and u ∈ Ru (Pλ ) \ Ru (Pλ )(k) are
26
chosen so that H = u · S is a torus which is no longer k-defined. Clearly, H ⊆ Pλ . But H
is not contained in any k-defined Levi subgroup. Indeed, if L = Lx·λ with x ∈ Ru (Pλ )(k) is
such a Levi subgroup, then it must coincide with Lu·λ , which is the unique Levi subgroup of
Pλ containing H. But this implies that u = x ∈ Ru (Pλ )(k), a contradiction.
Remark 9.10. Following McNinch (see [37]), a subalgebra h of g is called G-completely reducible provided that whenever h is contained in Lie(P ) for P an R-parabolic subgroup of
G, there exists an R-Levi subgroup L of P with h ⊆ Lie(L). All of the concepts and results
of this section may be formulated and proved for Lie algebras as well. For a generic tuple
associated to h we simply take a generating tuple h ∈ hn for a suitable n. See also [9, §5.3].
10. The action of GL(W ) on End(W )
In this section we illustrate the notion of cocharacter-closedness in the classical context of
conjugacy of endomorphisms under the general linear group. Let W0 be a finite-dimensional
k-vector space with associated k-space W = k ⊗k W0 . Consider G = GL(W ) with its natural
action by conjugation on V = End(W ), and let both G and V be endowed with the kstructures induced from W0 , so that the action is k-defined. For f ∈ V (k) = End(W0 ), let
µf ∈ k[T ] be the minimal polynomial of f , and χf ∈ k[T ] the characteristic polynomial.
Note that both χf and µf remain invariant under field extensions of k. For any µ ∈ k[T ],
there exists f ∈ V (k) having µ as its minimal polynomial and its characteristic polynomial
(e.g., via considering the companion matrix of µ, see [47, Thm. 7.12]).
In general, the minimal polynomial µf only encodes partial information about the rational
normal form of f and hence about the orbit G(k) · f = GL(W0 ) · f . However, our next result
shows that one can read off from µf whether G(k) · f is cocharacter-closed. It is convenient
to consider W0 as a k[T ]-module, where T acts via f .
Proposition 10.1. The following are equivalent:
(i) The orbit G(k) · f ⊆ V is cocharacter-closed over k;
(ii) µf is square-free in k[T ] (i.e., has no repeated irreducible factors);
(iii) W0 is semisimple as a k[T ]-module.
Proof. Suppose first that W0 = W1 ⊕ W2 decomposes as a k[T ]-module. Let fi = f |Wi and
let Gi = GL(k ⊗ Wi ) (i = 1, 2). Then by an application of Theorem 5.4(ii), G(k) · f is
cocharacter-closed if and only if G1 (k) · f1 and G2 (k) · f2 are cocharacter-closed. Moreover,
µf = lcm(µf1 , µf2 ) (cf. [47, Thm. 7.7]) is square-free if and only if both µf1 and µf2 are. Thus
we may assume from the outset that W0 is an indecomposable k[T ]-module.
In this case, it follows (e.g., from the primary cyclic decomposition of W0 , see [47, Thm.
7.6]) that χf = µf is the power of an irreducible polynomial in k[T ]. Hence µf is square-free
if and only if χf is irreducible, which is easily seen to be equivalent to W0 being irreducible
as a k[T ]-module. This proves the equivalence of (ii) and (iii).
To prove that (i) is equivalent to (iii), suppose first that W0 is irreducible. Then f
stabilizes no proper subspace of W0 , and hence f is not destabilized by any non-central kdefined cocharacter of G. Thus G(k) · f is cocharacter-closed. Conversely, suppose that W0
contains a proper f -stable subspace U. Let λ be a k-defined cocharacter such that Pλ (G)(k)
is the stabilizer of U in G(k). Then f ′ = lim λ(a) · f stabilizes a complement to U. As W0 is
indecomposable, f ′ is not G(k)-conjugate to f , hence G(k) · f is not cocharacter-closed.
27
Remark 10.2. (i). Note that f is a semisimple endomorphism if and only if µf is separable
(that is, square-free in k[T ]), see [47, Thm. 8.11]. Proposition 10.1 thus recovers the fact
that an endomorphism f is semisimple if and only if G · f is cocharacter-closed over k. An
endomorphism satisfying the equivalent conditions of Proposition 10.1 is called k-semisimple
in [22], where an independent proof of the equivalence of (ii) and (iii) may be found.
(ii). If k is not perfect, there exist irreducible polynomials in k[T ] which are not separable. Hence there exist f ∈ V (k) with G(k) · f cocharacter-closed over k and G(k) · f not
cocharacter-closed over k.
(iii). We have used Levi descent and ascent in the proof of Proposition 10.1. In terms
of the characterization by k[T ]-modules, Levi descent/ascent is just the fact that a module
W = W1 ⊕ W2 is semisimple if and only if both summands are.
All G = GL(W )-orbits on V = End(W ) are separable (as orbit stabilizers are principal
open subsets of the stabilizers in gl(W )). Hence we may apply Proposition 7.4 and Remark
7.5 and deduce the following results:
Corollary 10.3. Let k ′ /k be an algebraic field extension and let f ∈ V (k) as above. If
G(k ′ )·f is cocharacter-closed over k ′ , then G(k)·f is cocharacter-closed over k. The converse
holds provided that k ′ /k is separable.
Moreover, the preorder ≺ is antisymmetric on the set of G(k)-orbits on End(W0 ).
Remark 10.4. The first part of the above corollary may also be deduced from the following
two facts, which are even valid for non-algebraic field extensions: if µf is square-free in k ′ [T ],
it is also square-free in k[T ]; if k ′ /k is separable and µf is square-free in k[T ], then µf is
square-free in k ′ [T ] (see [16, Ch. V, §15, Cor. 1]).
We give an example for the failure of Galois ascent for inseparable field extensions.
Example 10.5. Let k = F2 (t) and suppose that µf = T 12 + t. This polynomial is irreducible
in k[T ] (e.g., by Eisenstein’s criterion), so in particular G(k) · f is cocharacter-closed over k,
by Proposition 10.1. Over separable field extensions k ′ /k, µf can only split up to the point
µf = (T 4 + s)(T 4 + ζs)(T 4 + ζ 2s),
where s3 = t and ζ 6= 1, ζ 3 = 1, and this expression is still square-free. In particular, G(k ′ ) · v
is then still cocharacter-closed over k ′ , by Proposition 10.1.
If a ∈ k satisfies a4 = s, we may further decompose µf by inseparable field extensions to
obtain
µf = (T 2 + a2 )2 (T 2 + ζ 2a2 )2 (T 2 + ζa2)2 = (T + a)4 (T + ζa)4 (T + ζ 2a)4 .
If for example k ′ = k(a2 ), then the factor (T 2 + a2 )2 = (T + a)4 in µf is a maximal prime
power in µf . In the primary cyclic decomposition (cf. [47, Thm. 7.6]) it hence corresponds
to a 4-dimensional indecomposable subspace of k ′ ⊗ W0 where the restriction g of f to this
subspace has minimal polynomial µg = χg = (T + a)4 . This implies (see [47, Thm. 7.12])
that a matrix representative for g is given by the companion matrix of (T + a)4 , which over
28
k ′ is conjugate to the matrix
0 a2
1 0
0 0
0 0
0 a2
0 0
.
0 a2
1 0
Clearly, we may pick a cocharacter which kills the entry a2 in the top right corner while
leaving the other entries untouched. In particular, the limit along such a cocharacter takes
this matrix to a matrix with minimal polynomial T 2 + a2 , so the orbit G(k ′ ) · g is no
longer cocharacter-closed over k ′ . Hence G(k ′ ) · f is not cocharacter-closed over k ′ , by
Theorem 5.4(ii).
Remark 10.6. (i). For a cocharacter λ ∈ Yk (G) and f ∈ V (k), the limit lim λ(a) · f exists
if and only if f stabilizes the flag associated to λ in W0 . Moreover, after choosing a basis
compatible with the flag, the limit f ′ is then the endomorphism corresponding to the block
diagonal of the corresponding matrix of f . In particular, χf = χf ′ and µf ′ divides µf .
(ii). As noted before Corollary 10.3 above, the assumptions of Proposition 7.4 are satisfied
for all f ∈ V (k), hence the assertion of Theorem 7.1 holds in our setup: if f stabilizes a
flag associated to λ as in (i), and if f ′ is G-conjugate to f , then it is Ru (Pλ )(k) conjugate.
After again choosing a basis compatible with the flag, this translates into the following nonelementary statement about matrices: a block upper triangular matrix A over k that is
conjugate (over k) to the corresponding block diagonal matrix A′ , can be conjugated to A′
by an block upper triangular matrix over k with identity matrices on the diagonal.
We use these facts to give an example motivated by classical invariant theory (see [30]),
concerning cocharacter-closures of sets.
Example 10.7. For χ ∈ k[T ] consider the set Cχ ⊂ V (k) of endomorphisms with prescribed
characteristic polynomial χ. By Remark 10.6(i), Cχ is cocharacter-closed. Moreover, f ∈ Cχ
has a cocharacter-closed orbit precisely when µf is the product of all distinct irreducible
factors of χ (Proposition 10.1). In this case, µf together with χf = χ uniquely determine
the orbit G(k) · f . Indeed, it follows—e.g., from the primary cyclic decomposition of W0 (cf.
[47, Thm. 7.2])—that the elementary divisors of f must in this case be given by the irreducible
factors of µ, counted with their multiplicities as factors of χ. Since the elementary divisors
determine a rational canonical form of f (cf. [47, Thm. 7.14]), they determine the orbit
G(k) · f . In particular, Cχ contains a unique cocharacter-closed orbit.
It is well known that the closed orbits in V —that is, the conjugacy classes of diagonalisable
n
matrices—are parametrized by k , where n := dim W . We now see that the cocharacterclosed orbits in V (k) are parametrized by the space k n . For let π : V (k) → k n be the
surjective function that maps f to the coefficients of 1, T, . . . , T n−1 in χf . The above discussion shows that each fibre of π contains a unique orbit that is cocharacter-closed over
k.
We finish this section by discussing representations of algebras. As in the case k = C
(see [30]), the setup in this section may be generalized by replacing the algebra k[T ] with
some finitely generated k-algebra A0 . Let A = k ⊗k A0 , where A0 is a finitely generated
k-algebra, and let V = modA,W be the variety of all A-module structures on W . We may
endow V with a k-structure such that V (k) = modA0 ,W0 . More specifically, if A0 has r
29
generators, we may identify V with a subvariety of End(W )r , with G := GL(W ) acting
by simultaneous conjugation. Here instead of considering orbits of a single endomorphism,
we are now studying orbits of tuples of endomorphisms, and these orbits correspond to
the isomorphism classes of A-module structures. It can be shown that the cocharacterclosed orbits (over k) in V (k) correspond to isomorphism classes of semisimple A0 -module
structures on W0 . Note that all GL(W )-orbits in W n are separable, hence the preorder
≺ is antisymmetric (cf. Corollary 10.3). Moreover, by applying Corollary 7.2(ii), we get a
geometric proof of the algebraic fact that an A0 -algebra is semisimple if it becomes semisimple
after extending scalars from k to k.
Suppose k is algebraically closed. We say that one module is a degeneration of another if
the tuple defining the first module is in the closure of the GL(W )-orbit of the tuple defining
the second. This yields a partial ordering ≤ on the set of modules. Zwara gave an algebraic
characterization of when a module is a degeneration of another [59, Thm. 1]. There is no
obvious geometric notion of degeneration if k is not algebraically closed. Our formalism
of cocharacter-closures and accessibility gives a way to approach this problem for arbitrary
fields: our accessibility relation ≺ gives a partial order on the set of modules. This does not
coincide with the partial ordering ≤ when k = k, for the same reason that the Zariski closure
of an orbit and its cocharacter-closure over k are not the same. It would be interesting to
take the algebraic condition from [59, Thm. 1] and give a geometric interpretation when
k 6= k.
11. Further examples
Already for k algebraically closed, the notions of cocharacter-closure and Zariski closure
differ, as illustrated by our first example. More precisely, there may exist non-accessible
orbits in the Zariski closure of an orbit. The fact that not every orbit in an orbit closure
need be 1-accessible was already noted by Kraft in [30, II, Rem. 4.6].
Example 11.1. We consider the unipotent classes in the simple group of type G2 over an
algebraically closed field k of characteristic p. Corresponding to each choice of maximal torus
T of G and Borel subgroup B of G containing T , we have two simple roots α (short) and β
(long) and then for each root γ in the root system of G we have a coroot γ ∨ , a root group
Uγ and a corresponding root group homomorphism uγ : k → Uγ . We have the following
information about conjugacy classes of unipotent elements (see [36, Table 22.1.5], and [55]
for the representatives):
Class label Representative
Centralizer
G2
uα (1)uβ (1)
2
G2 (a1 )
uβ (1)u2α+β (1)
4
(Ã1 )3 (p = 3) uβ (1)uα+β (1)
6
Ã1
uα (1)
A1 + 3 (p 6= 3), A1 + 5 (p = 3)
A1
uβ (1)
A1 + 5
Trivial
1
G2
The notation in the final column gives the reductive part of the centralizer plus a number
denoting the dimension of the unipotent radical of the centralizer.
We record the closure and 1-accessibility relations in Figure 11.1. Here the first and third
diagrams give the Hasse diagrams for the partial order of containment between orbit closures
30
G2
G2
G2
G2 (a1 )
G2
G2 (a1 )
G2 (a1 )
(Ã1 )3
G2 (a1 )
(Ã1 )3
Ã1
Ã1
A1
A1
1
1
1
1
(p 6= 3)
(p 6= 3)
(p = 3)
(p = 3)
Ã1
Ã1
A1
A1
Figure 1. Closure relation and 1-accessibility of unipotent classes in G2
taken from [52, II, Prop. 10.4]. In particular, the (Zariski) closure of a given conjugacy class
consists of the union of that class together with those classes below it which are linked to it
by a downwards path in the diagram. The second and fourth diagrams depict the directed
graphs corresponding to the reflexive relation given by 1-accessibility (see Definition 3.2).
In this case, it turns out that 1-accessibility is transitive (see Example 11.2 for an instance
where this property fails). The cocharacter-closure of a given class can be read off similarly
from these diagrams. In particular, it can be seen here that the cocharacter-closure of an
orbit may be strictly smaller than the closure: e.g., the regular unipotent class labelled G2
has the subregular class G2 (a1 ) in its closure, but not in its cocharacter-closure.
The correctness of the two graphs for 1-accessibility can be verified by an argument along
the following lines. First, orbits that are not in the closure cannot be 1-accessible. Therefore,
for p = 3, the class A1 is not 1-accessible from Ã1 (and vice versa for all p). Second, orbits
with a unipotent centralizer cannot be 1-accessible from other orbits (as a limit along a
cocharacter is fixed by that cocharacter). Thus the classes G2 (a1 ) and (Ã1 )3 are not 1accessible from any other class. It remains to check that all further possible 1-accessibility
relations do hold. The trivial class is clearly 1-accessible from every unipotent class. For the
regular class G2 , the representative uα (1)uβ (1) is destabilized by (3α + 2β)∨ to uα (1), and by
(2α + β)∨ to uβ (1). The element uβ (1)u2α+β (1) in G2 (a1 ) is also taken by (2α + β)∨ to uβ (1),
and by β ∨ to u2α+β (1), which is another representative for Ã1 . The representative uα (1)
of Ã1 is conjugate by u2α+β (1) to uα (1)u3α+β (±3), and the latter element is destabilized by
−(α + β)∨ to u3α+β (±3). If p 6= 3, this is a representative for the class labelled A1 . Finally,
the element uβ (1)uα+β (1) in (Ã1 )3 is destabilized by (2α + β)∨ to uβ (1), and by −(3α + β)∨
to uα+β (1), which represents Ã1 .
Example 11.2. We give an elementary example which shows that 1-accessibility is not a
transitive relation. Start with H = SL2 (k), where k is an algebraically closed field, and
let E denote the natural module for H, with standard basis {e1 , e2 }. Set W = S 2 (E), the
31
symmetric square of E, and let {x2 , xy, y 2} denote the standard basis for W . Let λ denote
the diagonal cocharacter of H which acts with weight 1 on e1 and weight −1 on e2 , and with
weights 2, 0 and −2 on x2 , xy and y 2 respectively.
Now let G := H × k ∗ and let µ : k ∗ → G be the cocharacter given by µ(a) = (1, a) ∈ G
for each a ∈ k ∗ . Then the images of λ and µ generate a maximal torus T of G, and λ and µ
span the cocharacter group YT . Let the k ∗ -factor of G act on E with weight −1 and on W
with weight 2. Since these actions of k ∗ commute with the H-actions, we get actions of G
on E and W , and we can combine these to get an action of G on V := W ⊕ E.
Consider the element v := xy + e1 ∈ V . Then λ(a) · v = xy + ae1 , so lima→0 λ(a) · v exists
and equals v ′ := xy. Now let u = (uh , 1) in G, where uh ∈ Ru (Pλ )(H) is the unipotent
element for which uh · xy = x2 + xy. Then u · v ′ = x2 + xy. Define another cocharacter
σ ∈ Y (T ) by σ = µ − λ. We have σ(a) · (x2 + xy) = x2 + a2 xy, so lima→0 σ(a) · (x2 + xy) = x2 .
Set v ′′ := x2 . The above shows that, in the language of this paper, we have a sequence of
orbits G(k) · v, G(k) · v ′ , G(k) · v ′′ , with G(k) · v ′ 1-accessible from G(k) · v and G(k) · v ′′
1-accessible from G(k) · v ′ . One can see by direct calculation that these are distinct orbits.
We claim that G(k) · v ′′ is not 1-accessible from G(k) · v, which shows that 1-accessibility
is not transitive. Since λ and µ generate the cocharacter group Y (T ) of the maximal torus
T of G, and every cocharacter of G is conjugate to one in Y (T ), to check that G(k) · v ′′ is
not 1-accessible from G(k) · v, it suffices to show that lima→0 (mλ + nµ)(a) · (g · v) 6∈ G(k) · v ′′
for any g ∈ G and m, n ∈ Z for which
limit exists. Let g = (h, b) ∈ G be an arbitrary
the
q r
element, and suppose h has matrix
with respect to the fixed basis for E. Then we
s t
can write
(mλ + nµ)(a) · (g · v) = (mλ + nµ)(a) · (b2 (qrx2 + (qt + rs)xy + sty 2 ) + b−1 (qe1 + se2 ))
= b2 (qra2m+2n x2 + (qt + rs)a2n xy + sta−2m+2n y 2 )
+ b−1 (qam−n e1 + sa−m−n e2 ).
Now any G-conjugate of v ′′ = x2 has zero component in E, hence to stand any chance of
the limit existing and lying in G(k) · v ′′ we need to kill off the e1 and e2 component in the
limit. For the element h to lie in SL2 (k), we can’t have q = s = 0, so we have three cases to
consider.
First suppose q 6= 0 and s 6= 0. Then we need m − n > 0 and −m − n > 0 to kill off the
e1 and e2 components. But −m − n > 0 implies 2m + 2n < 0, so looking at the coefficient
of x2 , we must have qr = 0 and hence r = 0. Also, m − n > 0 implies −2m + 2n < 0, so
looking at the coefficient of y 2 we must have st = 0 and hence t = 0. But then h is not an
invertible matrix, so this is impossible.
Second suppose q = 0 and s 6= 0. Then we need −m − n > 0, i.e., −m > n. For h to
be invertible, we must have r 6= 0 and then the coefficient of xy tells us that n ≥ 0. The
inequality −m > n now gives −m > 0, and hence −2m + 2n > 0. In this case the limit
does exist and it equals 0 if n > 0 or b2 rsxy if n = 0. In the first case, this is clearly not
′′
′
conjugate
to
v , and in the second case the limit is conjugate to v = xy (e.g., by the element
1
I, (b2 rs) 2 ∈ G). As v ′ 6∈ G · v ′′ , we therefore cannot get into the orbit of v ′′ when q = 0
and s 6= 0.
32
The case q 6= 0 and s = 0 is similar—the limit, when it exists, is either 0 or conjugate to
v ′ —and we see that G(k) · v ′′ is not 1-accessible from G(k) · v, as claimed.
Acknowledgments: The authors acknowledge the financial support of EPSRC Grant
EP/L005328/1 and of Marsden Grants UOC1009 and UOA1021. Part of the research for
this paper was carried out while the authors were staying at the Mathematical Research
Institute Oberwolfach supported by the “Research in Pairs” programme. Also, part of this
paper was written during mutual visits to Auckland, Bochum and York. We are grateful to
the referees for their careful reading of the paper and for helpful suggestions.
References
[1] P. Bardsley and R.W. Richardson, Étale slices for algebraic transformation groups in characteristic p,
Proc. London Math. Soc. (3) 51 (1985), no. 2, 295–317.
[2] M. Bate, Optimal subgroups and applications to nilpotent elements, Transformation Groups 14 (2009),
no. 1, 29–40.
[3] M. Bate, S. Herpel, B. Martin, G. Röhrle, Cocharacter-closure and spherical buildings, Pacific J. Math.
279 (2015), no. 1–2, 65–85.
[4] M. Bate, B. Martin, G. Röhrle, A geometric approach to complete reducibility, Invent. Math. 161, no.
1 (2005), 177–218.
, Complete reducibility and commuting subgroups, J. Reine Angew. Math. 621 (2008), 213–235.
[5]
[6]
, Complete reducibility and separable field extensions, C. R. Acad. Sci. Paris Ser. I Math. 348
(2010), 495–497.
, The strong Centre Conjecture: an invariant theory approach, J. Algebra 372 (2012), 505–530.
[7]
[8] M. Bate, B. Martin, G. Röhrle, R. Tange, Complete reducibility and separability, Trans. Amer. Math.
Soc. 362 (2010), no. 8, 4283–4311.
, Closed orbits and uniform S-instability in geometric invariant theory, Trans. Amer. Math. Soc.
[9]
365 (2013), no. 7, 3643–3673.
[10] G. Berhuy, An introduction to Galois cohomology and its applications. With a foreword by Jean-Pierre
Tignol. London Mathematical Society Lecture Note Series, 377. Cambridge University Press, Cambridge, 2010.
[11] D. Birkes, Orbits of linear algebraic groups, Ann. of Math. (2) 93 (1971), 459–475.
[12] A. Borel, Introduction aux groupes arithmétiques, Publications de l’Institut de Mathématique de
l’Université de Strasbourg, XV. Actualités Scientifiques et Industrielles, No. 1341 Hermann, Paris 1969.
[13]
, Linear algebraic groups, Graduate Texts in Mathematics, 126, Springer-Verlag 1991.
[14] A. Borel, J. Tits, Éléments unipotents et sous-groupes paraboliques des groupes réductifs, I,, Invent.
Math. 12 (1971), 95–104.
[15] S. Bosch, W. Lütkebohmert, M. Raynaud, Néron models, Ergebnisse der Mathematik und ihrer Grenzgebiete, 21. Springer-Verlag, Berlin, 1990.
[16] N. Bourbaki, Éléments de mathématique, Algèbre, Chapitres 4 à 7, Lecture Notes in Mathematics, 864,
Masson, Paris, 1981.
[17] R.J. Bremigan, Quotients for algebraic group actions over non-algebraically closed fields, J. Reine Angew.
Math. 453 (1994), 21–47.
[18] M. Clarke and A. Premet, The Hesselink stratification of nullcones and base change, Invent. Math. 191
(2013), 631–669.
[19] B. Conrad, Finiteness theorems for algebraic groups over function fields, Compositio Math. 148 (2012),
no. 2, 555–639.
[20] B. Conrad, O. Gabber, G. Prasad, Pseudo-reductive groups, New Mathematical Monographs, vol. 17,
Cambridge University Press, Cambridge, 2010.
[21] M. Demazure, P. Gabriel, Groupes algébriques. Tome I: Géométrie algébrique, généralités, groupes
commutatifs, Masson & Cie, Éditeur, Paris, 1970, Avec un appendice Corps de classes local par M.
Hazewinkel.
33
[22] C. Dietz, k-semisimple elements and pseudotori, Dissertation, Fakultät für Mathematik der Universität
Bielefeld, 2013.
[23] P. Gille, A. Quéguiner-Mathieu, Exemples de groupes semi-simples simplement connexes anisotropes
contenant un sous-groupe unipotent, Pure Appl. Math. Q. 9 (2013), no. 3, 487–492.
[24] T.L. Gómez, A. Langer, A.H.W. Schmitt, I. Sols, Moduli spaces for principal bundles in arbitrary
characteristic, Adv. Math. 219 (2008), no. 4, 1177–1245.
[25] J. Hausen, A general Hilbert-Mumford criterion, Ann. Inst. Fourier 53 (2003), no. 3, 701–712.
[26] S. Herpel, On the smoothness of centralizers in reductive groups, Trans. Amer. Math. Soc. 365 (2013),
no. 7, 3753–3774.
[27] W.H. Hesselink, Uniform instability in reductive groups, J. Reine Angew. Math. 303/304 (1978), 74–96.
[28] D. Huybrechts and M. Lehn, The geometry of moduli spaces of sheaves, 2nd ed., Cambridge Mathematical Library, Cambridge University Press, Cambridge, 2010.
[29] G.R. Kempf, Instability in invariant theory, Ann. Math. 108 (1978), 299–316.
[30] H. Kraft, Geometric Methods in Representation Theory. Representations of algebras (Puebla, 1980), pp.
180–258, Lecture Notes in Math. 944, Springer, Berlin-New York, 1982.
[31] V. Lafforgue, Chtoucas pour les groupes réductifs et paramétrisation de Langlands globale,
http://arxiv.org/abs/1209.5352v5.
[32] J. Levy, Rationality and orbit closures, Canad. Math. Bull. 46, (2003), 204–215.
, Rationality and the Jordan-Gatti-Viniberghi decomposition, Canad. Math. Bull. 57 (2014), no.
[33]
1, 97–104.
[34] M.W. Liebeck, G.M. Seitz, Reductive subgroups of exceptional algebraic groups. Mem. Amer. Math. Soc.
no. 580 (1996).
, Variations on a theme of Steinberg, Special issue celebrating the 80th birthday of Robert
[35]
Steinberg. J. Algebra 260 (2003), no. 1, 261–297.
, Unipotent and nilpotent classes in simple algebraic groups and Lie algebras, Mathematical
[36]
Surveys and Monographs 180, American Mathematical Society (2012).
[37] G. McNinch, Completely reducible Lie subalgebras, Transform. Groups 12 (2007), no. 1, 127–135.
, Linearity for actions on vector groups, J. Algebra 397 (2014), 666–688.
[38]
[39] K. Mulmuley, M. Sohoni, Geometric complexity theory. I. An approach to the P vs. NP and related
problems, SIAM J. Comput. 31 (2001), no. 2, 496–526.
[40]
, Geometric complexity theory. II. Towards explicit obstructions for embeddings among class
varieties, SIAM J. Comput. 38 (2008), no. 3, 1175–1206.
[41] D. Mumford, J. Fogarty, F. Kirwan, Geometric invariant theory. Third edition. Ergebnisse der Mathematik und ihrer Grenzgebiete, 34. Springer-Verlag, Berlin, 1994.
[42] H. Nakajima, Lectures on Hilbert schemes of points on surfaces, University Lecture Series 18, American
Mathematical Society, Providence, RI, 1999.
[43] A. Premet, An analogue of the Jacobson-Morozov theorem for Lie algebras of reductive groups of good
characteristics, Trans. Amer. Math. Soc. 347 (1995), 2961–2988.
, Nilpotent orbits in good characteristic and the Kempf-Rousseau theory, J. Algebra 260 (2003),
[44]
338–366.
[45] N. Ressayre, Geometric invariant theory and the generalized eigenvalue problem, Invent. Math. 180
(2010), 389–441.
[46] R.W. Richardson, On orbits of algebraic groups and Lie groups, Bull. Austral. Math. Soc. 25 (1982),
no. 1, 1–28.
[47] S. Roman, Advanced Linear Algebra, Third Edition, Graduate Texts in Mathematics, 135, Springer,
New York, 2008.
[48] G. Rousseau, Immeubles sphériques et théorie des invariants, C.R.A.S. 286 (1978), 247–250.
[49] A.H.W. Schmitt, A closer look at semistability for singular principal bundles, Int. Math. Res. Not. 62
(2004), 3327–3366.
[50] J.-P. Serre, La notion de complète réductibilité dans les immeubles sphériques et les groupes réductifs,
Séminaire au Collège de France, résumé (1997).
, Complète réductibilité, Séminaire Bourbaki, 56ème année, 2003–2004, no 932.
[51]
34
[52] N. Spaltenstein, Classes unipotentes et sous-groupes de Borel, Lecture Notes in Mathematics, 946,
Springer-Verlag, Berlin Heidelberg New York, 1982
[53] T.A. Springer, Linear algebraic groups, Second edition. Progress in Mathematics, 9. Birkhäuser Boston,
Inc., Boston, MA, 1998.
[54] D.I. Stewart, On unipotent algebraic G-groups and 1-cohomology, Trans. Amer. Math. Soc. 365 (2013),
no. 12, 6343–6365.
[55] U. Stuhler, Unipotente und nilpotente Klassen in einfachen Gruppen und Liealgebren vom Typ G2 ,
Indag. Math. 33 (1971), 365–378.
[56] G. Székelyhidi, An introduction to extremal Kähler metrics, Graduate Studies in Mathematics 152,
American Mathematical Society, Providence, RI, 2014.
[57] T. Uchiyama, Separability and complete reducibility of subgroups of the Weyl group of a simple algebraic
group of type E7 , J. Algebra 422 (2014), 357–372.
[58] A. Zamora, On the Harder-Narasimhan filtration for finite dimensional representations of quivers, Geom.
Dedicata 170 (2014), 185–194.
[59] G. Zwara, Degenerations of finite-dimensional modules are given by extensions, Compositio Math. 121
(2000), 205–218.
Department of Mathematics, University of York, York YO10 5DD, United Kingdom
E-mail address: [email protected]
Fakultät für Mathematik, Ruhr-Universität Bochum, D-44780 Bochum, Germany
E-mail address: [email protected]
Department of Mathematics, University of Aberdeen, King’s College, Fraser Noble
Building, Aberdeen AB24 3UE, United Kingdom
E-mail address: [email protected]
Fakultät für Mathematik, Ruhr-Universität Bochum, D-44780 Bochum, Germany
E-mail address: [email protected]
35
| 4math.GR
|
arXiv:1507.08020v4 [math.AG] 31 May 2017
Geometry & Topology XX (20XX) 1001–999
1001
Affine representability results in A1 –homotopy theory II:
principal bundles and homogeneous spaces
ARAVIND ASOK
MARC HOYOIS
MATTHIAS WENDT
We establish a relative version of the abstract “affine representability” theorem in
A1 –homotopy theory from Part I of this paper. We then prove some A1 –invariance
statements for generically trivial torsors under isotropic reductive groups over
infinite fields analogous to the Bass–Quillen conjecture for vector bundles. Putting
these ingredients together, we deduce representability theorems for generically
trivial torsors under isotropic reductive groups and for associated homogeneous
spaces in A1 –homotopy theory.
14F42; 14L10, 55R15, 20G15
1 Introduction
Suppose k is a fixed commutative unital base ring, and write H (k) for the Morel–
Voevodsky A1 –homotopy category over k [44]. The category H (k) is constructed
as a certain localization of the category of simplicial presheaves on Smk , the category
of smooth k–schemes. Write Smaff
for the subcategory of Smk consisting of affine
k
schemes. If X is a simplicial presheaf on Smk , by an “affine representability” result
defined by
for X , we will mean, roughly, a description of the presheaf on Smaff
k
U 7→ [U, X ]A1 := HomH (k) (U, X).
Here is a flavor of the description we provide: if X is a simplicial presheaf on
1
one can consider the simplicial set SingA X (U)
Smk , then for any U ∈ Smaff
k
[44, p. 87]. The 0–simplices of this simplicial set are morphisms U → X and
the 1–simplices are “naive” or “elementary” A1 –homotopies U × A1 → X . The
1
1
assignment U 7→ π0 (SingA X (U)) defines a presheaf π0 (SingA X ) of “naive” A1 –
homotopy classes of maps U → X . In [9], we gave conditions that allowed us
1
to identify π0 (SingA X )(U) ∼
= [U, X ]A1 , i.e., under which “naive” A1 –homotopy
classes coincide with “true” A1 –homotopy classes.
Published: XX Xxxember 20XX
DOI: 10.2140/gt.20XX.XX.1001
1002
Aravind Asok, Marc Hoyois and Matthias Wendt
In [9, Theorem 1], building on results of M Schlichting [56, Theorems 6.15 and 6.22],
we simplified and generalized F Morel’s affine representability result for vector bundles;
we encourage the reader to consult the introduction of [9] for a more detailed discussion
of these points. Our goal in this paper is to further extend the scope of these affine
representability results in A1 –homotopy theory. For example, the following result
provides a generalization of the representability result from vector bundles to torsors
under suitable reductive group schemes (the description in terms of naive homotopy
classes is hidden here).
Theorem 1 (See Theorem 4.1.3) Suppose k is an infinite field, and G is an isotropic
reductive k –group (see Definition 3.3.5). For every smooth affine k –scheme X , there
is a bijection
1
HNis
(X, G) ∼
= [X, BG]A1
that is functorial in X .
Remark 2 Theorem 1 is essentially the strongest possible representability statement
for which one could hope. First, one cannot expect the functor “isomorphism classes of
Nisnevich locally trivial G–torsors” to be representable on H (k) in general. Indeed,
if we do not restrict attention to the category Smaff
k , then this functor need not even be
1
A –invariant (see, e.g., Ramanathan [55] for a study of failure of homotopy invariance
in case X = P1 or the introduction to [9] for other ways in which A1 –invariance can
fail). Second, at least if k is infinite and perfect, then the hypothesis that G is isotropic
cannot be weakened. Indeed, if G is not an isotropic reductive k–group in the sense
mentioned above, then even affine representability for G–torsors fails in general; see
Remark 3.3.8 and Balwe–Sawant [14, Theorem 1] for more details. We do not know
if Theorem 1 holds if k is finite.
Remark 3 It has been known for some time that an analogue of Morel’s theorem
should hold for torsors under groups like SLn and Sp2n (for SLn this is mentioned,
e.g., in Asok–Fasel [5, Theorem 4.2]). Schlichting observed [56, Remark 6.23] that
his techniques also apply to torsors under groups like SLn or Sp2n . Combined with the
results of [9], one therefore expects affine representability results to hold for torsors
under such groups in the same generality as for vector bundles. For completeness, we
include such results here as Theorems 4.1.1 and 4.1.2.
We also establish affine representability results for various homogeneous spaces under
reductive groups.
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1003
Theorem 4 (See Theorem 4.2.10) Suppose k is an infinite field, and G is an isotropic
reductive k –group. If P ⊂ G is a parabolic k –subgroup possessing an isotropic Levi
k –subgroup, then for any smooth affine k –scheme X , there is a bijection
1
∼
π0 (SingA G/P)(X) −→ [X, G/P]A1
that is functorial in X .
Remark 5 As suggested prior to the statement, we actually establish representability
results with targets that are more general homogeneous spaces. In this direction,
observe that it is often possible to “explicitly” identify sets of naive homotopy classes
and thus, via Theorem 4 true A1 –homotopy classes. Barge and Lannes [15, Chapter 4]
provide such identifications in the case where the target is related to symmetric bilinear
forms. Cazanave [19] provides such identifications in the case where the target is Pn .
In addition, Fasel [26, Theorem 2.1] gives such an identification in the case where the
target is a Stiefel variety (various homogeneous spaces of GLn ).
Building on the ideas of Schlichting and Morel, the proofs of the results above are
established using the framework developed in [9]: affine representability follows from
affine Nisnevich excision and affine homotopy invariance. The restrictions on k that
appear in our results are imposed to guarantee that affine homotopy invariance holds
for Nisnevich locally trivial torsors under G.
While affine homotopy invariance for vector bundles is precisely the Bass–Quillen
conjecture (about which much is known), precise statements regarding affine homotopy
invariance for torsors under other groups are harder to find in the literature (in part
because such results are typically false for étale locally trivial torsors), but see Wendt
[65, Section 3]. The entirety of Section 3 is devoted to studying affine homotopy
invariance for torsors under reductive group schemes over a rather general base.
Theorem 1 is a straightforward consequence of our general representability result
(see Theorem 2.3.5) combined with affine homotopy invariance (see Theorem 3.3.7
for a precise statement of what we mean by this term). Theorem 4 follows from
Theorem 2.4.2 and affine homotopy invariance for isotropic reductive k–groups by
a reduction from P to a Levi factor of P (which by assumption is also an isotropic
reductive k–group). Again, for certain groups, significantly more general statements
can be made; see Theorem 4.2.12.
Our techniques also allow us to establish significant generalizations (with simpler
proofs) of some results of F Morel regarding when classifying spaces for split groups
are A1 –local [42, Theorems 1.3, 1.5 and A.2]. While Morel deduces these results
Geometry & Topology XX (20XX)
1004
Aravind Asok, Marc Hoyois and Matthias Wendt
from strong A1 –invariance of non-stable K1 –functors, which he establishes by appeal
to classical results regarding elementary matrices, we are, in sharp contrast, able to
deduce such strong A1 –invariance statements as a direct consequence of our general
representability result (see Theorem 4.3.3 for more details). As another sample application of these results, we adapt some classical ideas of G W Whitehead [66] to
establish nilpotence results for non-stable K1 –functors (see Theorem 4.4.3), along the
lines of the results of Bak [11] and Bak–Hazrat–Vavilov [12]. In particular, we are able
to resolve [12, Problem 6] in a number of new situations (see Remark 4.4.4 for more
details).
The representability results for homogeneous spaces are relevant when applying the
methods of obstruction theory to analyze algebraic classification problems. For example, if the base k is a perfect field, the A1 –fibration sequence
An \ {0} −→ BGLn−1 −→ BGLn
was used by F Morel [43, Chapter 8] to develop an obstruction-theoretic approach to
answering the question of when a vector bundle over a smooth affine variety splits off
a trivial rank 1 summand; this approach was further developed by the first author and
Fasel in [6, 7] to which we refer the interested reader for a more detailed discussion.
The results of this paper (specifically Theorem 2.2.5) open the possibility of studying
such questions over more general base rings, e.g., Z.
Our representability results also broaden the scope of geometric and algebraic applications of A1 –homotopy theory. We mention a few such directions here (though we do not
develop the applications). First, Theorem 1 allows one to give explicit classifications of
principal G–bundles on certain quadric hypersurfaces, see Asok–Fasel [5] and Asok–
Doran–Fasel [4]. Theorems 4.2.1 and 4.2.2 establish affine representability results for
“split” quadric hypersurfaces. The former result has relevance to questions regarding
unimodular rows (see [5]). Building on the ideas of Fasel [27], affine representability
results for even dimensional quadrics are a key tool in Asok–Fasel [8] to interpret Euler
class groups à la Bhatwadekar–Sridharan in terms of A1 –homotopy theory. In another
direction, since the homogeneous space G2 /SL3 is a 6–dimensional “split” smooth
affine quadric, we use our results in [10] to study questions regarding reductions of
structure group for “generically trivial” octonion algebras. In algebraic terms this can
be rephrased as follows: when is an octonion algebra a Zorn (“vector-matrix”) algebra
(see, e.g., Springer and Veldkamp [57, p. 19])?
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1005
Dependency of sections/prerequisites
Section 2 is devoted to extending our results from [9]; the proofs rely on ideas from
loc. cit, which we will use rather freely together with some basic properties of torsors
and homogeneous spaces collected in Sections 2.3 and 2.4. Section 3 is devoted to
establishing affine homotopy invariance results for torsors under reductive groups. The
results of this section rely on the basic properties of torsors and homogeneous spaces
recalled in Section 2 as well as the theory of (reductive) group schemes over a base;
regarding the latter: we review some of the main definitions and basic properties, but
we mainly provide pointers to the literature. At the very end of Section 3.3 we also
rely on the representability results from Section 2. Section 4 contains applications of
our main results and thus relies on all of the preceding sections. We refer the reader to
the beginning of each section for a more detailed description of its contents.
Acknowledgements
The authors would like to thank Brian Conrad for extremely helpful correspondence
regarding [21] and [22]. In particular, the proof of Lemma 3.1.5 in its current form was
a product of these discussions. The authors would also like to thank Chetan Balwe and
Anand Sawant for helpful discussions of [13], Philippe Gille and Anastasia Stavrova
for helpful comments and corrections on a previous version of this paper, and Marco
Schlichting and Jean Fasel for mentioning at some point the Zariski local triviality of the
torsor appearing in the proof of Lemma 3.1.7. Finally, this paper owes an intellectual
debt to Marco Schlichting: even if it is obscured in references to [9], the ideas of [56,
§6] served to focus our attention (for example, he established a representability result
for special linear groups or symplectic groups [56, Remark 6.23]).
Aravind Asok was partially supported by National Science Foundation Award DMS1254892. Marc Hoyois was partially supported by National Science Foundation
Award DMS-1508096. Matthias Wendt was partially supported by EPSRC grant
EP/M001113/1.
Preliminaries/Notation
All rings considered in this paper will be assumed unital. We use the symbol S for a
quasi-compact, quasi-separated base scheme, SmS for the category of finitely presented
smooth S–schemes, and Smaff
S ⊂ SmS for the full subcategory of affine schemes (in
the absolute sense). We also reuse some terminology and notation introduced in [9],
Geometry & Topology XX (20XX)
1006
Aravind Asok, Marc Hoyois and Matthias Wendt
e.g., the notion of affine Nisnevich excision [9, Example 2.1.2 and Definition 3.2.1],
the t–localization functor Rt [9, §3.1], the singular construction SingI [9, §4.1], etc.
2 Some general representability results
The goal of this section is to extend the affine representability results of [9]. In
particular, Theorem 2.2.4 provides a relative version of [9, Theorem 5.1.3]. We then
specialize this result to two cases of particular interest in Theorems 2.3.5 and 2.4.2.
2.1 Naive A1 –homotopy classes
Let F be a simplicial presheaf on SmS . Given X ∈ SmS , there is a canonical map
(2–1)
1
π0 (SingA F )(X) → [X, F ]A1 ,
where the right-hand side is the set of maps in the A1 –homotopy category H (S).
The left-hand side is the set of naive A1 –homotopy classes of maps from X to F :
it is the quotient of the set of maps X → F by the equivalence relation generated
by A1 –homotopies. For presheaves F of “geometric origin”, such as representable
presheaves, it is rare that (2–1) is a bijection for all X ∈ SmS (this happens for
example when F is represented by an A1 –rigid smooth scheme in the sense of Morel–
Voevodsky [44, §3 Example 2.4], e.g., a smooth curve of genus g > 0 or an abelian
variety). However, one of the main themes of this paper is that there are many examples
of presheaves F such that (2–1) is a bijection for every affine X . We formalize this
idea in the following definition.
Definition 2.1.1 Let F be a simplicial presheaf on SmS and let F˜ be a Nisnevich1
local A1 –invariant fibrant replacement of F . Then there is a canonical map SingA F →
F˜ , well-defined up to simplicial homotopy. We will say that F is A1 –naive if the
1
map SingA F (X) → F˜ (X) is a weak equivalence for every X ∈ Smaff
S .
Remark 2.1.2 If F is A1 –naive, then in particular (2–1) is a bijection for every
1
X ∈ Smaff
S . More generally, if F is A –naive and pointed, then
1
πn (SingA F )(X) ∼
= [Sn ∧ X+ , F ]A1 ,∗
for every X ∈ Smaff
S and n ≥ 0.
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1007
Proposition 2.1.3 If F is a simplicial presheaf on SmS , then F is A1 –naive if
1
and only if SingA F satisfies affine Nisnevich excision (see [9, §2.1]). In that case,
1
RZar SingA F is Nisnevich-local and A1 –invariant.
Proof Let F˜ be a Nisnevich-local A1 –invariant replacement of F . Suppose that
1
F is A1 –naive. Then the restriction of SingA F to Smaff
is (objectwise) weakly
S
1
equivalent to F˜ , and hence it is Nisnevich-local. But this implies that SingA F
satisfies affine Nisnevich excision, by [9, Theorem 3.2.5].
1
Conversely, suppose that SingA F satisfies affine Nisnevich excision. By [9, Theorem
3.3.4], the canonical map
1
1
SingA F (X) → RZar SingA F (X)
1
A
is a weak equivalence for every X ∈ Smaff
S , and RZar Sing F is Nisnevich-local. By
1
A1
1
[9, Lemma 5.1.2], RZar Sing F is also A –invariant. Hence, RZar SingA F ≃ F˜ and
F is A1 –naive.
2.2 The singular construction and homotopy fiber sequences
The notion of representable interval object was formulated in [9, Definition 4.1.1]. By
a homotopy fiber sequence of pointed simplicial presheaves, we mean a homotopy
Cartesian square in which either the top-right or bottom-left corner is a point.
Proposition 2.2.1 Let C be a small category and I a representable interval object in
C . Let
F −→ G −→ H
be a homotopy fiber sequence of pointed simplicial presheaves on C . If π0 (H ) is
I –invariant, then
SingI F −→ SingI G −→ SingI H
is a homotopy fiber sequence.
Proof For X ∈ C, consider the square of bisimplicial sets
F (X × I• )
/ G (X × I• )
/ H (X × I• )
∗
Geometry & Topology XX (20XX)
1008
Aravind Asok, Marc Hoyois and Matthias Wendt
which is degreewise homotopy Cartesian. Since π0 (H ) is I–invariant, the simplicial
set π0 H (X × I• ) is constant. By [9, Lemma 4.2.1], the diagonal of this square is
homotopy Cartesian, i.e.,
SingI F (X) −→ SingI G (X) −→ SingI H (X)
is a homotopy fiber sequence.
Corollary 2.2.2 Let C be a small category and I a representable interval object in C .
If F is a pointed simplicial presheaf on C such that π0 (F ) is I –invariant, then the
canonical map
SingI RΩF −→ RΩ SingI F
is a weak equivalence.
Proof This follows from Proposition 2.2.1 applied to the homotopy fiber sequence
RΩ(F ) → ∗ → F .
Lemma 2.2.3 Suppose C is a small category with an initial object and let P be a
cd-structure on C . If J is a small diagram and F : J → sPre(C) is a functor such that
F(j) satisfies P –excision for every j ∈ J , then holimJ F satisfies P –excision as well.
Proof This is a straightforward consequence of commutation of homotopy limits.
Theorem 2.2.4 Suppose
F −→ G −→ H
is a homotopy fiber sequence of pointed simplicial presheaves on SmS . If the following
conditions hold:
(i) G and H satisfy affine Nisnevich excision, and
(ii) π0 (G ) and π0 (H ) are A1 –invariant on affine schemes,
then F is A1 –naive.
Proof By Proposition 2.2.1, for every U ∈ Smaff
S , the sequence
(2–2)
1
1
1
SingA F (U) −→ SingA G (U) −→ SingA H (U)
1
1
is a homotopy fiber sequence. By [9, Corollary 4.2.4], both SingA G and SingA H
1
satisfy affine Nisnevich excision. Hence by Lemma 2.2.3, SingA F also satisfies affine
Nisnevich excision. In other words, by Proposition 2.1.3, F is A1 –naive.
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1009
The following result is not used in the sequel, but it fits the theme of this section. It is a
variant of a result of Morel [43, Theorem 6.53] that holds over arbitrary base schemes.
Theorem 2.2.5 Let F → G → H be a homotopy fiber sequence of pointed simplicial presheaves on SmS . Assume that:
(i) H satisfies affine Nisnevich excision;
(ii) π0 (H ) is A1 –invariant on affine schemes.
Then F → G → H is an A1 –fiber sequence, i.e., it remains a homotopy fiber
sequence after taking Nisnevich-local A1 –invariant replacements.
Proof As in Theorem 2.2.4, the sequence (2–2) is a homotopy fiber sequence for
aff
∗
every U ∈ Smaff
and Ri∗ its
S . Let i be the restriction functor from SmS to SmS
derived right adjoint. By [9, Lemma 3.3.2], there is a natural equivalence of functors
RZar ≃ Ri∗ RZar i∗ . Since Ri∗ and RZar preserve homotopy fiber sequences, we deduce
that
1
1
1
RZar SingA F −→ RZar SingA G −→ RZar SingA H
1
is a homotopy fiber sequence. By [9, Theorem 5.1.3], RZar SingA H is Nisnevich-local
and A1 –invariant. But it follows from the right properness of the Morel–Voevodsky
model structure [44, §2 Theorem 2.7] that every homotopy fiber sequence whose base
is Nisnevich-local and A1 –invariant is an A1 –fiber sequence.
2.3 Application to torsors
In this subsection we specialize the general representability result of [9, §5.1] to
simplicial presheaves classifying G–torsors for some group G. We start by recalling
some general facts about torsors.
Definition 2.3.1 Let C be a small category equipped with a Grothendieck topology
t, let G be a t–sheaf of groups on C, and let X ∈ C. A G–torsor over X is a triple
(P, π, a) where P is a t–sheaf on C, a : P × G → P is a right action of G on P ,
and π : P → X is a morphism that is G–equivariant for the trivial G–action on X ,
such that:
(i) the morphism P × G → P ×X P of components π1 and a is an isomorphism;
(ii) π is t–locally split, i.e., the collection of morphisms U → X in C such that
P ×X U → U has a section is a t–covering sieve of X .
Geometry & Topology XX (20XX)
1010
Aravind Asok, Marc Hoyois and Matthias Wendt
The collection of G–torsors over various X ∈ C can be assembled into a category
Torst (G) fibered in groupoids over C. We write BTorst (G) for the simplicial presheaf
whose value on U ∈ C is the nerve of the groupoid of sections of Torst (G) over C/U
(this groupoid is canonically equivalent to the groupoid of G–torsors over U , but is
strictly functorial in U , cf. Hollander [33, §3.3]). It is well-known that Torst (G) is
a stack for the topology t. As shown in [33, Theorem 3.9], this is equivalent to the
statement that BTorst (G) satisfies t–descent.
We denote by BG the pointed simplicial presheaf with n–simplices Gn and with the
usual face and degeneracy maps, and we let
Bt G := Rt BG
be its t–local replacement (see [9, §3]). There is a morphism BG → BTorst (G)
sending the unique vertex of BG(U) to the trivial G–torsor over U . Since BTorst (G)
is t–local, we obtain a morphism of simplicial presheaves
(2–3)
Bt G −→ BTorst (G).
Lemma 2.3.2 Let C be a small category, t a Grothendieck topology on C , and G a
t –sheaf of groups on C . Then:
(i) The map (2–3) is a weak equivalence of simplicial presheaves.
(ii) There is a natural isomorphism
π0 (Bt G)(−) ∼
= Ht1 (−, G).
(iii) There is a canonical weak equivalence RΩBt G ≃ G .
Proof It is clear that the map (2–3) induces an isomorphism on t–sheaves of homotopy
groups, so that it is a weak equivalence in the Jardine model structure. To deduce that it
is a weak equivalence, it therefore suffices to show that the source and target are fibrant
in the Jardine model structure. By Dugger{Hollander{Isaksen [25, Corollary A.8], it
suffices to show that, for every U ∈ C, the simplicial sets Bt G(U) and BTorst (G)(U)
have no homotopy in dimensions ≥ 2. This statement is clear for the latter as it is
the nerve of a groupoid. To treat the former case, we recall a fact from simplicial
homotopy theory: if X is a simplicial set, then X has no homotopy in dimensions ≥ k
if and only if the homotopy fibers of the diagonal map X → X ×h X have no homotopy
in dimensions ≥ k − 1; this can be checked by assuming X is a Kan complex and
studying homotopy groups. Thus, a simplicial set X has no homotopy in dimensions
≥ 2 if and only if its 3-fold diagonal
X −→ X ×hX×h
X×h X
Geometry & Topology XX (20XX)
X
X
Affine representability results in A1 –homotopy theory II
1011
is a weak equivalence. Since Rt preserves homotopy pullbacks, it also preserves the
property of having no homotopy in dimensions ≥ 2. This proves (i). Assertions (ii)
and (iii) are true essentially by definition if we replace Bt G by BTorst (G), so they both
follow from (i).
Torsors under S–group schemes
Our main interest is to representability results for torsors under group schemes, so
we now discuss that situation in greater detail. Let G be an S–group scheme and let
X be an S–scheme. By a G–torsor over X we will mean a G–torsor in the sense of
Definition 2.3.1, for C the category of S–schemes and t the fppf topology. In the sequel
G will always be affine over S, and in that case a G–torsor over X is automatically
representable by an S–scheme, by Milne [40, Theorem III.4.3 (a)] (note: the implicit
Noetherian hypothesis in Milne’s argument is unnecessary).
If moreover X and G belong to SmS , then taking C to be the category SmS with t
the étale topology one obtains an equivalent notion of torsor. Indeed, if π : P → X is
a G–torsor over X , then π is finitely presented and smooth by the following lemma.
Since smooth morphisms admit sections étale locally, π itself is a cover of X in the
étale topology which trivializes the torsor.
Lemma 2.3.3 Suppose G is an affine S –group scheme, X is an S –scheme, and
π : P → X is a G –torsor over X . If G → S is finitely presented, flat, or smooth, then
so is π : P → X .
Proof By definition, there exists an fppf cover {Ui → X}i∈I such that P ×X Ui → Ui
is isomorphic to G ×S Ui → Ui , which is finitely presented, flat, or smooth. We
conclude using the fact that each of these properties of a morphism is fppf-local
on the target, by [58, Tag 02L0 Lemma 34.19.11, Tag 02L2 Lemma 34.19.13, and
Tag 02VL Lemma 34.19.25].
Example 2.3.4 Let t be a topology on SmS in between the Zariski topology and the
étale topology and let n ≥ 1. The groupoid of GLn –torsors over a scheme is canonically
equivalent to the groupoid of rank n vector bundles. Since GLn is a smooth special
group, any GLn –torsor is t–locally trivial. In particular, by Lemma 2.3.2 (ii), we have
π0 (Bt GLn )(X) ∼
= Vn (X)
Geometry & Topology XX (20XX)
1012
Aravind Asok, Marc Hoyois and Matthias Wendt
for any X ∈ SmS , where Vn (X) denotes the set of isomorphism classes of rank n vector
bundles on X . Similarly, we have
π0 (Bt SLn )(X) ∼
= Vno (X) and
π0 (Bt Sp2n ) ∼
= H V 2n (X),
where Vno (X) (resp. H V 2n (X)) is the set of isomorphism classes of rank n oriented
(resp. rank 2n symplectic) vector bundles (see the beginning of Section 3.3 for reminders about oriented and symplectic vector bundles).
Affine representability for Nisnevich locally trivial G–torsors
Theorem 2.3.5 Suppose G is a finitely presented smooth S –group scheme.
1 (−, G) is A1 –invariant on Smaff , then
HNis
S
If
1
(i) The simplicial presheaf RZar SingA BNis G is Nisnevich-local and A1 –invariant.
(ii) For every affine X ∈ Smaff
S , the canonical map
1
HNis
(X, G) −→ [X, BG]A1
is a bijection that is functorial with respect to X .
Proof Since BNis G is Nisnevich-local by definition, it satisfies Nisnevich excision by
1 (−, G)
[9, Theorem 3.2.5]. Taking into account the identification π0 (BNis G) ∼
= HNis
from point (ii) of Lemma 2.3.2, we can apply [9, Theorem 5.1.3] to BNis G, which
implies (i) and (ii) (note also that [X, BNis G]A1 ∼
= [X, BG]A1 since BG → BNis G is a
Nisnevich-local equivalence).
2.4 Application to homogeneous spaces
Let C be a small category equipped with a Grothendieck topology t. Let G and H be
t–sheaves of groups on C with H ⊂ G. We then have a homotopy fiber sequence of
simplicial presheaves
G/H −→ BH −→ BG,
where G/H denotes the presheaf U 7→ G(U)/H(U). Applying the t–localization
functor Rt , we obtain a homotopy fiber sequence of t–local simplicial presheaves
(2–4)
at (G/H) −→ Bt H −→ Bt G.
We now restrict attention to C = SmS with the goal of applying Theorem 2.2.4. For
geometric applications, we need to better understand the sheaf at (G/H).
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1013
Homogeneous spaces: topologies and quotient sheaves
Write rX for the presheaf on the category of S–schemes represented by an S–scheme
X , and r′ X for the restriction of the presheaf rX to SmS . Suppose that G and H are
finitely presented smooth S–group schemes, and that H is a closed subgroup of G.
The right translation action of H on G is scheme-theoretically free and it follows from
a result of Artin [2, Corollary 6.3] that the sheaf afppf (rG/rH) is representable by an
S–algebraic space. Two questions naturally present themselves: first, when does the
fppf sheaf quotient coincide with the Zariski or Nisnevich sheaf quotient and second,
is the fppf-sheaf afppf (rG/rH) representable by a smooth scheme? We address the first
question here; we answer the second question in various cases in Section 3.1.
Lemma 2.4.1 Suppose G is a finitely presented S –group scheme and H ⊂ G is a
finitely presented closed S –subgroup scheme. Assume that H is flat over S and that
the quotient G/H exists as an S –scheme. Then G → G/H is an H –torsor, and the
following statements hold.
(i) If t is a subcanonical topology on S –schemes such that the map G → G/H is
t –locally split, then r(G/H) ∼
= at (rG/rH) .
(ii) If G is smooth over S , then G/H is smooth over S . Moreover, if t is a
subcanonical topology on SmS such that the map G → G/H is t –locally split,
then r′ (G/H) ∼
= at (r′ G/r′ H) .
Proof By a theorem of Anantharaman [1, Appendice I, Théorème 6], we have
r(G/H) ∼
= afppf (rG/rH). In particular, G → G/H is an H –torsor, and hence it is
flat by Lemma 2.3.3. If G is smooth, it follows from [32, Proposition 17.7.7] that G/H
is also smooth. If G → G/H is t–locally split, then rG → r(G/H) is an epimorphism
of t–sheaves. By [3, Proposition 4.3 (2)], this implies that r(G/H) is the coequalizer of
the equivalence relation rG ×r(G/H) rG ∼
= rG × rH ⇒ rG in the category of t–sheaves,
which exactly means that r(G/H) ∼
a
= t (rG/rH). The second statement is proved in
the same way.
Affine representability for homogeneous spaces
Theorem 2.4.2 Suppose G is a finitely presented smooth S –group scheme and H ⊂ G
is a finitely presented smooth closed S –subgroup scheme such that the quotient G/H
exists as an S –scheme. Suppose that G → G/H is Nisnevich locally split and that
Geometry & Topology XX (20XX)
1014
Aravind Asok, Marc Hoyois and Matthias Wendt
1 (−, G) and H 1 (−, H) are A1 –invariant on Smaff . Then G/H is A1 –naive. In
HNis
S
Nis
particular, for every X ∈ Smaff
S , there is a bijection
1
π0 (SingA G/H)(X) ∼
= [X, G/H]A1 .
Proof The assumption on G → G/H combined with Lemma 2.4.1 allow us to
conclude that r′ (G/H) ∼
= aNis (r′ G/r′ H) and thus the homotopy fiber sequence (2–4)
has the form r′ (G/H) → BNis H → BNis G. The simplicial presheaves BNis G and BNis H
are Nisnevich-local and hence satisfy Nisnevich excision by [9, Theorem 3.2.5]. The
result is now a direct application of Theorem 2.2.4, taking into account Lemma 2.3.2
(ii).
3 Homotopy invariance for torsors under group schemes
1 (−, G) for G
The main goal of this section is to study A1 –invariance of the functors HNis
a linear group. Section 3.1 reviews basic definitions about group schemes, torsors and
homogeneous spaces; it also collects a number of results that will be used later in the
text. Section 3.2 establishes an analog of the local-to-global principle (a.k.a. “Quillen
patching”) for torsors under linear group schemes under rather general hypotheses; the
main result is Theorem 3.2.5. Finally, Section 3.3 proves general homotopy invariance
results; the main results are Theorems 3.3.3 and 3.3.7. For simplicity, we assume
throughout this section that the base scheme S is the spectrum of a commutative ring
R. In general there is a tradeoff between generality of the group G under consideration
and the base ring R.
3.1 Reductive group schemes and homogeneous spaces: recollections
The goal of this section is to recall some basic definitions and properties of group
schemes, torsors and homogeneous spaces over rather general bases. Rather than
attempting to be exhaustive, we only aim to point the reader to places in the literature
where they can find the required results. The grouping of these results is slightly
eclectic: only a very small portion of the definitions and results established here will be
used in the remainder of Section 3. Many of the results we state here are significantly
easier to establish (or even unnecessary) if the base ring R is a field.
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1015
Linear and reductive group schemes
We write GLn,R for the general linear group scheme over R and Gm ,R for GL1,R . If R
is clear from context, we will drop it from the notation.
Definition 3.1.1 By a linear R–group scheme, we mean a group scheme G over R
admitting a finitely presented closed immersion group homomorphism G → GLn,R .
Later, the homotopy invariance results we establish will require much more stringent
hypotheses on G. We use the definition of reductive (resp. semi-simple) R–group
scheme of Demazure–Grothendieck [24, Exposé XIX Definition 2.7]: a reductive (resp.
semi-simple) R–group scheme is a smooth, affine R–group scheme with geometric
fibers that are connected reductive (resp. semi-simple) groups in the usual sense [24,
Exposé XIX 1.6], i.e., have trivial unipotent radical (resp. radical). Recall that a
reductive R–group scheme G is called split if it contains a split maximal torus [24,
Exposé XXII Définition 1.13]. Any split reductive group scheme is pulled back from
a unique “Chevalley” group scheme over Spec Z.
If R is a field, it is a well-known consequence of the classification of reductive groups
that reductive R–group schemes are linear R–group schemes. If R is no longer a field,
the connection between “reductive” and “linear” becomes more complicated, as the
following example demonstrates.
Example 3.1.2 Groups of multiplicative type need not be linear in general [23, Expose
IX Définition 1.1]. Indeed, [23, Exposé XI Remarque 4.6] explains that if R is
a Noetherian and connected ring, then a group G of multiplicative type admits an
embedding in GLn if and only if it is isotrivial.
Nevertheless, the following result shows that, assuming suitable hypotheses on the
base, reductive R–group schemes are always linear.
Proposition 3.1.3 (Thomason) Suppose G is a reductive R –group scheme. Assume
one of the following additional hypotheses holds:
(i) R is regular and Noetherian; or
(ii) G is split.
Then G is a linear R –group scheme.
Proof If G is split, we can assume that R = Z and in particular that R is regular
Noetherian. In that case, the result follows from Thomason [61, Corollary 3.2 (3)].
Geometry & Topology XX (20XX)
1016
Aravind Asok, Marc Hoyois and Matthias Wendt
Remark 3.1.4 Thomason actually gives a sufficient condition for a group scheme to
admit a closed immersion group homomorphism into the automorphism group scheme
of a vector bundle over an arbitrary base S [61, Theorem 3.1]. Since we have in mind
applications to homotopy invariance, we have restricted attention to spectra of regular
rings.
Homogeneous spaces for reductive groups
Suppose G is a reductive R–group scheme and λ : Gm → G is a homomorphism
of R–group schemes. Via λ, we may consider the Gm –action λ : Gm × G → G
defined pointwise by the formula λ(t, g) := λ(t)gλ(t)−1 . We can define a subfunctor
PG (λ) ⊂ G consisting of those points g ∈ G such that limt→0 λ(t, g) exists and a
sub-functor UG (λ) ⊂ G consisting of those points g ∈ G such that limt→0 λ(t, g) = 1
(see Conrad [21, Theorem 4.1.7] for precise definitions). By [21, Theorem 4.1.7] both
of these functors are representable by R–subgroup schemes of G; since we assumed
G reductive it follows also that PG (λ) and UG (λ) are smooth and connected. By
[21, Example 5.2.2] PG (λ) is parabolic, and UG (λ) is a closed normal R–subgroup
scheme whose geometric fibers correspond to unipotent radicals of the geometric fibers
of PG (λ) [21, Corollary 5.2.5]; we will abuse terminology and refer to UG (λ) as the
unipotent radical of PG (λ).
If ZG (λ) is the centralizer of λ, then by [21, Definition 5.4.2] and the subsequent
discussion, ZG (λ) is a Levi factor of PG (λ), i.e., ZG (λ) is a smooth reductive R–group
scheme, and there is a semi-direct product decomposition of the form ZG (λ)⋉UG (λ) ∼
=
PG (λ). This description of parabolics, their unipotent radicals and Levi factors is called
a “dynamic” description in [22, 21] (since it arises from a study of “flows” under an
action of Gm ). We use these ideas to establish the following result.
Lemma 3.1.5 Suppose R is a connected ring, G is a reductive R –group scheme,
P ⊂ G is a parabolic R –subgroup scheme and L is a Levi factor of P . The following
statements hold.
(i) The quotients G/L and G/P exist as smooth R –schemes.
(ii) The morphism G → G/L is a generically trivial L –torsor.
(iii) The morphism G/L → G/P is a composition of torsors under vector bundles.
Proof For later use, we observe that since R is assumed connected and L is presumed
to exist, by Gille [30, Théorème 9.3.1], there is a cocharacter λ : Gm → G such that
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1017
P = PG (λ) and L = ZG (λ). If S is the spectrum of a field, which is the case we will
use later, the fact that all pairs (P, L) consisting of a parabolic together with a Levi
factor, are of the form (PG (λ), ZG (λ)) for a suitable cocharacter λ is contained in [22,
Proposition 2.2.9].
For Point (i), begin by observing that since P is a parabolic subgroup of G it is a
self-normalizing subgroup [21, Corollary 5.2.8]. The quotients G/L and G/P exist
as smooth R–schemes by [21, Theorems 2.3.1 and 2.3.6] (and, by Lemma 2.4.1, the
morphisms G → G/L and G → G/P are an L–torsor and a P–torsor, respectively).
For Point (ii), set U − = UG (−λ), i.e., the “unipotent radical” of an opposite parabolic.
We know that there is a dense open subscheme of G isomorphic to U − × P [21,
Theorem 4.1.7] (here and below, we will refer to this as the “big cell”). The image of
this open subscheme in G/L, which is isomorphic to U − × P/L, is again open and
dense since G → G/L is smooth and surjective. The Levi decomposition yields an
isomorphism of schemes P ∼
= L × U , and thus an identification P/L ∼
= U . Under these
−
identifications, the unit map U → P provides a morphism U × U → U − × L × U ,
which yields the required generic trivialization.
For Point (iii), let U be the unique smooth closed normal R–subgroup scheme of P
whose geometric fibers coincide with the unipotent radicals of the geometric fibers of
P, which is guaranteed to exist by [21, Corollary 5.2.5]. By the uniqueness assertion,
U∼
= UG (λ) for the character whose existence we observed in the first paragraph. By [21,
Theorem 5.4.3], U admits a finite descending filtration by AutP/R –stable closed normal
smooth R–subgroup schemes Ui with successive subquotients Ui /Ui+1 isomorphic to
P–equivariant vector bundles over R. Moreover, the isomorphism P/L ∼
= U described
in Point (ii) is actually P–equivariant.
Now, the morphism G/L −→ G/P is G–equivariant by definition. The schemetheoretic fiber over the identity coset in G/P is isomorphic to the quotient P/L and
∼
there is an induced G–equivariant isomorphism G ×P P/L → G/L under which the
morphism G/L → G/P is sent to the projection onto the first factor. In particular,
since P/L ∼
= U is smooth, G × P/L → G is smooth and since smoothness is fppf local
on the base [58, Tag 02VL Lemma 34.19.25], we conclude that G/L → G/P is also
smooth. By discussion of the previous paragraph, the morphism G/L → G/P thus
factors successively through morphisms of the form
(3–1)
G ×P U/Ui+1 −→ G ×P U/Ui .
To finish the proof, it suffices to inductively establish that each morphism in (3–1) is a
torsor under a vector bundle.
Geometry & Topology XX (20XX)
1018
Aravind Asok, Marc Hoyois and Matthias Wendt
Each morphism U/Ui+1 → U/Ui is, by construction, a torsor under the vector bundle
Ui /Ui+1 and, as we observed above, provided with a P–equivariant structure. If
1 (X, E ) = H 1 (X, E ) by [58,
E is a quasi-coherent sheaf on a scheme X , then Hfppf
Zar
1 (X, E ) parameterizes fppf-torsors under
Tag 03DR Proposition 34.7.10]. Since Hfppf
the quasi-coherent sheaf E , the P–equivariant structure on Ui /Ui+1 allows us to
conclude, by fppf-descent, that G ×P Ui /Ui+1 is a torsor under a vector bundle on
G/P. In other words, each morphism in (3–1) is again a torsor under the vector bundle
Ui /Ui+1 .
Remark 3.1.6 A number of remarks are in order.
(1) Since R a connected ring, it is not necessary to assume in the statement above
that L exists; this follows from Conrad [21, Corollary 5.4.8]. If we were to
work over a non-affine base scheme, parabolics need not have Levi factors (see
[21, Example 5.4.9] for more details). By reorganizing the proof, the argument
presented in Point (iii) actually shows that the quotient G/L exists assuming we
know G/P to exist and the relevant results on the structure of U .
(2) By Lemma 2.3.3, since L is a smooth R–group scheme by assumption, G → G/L
is étale locally trivial. If R is Noetherian and regular, then the morphism
G → G/L being generically trivial is tantamount to G → G/L being Nisnevich
locally trivial. To prove this, it suffices to show that generically trivial L–torsors
over Henselian local rings are trivial. If G is split reductive, then L is as well,
and the asserted triviality follows from Białynicki-Birula [17, Proposition 2]. If
G is not necessarily split, then L can be an arbitrary reductive group and one can
appeal to Nisnevich [47, Théorème 4.5] to deduce the required triviality result
(Nisnevich makes a statement for semi-simple group schemes, but it is true more
generally, see Fedorov and Panin [28, §1.1]).
(3) If G is split, it is possible to use translation of the big cell by elements of the
Weyl group to produce an explicit Zariski local trivialization of G → G/L. In
fact, even if G is not split, to establish Zariski local triviality of G → G/L (or,
equivalently, G → G/P), it suffices to know that the G(R)–translates of the
big-cell form an open cover of G/L (or G/P). If R is an infinite field, this kind
of result follows from the fact that the image of G(R) in G/P(R) is Zariski dense
(via the unirationality of G).
(4) In contrast, if R is a finite field (and G is non-split), it is a priori not obvious
that G(R) translates of the big cell cover G/L (or G/P). Nevertheless, assuming
the Grothendick–Serre conjecture, one knows that G → G/L is Zariski locally
trivial. If R is the spectrum of a finite field, the Grothendieck–Serre conjecture
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1019
was established by Gabber for reductive groups coming from the ground field
(unpublished), but another proof of a more general case was recently given by
Panin [48] (see also [28]).
Write SOn for the split special orthogonal group over R. We restrict attention to the case
where 2 is a unit in R so we can view SOn as the R–subgroup scheme of GLn consisting
of automorphisms of the standard hyperbolic form qn with trivial determinant (see, e.g.,
Conrad [21, Definition C.1.2]); for more details on special orthogonal groups, see [21,
Appendix C]).
Lemma 3.1.7 If R is a ring in which 2 is invertible, then the following statements
hold.
(i) If n ≥ 3 , the quotient SOn /SOn−1 exists and is isomorphic to a quadric hypersurface in AnR defined by the equation qn = 1 .
(ii) If n ≥ 3 , the projection morphism SOn → SOn /SOn−1 makes SOn into a Zariski
locally trivial SOn−1 –torsor over the quotient.
Proof Without loss of generality, we can take R = Z[1/2], which is Noetherian of
dimension ≤ 1. Since SOn−1 is a closed R–subgroup scheme of SOn , the quotient
SOn /SOn−1 exists as a scheme [1, Théorème 4.C].
To identify this quotient with the quadric in the statement, we proceed as follows.
Since SOn−1 = SOn ∩ SLn−1 inside of SLn , the inclusion SOn ⊂ SLn induces a
monomorphism SOn /SOn−1 ֒→ SLn /SLn−1 . Note that if A is an R–algebra, the map
sending X ∈ SLn (A) to its first row and the first column of its inverse determines an
isomorphism SLn /SLn−1 ∼
= Spec R[x1 , . . . , x2n ]/(q2n − 1). If we restrict X ∈ SOn (A)
and if J is the symmetric matrix corresponding to the bilinear form associated with
qn , then the orthogonality condition imposes the relation X −1 = JX T . Using this
observation, it is straightforward to check that the image is isomorphic, in suitable
coordinates, to a sub-quadric given by the equation qn = 1.
For the second statement, observe that morphisms X → SOn /SOn−1 classify SOn−1 –
torsors which are trivial after stabilization to SOn –torsors. The Witt cancellation
theorem (see Milnor and Husemoller [41, Lemma 6.3]) implies that, over a local ring
in which 2 is invertible, such an SOn−1 –torsor is already trivial.
Geometry & Topology XX (20XX)
1020
Aravind Asok, Marc Hoyois and Matthias Wendt
3.2 The local-to-global principle for torsors under linear group schemes
In this section we establish a local-to-global principle or “Quillen patching” for torsors
under linear R–group schemes in the sense of Definition 3.1.1. The main result of this
section is Theorem 3.2.5, which is a multi-variable analog of a result of Quillen [53,
Theorem 1] along the lines of Lam [38, Theorem V.1.6]. As will be clear from the
presentation, the argument follows quite closely that for projective modules given in
[38, Chapter V.1].
That the local-to-global principle holds for torsors under linear group schemes is certainly “well-known to experts”, under suitable hypotheses. For example, Raghunathan
[54] states (without proof) that Quillen’s local-to-global principle holds for linear algebraic groups over a field and Bass–Connell–Wright developed an axiomatic method to
establish such results [16, Proposition 3.1]; in particular, the latter approach applies for
various classical groups [16, Remark 4.15.4] over a general base ring. Nevertheless,
since we could not find a suitable published reference for precisely what we needed, in
the interest of completeness, we decided to collect the necessary results here.
Modifying automorphisms
We begin by generalizing [53, Lemma 1] (also [38, Corollary V.1.2]) and [38, Corollary
V.1.3] to linear R–group schemes over an arbitrary commutative ring R. The following
pair of results are due to Moser [45, Lemmas 3.5.3–3.5.5] (though our hypotheses
differ slightly); we include them here for the convenience of the reader.
Lemma 3.2.1 Let R be a commutative ring, let G be a linear R –group scheme, let
f ∈ R , and let θ(t) ∈ G(Rf [t]) be such that θ(0) = 1 ∈ G(Rf ) . There exists an integer
s ≥ 0 such that for any a, b ∈ R with a − b ∈ f s R , there exists ψ ∈ G(R[t]) with
ψ(0) = 1 and such that ψf (t) = θ(at)θ(bt)−1 ∈ G(Rf [t]) .
Proof Since G is a linear R–group scheme, by definition there is a finitely presented
closed immersion G → GLn . For s ∈ N, set ψs (t, x, y) := θ((x + f s y)t)θ(xt)−1 ∈
G(Rf [t, x, y]). It suffices to show that there exists s such that ψs can be lifted to
an element ψs ∈ G(R[t, x, y]). Indeed, in that case, by specializing with x = b,
a = b + f s α, we see that θ(at)θ(bt)−1 = ψs (t, b, α) lifts as well. By the proof of
[53, Lemma 1], we know that there exists s such that ψs (t, x, y) lifts to an element of
GLn (R[t, x, y]) and such that ψs (0, x, y) = 1 (see also [38, Theorem V.1.1]). Observe
that, by definition, ψs (t, x, 0) = 1 and thus ψs (t, x, 0) ∈ G(R[x, t]).
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1021
It remains to show that there exists i ≥ 0 such that ψs (t, x, f i y) ∈ G(R[t, x, y]). We
first recast this in ring-theoretic terms. Set A := R[t, x], let B be the coordinate ring
of GLn , and let I ⊂ B be the finitely generated ideal defining G. The lift of ψs is
given by a homomorphism ϕ : B → A[y], and we want to show that, for some i ≥ 0,
ϕ(−)(f i y) vanishes on I . We claim that, for every r ∈ I , there exists an integer ir such
that ϕ(r)(f i y) = 0 for i ≥ ir . If J ⊂ I is a finite generating set and i = maxr∈J ir , then
i will have the desired property.
Note that ϕ has the following properties: if ev0 : A[y] → A is the evaluation homomorphism, then the composites ev0 ◦ ϕ : B → A and B → A[y] → Af [y] both vanish
on I . If r ∈ I and P := φ(r) ∈ A[y], these properties imply that P = yQ for some
Q ∈ A[y] and that f ir P = 0 for some ir ≥ 0. Combining these two observations, we
have 0 = f ir P = f ir yQ. Therefore, f ir Q = 0 as well. Thus, P(f i y) = f i yQ(f i y) = 0
for all i ≥ ir , which is what we wanted to show.
Lemma 3.2.2 Let R be a commutative ring and G a linear R –group scheme. Given
f0 , f1 ∈ R such that f0 R + f1 R = R , and θ ∈ G(Rf0 f1 [t]) with θ(0) = 1 , then we can
find τi ∈ G(Rfi [t]) with τi (0) = 1 such that θ = τ0 τ1−1 .
Proof Let θ(t) ∈ G(Rf0 f1 [t]). We can apply Lemma 3.2.1 to the localizations Rf0 →
Rf0 f1 and Rf1 → Rf1 f0 : pick an integer s that suffices for both localizations. For any
b ∈ R, we can write
θ(t) = [θ(t)θ(bt)−1 ]θ(bt).
If f0 R + f1 R = R, then the same thing is true for f0s and f1s . Thus, we can pick b ∈ f1s R
such that 1 − b ∈ f0s R. In that case, θ(t)θ(bt)−1 ∈ G(Rf1 [t])f0 and θ(bt) ∈ G(Rf0 [t])f1
lift to elements τ1 and τ0 with the stated properties.
Remark 3.2.3 Lemma 3.2.1 implies “Axiom Q” (in the sense of Bass, Connell, and
Wright [16, §1.1]) holds for the functor on R–algebras determined by G. Lemma 3.2.2
essentially corresponds to [16, Theorem 2.4].
The local-to-global principle
Let R be a commutative ring and suppose G is a linear R–group scheme. If A
is a commutative R–algebra, by a G–torsor over A we will mean a G–torsor over
Spec A; by assumption our G–torsors are locally trivial in the fppf-topology (see
Definition 2.3.1 and the discussion just prior to Lemma 2.3.3 for more details). A
G–torsor over A[t1 , . . . , tn ] that is pulled back from a G–torsor over A will be called
Geometry & Topology XX (20XX)
1022
Aravind Asok, Marc Hoyois and Matthias Wendt
extended from A. For the remainder of this section, we will essentially confine our
attention to a fixed G–torsor P , which will be important for subsequent applications.
Proposition 3.2.4 Let R be a commutative ring. If P is a G –torsor over R[t] , then
the set Q(P) consisting of g ∈ R such that P|Spec Rg [t] is extended from Rg is an ideal
in R .
Proof It is immediate that Q(P) is closed under multiplication by elements in R.
Thus, we have to show that if f0 , f1 ∈ Q(P), then f = f0 + f1 lies in Q(P) as well.
After replacing R by Rf , we can assume that f0 R + f1 R = R.
Write 0 : Spec R → A1R , and pr : A1R → Spec R for the zero section and the structure
morphism. Thus, suppose P is a G–torsor over R[t] and assume that the restrictions
Pi := P|Spec Rfi [t] are extended. We want to show that P ∼
= pr∗ 0∗ P .
∼ pr∗ 0∗ Pi over Rf [t]. By modifying
By assumption, there are isomorphisms ui : Pi =
i
ui if necessary, we may assume that 0∗ ui = 1. Let P01 be the restriction of P
to Rf0 f1 [t]. Then u0 and u1 restrict to give two isomorphisms (u0 )f1 , (u1 )f0 : P01 ∼
=
∈
G(R
[t]),
then
there
is
a
commutative
pr∗ 0∗ P01 . If we set θ = (u1 )f0 (u0 )−1
f
f
0 1
f1
diagram of the form
P0 o
u0
pr∗ 0∗ P0 o
s
ss
ss
s
s
yss (u0 )f1
pr∗ 0∗ P01
P01 ❑
θ
❑❑ (u1 )f
❑❑ 0
❑❑
❑❑
%
/ pr∗ 0∗ P01
/ P1
u1
/ pr∗ 0∗ P1 .
If θ is the identity, then by fppf descent for G–torsors, the isomorphisms u0 and u1 glue
to give an isomorphism P ∼
= pr∗ 0∗ P , as desired. If not, since 0∗ ui = 1, we see that
θ restricts along t = 0 to the identity. Then, Lemma 3.2.2 guarantees that we can find
τi ∈ G(Rfi [t]) such that τi (0) = 1 and such that θ = τ0 τ1−1 . Thus, (τ0 u0 )f1 = (τ1 u1 )f0
and replacing u0 by τ0 u0 and u1 by τ1 u1 , we can glue these isomorphisms to conclude
that P is extended.
Theorem 3.2.5 (Local-to-global principle) Let R be a commutative ring and suppose
G is a linear R –group scheme. If P is a G –torsor over R[t1 , . . . , tn ] , then
(An ) the set Q(P) consisting of g ∈ R such that P|Spec Rg [t1 ,...,tn ] is extended from
Rg is an ideal in R .
(Bn ) If P|Spec Rm [t1 ,...,tn ] is extended for every maximal ideal m ⊂ R , then P is
extended.
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1023
Proof We know that (A1 ) holds by Proposition 3.2.4.
We show (An ) =⇒ (Bn ). It suffices to check that for P satisfying the conditions in
(Bn ) that the ideal Q(P) is the unit ideal in R. To this end, let P|0 be the pullback of
P along the zero section Spec R → Spec R[t1 , . . . , tn ] and let P ′ be the pullback of
P|0 along the structure map Spec R[t1 , . . . , tn ] → Spec R.
For any maximal ideal m ⊂ R, since P|Spec Rm [t1 ,...,tn ] is by assumption extended, we
∼
know there is an isomorphism ϕ : P|Spec Rm [t1 ,...,tn ] → P ′ |Spec Rm [t1 ,...,tn ] . Since G–
torsors over affine bases are of finite presentation under our hypotheses by Lemma 2.3.3,
there exists g ∈ R \ m such that ϕ is the localization of an isomorphism of torsors
over Spec Rg [t1 , . . . , tn ]. It follows that g ∈ Q(P) \ m and therefore that Q(P) is not
contained in m, i.e., Q(P) = R.
We show (A1 ) =⇒ (An ). We proceed by induction on n. Assume therefore that (An−1 )
holds. By the conclusion of the previous step, this means (Bn−1 ) holds as well. Form
the set Q(P) as in (An ). It is straightforward to check that R · Q(P) ⊂ Q(P) and
therefore it suffices to show that if f0 , f1 ∈ Q(P), then f0 + f1 ∈ Q(P) as well.
Write f = f0 + f1 . Consider the quotient map R[t1 , . . . , tn ] → R[t1 , . . . , tn−1 ] and
set P|tn =0 to be the restriction of P under the corresponding morphism of schemes.
Likewise, write P|0 for the restriction of P along the zero section as in the previous
step. Applying (A1 ) to the map R[t1 , . . . , tn−1 ] → R[t1 , . . . , tn−1 ][tn ], we conclude
that Pf is extended from (P|tn =0 )f .
We claim that (P|tn =0 )f is itself extended from Rf . If that is the case, then Pf is
extended and so f ∈ Q(P). Since (Bn−1 ) holds, it suffices to show that (P|tn =0 )f is
extended upon restriction to every maximal ideal m ∈ Rf . Write m = pf where p is
the pre-image of m under the localization map R → Rf . Since f ∈
/ p it follows that
either f0 or f1 is not in p; without loss of generality, we can assume that f0 ∈
/ p. By
assumption, however, Pf0 is extended from (P0 )f0 so we conclude that the restriction
of (P|tn =0 )f to the maximal ideal m is extended from (P0 )p , which is what we wanted
to show.
Corollary 3.2.6 Let G be a reductive R –group scheme. If R is regular Noetherian or
G is split, then the local-to-global principle holds for G –torsors, i.e., a G –torsor over
R[t1 , . . . , tn ] is extended from R if and only if for every maximal ideal m ⊂ R , the
G –torsor on Rm [t1 , . . . , tn ] obtained by restriction is extended from Rm .
Proof Combine Proposition 3.1.3 and Theorem 3.2.5.
Geometry & Topology XX (20XX)
1024
Aravind Asok, Marc Hoyois and Matthias Wendt
3.3 Affine homotopy invariance for G–torsors
Let G be a smooth linear R–group scheme. In this section, we analyze when the
pullback map
1
1
HNis
(X, G) −→ HNis
(X × A1 , G)
is a bijection for X a smooth affine R–scheme.
Special linear groups
We begin by recalling some facts about oriented vector bundles over schemes. If X is
a scheme, then recall that an oriented vector bundle on X is a pair (E , ϕ) consisting
∼
of a vector bundle E on X equipped with an isomorphism ϕ : det E → OX . There
is a standard equivalence between the groupoid of oriented vector bundles on X and
that of SLn –torsors over X . Write Vno (X) for the set of isomorphism classes of rank n
oriented vector bundles on X .
Theorem 3.3.1 (Special linear homotopy invariance) Fix an integer n ≥ 1 and
suppose R is a ring such that, for every maximal ideal m ⊂ R , Rm is ind-smooth over
a Dedekind ring with perfect residue fields (for example, Rm is Noetherian and regular
over such a Dedekind ring). For every integer m ≥ 0 , the map
Vno (Spec R) −→ Vno (Spec R[t1 , . . . , tm ])
is a bijection.
Proof By [9, Theorem 5.2.1], every vector bundle on Spec R[t1 , . . . , tm ] is pulled
back from a vector bundle on Spec R. In particular, every oriented vector bundle
on Spec R[t1 , . . . , tm ] is pulled back from a vector bundle on Spec R with trivial
determinant. It remains to show that every automorphism of the trivial line bundle on
Spec R[t1 , . . . , tm ] is extended from Spec R. In other words, we must show that the
inclusion map R → R[t1 , . . . , tm ] induces an isomorphism on unit groups.
Observe that our assumptions guarantee that Rm is reduced for every maximal ideal
m ⊂ R, and therefore R must itself be reduced. Since R is reduced, the fact that
R → R[t1 , . . . , tm ] induces an isomorphism on unit groups follows from a straightforward induction argument, using the elementary observation that if A is a reduced
commutative ring, then the map A → A[t] induces an isomorphism A× → A[t]× .
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1025
Remark 3.3.2 In [43, Definition 4.3], Morel defines an orientation on a vector bundle
E to be an isomorphism between det E and the square of a line bundle. Oriented vector
bundles in this sense correspond to torsors under the metalinear group MLn defined by
the pullback square
/ Gm
MLn
2
GLn
det
/ Gm .
This more general notion of orientation is very natural in Morel’s theory of the Euler
class, since the latter only depends on an orientation in this sense. Theorem 3.3.1 is
also true for MLn –torsors instead of SLn –torsors, with a very similar proof.
Symplectic groups
We refer the reader to Knus [36, §I.4] for more details about symplectic spaces over
rings; we briefly fix notations in the scheme-theoretic context. If X is a scheme and B is
a quasi-coherent sheaf on X , an alternating bilinear form on B is a morphism of quasicoherent sheaves ϕ : B ⊗OX B → OX such that ϕ ◦ ∆ = 0, where ∆ : B → B ⊗OX B
is the (nonlinear) diagonal map. If (B, ϕ) is a quasi-coherent sheaf equipped with
an alternating bilinear form, then we will say that ϕ is non-degenerate if ϕ induces
an isomorphism B → B ∨ := HomOX (B, OX ). By a symplectic bundle (of rank 2n)
we will mean a pair (B, ϕ) consisting of a (rank 2n) vector bundle B on X equipped
with a non-degenerate alternating bilinear form ϕ. Write H V 2n (X) for the set of
isomorphism classes of rank 2n symplectic bundles on X .
We briefly recall the standard equivalence between the groupoid of symplectic vector
bundles and that of Sp2n –torsors on X . In one direction, send a symplectic vector
bundle (B, ϕ) to its bundle of “symplectic frames”; by [36, Proposition I.4.1.4] this
construction yields an fppf torsor under Sp2n . In the other direction, given an Sp2n –
torsor P on X , consider the vector bundle associated with the standard 2n–dimensional
representation of Sp2n , which comes equipped with a reduction of structure group to
Sp2n , i.e., an alternating form on the bundle. By [36, Corollary 4.1.2] any symplectic
bundle on a scheme X is Zariski locally on X isometric to the hyperbolic space of
a trivial vector bundle [36, I.3.5]. Combining these observations, we see that Sp2n –
torsors are Zariski locally trivial and that there is an equivalence between the groupoid
of symplectic vector bundles over X and that of Nisnevich locally trivial Sp2n –torsors
(as mentioned in Example 2.3.4).
Geometry & Topology XX (20XX)
1026
Aravind Asok, Marc Hoyois and Matthias Wendt
Theorem 3.3.3 (Symplectic homotopy invariance) Fix an integer n ≥ 1 and suppose
R is a ring such that, for every maximal ideal m ⊂ R , Rm is ind-smooth over a Dedekind
ring with perfect residue fields (for example, Rm is Noetherian and regular over such
a Dedekind ring). For every integer m ≥ 0 , the map
H V 2n (Spec R) −→ H V 2n (Spec R[t1 , . . . , tm ])
is a bijection.
Proof For any integer n ≥ 1, the group Sp2n is a split reductive R–group scheme (and,
by definition, linear). Therefore, applying Theorem 3.2.5, it suffices to demonstrate
the result with R replaced by Rm . Since Rm is local, every finitely generated projective
module over Rm is free. By the assumption on R and [9, Theorem 5.2.1], we know
that, for any integer m, every finitely generated projective Rm [t1 , . . . , tm ]–module is
free. Applying [36, Corollary I.4.1.2], we conclude that every symplectic space over
Rm [t1 , . . . , tm ] is isometric to the hyperbolic space of a free module. In particular, every
symplectic space over Rm [t1 , . . . , tm ] is extended from Rm .
A formalism for homotopy invariance
We recall a formalism introduced by Colliot-Thélène–Ojanguren; the following result
is a slight extension of [20, Théorème 1.1].
Proposition 3.3.4 Fix an infinite base field k . Suppose F is a functor from the
category of k –algebras to the category of pointed sets with the following properties:
P1 The functor F commutes with filtered inductive limits of rings with flat transition
morphisms.
P2 For every extension field L/k and every integer n ≥ 0 , the restriction map
F(L[t1 , . . . , tn ]) −→ F(L(t1 , . . . , tn ))
has trivial kernel.
P3 The functor F has weak affine Nisnevich excision, i.e., for any smooth k –algebra
A , any étale A –algebra B , and any element f ∈ A such that A/fA ∼
= B/fB the
map
ker(F(A) → F(Af )) −→ ker(F(B) −→ F(Bf ))
is a surjection.
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1027
If B is the localization of a smooth k –algebra at a maximal ideal, then, setting KB =
Frac(B) , for any integer n ≥ 0 the restriction map
F(B[t1 , . . . , tn ]) −→ F(KB (t1 , . . . , tn ))
has trivial kernel.
Proof Set d := dim B and write m for the maximal ideal of B. Suppose that
ξ0 ∈ ker(F(B[t1 , . . . , tn ]) −→ F(KB (t1 , . . . , tn ))).
Let ξ be the image of ξ0 in F(KB [t1 , . . . , tn ]). Then, by assumption, ξ lies in the kernel
of F(KB [t1 , . . . , tn ]) → F(KB (t1 , . . . , tn )). By P2, we conclude that ξ is trivial.
By using P1, we conclude that there is an element g ∈ m \ 0 such that ξ0 restricts trivially to F(Bg [t1 , . . . , tn ]). Then, by Knus [36, Corollary VIII.3.2.5], there exist a polynomial ring L[x1 , . . . , xd ], a maximal ideal n ⊂ L[x1 , . . . , xd ], a local essentially étale morphism ϕ : A → B (where A = L[x1 , . . . , xd ]n ), and an element f ∈ m such that ϕ(f ) =
∼
ug for u a unit in Bm and ϕ induces an isomorphism A/fA → B/gB. By P3, we conclude that there exists an element ξ0′ ∈ ker(F(A[t1 , . . . , tn ]) → F(Af [t1 , . . . , tn ])) mapping to ξ0 . However, ξ0′ is also evidently in ker(F(A[t1 , . . . , tn ]) → F(KA (t1 , . . . , tn ))).
Thus, it suffices to establish the result in the case where B is the localization of a
polynomial ring at a maximal ideal, which is precisely [20, Proposition 1.5].
Isotropic reductive groups
If k is a field, a reductive k–group scheme will be called anisotropic if it contains
no k–subgroup isomorphic to Gm . We take the following definition for isotropic
reductive k–group, but we caution the reader that our definition differs from that of
Borel [18, Definition V.20.1]; we choose this definition because it better suits our
eventual applications.
Definition 3.3.5 If k is a field, a reductive k–group scheme G will be called isotropic
if each of the almost k–simple components of the derived group of G contains a
k–subgroup scheme isomorphic to Gm .
Remark 3.3.6 See Borel [18, §V.20] or Gille [30, §9.1] for further discussion of
isotropic reductive groups. In general, the existence of a non-central split multiplicative
k–subgroup is equivalent to the existence of a parabolic k–subgroup by the dynamic
construction described just before Lemma 3.1.5. In particular, isotropic reductive
k–groups admit proper parabolic subgroups.
Geometry & Topology XX (20XX)
1028
Aravind Asok, Marc Hoyois and Matthias Wendt
Theorem 3.3.7 If k is an infinite field, and G is an isotropic reductive k –group (see
Definition 3.3.5), then for any smooth k –algebra A and any integer n ≥ 0 , the map
1
1
HNis
(Spec A, G) −→ HNis
(Spec A[t1 , . . . , tn ], G)
is a bijection.
Proof We have to show that every Nisnevich locally trivial G–torsor P over A[t1 , . . . , tn ]
is extended from A. After Corollary 3.2.6, it suffices to show that, for every maximal
ideal m of A, the G–torsor Pm over Am [t1 , . . . , tn ] is extended from Am ; we will
show that in fact Pm is trivial.
1 (Spec A, G) from k–algebras to pointed sets satisWe claim that the functor A 7→ HNis
fies the axioms P1 − P3 of Proposition 3.3.4. Axiom P1 is a consequence of our finite
presentation hypotheses by way of Lemma 2.3.3. Axiom P2 uses the hypothesis that G
is isotropic and follows from [20, Proposition 2.4 and Theorem 2.5] (note that our definition of isotropic reductive k–group coincides with that used in [20, §2 p. 103]). Axiom
1 (−, G) ∼ π (BTors (G)) where
P3 is a formal consequence of the fact that HNis
= 0
Nis
BTorsNis (G) satisfies affine Nisnevich excision (see Section 2.3). By the conclusion of
Proposition 3.3.4, it suffices to show Pm becomes trivial over Frac(Am )(t1 , . . . , tn ), but
this follows immediately from the fact that a field has no nontrivial Nisnevich covering
sieves.
Remark 3.3.8 At least if k is an infinite perfect field, Theorem 3.3.7 admits a converse:
1 (−, G) is A1 –invariant on Smaff , then G
if G is a reductive k–group such that HNis
k
is isotropic, see Balwe and Sawant [14, Theorem 1]. In fact, for G reductive, the
following three conditions are equivalent:
(i) G is isotropic (in the sense of Definition 3.3.5);
1 (−, G) is A1 –invariant on smooth affine k–schemes;
(ii) HNis
1
(iii) RNis SingA G is A1 –invariant.
The implication (i) ⇒ (ii) is Theorem 3.3.7, (ii) ⇒ (iii) is a special case of Theorem 2.4.2, and (iii) ⇒ (i) is [14, Theorem 4.7].
4 Applications
In this section, we collect a number of applications of the results established so far.
Section 4.1 discusses representability results for Nisnevich locally trivial torsors. As
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1029
mentioned in Remark 3, that representability results should hold for torsors under SLn
and Sp2n was observed by Schlichting [56, Remark 6.23]; we simply observe in these
cases that the expected classical geometric objects yield models for the representing
spaces. In Section 4.2 we establish that for various classes of homogeneous spaces
for reductive groups applying the singular construction produces an A1 –local space.
Section 4.3 establishes strong A1 –invariance of homotopy sheaves of the singular
construction of a reductive group under suitable additional hypotheses. Finally, Section 4.4 studies a purely algebraic problem using our techniques, namely nilpotence of
non-stable K1 –functors.
4.1 Affine representability results for torsors
Let Grn,n+N be the usual Grassmannian parameterizing n–dimensional subspaces of an
f n,n+N be the complement of the zero section
(n + N)–dimensional vector space. Let Gr
in the total space of the determinant of the tautological vector bundle on Grn,n+N . The
f n,n+N parameterizes rank n subspaces of the (n + N)–dimensional vector
space Gr
fn :=
space equipped with a specified trivialization of their determinant. We set Gr
f
colimN Grn,n+N where the transition maps are the same as those in the definition Grn .
With these definitions, we can establish a geometric representability result for oriented
vector bundles.
Theorem 4.1.1 Suppose k is ind-smooth over a Dedekind ring with perfect residue
fields. Then, for any X ∈ Smaff
k , and any integer n ≥ 1 , there is a bijection
f n ]A1
Vno (X) ∼
= [X, Gr
that is functorial in X .
Proof Recall from Example 2.3.4 and the discussion preceding Theorem 3.3.1 that,
1 (X, SL ).
for any integer n ≥ 1, there is a functorial bijection of the form Vno (X) ∼
= HNis
n
Combining Theorems 2.3.5 and 3.3.1, we conclude that, under the stated hypotheses
1 (X, SL ) ∼ [X, BSL ] .
on k, for any smooth affine k–scheme X , HNis
n =
n A1
Using the notation of Morel and Voevodsky [44, §4.2], the space Bgm (SLn , i) (attached
fn . Therefore
to the defining inclusion i : SLn ֒→ GLn ) is precisely the space Gr
combining the results of [44, §4.2], and using the fact that all SLn –torsors are Zariski
f n → BSLn classifying
(and thus Nisnevich) locally trivial we conclude that the map Gr
1
fn is an A –weak equivalence.
the universal SLn –torsor over Gr
Geometry & Topology XX (20XX)
1030
Aravind Asok, Marc Hoyois and Matthias Wendt
If we let H be the standard 2–dimensional hyperbolic space, then we can consider
the symplectic vector space H⊕N . Panin and Walter construct a scheme HGrn,n+N
that parameterizes rank 2n symplectic subspaces of H⊕(n+N) and we set HGrn :=
colimN HGrn,n+N [50]. Alternatively, HGr can be described as the colimit colimN Sp2(n+N) /(Sp2n ×
Sp2N ). Using these definitions, we are now able to establish a geometric representability
theorem for symplectic vector bundles.
Theorem 4.1.2 Suppose k is ind-smooth over a Dedekind ring with perfect residue
fields. Then, for any X ∈ Smaff
k , there is a bijection
H V 2n (X) ∼
= [X, HGrn ]A1
that is functorial in X .
Proof Proceeding as in the proof of Theorem 4.1.1, we combine Example 2.3.4 and
the discussion preceding Theorem 3.3.3 to conclude that there is a functorial bijection
1 (X, Sp ). Combining Theorems 2.3.5 and 3.3.3, we
of the form H V 2n (X) ∼
= HNis
2n
conclude that, under the stated hypotheses on k, for any smooth affine k–scheme
1 (X, Sp ) ∼ [X, BSp ] . Finally, by the proof of [49, Theorem 8.2], we
X , HNis
2n =
2n A1
can conclude that HGrn is A1 –weakly equivalent to BSp2n , and thus for any smooth
k–scheme X , [X, HGrn ]A1 ∼
= [X, BSp2n ]A1 .
We now establish Theorem 1.
Theorem 4.1.3 Suppose k is an infinite field, and G is an isotropic reductive k –group
(see Definition 3.3.5). For any smooth affine k –scheme X , there is a functorial bijection
1
HNis
(X, G) ∼
= [X, BG]A1 .
Proof Combine Theorems 2.3.5 and 3.3.7.
Remark 4.1.4 In Theorem 4.1.3, the isotropy condition on G cannot be weakened,
cf. Remark 3.3.8.
4.2 Affine representability results for some homogeneous spaces
P
Let Q2n−1 be the smooth affine quadric over Z defined by i xi yi = 1. There is a
∼
standard identification SLn /SLn−1 → Q2n−1 . Let Q2n be the smooth affine quadric
P
over Z defined by i xi yi = z(z + 1) (in Asok–Doran–Fasel [4], it is shown that Q2n
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1031
∧n
is A1 –weakly equivalent to P1 over Spec Z). In particular, there are isomorphisms
Q2 ∼
= SL2 /Gm and Q4 ∼
= Sp4 /(Sp2 × Sp2 ) over Spec Z. If R is a ring in which 2
is invertible, then Q2n is isomorphic over R to the quadric defined by the standard
P
2
hyperbolic form
i xi yi + z = 1. It then follows from Lemma 3.1.7 that Q2n is
isomorphic over R to the homogeneous space SO2n+1 /SO2n .
Theorem 4.2.1 If R is a ring that is ind-smooth over a Dedekind ring with perfect
residue fields, then Q2n−1 is A1 –naive. In particular, for any smooth affine R –scheme
X , there is a functorial bijection
∼
1
π0 (SingA Q2n−1 )(X) −→ [X, Q2n−1 ]A1 .
Proof The scheme Q2n−1 is isomorphic over Spec Z to the homogeneous space
GLn /GLn−1 . Since all torsors for GLn−1 are Zariski locally trivial, it follows that
GLn → Q2n−1 is Zariski locally trivial (in fact, one can just write down an explicit
trivialization). Using [9, Theorem 5.2.1] we may apply Theorem 2.4.2 to conclude.
Theorem 4.2.2 If either (a) n ≤ 2 , and R is a ring that is ind-smooth over a Dedekind
ring with perfect residue fields, or (b) n ≥ 3 and R is an infinite field having characteristic unequal to 2 , then Q2n is A1 –naive. In particular, under either set of hypotheses,
for any smooth affine R –scheme X , there is a functorial bijection
1
∼
π0 (SingA Q2n )(X) −→ [X, Q2n ]A1 .
Proof For n = 1 consider the identification Q2 ∼
= SL2 /Gm . Affine homotopy
invariance holds for Gm –torsors over an arbitrary regular base, and for torsors under
SL2 ∼
= Sp2 by assumption. The result follows immediately from Theorem 2.4.2.
Similarly, for n = 2 consider the identification Q4 ∼
= Sp4 /(Sp2 × Sp2 ). Again, by
assumption we may combine Theorems 3.3.3 and 2.4.2 to conclude.
For n ≥ 3 we proceed slightly differently. The SO2n –torsor SO2n+1 → Q2n is still
Zariski locally trivial by Lemma 3.1.7. Since SOm is split for m ≥ 3, we may apply
1 (−, SO ) is A1 –invariant on Smaff for any integer
Theorem 3.3.7 to conclude that HNis
m
R
m ≥ 3. Then, we apply Theorem 2.4.2 to conclude.
Remark 4.2.3 If X = Spec A, then a map f : X → Q2n yields an ideal I ⊂ A
and a surjection ω : (A/I)⊕n → I/I 2 ; the ideal I is the ideal generated by the
images of x1 , . . . , xn , z in the coordinate presentation of the quadric. The class of f
1
in π0 (SingA Q2n )(X) depends only on the pair (I, ω) and is called the “Segre class”
of (I, ω), see Fasel [27, Theorem 2.0.2]. When X is smooth over an infinite field, the
Segre class provides an obstruction to lifting ω to a surjection A⊕n → I [27, Theorem
3.2.8].
Geometry & Topology XX (20XX)
1032
Aravind Asok, Marc Hoyois and Matthias Wendt
Zariski fiber bundles with affine space fibers
If F is a fixed S–scheme, we will say that an S–morphism π : E → B is a Zariski
fiber bundle of S–schemes with fibers isomorphic to F if there exist an S–scheme U , a
∼
Zariski covering morphism U → B and an isomorphism ϕ : U ×B E → U ×S F over U .
The following result, which generalizes a result of Morel [43, Theorem 8.9(2)], applies
to affine vector bundle torsors (a.k.a. Jouanolou–Thomason devices, see Weibel [63,
Definition 4.2 and Proposition 4.4]).
Lemma 4.2.4 Suppose B ∈ SmS , and π : E → B is a Zariski fiber bundle of S –
schemes with fibers isomorphic to AnS . For any X = Spec R ∈ Smaff
S , the induced
map
1
1
SingA E(X) −→ SingA B(X)
is an acyclic Kan fibration. Moreover, E is A1 –naive if and only if B is A1 –naive.
Proof By Goerss and Jardine [31, Theorem I.11.2], it suffices to show that for any
integer n ≥ 0, given a diagram of the form
∂∆nR
/E
π
∆nR
/B
there is a morphism ∆nR → E making both resulting triangles commute.
Given a diagram as above, there is an induced map ∂∆nR → ∆nR ×B E . By the
assumption on π , the pullback π ′ : ∆nR ×B E → ∆nR makes the ring of functions on
∆nR ×B E into a locally polynomial algebra over R[t1 , . . . , tn ] in the sense of Bass–
Connell–Wright [16, Theorem 4.4]. Therefore, by [16, Theorem 4.4] we conclude that
π ′ is a geometric vector bundle over ∆nR , i.e., the spectrum of a symmetric algebra
over ∆nR . Now, if E → ∆nR is a geometric vector bundle, then the inclusion map
∂∆nR → ∆nR induces a surjective map Hom(∆nR , E ) → Hom(∂∆nR , E ). Therefore, the
lift we hoped to construct is guaranteed to exist.
For the second statement, let Ẽ and B̃ be Nisnevich-local A1 –invariant replacements
of E and B, respectively, and consider the commutative square of simplicial presheaves
1
SingA E
/ Ẽ
/ B̃.
1
SingA B
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1033
Since the left vertical map is a weak equivalence on affines, the right vertical map is a
weak equivalence. It follows that the upper horizontal map is a weak equivalence on
affines if and only if the lower horizontal map is.
Example 4.2.5 If X ∈ Smaff
S is an affine scheme, then any finitely presented Zariski
fiber bundle of S–schemes π : E → X with fibers isomorphic to affine spaces is
actually a vector bundle by the result of Bass–Connell–Wright mentioned above [16];
this result was obtained independently by Suslin [60]. On the other hand, if X is not
affine, then even if π admits a section, it may not be isomorphic to a vector bundle:
see Iarrobino [34, Theorem 1] for an example with X = P1 .
Homogeneous spaces with non-reductive stabilizers
The following result extends and simplifies the proof of a theorem of Morel [43,
Theorem 8.9] (in particular, we allow the case n = 2).
Corollary 4.2.6 If R is a ring that is ind-smooth over a Dedekind ring with perfect
residue fields, then An \ 0 is A1 –naive. In particular, for any smooth affine R –scheme
X , there is a canonical bijection
1
∼
π0 SingA (An \ 0)(X) −→ [X, An \ 0]A1 .
Proof The map SLn → An \ 0 given by “projection onto the first column” factors
through a map SLn /SLn−1 → An \ 0; this map is a Zariski fiber bundle with fibers
isomorphic to affine spaces. By Lemma 4.2.4, it suffices to show that SLn /SLn−1 is A1 –
naive. This follows from Theorem 4.2.1 via the standard isomorphism SLn /SLn−1 ∼
=
Q2n−1 (send a matrix in SLn to the first row and first column of its inverse).
Lemma 4.2.7 Let X be a simplicial set and k ≥ 0 . If X has the right lifting property
with respect to the inclusion ∂∆m ⊂ ∆m for every m ≤ k + 1 , then X is k –connected.
Proof A simplicial set X is k–connected if and only if the Kan complex coskk+1 Ex∞ X
is contractible, or equivalently has the right lifting property with respect to ∂∆m ⊂ ∆m
for all m. By adjunction, this is the case if and only if Ex∞ X has the right lifting
property with respect to ∂∆m ⊂ ∆m for m ≤ k + 1. By definition of Ex∞ , it suffices
to show that X itself has the right lifting property with respect to sdr (∂∆m ) ⊂ sdr (∆m )
for all r and all m ≤ k + 1. In fact, X has the right lifting property with respect to any
monomorphism between (k + 1)–skeletal simplicial sets, since such a monomorphism
is a transfinite composition of pushouts of ∂∆m ⊂ ∆m for m ≤ k + 1.
Geometry & Topology XX (20XX)
1034
Aravind Asok, Marc Hoyois and Matthias Wendt
Proposition 4.2.8 Let n, k ≥ 0 and let R be a commutative ring such that the Bass
1
stable range of R[t0 , . . . , tk ] is at most n . Then the simplicial set SingA (An \0)(R) is k –
1
connected. In particular, if R is Noetherian of Krull dimension d , then SingA (An \0)(R)
is (n − d − 2) –connected.
Proof By Lemma 4.2.7, it suffices to show that the map
m
Umn (∆m
R ) → Umn (∂∆R )
is surjective for all m ≤ k+1, where Umn (X) = Hom(X, An \0) is the set of unimodular
rows of length n in O(X). By assumption, the Bass stable range of ∆k+1
is at most n. It
R
m
follows that the Bass stable range of ∆R is at most n, for all m ≤ k + 1. Now the result
is a special case of the following more general statement, which follows easily from
the definition of Bass stable range: if X is an affine scheme of Bass stable range ≤ n
and Y ⊂ X is a finitely presented closed subscheme, then the map Umn (X) → Umn (Y)
is surjective.
1
Remark 4.2.9 Under the assumption of Corollary 4.2.6, if n ≥ 3, the set π0 SingA (An \
0)(X) has a concrete description due to Fasel [26, Theorem 2.1]. Indeed, it is the quotient of the set Umn (X) of unimodular rows of length n by the action of the subgroup
En (X) ⊂ SLn (X) generated by elementary shearing matrices. In loc. cit., it is assumed
that R is a field, but the proof works more generally using a result of Lindel–Popescu
[52, Proposition 2.1]. Taking X = Q2n−1 , we obtain a bijection
[An \ 0, An \ 0]A1 ∼
= Umn (Q2n−1 )/En (Q2n−1 ).
1
By Corollary 4.2.6, we have [S1 , An \ 0]A1 ,∗ ∼
= π1 SingA (An \ 0)(R), and Proposition 4.2.8 shows that this group is trivial if n is at least the Bass stable range of R[t0 , t1 ].
In that case, we may therefore identify [An \ 0, An \ 0]A1 with the set of maps in the
pointed A1 –homotopy category. Note that colimn [An \ 0, An \ 0]A1 ,∗ is the set of
endomorphisms of the motivic sphere spectrum over the ring R.
The following result is Theorem 4.
Theorem 4.2.10 If k is an infinite field, G is an isotropic reductive k –group (see
Definition 3.3.5) and P ⊂ G is a parabolic k –subgroup possessing an isotropic Levi
factor (e.g., if G is split), then G/P is A1 –naive. In particular, for any smooth affine
k –scheme X , there is a functorial bijection
1
∼
π0 (SingA G/P)(X) −→ [X, G/P]A1 .
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1035
Remark 4.2.11 Given a reductive k–group and a non-trivial parabolic subgroup P ⊂
G, it is not obvious that P has a Levi factor. Nevertheless, as mentioned in Remark 3.1.6,
our hypotheses guarantee that P has a Levi factor. If L is a Levi factor for P, then L
may itself be anisotropic.
Proof Lemma 3.1.5(ii) implies that G → G/L is generically trivial. Since k is
assumed infinite and L is reductive, we claim G → G/L is actually Zariski locally
trivial. An elementary argument for Zariski local triviality of G → G/L sketched in Remark 3.1.6(2), but alternatively we can use [20, Théorème 2.1], to which, momentarily,
implicit appeal will be made.
By Theorem 2.4.2, whose hypotheses hold by Theorem 3.3.7, we conclude that G/L
is A1 –naive. By Lemma 3.1.5(iii), G/L → G/P is a composition of Zariski fiber
bundles with affine space fibers. Hence, G/P is also A1 –naive by Lemma 4.2.4.
The above result can be significantly strengthened at the expense of further restrictions
on the groups under consideration.
Theorem 4.2.12 Suppose R is ind-smooth over a Dedekind ring with perfect residue
fields (for example, R is Noetherian and regular over such a Dedekind ring). If G ∼
= GLn
1
or Sp2n , and if P ⊂ G is a standard parabolic subgroup, then G/P is A –naive. In
particular, for any smooth affine R –scheme X , there is a functorial bijection
1
∼
π0 (SingA G/P)(X) −→ [X, G/P]A1 .
Proof Assume first that R = Z. If P ⊂ G is a standard parabolic with Levi factor L,
then L is itself a special group in the sense of Grothendieck–Serre, i.e., all étale locally
trivial torsors are Zariski locally trivial. Thus, the map G → G/L in Lemma 3.1.5(ii)
is automatically Zariski locally trivial. One sees that the map G/L → G/P is a Zariski
fiber bundle with affine space fibers by combining Lemma 3.1.5(iii) with the fact that
all finitely generated projective Z–modules are free. By extending scalars to R, it
follows that corresponding statements hold for the resulting group scheme over R.
With these modifications, the proof is essentially identical to that of Theorem 4.2.10;
however, instead of appealing to Theorem 3.3.7, we use Theorem 3.3.3 or [9, Theorem
5.2.1] to establish the necessary homotopy invariance statement.
Example 4.2.13 Theorem 4.2.12 applies if P ⊂ GLn is a maximal parabolic subgroup,
in which case G/P ∼
= Grm,n for some integer m ≤ n.
Geometry & Topology XX (20XX)
1036
Aravind Asok, Marc Hoyois and Matthias Wendt
4.3 Affine representability for non-stable K-theory and strong A1 –invariance
results
Suppose G is a smooth linear R–group scheme. For any integer i ≥ 1, one can define
Karoubi–Villamayor-style non-stable K-theory functors attached to G by means of the
formula:
1
G
KVi+1
(U) := πi (SingA G)(U)
In this form, the definition goes back to Jardine [35, Theorem 3.8], but had precursors
in the work of Krusemeyer [37, §3]; see Wendt [64] for a more detailed analysis of
such functors in the context of A1 –homotopy theory. As a straightforward application
of our results, we obtain A1 –representability results for non-stable KV –functors.
Theorem 4.3.1 If k is an infinite field, and G is an isotropic reductive k –group (in
the sense of Definition 3.3.5), then G is A1 –naive. In particular, for any smooth affine
k –scheme U , there are canonical isomorphisms
G
KVi+1
(U) ∼
= [Si ∧ U+ , G]A1 ,∗ .
Proof Apply Theorem 2.4.2 with H = e (hypotheses being satisfied by Theorem 3.3.7).
Remark 4.3.2 Results such as the above were studied initially by Morel [43, Theorem
8.1] and Moser [46] (see also [65, Theorem 5.3]) for G a general split group, and by
the third author and K Völkel in the isotropic reductive case [62]. These results depend
crucially on first establishing homotopy invariance for non-stable K1 –functors via
“elementary matrix” techniques. As a consequence these proofs do not easily extend
to the important case where G has semi-simple rank 1, which was treated separately
by Moser. Our proof above makes no such assumption on the homotopy invariance of
non-stable K1 –functors. As a consequence, Theorem 4.3.1 can also be used to slightly
uniformize the proof of [13, Theorem 3.4].
We can also establish the strong A1 –invariance of the sheafifications of the non-stable
K1 –presheaves attached to arbitrary isotropic reductive k–groups with k infinite.
Theorem 4.3.3 Suppose k is an infinite field, and G is an isotropic reductive k –group
(in the sense of Definition 3.3.5). For any integer n ≥ 0 , the following statements
hold.
1
(i) The Zariski sheaf aZar πn (SingA G) is a Nisnevich sheaf.
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1037
1
(ii) The sheaf aZar πn (SingA G) is strongly A1 –invariant.
Proof We begin by recalling some key results of Morel [43, Chapter 6]. If X is
a Nisnevich-local and A1 –invariant pointed simplicial presheaf on Smk , the sheaf
aNis π1 (X ) is strongly A1 –invariant by [43, Theorem 6.1]. Moreover, the map
aZar π1 (X ) → aNis π1 (X ) is an isomorphism by [43, Corollary 6.9(2)] (the standing assumption that aNis π0 (X ) is trivial is not used in the proof).
By Theorems 2.3.5(i) and 3.3.7, under the stated hypotheses on k, the simplicial
1
presheaf RZar SingA BNis G is Nisnevich-local and A1 –invariant. Applying the results
of the previous paragraph to the simplicial presheaf
1
X = RΩn RZar SingA BNis G,
we conclude that
1
aZar πn+1 (SingA BNis G)
is a strongly A1 –invariant Nisnevich sheaf of groups for any n ≥ 0. By Corollary 2.2.2,
the map
1
1
πn (SingA RΩBNis G) −→ πn+1 (SingA BNis G)
is an isomorphism on affines, and hence it becomes an isomorphism after Zariski
sheafification. Finally, we conclude the proof by observing that G ≃ RΩBNis G by
Lemma 2.3.2 (iii).
Remark 4.3.4 We note that the results from [43, Chapter 6] used in the proof of
Theorem 4.3.3 do not require k to be perfect. If the base field k is in addition perfect,
1
then, provided aZar πn (SingA G) is abelian, we can use [43, Theorem 5.46] to conclude
that it is strictly A1 –invariant. The assumption that k is infinite in the above statement
only appears because of our appeal to Theorem 3.3.7. To remove this restriction, we
would need homotopy invariance for torsors under isotropic reductive groups over finite
fields.
If G is a reductive k–group, we can define G(k)+ to be the normal subgroup of G(k)
generated by the k–points of subgroups of G isomorphic to Ga . The Whitehead group
of G is defined by the formula
W(k, G) := G(k)/G(k)+ ;
we refer the reader to P Gille’s survey [29] for more details about Whitehead groups.
In particular, Tits showed that W(k, G) detects whether G(k) is projectively simple.
Results of Margaux allow us to connect non-stable K1 –functors (as above) with Whitehead groups. More precisely, one has the following result.
Geometry & Topology XX (20XX)
1038
Aravind Asok, Marc Hoyois and Matthias Wendt
Proposition 4.3.5 Suppose k is an infinite field, and G is an isotropic reductive k –
group (in the sense of Definition 3.3.5). For any finitely generated separable extension
L/k , there are canonical isomorphisms
1
π0 (SingA G)(L) ∼
= W(L, G).
functorial with respect to field extensions. Moreover, the assignment L 7→ W(L, G)
extends to a strongly A1 –invariant sheaf on Smk .
Proof The first statement follows from Margaux [39, Theorem 3.10] (see also Gille
[29, §4.3]) and only requires G to be isotropic in the sense of Borel [18, Definition V.20.1]. The second statement follows from the strong A1 –invariance of
1
aZar π0 (SingA G) established in Theorem 4.3.3(2).
Whitehead groups are also related to arithmetic questions, e.g., regarding R–equivalence
in G(k) (see Gille [29, §7] for a discussion of R–equivalence in the context under consideration).
Corollary 4.3.6 Let k be an infinite field and G a semisimple simply-connected
1
absolutely almost simple isotropic k –group, and set G := aZar π0 (SingA G) . The
following statements hold:
(i) for any finitely generated separable extension L/k , there is an isomorphism of
the form G(L) ∼
= G(L)/R ,
(ii) the contracted sheaf G−1 is trivial, and
(iii) if k is furthermore perfect, and G has classical type, then G is strictly A1 –
invariant.
Proof The first statement follows from Proposition 4.3.5 and [29, Théorème 7.2].
For the second statement, recall that G−1 (U) = ker((id, 1)∗ : G(U × Gm ) → G(U)).
As G is strongly A1 –invariant by Theorem 4.3.3, G−1 is also strongly A1 –invariant
by Morel [43, Lemma 2.32]. In particular, it is an unramified sheaf, which implies that
the map G(X) → G(k(X)) is injective for any irreducible smooth scheme X . By [29,
Theorem 5.8], we conclude that G(k(U)) → G(k(U × Gm )) is a bijection and thus that
G−1 (U) is trivial, for any U ∈ Smk .
For the final statement, if k is furthermore perfect, it suffices by [43, Theorem 5.46] to
show that G is an abelian group valued functor. Because G is unramified, it suffices
to check abelianness on sections over extensions of the base field. By Point (i), if G
has classical type, this follows from a result of Chernousov–Merkurjev [29, Théorème
7.7].
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1039
Remark 4.3.7 The statement G−1 = 0 of Corollary 4.3.6(ii) is equivalent to the
assertion that G is a birational sheaf. If G is not simply-connected, then the sheaf G is
not, in general, birational. For example suppose G is a split semisimple group having
non-trivial algebraic fundamental group Π (in the sense of Chevalley groups). If we
let Hét1 (Π) be the Nisnevich sheaf associated with the presheaf U 7→ Hét1 (U, Π), then
G∼
= Hét1 (Π), which is not birational.
4.4 Nilpotence for non-stable K1 functors
In this section, we include one more sample application of our results: we give a uniform
proof of some nilpotence results for non-stable K1 –functors discussed in the previous
section; such nilpotence results have been studied for instance by Bak [11] and Bak–
Hazrat–Vavilov [12]. The main result of this section is Theorem 4.4.3 which solves
[12, Problem 6] in a number of cases of interest (see Remark 4.4.4 for more details).
The approach we pursue has the benefit that it is conceptually simple (modelled on
classical topological results) and applies to rather general isotropic reductive k–groups.
The tradeoff to this generality is that unlike [12] we are forced to restrict attention to
smooth k–algebras with k an infinite field.
We use the following notation/terminology. If (X , x) is a pointed simplicial presheaf
on Smk , we will say that X is Nisnevich-connected if aNis π0 (X ) is trivial and, given
an integer n ≥ 1, we will say that X is Nisnevich n–connected if aNis πi (X , x) is
trivial for i ≤ n.
Now, suppose G is a simplicial presheaf of group-like h–spaces on Smk (h–group for
short) pointed by the identity. In that case, there is an induced morphism G → aNis π0 G ;
this morphism is a morphism of h–groups. Write G 0 for the homotopy fiber of G , so
that there is a homotopy fiber sequence of the form
G 0 −→ G −→ aNis π0 G .
By construction, G 0 is a Nisnevich-connected h–group. Using this notation, we can
adapt arguments of Whitehead [66, Corollary 2.12] to establish an abstract nilpotence
result.
Proposition 4.4.1 Assume k is a Noetherian ring of finite Krull dimension, and
suppose G is a Nisnevich-local simplicial presheaf of h –groups on Smk (pointed by
the identity).
(i) For any X ∈ Smk , there is an exact sequence of groups of the form
1 −→ [X, G 0 ] −→ [X, G ] −→ aNis π0 (G )(X).
Geometry & Topology XX (20XX)
1040
Aravind Asok, Marc Hoyois and Matthias Wendt
(ii) If X ∈ Smk has Krull dimension ≤ d , then [X, G 0 ] is nilpotent of class ≤ d .
Proof Point (i) is immediate from the long exact sequence of maps into a homotopy
fiber sequence and the fact that aNis π0 (G ) is 0–truncated.
For Point (ii), it suffices to assume G = G 0 is Nisnevich-connected. In that case, G ∧n
is Nisnevich n–connected. Indeed, this follows from the corresponding connectivity
estimate for smash products of simplicial sets by checking on stalks. Now, a straightforward obstruction theory argument (see Morel [43, Lemma B.5]) using the connectivity
estimate we just mentioned shows that [X, G ∧n ] = ∗ if dim X ≤ n. To conclude, we
simply observe that every n–fold commutator in [X, G ] factors as X → G ∧n → G
(here, we use the assumption that G is an h–group and thus has a strict identity).
Remark 4.4.2 The result above is rather general. Indeed, as is evident from the proof,
it holds for simplicial h–group objects in the local homotopy theory of simplicial
presheaves on a site for which Postnikov towers converge.
Now, suppose G is an isotropic reductive k–group in the sense of Definition 3.3.5. Following Petrov and Stavrova [51], for any commutative k–algebra R and any parabolic
k–subgroup P ⊂ G, we define the elementary subgroup EP (R) as the subgroup of
G(R) generated by the R–points of the unipotent radical of P and the R–points of the
unipotent radical of its opposite. A priori EP (R) depends on P and EP (R) need not be a
normal subgroup of G. However, [51, Theorem 1] guarantees that if each semi-simple
normal subgroup of G has rank ≥ 2, then EP (R) is both independent of P and normal
in G(R); under these hypotheses we define E(R) := EP (R) for any choice of proper
parabolic and define K1G (R) := G(R)/E(R).
We can also consider G0 (R) ⊂ G(R), the subset of G(R) consisting of matrices g
for which there exists g(t) ∈ G(R[t]) with g(0) = 1 and g(1) = g; this subgroup
is evidently normal. By construction EP (R) ⊂ G0 (R) and KV1G (R) = G(R)/G0 (R).
Therefore there is a short exact sequence of groups
1 −→ G0 (R)/E(R) −→ K1G (R) −→ KV1G (R) −→ 1.
Theorem 4.4.3 Suppose k is an infinite field, G is an isotropic reductive k –group in
the sense of Definition 3.3.5 and R is a smooth k –algebra of dimension d .
(i) If for every finitely generated separable extension L/k the Whitehead group
W(L, G) is trivial (abelian), then KV1G (R) is (an extension of an abelian group
by) a nilpotent group of class ≤ d .
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1041
(ii) If furthermore k is perfect, for every finitely generated separable extension L/k
the Whitehead group W(L, G) is trivial (abelian), and every semi-simple normal
subgroup of G has rank ≥ 2 , then K1G (R) is (an extension of an abelian group
by) a nilpotent group of class ≤ d .
1
Proof Let G = RZar SingA G. According to Theorem 4.3.3, the Nisnevich sheaf
aNis π0 (G ) is strongly A1 –invariant. By Proposition 4.3.5 the group of sections
aNis π0 (G )(L) over finitely generated extensions L/k coincides with W(L, G). In particular, the assumption that W(L, G) is trivial (abelian) for every finitely generated
separable extension L/k implies that the sheaf aNis π0 (G ) is trivial (abelian).
By Theorem 4.3.1 and Proposition 2.1.3, G is Nisnevich-local and KV1G (R) = [Spec R, G ].
Point (i) then follows immediately from Proposition 4.4.1.
Consider the exact sequence appearing before the statement gives a surjective map
K1G (R) → KV1G (R). Under the additional hypotheses in Point (ii), it follows immediately from a result of Stavrova [59, Theorem 1.3] that this surjection is an isomorphism
and Point (ii) follows from Point (i).
Remark 4.4.4 Combined with known structural results about W(−, G) (viewed as a
functor on the category of finitely generated extensions of the base field), the above
result solves a problem posed by Bak, Hazrat, and Vavilov [12, Problem 6] in a number of new cases. For example, in [29, Théorème 6.1], Gille summarizes results of
Chernousov–Platonov that detail situations where W(−, G) is trivial for all finitely
generated separable extensions L/k. See Corollary 4.3.6(iii) for hypotheses that guarantee W(−, G) is an abelian group valued functor on the category of (e.g., if G has
classical type). Furthermore, it has been conjectured that W(−, G) always takes values
in abelian groups.
References
[1] S Anantharaman, Schémas en groupes, espaces homogènes et espaces algébriques
sur une base de dimension 1, volume 33 of Mémoires de la Société Mathématique de
France (1973)
[2] M Artin, Versal deformations and algebraic stacks, Invent. Math. 27 (1974) 165–189
[3] M Artin, A Grothendieck, J-L Verdier, Théorie des topos et cohomologie étale
des schémas. Tome 1: Théorie des topos, Lecture Notes in Mathematics, Vol. 269,
Springer-Verlag, Berlin-New York (1972)Séminaire de Géométrie Algébrique du BoisMarie 1963–1964 (SGA 4), Dirigé par M. Artin, A. Grothendieck, et J. L. Verdier. Avec
la collaboration de N. Bourbaki, P. Deligne et B. Saint-Donat
Geometry & Topology XX (20XX)
1042
Aravind Asok, Marc Hoyois and Matthias Wendt
[4] A Asok, B Doran, J Fasel, Smooth models for motivic spheres and
the clutching construction, Int. Math. Res. Not. IMRN (2016)Available at
http://dx.doi.org/10.1093/imrn/rnw065
[5] A Asok, J Fasel, Algebraic vector bundles on spheres, J. Topol. 7 (2014) 894–926
[6] A Asok, J Fasel, A cohomological classification of vector bundles on smooth affine threefolds,
Duke Math. J. 163 (2014) 2561–2601
[7] A Asok, J Fasel, Splitting vector bundles outside the stable range and A1 -homotopy sheaves of punctured affine spac
J. Amer. Math. Soc. 28 (2015) 1031–1062
[8] A Asok, J Fasel, Euler class groups and motivic stable cohomotopy (2016)Preprint,
available at http://arxiv.org/abs/1601.05723
[9] A Asok, M Hoyois, M Wendt, Affine representability results in A1 -homotopy theory
I: vector bundles, Duke Math. J. (2017)Advance publication, 18 March 2017. doi:
http://dx.doi.org/10.1215/00127094-0000014X
[10] A Asok, M Hoyois, M Wendt, Generically split octonion algebras in A1 -homotopy
theory (2017)Preprint, available at https://arxiv.org/abs/1704.03657
[11] A Bak, Nonabelian K -theory: the nilpotent class of K1 and general stability,
Theory 4 (1991) 363–397
K-
[12] A Bak, R Hazrat, N Vavilov, Localization-completion strikes again: relative K1 is nilpotent by abelian,
J. Pure Appl. Algebra 213 (2009) 1075–1085
[13] C Balwe, A Sawant, R-equivalence and A1 -connectedness in anisotropic groups, Int.
Math. Res. Not. IMRN (2015) 11816–11827
[14] C Balwe, A Sawant, A1 -connectedness in reductive algebraic groups (2016)Preprint,
available at http://arxiv.org/abs/1605.04535
[15] J Barge, J Lannes, Suites de Sturm, indice de Maslov et périodicité de Bott, volume
267 of Progress in Mathematics, Birkhäuser Verlag, Basel (2008)
[16] H Bass, E H Connell, D L Wright, Locally polynomial algebras are symmetric algebras, Invent. Math. 38 (1976/77) 279–299
[17] A Białynicki-Birula, Rationally trivial homogeneous principal fibrations of schemes,
Invent. Math. 11 (1970) 259–262
[18] A Borel, Linear algebraic groups, volume 126 of Graduate Texts in Mathematics,
second edition, Springer-Verlag, New York (1991)
[19] C Cazanave, Algebraic homotopy classes of rational functions, Ann. Sci. Éc. Norm.
Supér. (4) 45 (2012) 511–534 (2013)
[20] J-L Colliot-Thélène, M Ojanguren, Espaces principaux homogènes localement triviaux,
Inst. Hautes Études Sci. Publ. Math. (1992) 97–122
[21] B Conrad, Reductive group schemes, from: “Autour des schémas en groupes. Vol. I”,
Panor. Synthèses 42/43, Soc. Math. France, Paris (2014) 93–444
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1043
[22] B Conrad, O Gabber, G Prasad, Pseudo-reductive groups, volume 17 of New Mathematical Monographs, Cambridge University Press, Cambridge (2010)
[23] M Demazure, A Grothendieck, Schémas en groupes. II: Groupes de type multiplicatif,
et structure des schémas en groupes généraux, Séminaire de Géométrie Algébrique du
Bois Marie 1962/64 (SGA 3). Dirigé par M. Demazure et A. Grothendieck. Lecture
Notes in Mathematics, Vol. 152, Springer-Verlag, Berlin-New York (1970)
[24] M Demazure, A Grothendieck, Schémas en groupes. III: Structure des schémas en
groupes réductifs, Séminaire de Géométrie Algébrique du Bois Marie 1962/64 (SGA
3). Dirigé par M. Demazure et A. Grothendieck. Lecture Notes in Mathematics, Vol.
153, Springer-Verlag, Berlin-New York (1970)
[25] D Dugger, S Hollander, D Isaksen, Hypercovers and simplicial presheaves, Math.
Proc. Cambridge Philos. Soc. 136 (2004) 9–51
[26] J Fasel, Some remarks on orbit sets of unimodular rows, Comment. Math. Helv. 86
(2011) 13–39
[27] J Fasel, On the number of generators of ideals in polynomial rings, Ann. of Math. (2)
184 (2016) 315–331
[28] R Fedorov, I Panin, A proof of the Grothendieck–Serre conjecture on principal bundles over regular local rings conta
Publ. Math. Inst. Hautes Études Sci. 122 (2015) 169–193
[29] P Gille, Le problème de Kneser-Tits, from: “Séminaire Bourbaki, volume 2007/2008,
Exposés 982–996”, Astérisque 326 (2009) 39–81
[30] P Gille, Sur la classification des schémas en groupes semi-simples, from: “Autour des
schémas en groupes. Vol. III”, Panor. Synthèses 47, Soc. Math. France, Paris (2016) 39–
110Preprint available at https://hal.archives-ouvertes.fr/hal-01063601v2
[31] P G Goerss, J F Jardine, Simplicial homotopy theory, Modern Birkhäuser Classics,
Birkhäuser Verlag, Basel (2009)Reprint of the 1999 edition [MR1711612]
[32] A Grothendieck, Éléments de géométrie algébrique. IV. Étude locale des schémas et
des morphismes de schémas. IV, Inst. Hautes Études Sci. Publ. Math. (1967) 361
[33] S Hollander, A homotopy theory for stacks, Israel J. Math. 163 (2008) 93–124
[34] A Iarrobino, An algebraic fibre bundle over P1 that is not a vector bundle, Topology
12 (1973) 229–232
[35] J F Jardine, On the homotopy groups of algebraic groups, J. Algebra 81 (1983) 180–
201
[36] M-A Knus, Quadratic and Hermitian forms over rings, volume 294 of Grundlehren
der Mathematischen Wissenschaften, Springer-Verlag, Berlin (1991)With a foreword
by I. Bertuccioni
[37] M I Krusemeyer, Fundamental groups, algebraic K -theory, and a problem of Abhyankar, Invent. Math. 19 (1973) 15–47
Geometry & Topology XX (20XX)
1044
Aravind Asok, Marc Hoyois and Matthias Wendt
[38] T Y Lam, Serre’s problem on projective modules, Springer Monographs in Mathematics, Springer-Verlag, Berlin (2006)
[39] B Margaux, The structure of the group G(k[t]): variations on a theme of Soulé,
Algebra Number Theory 3 (2009) 393–409
[40] J Milne, Étale Cohomology, Princeton University Press (1980)
[41] J W Milnor, D Husemoller, Symmetric bilinear forms, Springer-Verlag (1973)Ergebnisse der Mathematik und ihrer Grenzgebiete 73
[42] F Morel, On the Friedlander-Milnor conjecture for groups of small rank, from: “Current developments in mathematics, 2010”, Int. Press, Somerville, MA (2011) 45–93
[43] F Morel, A1 -algebraic topology over a field, volume 2052 of Lecture Notes in Mathematics, Springer, Heidelberg (2012)
[44] F Morel, V Voevodsky, A1 -homotopy theory of schemes, Inst. Hautes Études Sci. Publ.
Math. (1999) 45–143 (2001)
[45] L-F
Moser,
Rational
triviale
Torseure
und
die
SerreGrothendiecksche
Vermutung
(2008)Diplomarbeit,
available
at
http://www.mathematik.uni-muenchen.de/~lfmoser/da.pdf
[46] L-F Moser, A1 -locality results for linear algebraic groups (2011)In preparation
[47] Y A Nisnevich, Espaces homogènes principaux rationnellement triviaux et
arithmétique des schémas en groupes réductifs sur les anneaux de Dedekind, C. R.
Acad. Sci. Paris Sér. I Math. 299 (1984) 5–8
[48] I Panin, Proof of the Grothendieck-Serre conjecture on principal bundles
over regular local rings containing a finite field (2015)Preprint available at
http://arxiv.org/abs/1406.0241
[49] I Panin, C Walter, On the motivic commutative ring spectrum BO (2010)Preprint
available at http://arxiv.org/abs/1011.0650
[50] I Panin, C Walter, Quaternionic Grassmannians and Pontryagin classes in algebraic
geometry (2010)Preprint available at http://arxiv.org/abs/1011.0649
[51] V A Petrov, A K Stavrova, Elementary subgroups in isotropic reductive groups, Algebra i Analiz 20 (2008) 160–188
[52] D Popescu, Polynomial rings and their projective modules, Nagoya Math. J. 113
(1989) 121–128
[53] D G Quillen, Projective modules over polynomial rings, Invent. Math. 36 (1976) 167–
171
[54] M S Raghunathan, Principal bundles on affine space, from: “C. P. Ramanujam—a
tribute”, Tata Inst. Fund. Res. Studies in Math. 8, Springer, Berlin-New York (1978)
187–206
[55] A Ramanathan, Deformations of principal bundles on the projective line, Invent.
Math. 71 (1983) 165–191
Geometry & Topology XX (20XX)
Affine representability results in A1 –homotopy theory II
1045
[56] M Schlichting, Euler class groups, and the homology of elementary and special linear
groups (2015)Preprint, available at http://arxiv.org/abs/1502.05424
[57] T A Springer, F D Veldkamp, Octonions, Jordan algebras and exceptional groups,
Springer Monographs in Mathematics, Springer-Verlag, Berlin (2000)
[58] T Stacks Project Authors, Stacks Project, http://stacks.math.columbia.edu
(2015)
[59] A Stavrova, Homotopy invariance of non-stable K1 -functors, J. K-Theory 13 (2014)
199–248
[60] A A Suslin, Locally polynomial rings and symmetric algebras, Izv. Akad. Nauk SSSR
Ser. Mat. 41 (1977) 503–515, 717
[61] R W Thomason, Equivariant resolution, linearization, and Hilbert’s fourteenth problem over arbitrary base schemes
Adv. in Math. 65 (1987) 16–34
[62] K Völkel, M Wendt, On A1 -fundamental groups of isotropic reductive groups, C. R.
Math. Acad. Sci. Paris 354 (2016) 453–458
[63] C A Weibel, Homotopy algebraic K -theory, from: “Algebraic K -theory and algebraic
number theory (Honolulu, HI, 1987)”, Contemp. Math. 83, Amer. Math. Soc., Providence, RI (1989) 461–488
[64] M Wendt, A1 -homotopy of Chevalley groups, J. K-Theory 5 (2010) 245–287
[65] M Wendt, Rationally trivial torsors in A1 -homotopy theory, J. K-Theory 7 (2011)
541–572
[66] G W Whitehead, On mappings into group-like spaces, Comment. Math. Helv. 28
(1954) 320–328
Department of Mathematics, University of Southern California, Los Angeles, CA 90089-2532,
United States
Department of Mathematics, Massachusetts Institute of Technology, Cambridge, MA 021394307, United States
Fakultät für Mathematik, Universität Duisburg-Essen, Thea-Leymann-Strasse 9, 45127 Essen,
Germany
[email protected], [email protected], [email protected]
Geometry & Topology XX (20XX)
| 4math.GR
|
Neural Task Programming:
Learning to Generalize Across Hierarchical Tasks
arXiv:1710.01813v2 [cs.AI] 14 Mar 2018
Danfei Xu∗1 , Suraj Nair∗2 , Yuke Zhu1 , Julian Gao1 , Animesh Garg1 , Li Fei-Fei1 , Silvio Savarese1
I. I NTRODUCTION
Autonomy in complex manipulation tasks, such as object
sorting, assembly, and de-cluttering, requires sequential
decision making with prolonged interactions between the
robot and the environment. Planning for a complex task and,
vitally, adapting to new task objectives and initial conditions
is a long-standing challenge in robotics [6, 12].
Consider an object sorting task in a warehouse setting –
it requires sorting, retrieval from storage, and packing for
shipment. Each task is a sequence of primitives – such
as pick_up, move_to, and drop_into – that can be
composed into manipulation sub-tasks such as grasping and
placing. This problem has an expansive space of variations
– different objects-bin maps in sorting, permutations of subtasks, varying length order lists – resulting in a large space of
tasks. As a concrete example, Figure 1(C) shows a simplified
setup of the object sorting task. The task is to transport
objects of four categories to four shipping containers. There
is a total of 256 possible mappings between object categories
and containers, and the variable number of object instances
further increases the complexity. In this paper, we attempt to
address two challenges in complex task planning domains,
namely (a) learning policies that generalize to new task
objectives, and (b) hierarchical composition of primitives for
long-term environment interactions.
We propose Neural Task Programming (NTP), a unified,
task-agnostic learning algorithm that can be applied to a
variety of tasks with latent hierarchical structure. The key
underlying idea is to learn reusable representations shared
across tasks and domains. NTP interprets a task specification
(Figure 1 left) and instantiates a hierarchical policy as a neural
program (Figure 1 middle), where the bottom-level programs
are primitive actions that are executable in the environment.
A task specification is defined as a time-series that describes
* These authors contributed equally to the paper
1 Stanford Vision & Learning Lab, 2 CS, Caltech.
Task Demonstration
Task(i)
Abstract— In this work, we propose a novel robot learning
framework called Neural Task Programming (NTP), which
bridges the idea of few-shot learning from demonstration and
neural program induction. NTP takes as input a task specification (e.g., video demonstration of a task) and recursively
decomposes it into finer sub-task specifications. These specifications are fed to a hierarchical neural program, where bottomlevel programs are callable subroutines that interact with the
environment. We validate our method in three robot manipulation tasks. NTP achieves strong generalization across sequential
tasks that exhibit hierarchal and compositional structures. The
experimental results show that NTP learns to generalize well towards unseen tasks with increasing lengths, variable topologies,
and changing objectives. stanfordvl.github.io/ntp/
Robot API
NTP
```
A
Task Final State
Env.
B
State obs.
Task-Conditional
Policy
C
D
Fig. 1: (top) At test time, NTP instantiates a task-conditional policy
(a neural program) that performs the specified task by interpreting a
demonstration of a task. The policy interacts with the environment
through robot APIs. (bottom) We evaluate NTP on Block Stacking
(A,B), Object Sorting (C, D) and Table Clean-up (Figure 8) tasks in
both simulated and real environment.
the procedure and the final objective of a task. It can either be
a task demonstration recorded as a state trajectory, first/thirdperson video demonstrations, or even a list of language
instructions. In this work, we use task demonstration as
the task specification. We experiment with two forms of
task demonstration: location trajectories of objects that are
involved in a task, and a third-person video demonstration
of a task. NTP decodes the objective of a task from the
input specification and factorizes it into sub-tasks, interacting
with the environment with closed-loop feedback until the
goal is achieved (Figure 1 right). Each program call takes as
input the environment observation and a task specification,
producing the next sub-program and a corresponding sub-task
specification. The lowest level of the hierarchy is symbolic
actions captured through a Robot API.
This hierarchical decomposition encourages information
hiding and modularization, as lower-level modules only
access their corresponding sub-task specifications that pertain
to their functionality. It prevents the model from learning
spurious dependencies on training data, resulting in better
reusability. Essentially, NTP addresses the key challenges in
task generalization: meta-learning for cross-task transfer and
hierarchical model to scale to more complex tasks. Hence,
NTP builds on the strengths of neural programming and
hierarchical RL while compensating for their shortcomings.
We demonstrate that NTP generalizes to three kinds of
variations in task structure: 1) Task Length: varying number
of steps due to the increasing problem size (e.g., having
more objects to transport); 2) Task Topology: the flexible
permutations and combinations of sub-tasks to reach the same
end goal (e.g., manipulating objects in different orders); and
3) Task Semantics: the varying task definitions and success
conditions (e.g., placing objects into a different container).
We evaluate NTP on three table-top manipulation tasks
State
observation
Observation
Encoder
Task Spec.
Encoder
Input
“pick_and_place”
program
Input task
specification
Task Spec.
Interpreter
NTP
Output program
Core
Network
Is output Yes
program
Primitive
?
No
End-of-program
probability
API
Decoder
API arguments
Task Spec.
Selection
Output task
specification
that require long-term interactions: Block Stacking, Object
Sorting, and Table Clean-up. We evaluate each task in both
simulated and real-robot setups.
Summary of Contributions:
1) Our primary contribution is a novel modeling framework:
NTP that enables meta-learning on hierarchical tasks.
2) We show that NTP enables knowledge transfer and oneshot demonstration based generalization to novel tasks with
increasing lengths, varying topology, and changing semantics
without restriction on initial configurations.
3) We also demonstrate that NTP can be trained with visual
input (images and video) end-to-end.
II. BACKGROUND & R ELATED WORK
Skill Learning: The first challenge is learning policies that
adapt to new task objectives. For learning a single task
policy, traditional methods often segment a complex task
into hand-engineered state machine composed of motion
primitives [6, 12, 27]. Although the model-based approaches
are well-founded in principle, they require meticulous model
specification and task-specific treatment leading to challenges
in scaling. Contrarily, learning-based methods such as reinforcement learning (RL) have yielded promising results
using end-to-end policy learning that obviates the need for
manually designed state representations through data-driven
task-salient features [20, 33]. Yet these methods fall short
because they need task-specific reward functions [21].
Learning from Demonstrations: LfD fills these gaps by
avoiding the need to define state machines or reward functions.
The objective in LfD is to learn policies that generalize beyond
the provided examples and are robust to perturbations [3, 17]
A common treatment to LfD is to model data as samples
from an expert policy for a fixed task, and use behavior
cloning [16, 25] or reward function approximation [22] to
output an expert-like policy for that task. However, learning
policies that generalize to new objectives with LfD remains
largely an unexplored problem.
Few-Shot Generalization in LfD: Our work is an instantiation of the decades-old idea of meta-learning with few
examples [11, 30]. It has seen a recent revival in deep learning
in part because it can address the problems above [31].
Our setting resembles learning by demonstration (LfD) in
robotics [5], particularly one-shot imitation [10, 32]. Our
method learns to learn from an input task specification during
training. At test time, it generates a policy conditioned on a
single demonstration provided as a time-series showing the
task execution. While similar in these aspects, existing works
in both skill learning and LfD are inept at tasks with sparse
reward functions and complex hierarchical structures such as
Montezuma’s Revenge [19].
Fig. 2: Neural Task Programming
(NTP): Given an input program, a task
specification, and the current environment observation, a NTP model predicts
the sub-level program to run, the subsequence of the task specification that the
sub-level program should take as input,
and if the current program should stop.
Hierarchical Skill Composition: The second challenge we
consider is the hierarchical composition of primitives to
enable long-term robot-environment interaction. The idea of
using hierarchical models for complex tasks has been widely
explored in both reinforcement learning and robotics [17, 28].
A common treatment to manage task structure complexity
is to impose hierarchy onto the learned policy. The options
framework composes primitive actions into multi-step actions,
which facilitates policy learning at higher-level semantic
and/or temporal abstraction [13, 29]. Notable examples
include structured reinforcement learning methods, especially
hierarchical variants of RL that handle decomposition through
multi-stage policies operating over options [4, 19, 23].
However, the naïve use of a hierarchical RL model with
"sub-policies" or options optimized for a specific task doesn’t
guarantee modularity or reusability across task objectives.
The core idea of NTP resonates with recent works on
dynamic neural networks, which aim to learn and reuse primitive network modules. These methods have been successfully
applied to several domains such as robot control [1] and
visual question answering [2]. However, they have exhibited
limited generalization ability across tasks. In contrast, we
approach the problem of hierarchical task learning via neural
programming to attain modularization and reusability [24]. As
a result, our model achieves significantly better generalization
results than non-hierarchical models such as [10].
FSMs and Neural Program Induction: An exciting and
non-intuitive insight of this paper is that the well-studied
Finite State Machine (FSM) model lends itself to learning
reusable hierarchical policies thereby addressing the problem
of composability without the need for hand-crafting state
transitions. There have been a few studies learning FSMs
from data [14, 18]. In line with the idea, recent works in neural
programming using deep models have enabled symbolic
reasoning systems to be trained end-to-end, which have shown
potential to handle multi-modal and raw input/out data [9,
24] and achieve symbolic generalization [7].
NTP belongs to a family of neural program induction
methods, where the goal is to learn a latent program
representation that generates program outputs [9, 24]. While
these models have been shown to generalize on task length,
they are tested on basic computational tasks only with limited
generalization to task semantics and topology. Similar to NTP,
Neural Programmer-Interpreter (NPI) [24] has proposed to
use a task-agnostic recurrent neural network to represent
and execute programs. In contrast to previous work on
neural program induction, NPI-based models are trained with
richer supervision from the full program execution traces and
can learn semantically meaningful programs with high data
efficiency. However, program induction, including NPI, is not
…
…
Input Task Spec.
Env. Observation
EOP: False
Pin: block_stacking
EOP: False
Pout: pick_and_place
Output Task Spec.
Pout: pick_and_place
Output Task Spec.
Env. Observation
EOP: False
Pin: pick_and_place
EOP: True
Pout: pick
Output Task Spec.
Pout: place
Output Task Spec.
Env. Observation
Input Task Spec.
Env. Observation
Input Task Spec.
Pin: pick
EOP: False
Pin: pick
EOP: True
Args: block_E
Pout: grip
Pout: move_to
move_to (block_E)
return
Args: block_E
grip (block_E)
Env. Observation
Input Task Spec.
Env. Observation
Input Task Spec.
Pin: place
EOP: False
Pin: place
EOP: True
Args: block_B
Pout: release
Pout: move_to
return
Input Task Spec.
Input Task Spec.
move_to (block_B)
return
return
Input Task Spec.
Pin: pick_and_place
return
Env. Observation
return
Env. Observation
Pin: block_stacking
Args: N/A
release()
return
Fig. 3: Sample execution trace of NTP on a block stacking task. The task is to stack lettered blocks into a specified configuration (block_D
on top of block_E, block_B on top of block_D, etc). Top-level program block_stacking takes in the entire demonstration as
input (red window), and predicts the next sub-program to run is pick_and_place, and it should take the part of task specification
marked by the orange window as the input specification. The bottom-level API call moves the robot and close / open the gripper. When
End of Program (EOP) is True, the current program stops and return its caller program.
capable of generalizing to novel programs without training.
NTP is a meta-learning algorithm that learns to instantiate
neural programs given demonstrations of tasks, thereby
generalizing to unseen tasks/programs. Intuitively, NTP
decomposes the overall objective (e.g., object sorting) into
simpler objectives (e.g., pick and place) recursively. For
each of such sub-tasks, NTP delegates a neural program
to perform the task. The neural programs, together with the
task decomposition mechanism, are trained end-to-end.
While previous work has largely focused on executing a
pre-defined task one at a time NTP not only exhibits one-shot
generalization to tasks with longer lengths as NPI, but also
generalizes to sub-task permutations (topology) and success
conditions (semantics).
III. P ROBLEM F ORMULATION
We consider the problem of an agent performing actions
to interact with an environment to accomplish tasks. Let T
be the set of all tasks, S be the environment state space,
and A be the action space. For each task t ∈ T, the Boolean
function g : S × T → {0, 1} defines the success condition of
the task. Given any state s ∈ S, g(s,t) = 1 if the task t is
completed in the state s, and g(s,t) = 0 otherwise. The task
space T can be infinite. We thus need a versatile way to
describe the task semantics. We describe each task using a
task specification ψ(t) ∈ Ψ, where Ψ is the set of all possible
task specifications. Formally, we consider a task specification
as a sequence of random variables ψ(t) = {x1 , x2 , . . . , xN }.
NTP takes a task specification ψ(t) as input in order to
instantiate a policy. ψ(t) is defined as a time series that
describes the procedure and the final objective of the task. In
experiments, we consider two forms of task specifications:
trajectories of object locations and raw video sequences.
In many real-world tasks, the agent has no access to the
underlying environment states. It only receives a sample of
environment observation o ∈ O that corresponds to the state
s, where O is the observation space. Our goal is to learn
a “meta-policy” that instantiates a feedback policy from a
specification of a task, π̃ : Ψ → (O → A). At test time, a
specification of a new task ψ(t) is input to NTP. The metapolicy then generates a policy π(a|o; ψ(t)) : O → A, to reach
task-completion state sT where g(sT ,t) = 1.
Why use Neural Programming for LfD? Previous work has
mostly used a monolithic network architecture to represent a
goal-driven policy [10, 26, 33]. These methods cannot exploit
the compositional task structures to facilitate modularization
and reusability. Instead, we represent our policy π̃ as a
neural program that takes a task specification as its input
argument. As illustrated in Figure 2, NTP uses a task-agnostic
core network to decide which sub-program to run next and
adaptively feeds a subset of the task specification to the
next program. Intuitively, NTP recursively decomposes a
task specification and solves a hierarchical task by divideand-conquer. Figure 3 highlights this feature with a sample
execution of a task. Our method extends upon a special type of
neural programming architecture named Neural ProgrammerInterpreter (NPI) [7, 24]. NPI generalizes well to input size but
cannot generalize to unseen task objectives. NTP combines
the idea of meta-learning and NPI. The ability to interpret
task specifications and instantiate policies accordingly makes
NTP generalize across tasks.
A. Neural Programmer-Interpreter (NPI)
Before introducing our NTP model, it is useful to briefly
overview the NPI paradigm [24]. NPI is a type of neural
program induction algorithm, in which a network is trained
to imitate the behavior of a computer program, i.e., the
network learns to invoke programs recursively given certain
context or stop the current program and return to upperlevel programs. The core of NPI is a long-short memory
(LSTM) [15] network. At the i-th time step, it selects the
next program to run conditioned on the current observation
oi and the previous LSTM hidden units hi−1 . A domainspecific encoder is used to encode the observation oi into a
state representation si . The NPI controller takes as input the
state si , the program embedding pi retrieved from a learnable
key-value memory structure [M key ; M prog ], and the current
arguments ai . It generates a program key, which is used to
invoke a sub-program pi+1 using content-based addressing,
the arguments to the next program ai+1 , and the end-ofprogram probability ri . The NPI model maintains a program
call stack. Each time a sub-program is called, the caller’s
LSTM hidden units embedding and its program embedding is
pushed to the stack. Formally, the NPI core has three learnable
components, a domain-specific encoder fenc , an LSTM flstm ,
and an output decoder fdec . The full update being: si =
fenc (oi , ai ) hi = flstm (si , pi , hi−1 ) ri , pi+1 , ai+1 = fdec (hi ).
When executing a program with the NPI controller, it performs
one of the following three things: 1) when the end-of-program
probability exceeds a threshold α (set to 0.5), this program is
popped up from the stack and control is returned to the called;
2) when the program is not primitive, a sub-program with its
arguments is called; and 3) when the program is primitive, a
low-level basic action is performed in the environments. The
LSTM core is shared across all tasks.
IV. N EURAL TASK P ROGRAMMING
Overview. NTP has three key components: Task Specification
Interpreter fT SI , Task Specification Encoder fT SE , and a
core network fCN (Figure 2. The Task Specification Encoder
transforms a task specification ψi into a vector space. The
core network takes as input the state si , the program pi , and
the task specification ψi , producing the next sub-program
to invoke pt+1 and an end-of-program probability rt . The
program returns to the caller when rt exceeds a threshold α
(set to 0.5). We detail the inference procedure in Algorithm 1.
NTP vs NPI: We highlight three main differences of NTP
than the original NPI: (1) NTP can interpret task specifications
and perform hierarchical decomposition and thus can be
considered as a meta-policy; (2) it uses APIs as the primitive
actions to scale up neural programs for complex tasks; and (3)
it uses a reactive core network instead of a recurrent network,
making the model less history-dependent, enabling feedback
control for recovery from failures. In addition to the three key
components, NTP implements two modules similar to the NPI
architectures [7, 24]: (1) domain-specific task encoders that
map an observation to a state representation si = fENC (oi ), and
(2) a key-value memory that stores and retrieves embeddings:
prog
j∗ = arg max j=1...N (M key
j,: ki ) and pi = M j∗ ,: , where ki is the
program key predicted by the core network.
Scaling up NTP with APIs. The bottom-level programs in
NPI correspond to primitive actions that are executable in the
environment. To scale up neural programs in coping with the
complexity of real-world tasks, it is desirable to use existing
tools and subroutines (i.e., motion planner) such that learning
can be done at an abstract level. In computer programming,
application programming interfaces (APIs) have been a
standard protocol for developing software by using basic
modules. Here we introduce the concept of API to neural
programming, where the bottom-level programs correspond to
a set of robot APIs, e.g., moving the robot arm using inverse
kinematics. Each API takes specific arguments, e.g., an object
category or the end effector’s target position. NTP jointly
learns to select APIs functions and to generate their input
arguments. The APIs that are used in the experiments are
move_to, grip, and release. move_to takes an object
Algorithm 1 NTP Inference Procedure
Inputs: task specification ψ, program id i, and environment
observation o
function RUN(i, ψ)
r ← 0, p ← Mi,:prog , s ← fENC (o), c ← fT SE (ψ)
while r < α do
k, r ← fCN (c, p, s), ψ2 ← fT SI (ψ, p, s)
i2 ← arg max j=1...N (M key
j,: k)
if program i2 is primitive then . if i2 is an API
a ← fT SI (ψ2 , i2 , s)
. decode API args
RUN_API(i2 , a)
. run API i2 with args a
else
RUN(i2 , ψ2 ) . run program i2 w/ task spec ψ2
end if
end while
end function
index as the API argument and calls external functions to
move the gripper to above the object whose position is either
given by the simulator or predicted by an object detector.
grip closes the gripper and release opens the gripper.
Task Specification Interpreter. The Task Specification Interpreter, taking a task specification as input, chooses to perform
one of the two operations: (1) when the current program p
is not primitive, it predicts the sub-task specification for the
next sub-program; and (2) when p is primitive (i.e., an API),
it predicts the arguments of the API.
Let ψi be the task specification of the i-th program
call, where ψi is a sequence of random variables ψi =
{x1 , x2 , . . . , xNi }. The next task specification ψi+1 is determined by three inputs: the environment state si , the current
program pi , and the current specification ψi . When pi is a
primitive, TSI uses an API-specific decoder (i.e., an MLP)
to predict the API arguments from the tuple (si , pi , ψi ).
We focus on the cases when pi is not primitive. In this case,
TSI needs to predict a sub-task specification ψi+1 for the
next program pi+1 . This sub-task specification should only
access relevant information to the sub-task. To encourage
information hiding from high-level to low-level programs, we
enforce the scoping constraint, such that ψi+1 is a contiguous
subsequence of ψi . Formally, given ψi = {x1 , x2 , . . . , xNi }, the
goal is to find the optimal contiguous subsequence ψi+1 =
{x p , x p+1 , . . . , xq−1 , xq }, where 1 ≤ p ≤ q ≤ Ni .
Subsequence Selection (Scoping). We use a convolutional
architecture to tackle the subsequence selection problem.
First, we embed each input element ψi = {x1 , x2 , . . . , xNi }
into a vector space φi = {w1 , w2 , . . . , wNi }, where each wi ∈
Rd . We perform temporal convolution at every temporal
location j of the sequence φi , where each convolutional
kernel is parameterized by W ∈ Rm×dk and b ∈ Rm , which
takes a concatenation of k consecutive input elements and
produces a single output yij ∈ Rm . We use relu as the
nonlinearities. The outputs from all convolutional kernels
yij are concatenated with the program embedding pi and the
encoded states si into a single vector h j = [pi ; yij ; si ]. Finally,
Task Semantics Variation
A B C D E
Task Topology Variation
C
A B
B
A
D E
C D E
D
A B C
C D
A B
B
A
E
D
C
E
E
Task Length Variation
Fig. 4: The variability of a task structure consists of changing
success conditions (task semantics), variable subtask permutations
(task topology), and larger task sizes (task length). We evaluate the
ability of our proposed model in generalizing towards these three
types of variations.
we compute the softmax probability of four scoping labels
Pr j (l ∈ {Start, End, Inside, Outside}). These scoping
labels indicate whether this temporal location is the start/end
point of the correct subsequence, or if it resides inside/outside
the subsequence. We use these probabilities to decode the
optimal subsequence as the output sub-task specification ψi+1 .
The decoding process can be formulated as the maximum
contiguous subsequence sum problem, which can be solved
optimally in linear time. However in practice, taking the
start and end points with the highest probabilities results in
a good performance. In our experiments, we set ψi+1 =
{xst , xst+1 , . . . , xed }, where st = arg max j=1...Ni Pr j (Start)
and ed = arg max j=1...Ni Pr j (End). This process is illustrated
in Figure 3, wherein the model factorizes a video sequence
which illustrates the procedure of pick_and_place into a
fraction that only illustrates pick. This convolutional TSI
architecture is invoked recursively along the program execution trace. It decomposes a long task specification into
increasingly fine-grained pieces from high-level to low-level
tasks. This method naturally enforces the scoping constraint.
Our experimental results show that such information hiding
mechanism is crucial to good generalization.
Model Training. We train the model using rich supervision
from program execution traces. Each execution trace is a
list of tuples {ξt | ξt = (ψt , pt , st ),t = 1 . . . T }, where T is the
length of the execution trace. Our training objective is to
maximize the probability of the correct executions over all
the tasks in the dataset D = {(ξt , ξt+1 )}, such that θ ∗ =
ΣD log Pr[ξt+1 |ξt ; θ ].
We collect a dataset that consists of execution traces from
multiple types of tasks and their task specifications. For
each specification, we provide the ground-truth hierarchical
decomposition of the specification for training by rolling
a hard-coded expert policy. We use cross-entropy loss at
every temporal location of the task specification to supervise
the scoping labels. We also adopted the idea of adaptive
curriculum from NPI [24], where the frequency of each minibatch being fetched is proportional to the model’s prediction
error with respect to the corresponding program.
V. E XPERIMENTAL S ETUP
The goal of our experimental evaluation is to answer the
following questions: (1) Does NTP generalize to changes in all
three dimensions of variation: length, topology, and semantics,
as illustrated in Figure 4, (2) Can NTP use image-based input
without access to ground truth state, and (3) Would NTP
also work in real-world tasks which have combinations of
these variations. We evaluate NTP in three robot manipulation
tasks: Object Sorting, Block Stacking, and Table Clean-up.
Each of these tasks requires multiple steps to complete and
can be recursively decomposed into repetitive sub-tasks.
Input State Representation. We use an expert policy to
generate program execution traces as training data. An expert
policy is an agent with hard-coded rules that call programs
(move_to, pick_and_drop, etc.) to perform a task. In
our experiment, we use the demonstration of a robot carrying
out a task as the task specification. For all experiments, unless
specified, the state representation in the task demonstrations
is in the form of object position trajectories relative to the
gripper frame. In the Block Stacking experiment, we also
report the results of using a learned object detector to predict
object locations and the results of directly using RGB video
sequence as state observations and task demonstrations.
Simulator Setup. We conduct our experiments in a 3D
environment simulated using the Bullet Physics engine [8].
We use a disembodied PR2 gripper for both gathering
training data and evaluation. We also evaluate NTP on a
simulated 7-DoF Sawyer arm with a parallel-jaw gripper as
shown in Figure 1 and Figure 8. Since NTP only considers
end-effector pose, the choice of robot does not affect its
performance in the simulated environment.
Real-Robot Setup. We also demonstrate NTP’s performance
on the Block Stacking and the Object Sorting tasks on a 7DoF Sawyer arm using position control. We use NTP models
that are trained with simulated data. Task demonstration are
obtained in the simulator, and the instantiated NTP models
are executed on the robot. All real-robot experiments use
object locations relative to the gripper as state observations.
A Kinect2 camera is used to localize objects in the 3D scene.
Evaluation Metrics. We evaluate NTP on three variations
of task structure as illustrated in Figure 4: 1) task length:
varying number of steps due to the increasing problem size
(e.g., having more objects to transport); 2) task topology:
variations in permutations of steps of sub-tasks to reach the
same end goal (e.g., manipulating objects in different orders);
and 3) task semantics: the unseen task objectives and success
conditions (e.g., placing objects into a different container).
We evaluate Task length on the Object Sorting task varying
the number of objects instances from 1 to 10 per category.
Further, we evaluate Task Topology on the Block Stacking task
with different permutations of pick-and-place sub-tasks that
lead to the same block configurations. Finally, we evaluate
Task Semantics on Block Stacking on a held-out set of task
demonstrations that lead to unseen block configurations as
task objectives.We report success rates for simulation tasks,
and we analyze success rates, causes of failure, and proportion
of task completed for real-robot evaluation. All objects are
randomly placed initially in all of the evaluation tasks in both
the simulated and the real-robot setting.
Baselines. We compare NTP to four baselines architecture
Object Sorting: Task Length
failures: a grasp failure and a collision checking failure.
VII. E XPERIMENT 2: B LOCK S TACKING
Fig. 5: Task Length: Evaluation of the Object Sorting in simulation.
The axes represent mean success rate (y) with 100 evaluations
each and the number of objects in unseen task instances (x). NTP
generalizes to increasingly longer tasks while baselines do not.
variations. (1) Flat is a non-hierarchical model, similar to [10],
that takes as input task demonstration and current observation,
and directly predicts the primitive APIs instead of calling
hierarchical programs. (2) Flat (GRU) is the Flat model with
a GRU cell. (3) NTP (no scope) is a variant of the NTP
model that feeds the entire demonstration to the subprograms,
thereby discarding the scoping constraint. (4) NTP (GRU) is a
complete NTP model with a GRU cell. This is to demonstrate
that the reactive core network in NTP can better generalize
to longer tasks and recover from unexpected failures due to
noise, which is crucial in robot manipulation tasks.
VI. E XPERIMENT 1: O BJECT S ORTING
Setup. The goal of Object Sorting is to transport objects
randomly scattered on a tabletop into their respective shipping
containers stated in the task demonstration. We use 4 object
categories and 4 containers in evaluating the Object Sorting
task. In the real robot setup, a toy duck, toy frog, lego block,
and marker are used as the objects for sorting, and are sorted
into 4 black plastic bins. This results in a total of 44 = 256
category-container combinations (multiple categories may be
mapped to the same container). However, as each category
can be mapped to 4 possible containers, a minimum of 4
tasks can cover all possible category-container pairs. We
select these 4 tasks for training and the other 252 unseen
tasks for evaluation. We train all models with 500 trajectories.
Each test run is on 100 randomly-selected unseen tasks.
Simulator. As shown in Figure 5, NTP significantly outperforms the flat baselines. We examine how the task size
affects its performance. We vary the numbers of objects to be
transported from 4 to 40 in the experiments. The result shows
that NTP retains a stable and good performance (over 90%) in
longer tasks. On the contrary, the flat models’ performances
decline from around 40% to around 25%, which is close to
random. The performance of the NTP (GRU) model also
declines faster comparing to the NTP model as the number of
objects increases. This comparison illustrates NTP’s ability
to generalize towards task length variations.
Real robot. Table I shows the results of the Object Sorting
task on the robot. We use 4 object categories with 3 instances
of each category. We carried out a total of 10 evaluation trials
on randomly selected unseen Object Sorting tasks. 8 trials
completed successfully, and 2 failed due to of manipulation
Setup. The goal of Block Stacking is to stack a set of blocks
into a target configuration, similar to the setup in [10]. We use
8, 5×5 cm wooden cubes of different colors both in simulation
and with real-robot. We randomly generate 2000 distinct
Block Stacking task instances. Two tasks are considered
equivalent if they have the same end configuration. We use a
maximum of 1000 training tasks and 100 trials for each task,
leaving the remaining 1000 task instances as unseen test cases.
A task is considered successful if the end configuration of the
blocks matches the task demonstration. We evaluate both seen
and unseen tasks, i.e., whether the end configuration appears
in training set. We use N = 8 blocks in our evaluation.
Simulator. Figure 6 shows that all models except the Flat
baseline are able to complete the seen tasks at around 85%
success rate. The performance of the Flat baseline decreases
dramatically when training with more than 400 tasks. It is
because the Flat model has very limited expressiveness power
to represent complex tasks. The Flat (GRU) model performs
surprisingly well on the seen tasks. However, as shown in
Figure 6, both Flat and Flat (GRU) fail to generalize to unseen
tasks. We hypothesize that the Flat (GRU) baseline simply
memorizes the training sequences. On the other hand, NTP
achieves increasingly better performances when the diversity
of the training data increases.
We evaluate task topology generalization on random
permutations of the pick-and-place sub-tasks that lead to
the same end configuration. Specifically, the task variations
are generated by randomly shuffling the order that the "block
towers" are built in the training tasks. Figure 6 illustrates
that NTP generalizes better towards variable topologies when
trained on a larger variety of tasks. We find that increasing
the diversity of training data facilitates NTP to learn better
generalizable modules.
Next, we evaluate task semantics generalization. The
variability of real-world environments prevents any taskspecific policy learning method from training for every
possible task. Figure 6(A) illustrates that NTP generalizes
well to novel task demonstrations and new goals. As the
number of training tasks increases, both NTP and its recurrent
variation steadily improve their performance on the unseen
tasks. When trained with 1000 tasks, their performances on
unseen tasks are almost on par with that of seen tasks.
The performance gaps between NTP (no scope) and NTP
highlight the benefit of the scoping constraint. NTP (no scope)
performance drops gradually as the task size grows implying
that the programs in NTP learn modularized and reusable
semantics due to information hiding, which is crucial to
achieving generalization towards new tasks.
Real robot. Table I shows the results of the Object Sorting
task in the real world setting. We carried out 20 trails of
randomly selected unseen Block Stacking tasks. Out of the 2
failure cases, one is caused by an incorrect placing; the other
is caused by the gripper knocking down a stacked tower and
not able to recover from the error.
A. Seen Task Objectives
Block Stacking: Task Semantics
B. Unseen Task Objectives
Block Stacking: Task Topology
Fig. 6: Task Semantics: Simulated evaluation of the Block Stacking. The x-axis is the number of tasks used for training. The y-axis is
the overall success rate. (A) and (B) show that NTP and its variants generalize better to novel task demonstrations and objectives as the
number of training tasks increases.
Task Topology: Simulated evaluation of the Block Stacking. NTP shows better performance in task topology generalization as the number
of training tasks grows. In contrast, the flat baselines cannot handle topology variability.
B. Unseen Task Objectives
N/A
N/A
Block Stacking: Visual State
N/A
N/A
A. Seen Task Objectives
A. Adversarial Dynamics
We show that the reactive core network in NTP enables
it to better recover from failures compared to its recurrent
variation. We demonstrate this by performing Block Stacking
under an adversary. Upon stacking each block, an adversary
applies a force to the towers with a probability of 25%. The
force can knock down the towers. We evaluate NTP and its
recurrent variant on the 1000 unseen tasks. Table II shows that
under the same adversary, the success rate of NTP with the
GRU core decreases by 46%, whereas the success rate of NTP
only decreases by 20%. This indicates that a reactive model
is more robust against unexpected failures as its behavior is
less history dependent than the recurrent counterpart. We also
demonstrate this feature in the supplementary video in the
real world setting.
B. NTP with Visual State
This experiment examines the ability of NTP to learn when
demonstrations come in the form of videos and the state is
a single image. Unlike the full state information used in
experiments thus far, we train an NTP model NTPVID (E2E)
to jointly learn a policy and task-relevant features without
explicit auxiliary supervision. An alternative is to use a 2phase pipeline with an object detector as state preprocessor
for NTP, termed as NTPVID (Detector). The detector is a
separately trained CNN to predict object position in R3 .
We explore these results in Figure 7, where we see compare
the visual models (NTPVID (E2E) and NTPVID (Detector))
against the best full state model (NTP), all trained on 100
demonstrations per task, for a varying number of tasks. For
NTPVID (E2E) we use a 7-layer convolutional network,
which takes as input a 64 × 64 image and outputs a length
128 feature vector. For NTPVID (Detector) , we use a VGG16
Fig. 7: NTP with Visual State: NTPVID (Detector) uses an
object detector on images which is subsequently used as state
in NTP. NTP (E2E) is an end-to-end model trained completely
on images with no low-level state information. We note that
in the partial observation case (only video), similar learning
trends were observed as compared to fully observed case (NTP
(Full State)), albeit with a decrease in performance.
based architecture, predicting the position of the N-task
objects from an input image of size 224 × 224.
We note that NTPVID (E2E) outperforms NTPVID (Detector) and achieves a higher success rate despite only having
partial state information. Both of these methods are inferior
to the full-state NTP version. NTPVID (Detector) does
not generalize due to task-specific state representation, and
cascading errors in detection propagate to NTP reducing
performance even when using a very deep network for the
detection. The detector errors are Gaussian with standard
deviation of 2 cm. However, this performance comes at a
computational cost. NTPVID (E2E) was trained on 1000
training tasks for 10 days on 8 Nvidia Titan X GPUs. NTPVID
(Detector) was trained for 24 hours on a single GPU. Due
to computational cost, we only evaluated NTPVID (E2E) on
400 and 1000 training tasks.
VIII. E XPERIMENT 3: TABLE C LEAN - UP
Setup. We also evaluate NTP on the Table Clean-up task,
which exemplifies a practical real-world task. Specifically, the
goal of the task is to clear up to 4 white plastic bowls and
20 red plastic forks into a bin such that the resulting stack of
bowls and forks can be steadily carried away in a tray. Task
variation comes in task length, where the number of utensils
varies, and task topology, where the ordering in which bowls
are stacked can vary. Using trajectories as demonstrations
and object positions as state space, a model is trained using
1000 task instances.
Simulator. We observe that performance varies between
55%-100% where increasing errors with more objects are
attributed to failures in collision checking, not incorrect
decisions from NTP. The result shows that NTP retains
(e.g. grasping failures and collisions).
Tasks
Blk. Stk.
Sorting
# Trials
Tasks
20Stk.
Blk.
10
Sorting
Success
# Trials
0.9
20
0.8
10
NTP Fail NTP
Manip.
Success
Fail Fail
Manip. Fail
0.05
0.9
0.050.05
0.05
0.80
0 0.20
0.20
3 B, 2 F
3 B, 3 F
0.60
0.55
Fig. 8: Table Cleanup: in simulated and real environment.The table
shows success rates for varying numbers of forks and bowls in
simulated evaluation.
II: TABLE
RecoverII:from
failure:
results of
the Block
Recover
fromEvaluation
failure: Evaluation
results
of the Block
TABLE
Real Robot
Evaluation:
Results
20find
unseen
Stacking
Task
in a I:
simulated
adversarial
environment.
We
that Block
Task
in a simulated
adversarial
environment.
We findof
that
# Bowls,
Success
Stacking
evaluations
andwith
10
unseen
evaluations
on
Sawyer
Fig.
8: Table Cleanup: in simulated and real environment.The
table
NTP with
GRU
performs
markedly
worse
withsorting
intermittent
failures.
h GRU
performs
markedly
worse
intermittent
failures.
# Forks
robot for the NTP model trained on simulator. NTP Fail
denotes
shows
success rates for varying numbers of forks
and
bowls
in
2
B,
1
F
1.00
No failure
With
failures
algorithmic
mistake,
while
Manip.
Fail denotes a mistake
in evaluation.
Model an Model
No failure
With
failures
simulated
2
B,
2
F
0.95
NTP
0.863
0.663
physical interaction
(e.g. grasping
NTP
0.863
0.663 failures and collisions).
0.884 0.422
0.422
NTP (GRU)NTP (GRU)
0.884
Tasks
# Trials Success NTP Fail Manip. Fail [5]
3 B, 2 F
0.75
P. Bacon, J. Harb, and D. Precup, “The option-critic
3 B, 3architecture”,
F
0.55
arXiv preprint arXiv:1609.05140, 2016.
Blk.aStk.
20
0.9
0.05 instead
0.05 of a
Simulator.
We
use
full
Sawyer
robot
model
putational cost, we Sorting
only evaluated
NTPVID (E2E)
on0.20 [6] A. Billard,
Fig. 8:
Clean-up:
in simulated
and real
environment.The
S. Table
Calinon,
R. Dillmann,
and S. Schaal,
“Robot
program10
0
gripper
only, and
we evaluate
NTP 0.8
performance
on this
task
table
shows
success
rates
for
varying
numbers
of
forks and bowls
ming
by
demonstration”,
in
Handbook
of
robotics,
2008.
d 1000
training
tasks.
TABLE II: Adversarial Dynamics: Evaluation results of the Block
in the “A
simulated
in a simulator.
WeTask
observe
that performance
varies between
[7] that
R. Brooks,
robust evaluation.
layered control system for a mobile robot”,
Stacking
in a simulated
adversarial environment.
We find
55%-90%
where
increasing
errors
with
more
objects
can
IEEE
journal
on
roboticsinand
automation,
1986.
VIII. E XPERIMENT
3:performs
TABLEmarkedly
C LEANworse
U P with intermittent failures.
NTP with GRU
Fig. 9: Table Cleanup:
simulated
and real
environment.The table
[11]
L.
Fei-Fei,
R.
Fergus,
and
P.
Perona,
“One-shot learning of object
[8]
Bullet
Physics
Library,
http://bulletphysics.org/.
be attributed to failures
in
collision
checking,not
incorrect
shows
success
rates
for
varying
numbers
of
forks and and
bowls
in
Model on the
No practical
failure With
failures
categories”,
transactions
on pattern
machine
p. We also evaluate NTP
task
of
[9] simulated
J. Cai, R.evaluation.
Shin, and IEEE
D. Song,
“Making
neuralanalysis
programming
decisions from NTP. NTP
0.863
0.663
intelligence,
vol.
28,
no.
4,
pp.
594–611,
2006.
architectures generalize via recursion”, ICLR, 2017.
g the table, which combines
the features
stacking
(GRU)
0.884theof
0.422model on
[12]
R. E.
P. E. Hart,
and N.and
J. Nilsson,
“Learning and executing
Real robot. We haveNTP
also
transferred
trained
[14]
R. Fox,
S. Fikes,
Krishnan,
I. Stoica,
K. Goldberg,
[10]
C. Devin,
A.generalized
Gupta,
T. robot
Darrell,
P. Abbeel,
and
S. Levine, “Multi-Level
“Learning
plans”,
ARTIFICIAL
INTELLIGENCE,
1972.
ting. Specifically,
the
goal
of
the
task
is
to
clear
upto
Discovery of
Deep Options”,
preprint
arXiv:1703.08294,
2017.
its generalization
abilitythe
in feasibility
a task thatasrequires
the real-Sawyer
arm to evaluate
shown multiple
in
Modular
Network
PoliciesI. in
for
Multi-Task
and Multi-Robot
[13]Neural
R. Fox,
S. Krishnan,
Stoica,
and K. Goldberg,
“Multi-Level
[15]
C.
L.
Giles,
C.
B.
Miller,
D.
Chen,
H.-H.
Chen,
G.-Z.
Sun,
and
plastic
bowls
and
20
red
plastic
forks
into
a
bin
such
Transfer”, arXiv
preprint
arXiv:1609.07088,
2017.
dimensions
of generalization.
Discovery
of Deep
Options”, in Preprint
arXiv:1703.08294, 2017.
Figure 9. We
also demonstrate
this task in the supplementary
Y.-C.
Lee,
“Learning
and
extracting
finite
state
automata
with
second[11]
J.
Devlin,
J.
Uesato,
S.
Bhupatiraju,
R.
Singh,
A.-r.
Mohamed,
and
C.
L.
Giles,
C.
B.
Miller,
D.
Chen,
H.-H.
Chen,
G.-Z.
Sun,
and
[14]
resulting
ofworld
bowls
and forks
can be steadily
Real
robot.
Wesetting.
have
also transferred
the trained model on
video in stack
the
real
order
recurrent
networks”,
Learning,
vol.
4,
no.
3, with
2008.
Y.-C. Lee, neural
“Learning
and extracting
finite state
automata
secondP.
Kohli,
“RobustFill:
Neural
Program
Learning
under
Noisy
I/O”,
the real-Sawyer
arm tocomes
evaluate
in
away in a tray.
Task variation
in the
taskfeasibility
length, as shown [16]
A. Graves,
G.
Wayne,neural
and I.networks”,
Danihelka,
“Neuralvol.
Turing
order
recurrent
4, no.Machines”,
3, 2008.
arXiv preprint
arXiv:1703.07469,
2017. Learning,
IX.
D ISCUSSION
& task
Fthis
UTURE
Wthe
ORK
arXiv preprint
arXiv:1410.5401,
2014.“Long short-term memory”, Neural
8.
Wevaries,
demonstrate
task in
supplementary
video
S.
Hochreiter
and
J.
Schmidhuber,
[15]
he number ofFigure
utensils
and
topology,
where
[12] A. Dosovitskiy and V. Koltun, “Learning to act by predicting the
[17] S. Hochreiter
and J. vol.
Schmidhuber,
“Long
short-term
memory”, Neural
computation,
9, no. 8, pp.
1735–1780,
1997.
in thebowls
real
world
setting.
ICLR,
2017.
Neural
Programming
eringWe
in introduced
which
areTask
stacked
can vary.(NTP),
Usinga meta- future”,[16]
T. M.
Jochem,
and C. E.
Thorpe, “Maniac: A next
computation,
vol. 9, D.
no.A.8,Pomerleau,
pp. 1735–1780,
1997.
[13] [18]
Y. Duan,
Andrychowicz,
B. C.
Stadie,
J. Ho,
J. Schneider,
generation
based
autonomous
follower”,
in Int’l
Conf.
framework that
learns
modular
reusable
neural
T. M.M.Jochem,
D.neurally
A. Pomerleau,
and
C. E. road
Thorpe,
“Maniac:
A next
rieslearning
as demonstrations
and
object
positions
as state
IX.
D ISCUSSION
& Fand
UTURE
W ORK
I. Sutskever,
Abbeel,based
and autonomous
W. Zaremba,
“One-Shot
Imitation
onP.Intelligent
Autonomous
Systems,
IOS
Publishers,
Amsterdam.,
generation
neurally
road
follower”,
in
Int’l Conf.
programs
for
hierarchical
tasks.
We
demonstrate
NTP’s
a model is trained
using 1000
demonstrations.
Learning”,
arXiv
preprint
arXiv:1703.07326,
Pittsburgh,
PA, 1993,
pp.
15–18. IOS 2017.
We introduced
Neural
Task Programming (NTP), a metaon Intelligent
Autonomous
Systems,
Publishers, Amsterdam.,
in two
thatinstead
require
pro[17]
J. Kober,
A.
Peters, “Reinforcement
[14]
Y. Duan,
J. Schulman,
X. Bagnell,
Chen,
P.and
L.J.Bartlett,
I. Sutskever, learning
and P. in
tedstrengths
world. We
use robot
aframework
full manipulation
Sawyer
model
Pittsburgh,
PA, J.
1993,
pp. 15–18.
learning
that robot
learnstasks
modular
and
reusable
neural
2robotics:
AFern,
survey”,
The Int’l
Journal
ofGoetschalckx,
Robotics
Research,
vol. 32,
Abbeel,
“RL
:
Fast
Reinforcement
Learning
via
Slow
Reinforcement
longed
and
complex
interactions
with
the
environment.
NTP
[19]
K.
Judah,
A.
P.
P.
Tadepalli,
and
R.
“Imitation
programs
for
hierarchical
tasks.
We
demonstrate
NTP’s
ipper only, and we evaluate NTP performance on
no. 11,preprint
pp. 1238–1274,
2013.
Learning”,
arXiv
arXiv:1611.02779,
2016.
learning
with
demonstrations
and
shaping
rewards”,
in
AAAI,
2014.
strong
generalization
results
towards varies
task
[18] S. Krishnan*, A. Garg*, S. Patil, C. Lea, G. Hager, P. Abbeel, and
strengths
in
three that
robot
manipulation
tasks length,
that require
k inachieves
a simulator.
We
observe
performance
J. Kober,
J. A. Bagnell,
and J. state
Peters,
“Reinforcement
learning
in
[15] [20]
L. Fei-Fei,
R.
and“Transition
P. Perona,
“One-shot
of
object
K. Fergus,
Goldberg,
clustering:learning
Unsupervised
surgical
topology, and
semantics.
This work
opens up with
the opportunity
prolonged
and
complex
interactions
the
environment.
robotics:
A
survey”,
The
Int’l
Journal
of
Robotics
Research,
vol.
32,
categories”,
IEEE
transactions
on
pattern
analysis
and
machine
n 55%-90% where increasing errors with more objects
trajectory segmentation for robot learning”, IJRR, 2018.
to use generalizable
neural
programs fortowards
modeling
no. 11,T.
pp.
1238–1274,
NTP achieves
generalization
taskhierarchical
length, topology,
intelligence,
vol.
no. 4, pp.
594–611, 2006.
[19]
D.28,
Kulkarni,
K.2013.
Narasimhan,
A. Saeedi, and J. Tenenbaum,
attributed to failures in collision checking,not incorrect
[21]
S. Krishnan*,
A. Garg*,
C. Lea,
G. Hager,
P. executing
Abbeel,
and
“Hierarchical
deepN.S.
reinforcement
learning:
Integrating
temporal
[16]
R. E. Fikes,
P. E. Hart,
and
J.Patil,
Nilsson,
“Learning
and
semantics.
This
opens uptothe
opportunity
to use
tasks. We and
intend
to extend
thiswork
framework
tackle
a richer
K. Goldberg,
“Transition
state
clustering:
Unsupervised
surgical
ns from
NTP.
abstraction
and Artificial
intrinsic
motivation”,
in NIPS,
2016.
generalized
robot
plans”,
Intelligence,
1972.
generalizable
neural
programs
for
modeling
hierarchical
set of complex tasks on real robots in future work.
trajectory
segmentation
for robot
2018.
[20]
Mnih,
K. I.
Kavukcuoglu,
D.learning”,
Silver,
A. IJRR,
A. Rusu,
J. Veness, M. G.
[17] state
R. Fox,
S. V.
Krishnan,
Stoica,
and
K. Goldberg,
“Multi-Level
orld. We have
also
model
on
tasks.
Fortransferred
future work,the
wetrained
intend to
1) improve
the
[22]
T.
D.
Kulkarni,
K.
Narasimhan,
A.
Saeedi,
and
J. Tenenbaum,
Bellemare,
A.
Graves,
M.
Riedmiller,
A.
K.
Fidjeland,
G.
Ostrovski,
Discovery
of
Deep
Options”,
in
preprint
arXiv:1703.08294,
2017.
Rthe
EFERENCES
-Sawyer arm
to evaluate
feasibility
shown insuch as object “Hierarchical
et al., “Human-level
control through
deep Integrating
reinforcementtemporal
learning”,
encoder
to extract
more
task-salientasinformation
deep reinforcement
learning:
[18] C. L. Giles, Nature,
C. B. vol.
Miller,
D.
Chen,
H.-H.
Chen,
G.-Z.
Sun,
and
518, no.motivation”,
7540, pp. 529–533,
abstraction and intrinsic
in NIPS,2015.
2016.
[1] also
J. Andreas,
D. Klein,
S. Levine,
“Modular
reinforcerelationships,
2)and
devise
a richer
set ofmultitask
APIs such
as velocity
8. We
demonstrate
this
task
in
the
supplementary
Y.-C.
Lee,
“Learning
and
extracting
finite
state
automata
with
secondA. Y.
D. Harada, D.
andSilver,
S. J. Russell,
“Policy
[23] [21]
V. Mnih,
K. Ng,
Kavukcuoglu,
A. A. Rusu,
J. invariance
Veness, M.under
G.
mentand
learning
with policycontrollers,
sketches”, in and
ICML,
torque-based
3) 2017.
extend this framework
order recurrent
neural
networks”,
Learning,
vol.
4, no.to3,reward
2008.
n the[2]realJ. world
reward
Theory
and
shaping”,
Bellemare,
A.transformations:
Graves,
M. Riedmiller,
A. application
K. Fidjeland,
G.
Ostrovski,
Andreas,setting.
M. Rohrbach, T. Darrell, and D. Klein, “Neural module
in
ICML, 1999,
pp.
278–287.
[19] A. Graves,
Wayne,
and control
I. Danihelka,
Turing Machines”,
to tackle
more 2016.
complex tasks on real robots.
et al.,G.
“Human-level
through“Neural
deep reinforcement
learning”,
networks”,
in CVPR,
[22]
Y. Ng
J. Russell,
“Algorithms
for inverse reinforcement
arXiv preprint
arXiv:1410.5401,
2014.
IX. B.DD.
ISCUSSION
& F UTURE
W
ORK
Nature,A.vol.
518,and
no.S.7540,
pp.
529–533,
2015.
[3]
Argall, S. Chernova,
M. Veloso,
and B. Browning, “A Survey
ACKNOWLEDGMENT
learning.”,
in ICML, 2000.
[20] [24]
S. Hochreiter
andS.J. Sen,
Schmidhuber,
“Long
memory”,
Neural
A. Murali,
B. Kehoe, A.
Garg,short-term
S. McFarland,
et al., “Learning
of Robot Learning from Demonstration”, Robotics and Autonomous
[23] R.vol.
Parr9,and
S.8,J. pp.
Russell,
“Reinforcement
learning with hierarchies
ntroduced
Neural
Task
(NTP),
a metacomputation,
no.
1735–1780,
1997.
by observation
for
surgical
subtasks:
Multilateral
cutting
of 3d
This
was
performed
at 2009.
the
SVL at Stanford
in affiliation with
Systems,
vol.research
57, no.Programming
5, pp.
469–483,
of machines”, in NIPS, 1998.
[21] T. M. Jochem,
A. and
Pomerleau,
andtissue
C.“Neural
E.phantoms”,
Thorpe,
“Maniac:
A next
viscoelastic
and
2d N.
orthotropic
in IEEE Int’l
Conf.in
[4] P. Bacon,
J. learns
Harb,
D.Stanford-Toyota
Precup,and
“Thereusable
option-critic
architecture”,
g framework
that
modular
the
Stanford
AIand
Lab,
AI
Center. neural
[24]
S. D.
Reed
de Freitas,
programmer-interpreters”,
on Robotics
and
Automation,
ICRA,
2015.
generation
neurally
based
autonomous
road
follower”,
in
Proceedings
arXiv preprint arXiv:1609.05140, 2016.
ICLR, 2016.
ms for
tasks.R. Dillmann,
We Rdemonstrate
[25]
A. Y. S.
Ng,Ross,
D. Harada,
and on
S.and
J.
Russell,
“Policy
invariance
under
[5] hierarchical
A. Billard, S. Calinon,
and S. Schaal, NTP’s
“Robot programof the [25]
International
Conference
Intelligent
Autonomous
Systems,
EFERENCES
G. J. Gordon,
D.
Bagnell,
“A reduction
of imitation
rewardlearning
transformations:
Theory
and application
to
reward
shaping”,in
hs in twoming
robot
manipulation
tasks
that
require
proby
demonstration”,
in
Handbook
of
robotics,
2008.
and
structured
prediction
to
no-regret
online
learning”,
IOS
Publishers,
Amsterdam.,
Pittsburgh,
PA,
1993,
pp.
15–18.
[1] J. Andreas, D. Klein, and S. Levine, “Modular multitask reinforcein ICML,
278–287.Intelligence
R. Brooks,
“A robust
layered
control
system in
forICML,
a mobile
Conf. pp.
on Tadepalli,
Artificial
and Statistics, 2011.
[22] K. Judah,
A.Int’l
P. 1999,
Fern,
P.
and R. Goetschalckx,
“Imitation
ment
learning
with the
policy
sketches”,
2017.robot”,
and [6]
complex
interactions
with
environment.
NTP
[26]
A.with
Y. Ng
and S. D.
J. Russell,
“Algorithms
reinforcement
[26]
T.demonstrations
Schaul,
Horgan,
Gregor, rewards”,
andfor
D.inverse
Silver,
“Universal
value
IEEE journal
on robotics
automation,
1986.
learning
andK.
shaping
in AAAI,
2014.
[2] J. Andreas,
M. and
Rohrbach,
T. Darrell,
and D. Klein, “Neural module
s strong
generalization
results
task length, [23] Ł. Kaiser
approximators”,
ICML, learn
2015. algorithms”, arXiv
learning.”,
in ICML,
2000.
[7] Bullet
Physics
Library,
networks”,
inhttp://bulletphysics.org/.
CVPR, towards
2016.
andfunction
I. Sutskever,
“NeuralinGPUs
[27]
Sen*,
D.“Reinforcement
Gealy, S. McKinley,
Y. Jen,
andhierarchies
K. Goldberg,
[27]
R.arXiv:1511.08228,
ParrS.and
S. A.
J. Garg*,
Russell,
learning
with
[8] semantics.
J. Cai,[3]R. B.
Shin,
and D.
Song, up
“Making
neural
D. Argall,
S. opens
Chernova,
M.the
Veloso,
and B.programming
Browning, “A Survey
y, and
This
work
opportunity
preprint
2015.
“Autonomous
Multiple-Throw
Multilateral Surgical Suturing with a
of machines”,
in NIPS,
1998.
architecturesofgeneralize
via recursion”,
ICLR, 2017.
Robot Learning
from Demonstration”,
Robotics and Autonomous
[24] S. Krishnan,Mechanical
A. Garg, Needle
S. Patil,
C. and
Lea,
G. Hager,based
P. Abbeel,
and
eneralizable
neural
programs
hierarchical
Guide
Optimization
Needle Planning”,
[28] S. Reed and N. de
Freitas,
“Neural
programmer-interpreters”,
in
[9] J. Devlin,
J.Systems,
Uesato,
S. Bhupatiraju,
Singh,
A.-r.
vol.
57,for
no. modeling
5, pp.R.
469–483,
2009. Mohamed, and
K. Goldberg,
“Transition
state
clustering:
Unsupervised
surgical
in
IEEE
Int.
Conf.
Robotics
and
Automation
(ICRA),
2016.
[4]
P.
Bacon,
J.
Harb,
and
D.
Precup,
“The
option-critic
architecture”,
ICLR, 2016.
Kohli,
“RobustFill:
Neural Program
Noisy I/O”,
We intendP. to
extend
this framework
to Learning
tackle aunder
richer
[28]
J. Sung,
Selman,
andD.
A.Bagnell,
Saxena,IJRR,
“Learning
of contrajectory
segmentation
for robot
learning”,
2018.sequences
ArXiv
preprint arXiv:1609.05140,
2016.
[29]
S. Ross,
G. J. B.
Gordon,
and
“A
reduction
of imitation
arXiv preprint
arXiv:1703.07469,
2017.
trollersK.
complex manipulation
in International
Conference
omplex
tasks
on A.
real
robotsCalinon,
in future
work.
[25]
T. D. learning
Kulkarni,
Narasimhan,
A. Saeedi,
and
J. Tenenbaum,
[5]
Billard,
R. Dillmann,
“Robot
programand for
structured
prediction
to tasks”,
no-regret
online
learning”,
in
[10] A. Dosovitskiy
and V.S.Koltun,
“Learning
to and
act S.
bySchaal,
predicting
the
ondeep
Machine
Learning, Citeseer,
2013. Integrating temporal
“Hierarchical
reinforcement
learning:
ming
by demonstration”, in Handbook of robotics, 2008.
Int’l
Conf.
on
Artificial
Intelligence
and
Statistics,
2011.
future”, ICLR,
2017.
[29] and
R. S.intrinsic
Sutton, D.motivation”,
Precup, and S.in
Singh,
“Between
and semi-mdps:
R.
Brooks, “A robustB.layered
controlJ. system
a mobile robot”,
[6] R
EFERENCES
abstraction
NIPS,
2016. mdps
[30]
T. Schaul,
D. Horgan,
Gregor,
and
D. Silver,
“Universallearning”,
value
[11] Y. Duan,
M.
Andrychowicz,
C. Stadie,
Ho, J.forSchneider,
A framework
for K.
temporal
abstraction
in reinforcement
IEEE
journal
on
robotics
and
automation,
1986.
[26]
K.
Kurach,
M.
Andrychowicz,
and
I.
Sutskever,
“Neural
randomfunction
approximators”,
invol.
ICML,
2015.
P.S.Abbeel,
and
W. Zaremba,
“One-Shot
Imitation
Andreas, I.
D. Sutskever,
Klein,
and
Levine,
“Modular
multitask
reinforceArtificial
intelligence,
112,
no.
1-2,
pp.
181–211,
1999.
[7] J. Cai, R. Shin, and D. Song, “Making neural programming
access [30]
machines”,
arXiv
preprint
[31]
S.
Sen*,
Garg*,
S.
Jen,2015.
and survey
K. Goldberg,
Learning”,
arXivsketches”,
preprint arXiv:1703.07326,
R. A.
Vilalta
andD.
Y.Gealy,
Drissi,arXiv:1511.06392,
“AMcKinley,
perspectiveY.view
and
of metaent learning
with policy
in ICML,
2017. 2017.
architectures generalize
via recursion”,
ICLR, 2017.
“Autonomous
Multiple-Throw
Multilateral
Surgical
SuturingM.with
[27] V. Mnih,
K. learning”,
Kavukcuoglu,
D.Intelligence
Silver,
A. Review,
A. Rusu,
J. Veness,
G. a
[12]
L.
Fei-Fei,
and
P.Y.Perona,
“One-shot
learning
of object
Artificial
2002.
Andreas, M. Rohrbach,
T. Darrell,
D.
“Neural
module
[8] R.
E. Fergus,
Coumans
andand
Bai,Klein,
Pybullet,
a python
module
for physics
Mechanical
Needle
Guide
and
Optimization
based
Needle
Planning”,
categories”,
IEEE
transactions
on
pattern
analysis
and
machine
[31]
O.
Vinyals,
C.
Blundell,
T.
Lillicrap,
D.
Wierstra,
et
al.,
“Matching
Bellemare,
A.
Graves,
M.
Riedmiller,
A.
K.
Fidjeland,
G.
Ostrovski,
simulation, games, robotics and machine learning, http : / /
etworks”, in CVPR, 2016.
IEEE
Int. Conf.
Robotics
and Automation
(ICRA),
2016.
networks
for
one
shot
learning”,
in reinforcement
Advances
in Neural
Information
intelligence,pybullet.org/,
vol. 28, no. 4, pp.2016–2017.
594–611, 2006.
et al., in
“Human-level
control
through
deep
learning”,
M. Andrychowicz,
M. Denil,
S. G. Colmenarejo,
M. W. Hoffman,
Processing
Systems,
2016,
pp. 3630–3638.
[13] R. E. Fikes,
P.Devlin,
E. Hart,
and N. S.
J. Nilsson,
“Learning
and
executing
Nature,
vol.
518,
no.
7540,
pp.
529–533,
2015.
[9]
J.
J.
Uesato,
Bhupatiraju,
R.
Singh,
A.-r.
Mohamed,
and
. Pfau, T. Schaul, and N. de Freitas, “Learning to learn by gradient
[32] S.
Y. Wu
Demiris, A.
“Towards
learning byS.imitation
generalized P.
robot
plans”,
Artificial
Intelligence,
Kohli,
Neural
Program 1972.
Learning under Noisy
[28] I/O”,
A. Murali,
Sen,andB.Y. Kehoe,
Garg,one
S. shot
McFarland,
Patil, for
escent by gradient descent”,
in“RobustFill:
NIPS, 2016.
humanoid robots”, in ICRA, 2010.
ArXiv preprint arXiv:1703.07469, 2017.
W.
D.
Boyd,
S.
Lim,
P.
Abbeel,
and
K.
Y.
Goldberg,
“Learning
. D. Argall, S. Chernova,
M. Veloso,
and B. Browning,
“A Survey
[33] Y. Zhu, R. Mottaghi, E. Kolve, J. J. Lim, A. Gupta, L. Fei-Fei, and
[10] Y. Duan,
M. Andrychowicz,
B. C. Stadie,
J. Ho, J. Schneider,
by observation
for surgical
subtasks:
Multilateral
of Scenes
3d visA. Farhadi,
“Target-driven
Visual
Navigationcutting
in Indoor
using
Robot Learning fromI. Demonstration”,
Robotics
andZaremba,
Autonomous
Sutskever, P. Abbeel,
and W.
“One-Shot Imitation
coelastic andDeep
2d orthotropic
tissue
phantoms”,
in IEEE
Reinforcement
Learning”,
in ICRA,
2017. International
ystems, vol. 57, no. 5,Learning”,
pp. 469–483,
ArXiv 2009.
preprint arXiv:1703.07326, 2017.
| 2cs.AI
|
On One Generalization of LRC Codes with
Availability
Stanislav Kruglik∗†‡ , Marina Dudina∗, Valeriya Potapova∗† and Alexey Frolov∗†
∗ Skolkovo Institute of Science and Technology
Moscow, Russia
arXiv:1705.11095v1 [cs.IT] 30 May 2017
† Institute for Information Transmission Problems
Russian Academy of Sciences
Moscow, Russia
‡ Moscow Institute of Physics and Technology
Moscow, Russia
[email protected], [email protected], [email protected], [email protected]
Abstract—We investigate one possible generalization of locally
recoverable codes (LRC) with all-symbol locality and availability
when recovering sets can intersect in a small number of coordinates. This feature allows us to increase the achievable code
rate and still meet load balancing requirements. In this paper we
derive an upper bound for the rate of such codes and give explicit
constructions of codes with such a property. These constructions
utilize LRC codes developed by Wang et al.
I. I NTRODUCTION
A locally recoverable code (LRC) is a code over finite
alphabet such that each symbol is a function of small number
of other symbols that form a recovering set [1], [2], [3], [4],
[5]. These codes are important due to their applications in
distributed and cloud storage systems. LRC codes are wellinvestigated in the literature. The bounds on the rate and
minimum code distance are given in [1], [3] for the case
of large alphabet size. The alphabet-dependent shortening
bound (see [6] for the method explanation) is proposed in [7].
Optimal code constructions are given in [8] based on rankmetric codes (for large alphabet size, which is an exponential
function of the code length) and in [9] based on Reed-Solomon
codes (for small alphabet, which is a linear function of the
code length).
The natural generalization of an LRC code is an LRC
code with availability (or multiple disjoint recovering sets).
Availability allows us to handle multiple simultaneous requests
to erased symbol in parallel. This property is very important
for hot data that is simultaneously requested by a large number
of users. The case of LRC codes with availability is much
less investigated. Bounds on parameters of such codes and
constructions are given in [4], [10], [11], [12]. Most of the
papers focused on information-symbol locality and availability.
In what follows we are interested in all-symbol locality and
availability that is preferable in applications as it permits a
uniform approach system design.
The property of availability decreases maximum achievable
code rate [10]. In this paper we propose a new generalization
of LRC codes with availability. Namely, we assume that
recovering sets can intersect in a small number of coordinates.
This feature allows us to increase the achievable code rate and
still meet load balancing requirements.
Our contribution is as follows. We investigate one possible
generalization of locally recoverable codes (LRC) with allsymbol locality and availability when recovering sets can
intersect in a small number of coordinates. We derive an upper
bound for the rate of such codes and give explicit constructions
of codes with such a property. These constructions utilize LRC
codes developed in [13].
II. P RELIMINARIES
A. LRC codes
Let us denote by Fq a field with q elements. Let [n] =
{1, 2, . . . , n}. The code C ⊂ Fnq has locality r if every symbol
of the codeword c ∈ C can be recovered from a subset of
r other symbols of c [1]. In other words, this means that,
given c ∈ C, i ∈ [n], there exists a subset of coordinates Ri ⊂
[n]\i, |Ri | ≤ r such that the restriction of C to the coordinates
in Ri enables one to find the value of ci . The subset Ri is
called a recovering set for the symbol ci .
B. LRC codes with availability
Generalizing this concept, assume that every symbol of the
code C can be recovered from t disjoint subsets of symbols of
size r. More formally, denote by CI the restriction of the code
C to a subset of coordinates I ⊂ [n]. Given a ∈ Fq define the
set of codewords C(i, a) = {c ∈ C : ci = a}, i ∈ [n].
Definition 1: A code C is said to have t disjoint recovering
sets if for every i ∈ [n] there are t pairwise disjoint subsets
R1i , . . . , Rti ⊂ [n]\i such that for all j = 1, . . . , t and every
pair of symbols a, a′ ∈ Fq , a 6= a′
C(i, a)Rj ∩ C(i, a′ )Rj = ∅.
i
i
In what follows we refer these codes as (r, t)-LRC codes.
We briefly list the existing results below. The first bound for
(r, t)-LRC codes was given in [14], [15]
t(k − 1) + 1
.
d≤n−k+2−
t(r − 1) + 1
An improvement of this bound was obtained in [10]
t
X
k−1
d≤n−
.
ri
i=0
III. A N
min
k−1
1≤x≤⌈ (r−1)t+1
⌉;1≤yj ≤t;j∈[x]
A. The recovery graph
dql−opt [n − B, k − A],
A<k;x,yj ∈Z +
P
P
where A = xj=1 (r −1)yj +x, B = xj=1 ryj +x and dql−opt
denote the largest possible minimum distance of a code over
Fq .
The bound on the rate of (r, t)-LRC codes was given in
[10]
t
Y
1
k
(1)
≤ R∗ (r, t) =
1 .
n
1
+
ir
i=1
This bound was improved in [11] for t = 2.
In [13] a recursive construction of binary (r, t)-LRC codes
was proposed.
The parameters of these codes are as follows:
r
n = r+t
,
R
=
t
r+t and d = t + 1. We refer these codes as
WZL codes. WZL code is defined by its’ parity-check matrix.
Let m = r + t. Let us define matrix H(m, t) as follows.
Each row of H(m, t) is associated with (t − 1)-subset of [m]
sorted in lexicographical order, each column – with t-subset
of [m] also sorted in lexicographical order. In this case the
element (i, j) of H(m, t) is equal to 1 if Ei ⊆ Fj , where
Ei is (t − 1)-subset of [m] associated with i-th row and Fj
is t-subset of [m] associated with
j-th column.
It must be
m
mentioned that H(m, t) has t−1
rows and m
t columns and
has the following structure:
H(m, t) =
(r, t, x)-LRC
CODES
An alphabet-dependent bound was proposed in [12] and has
form
d≤
UPPER BOUND ON THE RATE OF
Based on the original idea from [10] we represent locally
recoverable codes with locality r and availability t as a graph
G in the following way. In accordance to the Definition 2 a
coordinate i has t recovering sets R1i , ...Rti , each of size r,
where Rji ⊂ [n]\i. Define a directed graph G as follows. The
set of vertices V = [n] corresponds to the set of n coordinates
of the LRC code. The ordered pair of vertices (i, j) forms a
directed edge i → j if j ∈ Rli for some l ∈ [t]. We color the
edges of the graph with t distinct colors in order to differentiate
between the recovering sets of each coordinate. Note, that as
the recovering sets can intersect, then some edges may have
several colors. We call G the recovery graph of the code C.
In what follows we need the following lemma
Lemma 1: Let j ∈ [t] and s = min{j, ⌊r/x⌋ + 1}, then
j
[
(2r − (s − 1)x)
Rli ≤ N (r, j, x) = jr.
s = N (r, j, x) ≤
2
l=1
Proof: The upper bound is trivial and correspond to the
case, when recovering sets do not intersect. To prove the lower
bound assume, that any two recovering sets intersect in exactly
x positions, we have
t
[
Rli
≥
r + (r − x) + (r − 2x) + . . . + (r − (s − 1)x)
=
(2r − (s − 1)x)
s.
2
l=1
Corollary 1: The out-degree of each vertex i ∈ V = V (G)
is upper bounded with tr and lower bounded with N (r, t, x).
!
H(m − 1, t − 1)
0
I(m−1)
H(m − 1, t) ,
t−1
where H(m, m) = H(m, 1)T
=
dim(H(m, 1)) = dim(H(m, t)) = m.
(1, ..., 1)T
and
C. LRC codes with availability and intersection of recovering
sets
Let us give to recovering sets an ability to intersect in at
most x positions and define this code as (r, t, x)-LRC. More
formally, we can say
Definition 2: A code C is said to be (r, t, x)-LRC if for
every i ∈ [n] there are t subsets Ri1 , ..., Rit ⊂ [n] \ i such, that
the following relations follow
1) for every pair l, l′ ∈ [t], l 6= l′
′
|Ril ∩ Ril | ≤ x;
2) for all j = 1, . . . , t and every pair of symbols a, a′ ∈
Fq , a 6= a′
C(i, a)Rj ∩ C(i, a′ )Rj = ∅.
i
i
In what follows we investigate the parameters of such codes.
B. Upper bound on the rate
The proof is very similar to the proof from [10]. For the
simplicity of the reader we present the proof here in all the
details. Let us introduce the following function
f (r, t, x)
t
1
j N (r, j, x) + 1
j=1,j=1 mod 2
t
X
t
1
−
j N (r, j, x) + 1
=
t
X
j=1,j=0
mod 2
The following lemma will be used in the proof.
Lemma 2: There exists a subset of vertices U ⊆ V of size
at least
|U | ≥ nf (r, t, x),
such that for any U ′ ⊆ U , the induced subgraph GU ′ on the
vertices U ′ has at least one vertex v ∈ U ′ such that its set of
outgoing edges {(v, j), j ∈ U ′ )} is missing at least one color.
Proof: For a given permutation τ of the set of vertices
V = [n], we define the coloring of some of the vertices as
follows: The color j ∈ [t] is assigned to the vertex v if
τ (v) > τ (m)
for all m ∈ Rjv .
(2)
Now let Xv be the indicator random variable for the event
that v ∈ U , then
X
E(|U |) =
E(Xv )
v∈V
=
Rjv ,
If this condition is satisfied for several recovering sets
the
vertex v is assigned any of the colors j corresponding to these
sets. Finally, if this condition is not satisfied at all, then the
vertex v is not colored.
Let U be the set of colored vertices, and consider one of
its subsets U ′ ⊆ U .
Let GU ′ be the induced subgraph on U ′ . We claim that there
exists v ∈ U ′ such that its set of outgoing edges is missing
at least one color in GU ′ . Assume toward a contradiction that
every vertex of GU ′ has outgoing edges of all t colors. Choose
a vertex v ∈ U ′ and construct a walk through the vertices of
GU ′ according to the following rule. If the path constructed
so far ends at some vertex with color j, choose one of its
outgoing edges also colored in j and leave the vertex moving
along this edge. By assumption, every vertex has outgoing
edges of all t colors, so this process, and hence this path can
be extended indefinitely. Since the graph GU ′ is finite, there
will be a vertex, call it v1 , that is encountered twice. The
segment of the path that begins at v1 and returns to it has the
form
v1 → v2 → ... → vl ,
where v1 = vl . For any i = 1, ..., l − 1 the vertex vi and the
edge (vi , vi+1 ) are colored with the same color. Hence by the
definition of the set U we conclude that τ (vi ) > τ (vi+1 ) for
all i = 1, . . . , l − 1, a contradiction.
In order to show that there exists such a set U of large
cardinality, we choose the permutation τ randomly and uniformly among all the n! possibilities and compute the expected
cardinality of the set U.
Let Av,j be the event that (2) holds for the vertex v and the
color j. Since Pr(Av,j ) does not depend on v, we suppress
the subscript v, and write
Pr(v ∈ U ) = Pr(∪tj=1 Aj ).
Let us compute the probability of the event ∪tj=1 Aj . Note
that for any set S ⊆ [t] the probability of the event that all the
Aj , j ∈ S occur simultaneously can be estimated as follows
1
1
≤ P (∩j∈S Aj ) ≤
.
N (r, |S|, x) + 1
N (r, |S|, x) + 1
Hence by the inclusion exclusion formula we get
t
X
j−1 t
t
P (A1 ∩ ... ∩ Aj )
(−1)
Pr(∪j=1 Aj ) =
j
j=1
t
X
t
1
≥
j N (r, j, x) + 1
j=1,j=1 mod 2
t
X
t
1
−
j N (r, j, x) + 1
j=1,j=0
=
Pr(v ∈ U )
v∈V
= n Pr(∪tj=1 Aj )
≥ nf (r, t, x).
The proof is completed by observing that there exists at least
one choice of τ for which |U | ≥ E(|U |).
Theorem 1: The rate of an (r, t, x)-LRC code C satisfies
R(C) ≤ R∗ (r, t, x) = 1 − f (r, t, x).
Proof: The colored vertices can be viewed as check
symbols as they can be recovered from the rest symbols.
Thus, the number of information symbols can be estimated
as follows
k ≤ n(1 − f (r, t, x)).
IV. L OWER BOUNDS ON THE RATE OF (r, t, x)-LRC CODES
In this section we derive a lower bound on the rate of codes
with all symbol locality and availability in which recovering
sets can intersect. To find a lower bound we propose the
following rather simple code construction. In what follows
we explain how to construct a parity-check matrix of a linear
(r, t, x)-LRC code. We start with a parity-check of (r̃, x̃)-WZL
code. Let is denote the matrix by HWZL . The matrix H of
(r, t, x)-LRC code is constructed as follows
H = HWZL ⊗ [11 . . . 1],
| {z }
x+1
where ⊗ denotes a Kronecker product of matrices.
t̃
As a result, we have a matrix of length n = (x + 1) r̃+
t̃ .
It is obvious, that each row of the new matrix will have (x +
1)(r̃ + 1) ones and the number of positions, in which two
recovering sets intersects is equal to x. This construction will
have the same availability t as it was for the standard WZL
code. Thus, the parameters of the resulting code are as follows
r
= (r̃ + 1)(x + 1) − 1
t
= t̃
The matrix H has exactly the same
rank as the matrix
t̃−1
HWZL , the rank is equal to r̃+
.
Thus,
the rate of the
t̃−1
resulting code can be calculated as follows
R
t̃
(r̃ + t̃)(x + 1)
t
= 1−
r+1
x+1 − 1 + t (x + 1))
= 1−
t
r + t + (t − 1)x
r + (t − 1)x
.
r + t + (t − 1)x
= 1−
mod 2
f (r, t, x).
X
=
TABLE I
U PPER BOUNDS ON THE RATE OF (r, t, x)-LRC CODES
(r, t)
(4, 2)
(5, 2)
(6, 2)
(7, 2)
(4, 3)
(5, 3)
(6, 3)
(7, 3)
x=0
0.7111
0.7576
0.7912
0.8167
0.6564
0.7102
0.7496
0.7795
x=1
0.7250
0.7667
0.7976
0.8214
0.6981
0.7375
0.7688
0.7938
x=2
0.7429
0.7778
0.8052
0.8269
0.7516
0.7708
0.7915
0.8103
x=3
0.7667
0.7917
0.8143
0.8333
0.8231
0.8125
0.8188
0.8295
TABLE II
C OMPARISON OF UPPER AND LOWER BOUNDS ON THE RATE OF
(r, t, x)-LRC CODES
(r, t)
(3, 2)
(5, 2)
(7, 2)
(3, 3)
(5, 3)
(7, 3)
x = 0 (WZL)
0.6000
0.7143
0.7778
0.5000
0.6250
0.7000
R∗ (r, t, x = 0)
0.6429
0.7576
0.8167
0.5786
0.7102
0.7795
Example 1: Let us start from
0 0
0 1
HWZL =
1 0
1 1
and construct a
The matrix has
0
0
H=
1
1
x=1
0.6667
0.7500
0.8000
0.6250
0.7000
0.7500
R∗ (r, t, x = 1)
0.6667
0.7667
0.8214
0.6500
0.7375
0.7938
an (2, 2)-WZL code
0 1 1 1
1 0 0 1
1 0 1 0
0 1 0 0
parity-check matrix of an (5, 2, 1)-LRC code.
a rate R = 0.75 and shown below
0 0 0 0 0 1 1 1 1 1 1
0 1 1 1 1 0 0 0 0 1 1
1 0 0 1 1 0 0 1 1 0 0
1 1 1 0 0 1 1 0 0 0 0
V. N UMERICAL
RESULTS
In Table I we present the comparison of upper bounds on the
rate of (r, t, x)-LRC codes for different values of parameter
x. We see that the value of the upper bound increases with
the parameter x. In the Table II we present the comparison of
the code rate obtained by proposed code construction (x = 1)
and the code rate of WZL codes with the same locality and
availability. In addition, we include the values of the upper
bounds for the code rate from [10] and the upper bounds for
the code rate proposed in this paper. We see, that e.g. for
r = 3, t = 2 and x = 1 the lower bound is tight and it is
better, then the upper bound for the case of r = 3, t = 2 and
x = 0.
VI. C ONCLUSION
We investigated one possible generalization of locally recoverable codes (LRC) with all-symbol locality and availability
when recovering sets can intersect in a small number of
coordinates. This feature allows us to increase the achievable
code rate and still meet load balancing requirements. In this
paper we derived an upper bound for the rate of such codes
and gave explicit constructions of codes with such a property.
ACKNOWLEDGMENT
A. Frolov thanks A. Barg for introducing this problem to
him and for numerous fruitful discussions during his stay in
University of Maryland.
R EFERENCES
[1] P. Goplan, C. Huang, H. Simitci, and S. Yekhanin, “On the locality
of codeword symbols,” IEEE Trans. Inf. Theory, vol. 58, no. 11, pp.
6925–6934, Nov. 2011.
[2] P. Goplan, C. Huang, B. Jenkins, and S. Yekhanin, “Explicit maximally
recoverable codes with locality,” IEEE Trans. Inf. Theory, vol. 60, no. 9,
pp. 5245 –5256, Sep. 2014.
[3] D. S. Papailiopoulos and A. G. Dimakis, “Locally repairable codes,”
IEEE Trans. Inf. Theory, vol. 60, no. 10, pp. 5843–5855, Oct 2014.
[4] A. S. Rawat, O. O. Koyluoglu, N. Silberstein, and S. Vishwanath,
“Optimal locally repairable and secure codes for distributed storage
systems,” IEEE Trans. Inf. Theory, vol. 60, no. 1, pp. 212–236, Jan
2014.
[5] S. Yekhanin, “Locally decodable codes,” Found. Trends Theoretical
Comput. Sci., vol. 6, no. 3, pp. 139 –255, 2012.
[6] Y. Ben-Haim and S. Litsyn, “Upper bounds on the rate of ldpc codes
as a function of minimum distance,” IEEE Trans. Inf. Theory, vol. 52,
no. 5, pp. 2092 –2100, May 2006.
[7] V. R. Cadambe and A. Mazumdar, “Bounds on the size of locally
recoverable codes,” IEEE Trans. Inf. Theory, vol. 61, no. 11, pp. 5787
–5794, Nov. 2015.
[8] N. Silberstein, A. S. Rawat, O. Koyluogly, and S. Vishwanath, “Optimal
locally repairable codes via rank metric codes,” in Proceedings IEEE
International Symposium on Information Theory (ISIT), Jul. 2013, pp.
1819–1823.
[9] I. Tamo and A. Barg, “A family of optimal locally recoverable codes,”
IEEE Trans. Inf. Theory, vol. 60, no. 8, pp. 4661 –4676, Aug. 2014.
[10] I. Tamo, A. Barg, and A. Frolov, “Bounds on the parameters of locally
recoverable codes,” IEEE Trans. Inf. Theory, vol. 62, no. 6, pp. 3070
–3083, Jun. 2016.
[11] N. Prakash, V. Lalitha, and P. V. Kumar, “Codes with locality for two
erasures,” in Proceedings IEEE International Symposium on Information
Theory (ISIT), Jun. 2014, pp. 1962–1966.
[12] P. Huang, E. Yaakobi, H. Uchikawa, and P. H. Siegel, “Linear locally
repairable codes with availability,” in Proceedings IEEE International
Symposium on Information Theory (ISIT), Jun. 2015, pp. 1871–1875.
[13] A. Wang, Z. Zhang, and M. Liu, “Achieving arbitrary locality and availability in binary codes,” in Proceedings IEEE International Symposium
on Information Theory (ISIT), Jun. 2015, pp. 1866 – 1870.
[14] A. Wang and Z. Zhang, “Repair locality with multiple erasure tolerance,”
IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 6979–6987, Nov 2014.
[15] A. S. Rawat, D. S. Papailiopoulos, A. G. Dimakis, and S. Vishwanath,
“Locality and availability in distributed storage,” in Proceedings IEEE
International Symposium on Information Theory (ISIT), June 2014, pp.
681–685.
| 7cs.IT
|
Mask R-CNN
Kaiming He
Georgia Gkioxari
Piotr Dollár
Ross Girshick
arXiv:1703.06870v3 [cs.CV] 24 Jan 2018
Facebook AI Research (FAIR)
Abstract
class
box
We present a conceptually simple, flexible, and general
framework for object instance segmentation. Our approach
efficiently detects objects in an image while simultaneously
generating a high-quality segmentation mask for each instance. The method, called Mask R-CNN, extends Faster
R-CNN by adding a branch for predicting an object mask in
parallel with the existing branch for bounding box recognition. Mask R-CNN is simple to train and adds only a small
overhead to Faster R-CNN, running at 5 fps. Moreover,
Mask R-CNN is easy to generalize to other tasks, e.g., allowing us to estimate human poses in the same framework.
We show top results in all three tracks of the COCO suite
of challenges, including instance segmentation, boundingbox object detection, and person keypoint detection. Without bells and whistles, Mask R-CNN outperforms all existing, single-model entries on every task, including the
COCO 2016 challenge winners. We hope our simple and
effective approach will serve as a solid baseline and help
ease future research in instance-level recognition. Code
has been made available at: https://github.com/
facebookresearch/Detectron.
RoIAlign
conv
conv
Figure 1. The Mask R-CNN framework for instance segmentation.
segmentation, where the goal is to classify each pixel into
a fixed set of categories without differentiating object instances.1 Given this, one might expect a complex method
is required to achieve good results. However, we show that
a surprisingly simple, flexible, and fast system can surpass
prior state-of-the-art instance segmentation results.
Our method, called Mask R-CNN, extends Faster R-CNN
[36] by adding a branch for predicting segmentation masks
on each Region of Interest (RoI), in parallel with the existing branch for classification and bounding box regression (Figure 1). The mask branch is a small FCN applied
to each RoI, predicting a segmentation mask in a pixel-topixel manner. Mask R-CNN is simple to implement and
train given the Faster R-CNN framework, which facilitates
a wide range of flexible architecture designs. Additionally,
the mask branch only adds a small computational overhead,
enabling a fast system and rapid experimentation.
In principle Mask R-CNN is an intuitive extension of
Faster R-CNN, yet constructing the mask branch properly
is critical for good results. Most importantly, Faster RCNN was not designed for pixel-to-pixel alignment between network inputs and outputs. This is most evident in
how RoIPool [18, 12], the de facto core operation for attending to instances, performs coarse spatial quantization
for feature extraction. To fix the misalignment, we propose a simple, quantization-free layer, called RoIAlign, that
faithfully preserves exact spatial locations. Despite being
1. Introduction
The vision community has rapidly improved object detection and semantic segmentation results over a short period of time. In large part, these advances have been driven
by powerful baseline systems, such as the Fast/Faster RCNN [12, 36] and Fully Convolutional Network (FCN) [30]
frameworks for object detection and semantic segmentation, respectively. These methods are conceptually intuitive
and offer flexibility and robustness, together with fast training and inference time. Our goal in this work is to develop a
comparably enabling framework for instance segmentation.
Instance segmentation is challenging because it requires
the correct detection of all objects in an image while also
precisely segmenting each instance. It therefore combines
elements from the classical computer vision tasks of object detection, where the goal is to classify individual objects and localize each using a bounding box, and semantic
1 Following
common terminology, we use object detection to denote
detection via bounding boxes, not masks, and semantic segmentation to
denote per-pixel classification without differentiating instances. Yet we
note that instance segmentation is both semantic and a form of detection.
1
bus.99
umbrella.98
umbrella.98
person1.00
person1.00
person1.00
backpack1.00
person1.00
person1.00
person1.00
person1.00 person.94
person.99
person1.00
handbag.96
person1.00
person1.00
person.95
person1.00 person.89
person1.00
person1.00
person.98
person1.00
person.99
person1.00
sheep.99
backpack.99
backpack.93
sheep1.00
sheep.96
sheep.99
sheep.99
sheep.96
sheep.93
sheep.99
sheep1.00
sheep.91
sheep.86
sheep.96
sheep.82
sheep1.00
sheep.95 sheep.96
sheep.99
dining table.96
person.99
bottle.99
bottle.99
bottle.99
person1.00
person.99person1.00
tv.99
traffic light.96
chair.90
dining table.99
chair.98
chair.99
chair.86
elephant1.00
chair.96
person1.00
person1.00
traffic light.95
person1.00
motorcycle1.00
person.96
person.96person1.00
person.83
person.98
person.99person.91
person.90
person.92
truck1.00
person1.00
chair.96
wine glass.97
bottle.99wine glass.93
bowl.85
wine glass1.00
chair.99
wine glass.99
chair.99
fork.95
bowl.81
traffic light.92
motorcycle1.00
wine glass1.00
traffic light.84
person.87
car1.00
car.99
car.92
person.99
car.99
car.93
motorcycle.95
person1.00
person.85
person.99
knife.83
person.96
Figure 2. Mask R-CNN results on the COCO test set. These results are based on ResNet-101 [19], achieving a mask AP of 35.7 and
running at 5 fps. Masks are shown in color, and bounding box, category, and confidences are also shown.
2. Related Work
a seemingly minor change, RoIAlign has a large impact: it
improves mask accuracy by relative 10% to 50%, showing
bigger gains under stricter localization metrics. Second, we
found it essential to decouple mask and class prediction: we
predict a binary mask for each class independently, without
competition among classes, and rely on the network’s RoI
classification branch to predict the category. In contrast,
FCNs usually perform per-pixel multi-class categorization,
which couples segmentation and classification, and based
on our experiments works poorly for instance segmentation.
Without bells and whistles, Mask R-CNN surpasses all
previous state-of-the-art single-model results on the COCO
instance segmentation task [28], including the heavilyengineered entries from the 2016 competition winner. As
a by-product, our method also excels on the COCO object
detection task. In ablation experiments, we evaluate multiple basic instantiations, which allows us to demonstrate its
robustness and analyze the effects of core factors.
Our models can run at about 200ms per frame on a GPU,
and training on COCO takes one to two days on a single
8-GPU machine. We believe the fast train and test speeds,
together with the framework’s flexibility and accuracy, will
benefit and ease future research on instance segmentation.
Finally, we showcase the generality of our framework
via the task of human pose estimation on the COCO keypoint dataset [28]. By viewing each keypoint as a one-hot
binary mask, with minimal modification Mask R-CNN can
be applied to detect instance-specific poses. Mask R-CNN
surpasses the winner of the 2016 COCO keypoint competition, and at the same time runs at 5 fps. Mask R-CNN,
therefore, can be seen more broadly as a flexible framework
for instance-level recognition and can be readily extended
to more complex tasks.
We have released code to facilitate future research.
R-CNN: The Region-based CNN (R-CNN) approach [13]
to bounding-box object detection is to attend to a manageable number of candidate object regions [42, 20] and evaluate convolutional networks [25, 24] independently on each
RoI. R-CNN was extended [18, 12] to allow attending to
RoIs on feature maps using RoIPool, leading to fast speed
and better accuracy. Faster R-CNN [36] advanced this
stream by learning the attention mechanism with a Region
Proposal Network (RPN). Faster R-CNN is flexible and robust to many follow-up improvements (e.g., [38, 27, 21]),
and is the current leading framework in several benchmarks.
Instance Segmentation: Driven by the effectiveness of RCNN, many approaches to instance segmentation are based
on segment proposals. Earlier methods [13, 15, 16, 9] resorted to bottom-up segments [42, 2]. DeepMask [33] and
following works [34, 8] learn to propose segment candidates, which are then classified by Fast R-CNN. In these
methods, segmentation precedes recognition, which is slow
and less accurate. Likewise, Dai et al. [10] proposed a complex multiple-stage cascade that predicts segment proposals
from bounding-box proposals, followed by classification.
Instead, our method is based on parallel prediction of masks
and class labels, which is simpler and more flexible.
Most recently, Li et al. [26] combined the segment proposal system in [8] and object detection system in [11] for
“fully convolutional instance segmentation” (FCIS). The
common idea in [8, 11, 26] is to predict a set of positionsensitive output channels fully convolutionally. These
channels simultaneously address object classes, boxes, and
masks, making the system fast. But FCIS exhibits systematic errors on overlapping instances and creates spurious
edges (Figure 6), showing that it is challenged by the fundamental difficulties of segmenting instances.
2
Figure 3. RoIAlign: The dashed grid rep-
Another family of solutions [23, 4, 3, 29] to instance segmentation are driven by the success of semantic segmentation. Starting from per-pixel classification results (e.g.,
FCN outputs), these methods attempt to cut the pixels of
the same category into different instances. In contrast to the
segmentation-first strategy of these methods, Mask R-CNN
is based on an instance-first strategy. We expect a deeper incorporation of both strategies will be studied in the future.
resents a feature map, the solid lines an RoI
(with 2×2 bins in this example), and the dots
the 4 sampling points in each bin. RoIAlign
computes the value of each sampling point
by bilinear interpolation from the nearby grid
points on the feature map. No quantization is
performed on any coordinates involved in the
RoI, its bins, or the sampling points.
class label used to select the output mask. This decouples
mask and class prediction. This is different from common
practice when applying FCNs [30] to semantic segmentation, which typically uses a per-pixel softmax and a multinomial cross-entropy loss. In that case, masks across classes
compete; in our case, with a per-pixel sigmoid and a binary
loss, they do not. We show by experiments that this formulation is key for good instance segmentation results.
3. Mask R-CNN
Mask R-CNN is conceptually simple: Faster R-CNN has
two outputs for each candidate object, a class label and a
bounding-box offset; to this we add a third branch that outputs the object mask. Mask R-CNN is thus a natural and intuitive idea. But the additional mask output is distinct from
the class and box outputs, requiring extraction of much finer
spatial layout of an object. Next, we introduce the key elements of Mask R-CNN, including pixel-to-pixel alignment,
which is the main missing piece of Fast/Faster R-CNN.
Mask Representation: A mask encodes an input object’s
spatial layout. Thus, unlike class labels or box offsets
that are inevitably collapsed into short output vectors by
fully-connected (fc) layers, extracting the spatial structure
of masks can be addressed naturally by the pixel-to-pixel
correspondence provided by convolutions.
Specifically, we predict an m × m mask from each RoI
using an FCN [30]. This allows each layer in the mask
branch to maintain the explicit m × m object spatial layout without collapsing it into a vector representation that
lacks spatial dimensions. Unlike previous methods that resort to fc layers for mask prediction [33, 34, 10], our fully
convolutional representation requires fewer parameters, and
is more accurate as demonstrated by experiments.
This pixel-to-pixel behavior requires our RoI features,
which themselves are small feature maps, to be well aligned
to faithfully preserve the explicit per-pixel spatial correspondence. This motivated us to develop the following
RoIAlign layer that plays a key role in mask prediction.
Faster R-CNN: We begin by briefly reviewing the Faster
R-CNN detector [36]. Faster R-CNN consists of two stages.
The first stage, called a Region Proposal Network (RPN),
proposes candidate object bounding boxes. The second
stage, which is in essence Fast R-CNN [12], extracts features using RoIPool from each candidate box and performs
classification and bounding-box regression. The features
used by both stages can be shared for faster inference. We
refer readers to [21] for latest, comprehensive comparisons
between Faster R-CNN and other frameworks.
Mask R-CNN: Mask R-CNN adopts the same two-stage
procedure, with an identical first stage (which is RPN). In
the second stage, in parallel to predicting the class and box
offset, Mask R-CNN also outputs a binary mask for each
RoI. This is in contrast to most recent systems, where classification depends on mask predictions (e.g. [33, 10, 26]).
Our approach follows the spirit of Fast R-CNN [12] that
applies bounding-box classification and regression in parallel (which turned out to largely simplify the multi-stage
pipeline of original R-CNN [13]).
Formally, during training, we define a multi-task loss on
each sampled RoI as L = Lcls + Lbox + Lmask . The classification loss Lcls and bounding-box loss Lbox are identical as those defined in [12]. The mask branch has a Km2 dimensional output for each RoI, which encodes K binary
masks of resolution m × m, one for each of the K classes.
To this we apply a per-pixel sigmoid, and define Lmask as
the average binary cross-entropy loss. For an RoI associated
with ground-truth class k, Lmask is only defined on the k-th
mask (other mask outputs do not contribute to the loss).
Our definition of Lmask allows the network to generate
masks for every class without competition among classes;
we rely on the dedicated classification branch to predict the
RoIAlign: RoIPool [12] is a standard operation for extracting a small feature map (e.g., 7×7) from each RoI. RoIPool
first quantizes a floating-number RoI to the discrete granularity of the feature map, this quantized RoI is then subdivided into spatial bins which are themselves quantized, and
finally feature values covered by each bin are aggregated
(usually by max pooling). Quantization is performed, e.g.,
on a continuous coordinate x by computing [x/16], where
16 is a feature map stride and [·] is rounding; likewise, quantization is performed when dividing into bins (e.g., 7×7).
These quantizations introduce misalignments between the
RoI and the extracted features. While this may not impact
classification, which is robust to small translations, it has a
large negative effect on predicting pixel-accurate masks.
To address this, we propose an RoIAlign layer that removes the harsh quantization of RoIPool, properly aligning
the extracted features with the input. Our proposed change
is simple: we avoid any quantization of the RoI boundaries
3
Faster R-CNN
w/ ResNet [19]
class
or bins (i.e., we use x/16 instead of [x/16]). We use bilinear interpolation [22] to compute the exact values of the
input features at four regularly sampled locations in each
RoI bin, and aggregate the result (using max or average),
see Figure 3 for details. We note that the results are not sensitive to the exact sampling locations, or how many points
are sampled, as long as no quantization is performed.
RoIAlign leads to large improvements as we show in
§4.2. We also compare to the RoIWarp operation proposed
in [10]. Unlike RoIAlign, RoIWarp overlooked the alignment issue and was implemented in [10] as quantizing RoI
just like RoIPool. So even though RoIWarp also adopts
bilinear resampling motivated by [22], it performs on par
with RoIPool as shown by experiments (more details in Table 2c), demonstrating the crucial role of alignment.
7×7
7×7 ave
2048
RoI ×1024 res5 ×2048
14×14
×256
Faster R-CNN
w/ FPN [27]
class
box
RoI
7×7
×256
14×14
×80
RoI
14×14
14×14
×256 ×4 ×256
mask
1024
1024
28×28
×256
box
28×28
×80
mask
Figure 4. Head Architecture: We extend two existing Faster RCNN heads [19, 27]. Left/Right panels show the heads for the
ResNet C4 and FPN backbones, from [19] and [27], respectively,
to which a mask branch is added. Numbers denote spatial resolution and channels. Arrows denote either conv, deconv, or fc layers
as can be inferred from context (conv preserves spatial dimension
while deconv increases it). All convs are 3×3, except the output
conv which is 1×1, deconvs are 2×2 with stride 2, and we use
ReLU [31] in hidden layers. Left: ‘res5’ denotes ResNet’s fifth
stage, which for simplicity we altered so that the first conv operates on a 7×7 RoI with stride 1 (instead of 14×14 / stride 2 as in
[19]). Right: ‘×4’ denotes a stack of four consecutive convs.
Network Architecture: To demonstrate the generality of
our approach, we instantiate Mask R-CNN with multiple
architectures. For clarity, we differentiate between: (i) the
convolutional backbone architecture used for feature extraction over an entire image, and (ii) the network head
for bounding-box recognition (classification and regression)
and mask prediction that is applied separately to each RoI.
We denote the backbone architecture using the nomenclature network-depth-features. We evaluate ResNet [19]
and ResNeXt [45] networks of depth 50 or 101 layers. The
original implementation of Faster R-CNN with ResNets
[19] extracted features from the final convolutional layer
of the 4-th stage, which we call C4. This backbone with
ResNet-50, for example, is denoted by ResNet-50-C4. This
is a common choice used in [19, 10, 21, 39].
We also explore another more effective backbone recently proposed by Lin et al. [27], called a Feature Pyramid Network (FPN). FPN uses a top-down architecture with
lateral connections to build an in-network feature pyramid
from a single-scale input. Faster R-CNN with an FPN backbone extracts RoI features from different levels of the feature pyramid according to their scale, but otherwise the
rest of the approach is similar to vanilla ResNet. Using a
ResNet-FPN backbone for feature extraction with Mask RCNN gives excellent gains in both accuracy and speed. For
further details on FPN, we refer readers to [27].
For the network head we closely follow architectures
presented in previous work to which we add a fully convolutional mask prediction branch. Specifically, we extend the Faster R-CNN box heads from the ResNet [19]
and FPN [27] papers. Details are shown in Figure 4. The
head on the ResNet-C4 backbone includes the 5-th stage of
ResNet (namely, the 9-layer ‘res5’ [19]), which is computeintensive. For FPN, the backbone already includes res5 and
thus allows for a more efficient head that uses fewer filters.
We note that our mask branches have a straightforward
structure. More complex designs have the potential to improve performance but are not the focus of this work.
3.1. Implementation Details
We set hyper-parameters following existing Fast/Faster
R-CNN work [12, 36, 27]. Although these decisions were
made for object detection in original papers [12, 36, 27], we
found our instance segmentation system is robust to them.
Training: As in Fast R-CNN, an RoI is considered positive
if it has IoU with a ground-truth box of at least 0.5 and
negative otherwise. The mask loss Lmask is defined only on
positive RoIs. The mask target is the intersection between
an RoI and its associated ground-truth mask.
We adopt image-centric training [12]. Images are resized
such that their scale (shorter edge) is 800 pixels [27]. Each
mini-batch has 2 images per GPU and each image has N
sampled RoIs, with a ratio of 1:3 of positive to negatives
[12]. N is 64 for the C4 backbone (as in [12, 36]) and 512
for FPN (as in [27]). We train on 8 GPUs (so effective minibatch size is 16) for 160k iterations, with a learning rate of
0.02 which is decreased by 10 at the 120k iteration. We
use a weight decay of 0.0001 and momentum of 0.9. With
ResNeXt [45], we train with 1 image per GPU and the same
number of iterations, with a starting learning rate of 0.01.
The RPN anchors span 5 scales and 3 aspect ratios, following [27]. For convenient ablation, RPN is trained separately and does not share features with Mask R-CNN, unless specified. For every entry in this paper, RPN and Mask
R-CNN have the same backbones and so they are shareable.
Inference: At test time, the proposal number is 300 for the
C4 backbone (as in [36]) and 1000 for FPN (as in [27]). We
run the box prediction branch on these proposals, followed
by non-maximum suppression [14]. The mask branch is
then applied to the highest scoring 100 detection boxes. Although this differs from the parallel computation used in
training, it speeds up inference and improves accuracy (due
to the use of fewer, more accurate RoIs). The mask branch
4
person1.00
person1.00
person1.00
person1.00
person1.00
person.99
person1.00
person1.00
bench.76
person.93
person.99
person1.00
person.95
umbrella.97
umbrella.97
umbrella.96 umbrella.99
person.99
umbrella1.00
person1.00
person.99 person.98
umbrella.89
umbrella1.00
person1.00
person1.00person1.00
person.89
person1.00umbrella.98
person1.00 handbag.97
person.95
person.80
person1.00
backpack.98
backpack.95
backpack.96
skateboard.91
handbag.81
person1.00
person1.00 person.98
person1.00
surfboard1.00
person1.00
person.91
surfboard1.00 surfboard.98 surfboard1.00
person.74
surfboard1.00
person1.00
person.98
person1.00
person1.00
baseball bat.99
handbag.85
person.93
person1.00
person1.00
person.99
horse1.00
horse1.00 horse1.00
skateboard.83
bicycle.93
skateboard.82
baseball bat.85
baseball bat.98dog1.00
person.99
kite.72
kite.89
kite.81
kite1.00
kite.98
person1.00
person.99
kite.73
kite.99
kite.89
kite.88
zebra1.00
zebra.90
zebra.99
zebra.96
zebra.99
zebra.74
zebra.99
zebra.96
zebra.88
zebra1.00
person.82
zebra.76
zebra1.00
kite.82
kite.97
person1.00
person.87
person.95person.72
person.94
person.99
person.92
person.95
person.97 person.99
person.88
person.97
person.97 person.77
person.82
person.83
person.89
person.97
person.98
person.99 person.86
person.81
person.77
person.88
person.98
person.96
person.94 person.88
person.96person.99
person.86
person.99
frisbee1.00
person.80
person.91
chair.96
chair.78
cup.93
cup.79
dining
table.81
dining table.75
dining table.96
chair.85
chair.89
cup.75
cup.71
chair.99
chair.95
chair.99
chair.98
chair.92
wine glass.80
chair.85
chair.98
dining table.78
wine glass.80
cup.83
chair.95
cup.71
cup.98
chair.94
chair.87 chair.97
kite.99
kite.95
kite.86 kite.88
kite.84
wine glass.91
wine glass.93
cup.91
person.80
person.87person.71
person.98
person.78
person.89
person.77
person.98
person.98
person.99
person.94
person.81
person.72
person1.00
person.84
person.99
person.95
person.82
person.72
person1.00
person.94
person1.00
person.99
person.99
person.96
person.98 motorcycle.72
person1.00
person1.00
person1.00
person1.00
person1.00
tv.98
person.91
person1.00
person.99
person.99
person1.00person1.00
person.99
person.98
person.80
person1.00
handbag.80
bus1.00
car1.00
car.95truck.86
car.98
bus1.00
car.97
skateboard.98
person1.00
couch.82
person.99
person.90
car.93
car.82
car.99
car.99
car.99
person.99
car.98
car.96
car.91car.94
backpack.88
handbag.91
person.76
person1.00
person.78
person.98
person.78
person.86person1.00
potted plant.92
car.93
car.78
traffic light.73
car.93
car.98 truck.88
skateboard.99
wine
wineglass.94
glass.94
wine glass.83
person.88
person1.00
traffic light.87
car.87
car.95
car.95
car.97
car.99
kite.93
diningchair.83
table.91
cup.96
person1.00
kite.98
kite.95
kite.84
car.98
person1.00
person.98
tv.84
car.78
person1.00
elephant1.00
elephant1.00
person1.00
traffic light.99
elephant.97
elephant1.00
bottle.97
person1.00
person.92 person.74
person.87
person.98
person1.00
tie.85
elephant.99
bird.93
traffic light.71
person.98
person.97
person.95
person1.00
person.99
person1.00
person.99
person.95
person.95
person.73
person.98
person.99
person.95
person1.00
person1.00
person.80
person.99
person.95
person1.00
dining table.95
handbag.88
handbag.88
handbag.73
bench.97
wine glass.99
person.99
stop sign.88
person.77
wine glass1.00
person.87
wine glass1.00
person.97
person.81
person.90
cell clock.73
phone.77
handbag.99
person.96
chair.93
chair.81
chair.97
chair.99
chair.99
chair.81
chair.93
chair.94
chair.92
chair.81
chair.98
chair.83
chair.91
chair.80
person.71
person.99
person.94
suitcase.98
suitcase1.00
chair.71
person.98
chair.73
suitcase.93
suitcase.72
suitcase.96
person1.00
suitcase1.00
suitcase.88
person.98
traffic light.99
traffic light1.00
car.95
car.81
car.89
car.98
car.91
car.96
car.97
car.96
person.99
car.95
car.97
car.97
car.99
car.94
car.94
person.87
car.95
car.97
bicycle.86
car1.00
car.98
car.97
car.99
car.97
donut.95
donut.89
donut.90
donut1.00
donut.99
donut.98
donut.89 donut.97
donut.94
parking meter.98
donut.96
suitcase.99
person1.00
horse.97 person.96
person.96
person.97 person.98 horse.99
donut.89donut.89
donut.93donut.99
donut.96
donut.86
donut.95
donut.98
donut.81
donut.89
donut1.00
donut.96donut.98
donut.98
horse.77
person1.00
sports ball.99
person1.00
tennis racket1.00
cow.93
person.99
donut.95
donut.98
donut.95
car1.00
donut.86
truck.92
donut1.00
donut.99
truck.93
bus.90
bus.99
donut.90
truck.97
truck.99
truck.96
truck.99
donut1.00
donut.88
car.86
Figure 5. More results of Mask R-CNN on COCO test images, using ResNet-101-FPN and running at 5 fps, with 35.7 mask AP (Table 1).
MNC [10]
FCIS [26] +OHEM
FCIS+++ [26] +OHEM
Mask R-CNN
Mask R-CNN
Mask R-CNN
backbone
ResNet-101-C4
ResNet-101-C5-dilated
ResNet-101-C5-dilated
ResNet-101-C4
ResNet-101-FPN
ResNeXt-101-FPN
AP
24.6
29.2
33.6
33.1
35.7
37.1
AP50
44.3
49.5
54.5
54.9
58.0
60.0
AP75
24.8
34.8
37.8
39.4
APS
4.7
7.1
12.1
15.5
16.9
APM
25.9
31.3
35.6
38.1
39.9
APL
43.6
50.0
51.1
52.4
53.5
Table 1. Instance segmentation mask AP on COCO test-dev. MNC [10] and FCIS [26] are the winners of the COCO 2015 and 2016
segmentation challenges, respectively. Without bells and whistles, Mask R-CNN outperforms the more complex FCIS+++, which includes
multi-scale train/test, horizontal flip test, and OHEM [38]. All entries are single-model results.
4.1. Main Results
can predict K masks per RoI, but we only use the k-th mask,
where k is the predicted class by the classification branch.
The m×m floating-number mask output is then resized to
the RoI size, and binarized at a threshold of 0.5.
Note that since we only compute masks on the top 100
detection boxes, Mask R-CNN adds a small overhead to its
Faster R-CNN counterpart (e.g., ∼20% on typical models).
We compare Mask R-CNN to the state-of-the-art methods in instance segmentation in Table 1. All instantiations of our model outperform baseline variants of previous state-of-the-art models. This includes MNC [10]
and FCIS [26], the winners of the COCO 2015 and 2016
segmentation challenges, respectively. Without bells and
whistles, Mask R-CNN with ResNet-101-FPN backbone
outperforms FCIS+++ [26], which includes multi-scale
train/test, horizontal flip test, and online hard example mining (OHEM) [38]. While outside the scope of this work, we
expect many such improvements to be applicable to ours.
Mask R-CNN outputs are visualized in Figures 2 and 5.
Mask R-CNN achieves good results even under challenging conditions. In Figure 6 we compare our Mask R-CNN
baseline and FCIS+++ [26]. FCIS+++ exhibits systematic
artifacts on overlapping instances, suggesting that it is challenged by the fundamental difficulty of instance segmentation. Mask R-CNN shows no such artifacts.
4. Experiments: Instance Segmentation
We perform a thorough comparison of Mask R-CNN to
the state of the art along with comprehensive ablations on
the COCO dataset [28]. We report the standard COCO metrics including AP (averaged over IoU thresholds), AP50 ,
AP75 , and APS , APM , APL (AP at different scales). Unless noted, AP is evaluating using mask IoU. As in previous
work [5, 27], we train using the union of 80k train images
and a 35k subset of val images (trainval35k), and report ablations on the remaining 5k val images (minival).
We also report results on test-dev [28].
5
FCIS
umbrella.99
umbrella1.00
Mask R-CNN
person1.00
person1.00 person1.00
person1.00
person1.00
person1.00
person1.00
person1.00
person1.00
person1.00
person1.00
person.99
car.99 car.93
person1.00
person1.00
person1.00
person1.00
person1.00
person.99
giraffe1.00
person1.00
train1.00
train.99
giraffe1.00
train.80
person.95
handbag.93
skateboard.98
tie.95
tie1.00
sports ball.98
sports ball1.00
skateboard.99
Figure 6. FCIS+++ [26] (top) vs. Mask R-CNN (bottom, ResNet-101-FPN). FCIS exhibits systematic artifacts on overlapping objects.
net-depth-features
ResNet-50-C4
ResNet-101-C4
ResNet-50-FPN
ResNet-101-FPN
ResNeXt-101-FPN
AP
30.3
32.7
33.6
35.4
36.7
AP50
51.2
54.2
55.2
57.3
59.5
AP75
31.5
34.3
35.3
37.5
38.9
(a) Backbone Architecture: Better backbones bring expected gains: deeper networks
do better, FPN outperforms C4 features, and
ResNeXt improves on ResNet.
RoIPool
RoIAlign
AP
23.6
30.9
+7.3
AP50
46.5
51.8
+ 5.3
AP75
21.6
32.1
+10.5
softmax
sigmoid
AP
24.8
30.3
+5.5
AP50
44.1
51.2
+7.1
AP75
25.1
31.5
+6.4
align? bilinear? agg.
RoIPool [12]
RoIWarp [10]
RoIAlign
(b) Multinomial vs. Independent Masks
(ResNet-50-C4): Decoupling via perclass binary masks (sigmoid) gives large
gains over multinomial masks (softmax).
APbb
28.2
34.0
+5.8
APbb
50
52.7
55.3
+2.6
APbb
75
26.9
36.4
+9.5
(d) RoIAlign (ResNet-50-C5, stride 32): Mask-level and box-level
AP using large-stride features. Misalignments are more severe than
with stride-16 features (Table 2c), resulting in big accuracy gaps.
MLP
MLP
FCN
X
X
X
X
X
X
max
max
ave
max
ave
AP
26.9
27.2
27.1
30.2
30.3
AP50
48.8
49.2
48.9
51.0
51.2
AP75
26.4
27.1
27.1
31.8
31.5
(c) RoIAlign (ResNet-50-C4): Mask results with various RoI
layers. Our RoIAlign layer improves AP by ∼3 points and
AP75 by ∼5 points. Using proper alignment is the only factor that contributes to the large gap between RoI layers.
mask branch
fc: 1024→1024→80·282
fc: 1024→1024→1024→80·282
conv: 256→256→256→256→256→80
AP
31.5
31.5
33.6
AP50
53.7
54.0
55.2
AP75
32.8
32.6
35.3
(e) Mask Branch (ResNet-50-FPN): Fully convolutional networks (FCN) vs.
multi-layer perceptrons (MLP, fully-connected) for mask prediction. FCNs improve results as they take advantage of explicitly encoding spatial layout.
Table 2. Ablations. We train on trainval35k, test on minival, and report mask AP unless otherwise noted.
4.2. Ablation Experiments
mask per class. Interestingly, Mask R-CNN with classagnostic masks (i.e., predicting a single m×m output regardless of class) is nearly as effective: it has 29.7 mask AP
vs. 30.3 for the class-specific counterpart on ResNet-50-C4.
This further highlights the division of labor in our approach
which largely decouples classification and segmentation.
We run a number of ablations to analyze Mask R-CNN.
Results are shown in Table 2 and discussed in detail next.
Architecture: Table 2a shows Mask R-CNN with various
backbones. It benefits from deeper networks (50 vs. 101)
and advanced designs including FPN and ResNeXt. We
note that not all frameworks automatically benefit from
deeper or advanced networks (see benchmarking in [21]).
RoIAlign: An evaluation of our proposed RoIAlign layer is
shown in Table 2c. For this experiment we use the ResNet50-C4 backbone, which has stride 16. RoIAlign improves
AP by about 3 points over RoIPool, with much of the gain
coming at high IoU (AP75 ). RoIAlign is insensitive to
max/average pool; we use average in the rest of the paper.
Additionally, we compare with RoIWarp proposed in
MNC [10] that also adopt bilinear sampling. As discussed
in §3, RoIWarp still quantizes the RoI, losing alignment
with the input. As can be seen in Table 2c, RoIWarp performs on par with RoIPool and much worse than RoIAlign.
This highlights that proper alignment is key.
We also evaluate RoIAlign with a ResNet-50-C5 backbone, which has an even larger stride of 32 pixels. We use
the same head as in Figure 4 (right), as the res5 head is not
applicable. Table 2d shows that RoIAlign improves mask
AP by a massive 7.3 points, and mask AP75 by 10.5 points
Multinomial vs. Independent Masks: Mask R-CNN decouples mask and class prediction: as the existing box
branch predicts the class label, we generate a mask for each
class without competition among classes (by a per-pixel sigmoid and a binary loss). In Table 2b, we compare this to
using a per-pixel softmax and a multinomial loss (as commonly used in FCN [30]). This alternative couples the tasks
of mask and class prediction, and results in a severe loss
in mask AP (5.5 points). This suggests that once the instance has been classified as a whole (by the box branch),
it is sufficient to predict a binary mask without concern for
the categories, which makes the model easier to train.
Class-Specific vs. Class-Agnostic Masks: Our default instantiation predicts class-specific masks, i.e., one m×m
6
Faster R-CNN+++ [19]
Faster R-CNN w FPN [27]
Faster R-CNN by G-RMI [21]
Faster R-CNN w TDM [39]
Faster R-CNN, RoIAlign
Mask R-CNN
Mask R-CNN
backbone
APbb
APbb
50
APbb
75
APbb
S
APbb
M
APbb
L
ResNet-101-C4
ResNet-101-FPN
Inception-ResNet-v2 [41]
Inception-ResNet-v2-TDM
ResNet-101-FPN
ResNet-101-FPN
ResNeXt-101-FPN
34.9
36.2
34.7
36.8
37.3
38.2
39.8
55.7
59.1
55.5
57.7
59.6
60.3
62.3
37.4
39.0
36.7
39.2
40.3
41.7
43.4
15.6
18.2
13.5
16.2
19.8
20.1
22.1
38.7
39.0
38.1
39.8
40.2
41.1
43.2
50.9
48.2
52.0
52.1
48.8
50.2
51.2
Table 3. Object detection single-model results (bounding box AP), vs. state-of-the-art on test-dev. Mask R-CNN using ResNet-101FPN outperforms the base variants of all previous state-of-the-art models (the mask output is ignored in these experiments). The gains of
Mask R-CNN over [27] come from using RoIAlign (+1.1 APbb ), multitask training (+0.9 APbb ), and ResNeXt-101 (+1.6 APbb ).
4.4. Timing
(50% relative improvement). Moreover, we note that with
RoIAlign, using stride-32 C5 features (30.9 AP) is more accurate than using stride-16 C4 features (30.3 AP, Table 2c).
RoIAlign largely resolves the long-standing challenge of
using large-stride features for detection and segmentation.
Finally, RoIAlign shows a gain of 1.5 mask AP and 0.5
box AP when used with FPN, which has finer multi-level
strides. For keypoint detection that requires finer alignment,
RoIAlign shows large gains even with FPN (Table 6).
Inference: We train a ResNet-101-FPN model that shares
features between the RPN and Mask R-CNN stages, following the 4-step training of Faster R-CNN [36]. This model
runs at 195ms per image on an Nvidia Tesla M40 GPU (plus
15ms CPU time resizing the outputs to the original resolution), and achieves statistically the same mask AP as the
unshared one. We also report that the ResNet-101-C4 variant takes ∼400ms as it has a heavier box head (Figure 4), so
we do not recommend using the C4 variant in practice.
Although Mask R-CNN is fast, we note that our design
is not optimized for speed, and better speed/accuracy tradeoffs could be achieved [21], e.g., by varying image sizes and
proposal numbers, which is beyond the scope of this paper.
Mask Branch: Segmentation is a pixel-to-pixel task and
we exploit the spatial layout of masks by using an FCN.
In Table 2e, we compare multi-layer perceptrons (MLP)
and FCNs, using a ResNet-50-FPN backbone. Using FCNs
gives a 2.1 mask AP gain over MLPs. We note that we
choose this backbone so that the conv layers of the FCN
head are not pre-trained, for a fair comparison with MLP.
Training: Mask R-CNN is also fast to train. Training with
ResNet-50-FPN on COCO trainval35k takes 32 hours
in our synchronized 8-GPU implementation (0.72s per 16image mini-batch), and 44 hours with ResNet-101-FPN. In
fact, fast prototyping can be completed in less than one day
when training on the train set. We hope such rapid training will remove a major hurdle in this area and encourage
more people to perform research on this challenging topic.
4.3. Bounding Box Detection Results
We compare Mask R-CNN to the state-of-the-art COCO
bounding-box object detection in Table 3. For this result,
even though the full Mask R-CNN model is trained, only
the classification and box outputs are used at inference (the
mask output is ignored). Mask R-CNN using ResNet-101FPN outperforms the base variants of all previous state-ofthe-art models, including the single-model variant of GRMI [21], the winner of the COCO 2016 Detection Challenge. Using ResNeXt-101-FPN, Mask R-CNN further improves results, with a margin of 3.0 points box AP over
the best previous single model entry from [39] (which used
Inception-ResNet-v2-TDM).
As a further comparison, we trained a version of Mask
R-CNN but without the mask branch, denoted by “Faster
R-CNN, RoIAlign” in Table 3. This model performs better
than the model presented in [27] due to RoIAlign. On the
other hand, it is 0.9 points box AP lower than Mask R-CNN.
This gap of Mask R-CNN on box detection is therefore due
solely to the benefits of multi-task training.
Lastly, we note that Mask R-CNN attains a small gap
between its mask and box AP: e.g., 2.7 points between 37.1
(mask, Table 1) and 39.8 (box, Table 3). This indicates that
our approach largely closes the gap between object detection and the more challenging instance segmentation task.
5. Mask R-CNN for Human Pose Estimation
Our framework can easily be extended to human pose
estimation. We model a keypoint’s location as a one-hot
mask, and adopt Mask R-CNN to predict K masks, one for
each of K keypoint types (e.g., left shoulder, right elbow).
This task helps demonstrate the flexibility of Mask R-CNN.
We note that minimal domain knowledge for human pose
is exploited by our system, as the experiments are mainly to
demonstrate the generality of the Mask R-CNN framework.
We expect that domain knowledge (e.g., modeling structures [6]) will be complementary to our simple approach.
Implementation Details: We make minor modifications to
the segmentation system when adapting it for keypoints.
For each of the K keypoints of an instance, the training
target is a one-hot m × m binary mask where only a single
pixel is labeled as foreground. During training, for each visible ground-truth keypoint, we minimize the cross-entropy
loss over an m2 -way softmax output (which encourages a
7
Figure 7. Keypoint detection results on COCO test using Mask R-CNN (ResNet-50-FPN), with person segmentation masks predicted
from the same model. This model has a keypoint AP of 63.1 and runs at 5 fps.
CMU-Pose+++ [6]
G-RMI [32]†
Mask R-CNN, keypoint-only
Mask R-CNN, keypoint & mask
APkp
AP50
kp
AP75
kp
APM
kp
APL
kp
61.8
62.4
62.7
63.1
84.9
84.0
87.0
87.3
67.5
68.5
68.4
68.7
57.1
59.1
57.4
57.8
68.2
68.1
71.1
71.4
Faster R-CNN
Mask R-CNN, mask-only
Mask R-CNN, keypoint-only
Mask R-CNN, keypoint & mask
APbb
person
APmask
person
APkp
52.5
53.6
50.7
52.0
45.8
45.1
64.2
64.7
Table 5. Multi-task learning of box, mask, and keypoint about the
person category, evaluated on minival. All entries are trained
on the same data for fair comparisons. The backbone is ResNet50-FPN. The entries with 64.2 and 64.7 AP on minival have
test-dev AP of 62.7 and 63.1, respectively (see Table 4).
Table 4. Keypoint detection AP on COCO test-dev. Ours is a
single model (ResNet-50-FPN) that runs at 5 fps. CMU-Pose+++
[6] is the 2016 competition winner that uses multi-scale testing,
post-processing with CPM [44], and filtering with an object detector, adding a cumulative ∼5 points (clarified in personal communication). † : G-RMI was trained on COCO plus MPII [1] (25k images), using two models (Inception-ResNet-v2 for bounding box
detection and ResNet-101 for keypoints).
RoIPool
RoIAlign
APkp
AP50
kp
AP75
kp
APM
kp
APL
kp
59.8
64.2
86.2
86.6
66.7
69.7
55.1
58.7
67.4
73.0
Table 6. RoIAlign vs. RoIPool for keypoint detection on
minival. The backbone is ResNet-50-FPN.
single point to be detected). We note that as in instance segmentation, the K keypoints are still treated independently.
We adopt the ResNet-FPN variant, and the keypoint head
architecture is similar to that in Figure 4 (right). The keypoint head consists of a stack of eight 3×3 512-d conv layers, followed by a deconv layer and 2× bilinear upscaling,
producing an output resolution of 56×56. We found that
a relatively high resolution output (compared to masks) is
required for keypoint-level localization accuracy.
Models are trained on all COCO trainval35k images that contain annotated keypoints. To reduce overfitting, as this training set is smaller, we train using image
scales randomly sampled from [640, 800] pixels; inference
is on a single scale of 800 pixels. We train for 90k iterations,
starting from a learning rate of 0.02 and reducing it by 10 at
60k and 80k iterations. We use bounding-box NMS with a
threshold of 0.5. Other details are identical as in §3.1.
multaneously predict boxes, segments, and keypoints while
running at 5 fps. Adding a segment branch (for the person category) improves the APkp to 63.1 (Table 4) on
test-dev. More ablations of multi-task learning on
minival are in Table 5. Adding the mask branch to the
box-only (i.e., Faster R-CNN) or keypoint-only versions
consistently improves these tasks. However, adding the
keypoint branch reduces the box/mask AP slightly, suggesting that while keypoint detection benefits from multitask
training, it does not in turn help the other tasks. Nevertheless, learning all three tasks jointly enables a unified system
to efficiently predict all outputs simultaneously (Figure 7).
We also investigate the effect of RoIAlign on keypoint
detection (Table 6). Though this ResNet-50-FPN backbone
has finer strides (e.g., 4 pixels on the finest level), RoIAlign
still shows significant improvement over RoIPool and increases APkp by 4.4 points. This is because keypoint detections are more sensitive to localization accuracy. This again
indicates that alignment is essential for pixel-level localization, including masks and keypoints.
Given the effectiveness of Mask R-CNN for extracting
object bounding boxes, masks, and keypoints, we expect it
be an effective framework for other instance-level tasks.
Main Results and Ablations: We evaluate the person keypoint AP (APkp ) and experiment with a ResNet-50-FPN
backbone; more backbones will be studied in the appendix.
Table 4 shows that our result (62.7 APkp ) is 0.9 points higher
than the COCO 2016 keypoint detection winner [6] that
uses a multi-stage processing pipeline (see caption of Table 4). Our method is considerably simpler and faster.
More importantly, we have a unified model that can si8
InstanceCut [23]
DWT [4]
SAIS [17]
DIN [3]
SGN [29]
Mask R-CNN
Mask R-CNN
training data
fine + coarse
fine
fine
fine + coarse
fine + coarse
fine
fine + COCO
AP [val]
15.8
19.8
29.2
31.5
36.4
AP
13.0
15.6
17.4
20.0
25.0
26.2
32.0
AP50
27.9
30.0
36.7
38.8
44.9
49.9
58.1
person
10.0
15.1
14.6
16.5
21.8
30.5
34.8
rider
8.0
11.7
12.9
16.7
20.1
23.7
27.0
car
23.7
32.9
35.7
25.7
39.4
46.9
49.1
truck
14.0
17.1
16.0
20.6
24.8
22.8
30.1
bus
19.5
20.4
23.2
30.0
33.2
32.2
40.9
train
15.2
15.0
19.0
23.4
30.8
18.6
30.9
mcycle bicycle
9.3
4.7
7.9
4.9
10.3
7.8
17.1
10.1
17.7
12.4
19.1
16.0
24.1
18.7
Table 7. Results on Cityscapes val (‘AP [val]’ column) and test (remaining columns) sets. Our method uses ResNet-50-FPN.
Appendix A: Experiments on Cityscapes
person:1.00
rider:0.59
person:0.79
person:1.00
person:1.00
person:0.66
person:1.00
person:0.59
person:0.99person:0.67
person:0.82
We further report instance segmentation results on the
Cityscapes [7] dataset. This dataset has fine annotations for 2975 train, 500 val, and 1525 test images. It has
20k coarse training images without instance annotations,
which we do not use. All images are 2048×1024 pixels.
The instance segmentation task involves 8 object categories,
whose numbers of instances on the fine training set are:
person
17.9k
rider
1.8k
car
26.9k
truck
0.5k
bus
0.4k
train
0.2k
car:0.81
person:0.99
person:1.00
bus:1.00
person:0.99
car:0.98
person:1.00
car:0.98
person:1.00
bus:0.95
truck:0.66
person:0.98
person:0.94
person:0.94
person:0.98
car:0.95
car:1.00
person:0.99
person:0.98
person:0.73
person:1.00
person:1.00
person:1.00
person:1.00
person:1.00
person:1.00
person:1.00
person:0.98
person:0.82
car:0.52
car:0.95
car:0.57
person:0.92
car:1.00
car:0.68
car:1.00
car:0.68
car:0.52
bicycle:0.83
car:1.00
car:0.64
person:0.82
person:0.63car:1.00
car:0.99car:1.00
car:0.69
car:1.00
car:0.95
car:1.00
car:0.95
rider:0.68
person:0.72
car:1.00
bicycle:0.56
person:1.00
person:0.99
person:0.98
person:0.73
person:0.98
person:0.99
person:0.99
person:0.93
person:0.97
person:0.86
person:0.84
person:0.99
person:0.98
person:0.72
person:0.72
person:0.91
car:1.00
person:1.00
car:1.00
car:1.00
person:0.73
car:1.00
person:0.85
person:0.93
car:1.00
car:1.00
car:1.00
car:1.00
car:1.00 car:1.00
car:0.98 car:0.88
car:1.00
car:0.97
car:0.72
person:0.98
car:0.72
car:0.76
person:0.78
person:1.00
person:0.58
car:1.00
car:1.00
car:0.65
car:1.00
car:1.00
car:0.50
car:1.00
car:1.00
person:1.00
person:1.00
car:1.00
car:1.00
car:1.00
car:1.00
car:1.00 car:1.00
car:1.00
bus:0.75
car:1.00 car:0.99
car:0.67
person:1.00 person:1.00
person:0.82
car:0.89
mcycle bicycle
0.7k
3.7k
Instance segmentation performance on this task is measured
by the COCO-style mask AP (averaged over IoU thresholds); AP50 (i.e., mask AP at an IoU of 0.5) is also reported.
person:1.00
person:0.75
person:1.00
person:1.00
person:1.00
person:1.00
person:1.00
person:0.99
person:1.00
person:1.00person:1.00
person:1.00person:0.98
person:1.00
person:1.00
person:1.00
person:0.99
person:1.00
person:1.00
person:0.70
person:0.59
person:1.00
person:1.00person:1.00
person:0.92
person:1.00person:0.97
person:1.00
person:1.00
person:0.96
person:1.00
person:1.00
person:1.00
person:0.93
car:1.00
person:1.00
rider:0.94
car:0.99
person:0.88
person:0.89
car:0.89
bicycle:0.99
person:1.00
person:1.00 person:0.96
car:1.00
bicycle:0.97
person:0.91
Implementation: We apply our Mask R-CNN models with
the ResNet-FPN-50 backbone; we found the 101-layer
counterpart performs similarly due to the small dataset size.
We train with image scale (shorter side) randomly sampled
from [800, 1024], which reduces overfitting; inference is on
a single scale of 1024 pixels. We use a mini-batch size of
1 image per GPU (so 8 on 8 GPUs) and train the model
for 24k iterations, starting from a learning rate of 0.01 and
reducing it to 0.001 at 18k iterations. It takes ∼4 hours of
training on a single 8-GPU machine under this setting.
Figure 8. Mask R-CNN results on Cityscapes test (32.0 AP).
The bottom-right image shows a failure prediction.
ing samples each. To partially remedy this issue, we further
report a result using COCO pre-training. To do this, we initialize the corresponding 7 categories in Cityscapes from a
pre-trained COCO Mask R-CNN model (rider being randomly initialized). We fine-tune this model for 4k iterations
in which the learning rate is reduced at 3k iterations, which
takes ∼1 hour for training given the COCO model.
The COCO pre-trained Mask R-CNN model achieves
32.0 AP on test, almost a 6 point improvement over the
fine-only counterpart. This indicates the important role
the amount of training data plays. It also suggests that
methods on Cityscapes might be influenced by their lowshot learning performance. We show that using COCO pretraining is an effective strategy on this dataset.
Finally, we observed a bias between the val and test
AP, as is also observed from the results of [23, 4, 29]. We
found that this bias is mainly caused by the truck, bus,
and train categories, with the fine-only model having
val/test AP of 28.8/22.8, 53.5/32.2, and 33.0/18.6, respectively. This suggests that there is a domain shift on
these categories, which also have little training data. COCO
pre-training helps to improve results the most on these categories; however, the domain shift persists with 38.0/30.1,
57.5/40.9, and 41.2/30.9 val/test AP, respectively. Note
that for the person and car categories we do not see any
such bias (val/test AP are within ±1 point).
Example results on Cityscapes are shown in Figure 8.
Results: Table 7 compares our results to the state of the
art on the val and test sets. Without using the coarse
training set, our method achieves 26.2 AP on test, which
is over 30% relative improvement over the previous best entry (DIN [3]), and is also better than the concurrent work of
SGN’s 25.0 [29]. Both DIN and SGN use fine + coarse
data. Compared to the best entry using fine data only
(17.4 AP), we achieve a ∼50% improvement.
For the person and car categories, the Cityscapes dataset
exhibits a large number of within-category overlapping instances (on average 6 people and 9 cars per image). We
argue that within-category overlap is a core difficulty of instance segmentation. Our method shows massive improvement on these two categories over the other best entries (relative ∼40% improvement on person from 21.8 to 30.5 and
∼20% improvement on car from 39.4 to 46.9), even though
our method does not exploit the coarse data.
A main challenge of the Cityscapes dataset is training
models in a low-data regime, particularly for the categories
of truck, bus, and train, which have about 200-500 train9
description
original baseline
+ updated baseline
+ e2e training
+ ImageNet-5k
+ train-time augm.
+ deeper
+ Non-local [43]
+ test-time augm.
backbone
X-101-FPN
X-101-FPN
X-101-FPN
X-101-FPN
X-101-FPN
X-152-FPN
X-152-FPN-NL
X-152-FPN-NL
AP
36.7
37.0
37.6
38.6
39.2
39.7
40.3
41.8
AP50
59.5
59.7
60.4
61.7
62.5
63.2
64.4
66.0
AP75
38.9
39.0
39.9
40.9
41.6
42.2
42.8
44.8
APbb
39.6
40.5
41.7
42.7
43.5
44.1
45.0
47.3
APbb
50
61.5
63.0
64.1
65.1
65.9
66.4
67.8
69.3
APbb
75
43.2
43.7
45.2
46.6
47.2
48.4
48.9
51.5
description
original baseline
+ updated baseline
+ deeper
+ ResNeXt
+ data distillation [35]
+ test-time augm.
backbone
R-50-FPN
R-50-FPN
R-101-FPN
X-101-FPN
X-101-FPN
X-101-FPN
APkp
64.2
65.1
66.1
67.3
69.1
70.4
kp
AP50
86.6
86.6
87.7
88.0
88.9
89.3
kp
AP75
69.7
70.9
71.7
73.3
75.3
76.8
kp
APM
58.7
59.9
60.5
62.2
64.1
65.8
kp
APL
73.0
73.6
75.0
75.6
77.1
78.1
Table 9. Enhanced keypoint results of Mask R-CNN on COCO
minival. Each row adds an extra component to the above row.
Here we use only keypoint annotations but no mask annotations.
We denote ResNet by ‘R’ and ResNeXt by ‘X’ for brevity.
Table 8. Enhanced detection results of Mask R-CNN on COCO
minival. Each row adds an extra component to the above row.
We denote ResNeXt model by ‘X’ for notational brevity.
Train-time augmentation: Scale augmentation at train
time further improves results. During training, we randomly
sample a scale from [640, 800] pixels and we increase the
number of iterations to 260k (with the learning rate reduced
by 10 at 200k and 240k iterations). Train-time augmentation improves mask AP by 0.6 and box AP by 0.8.
Model architecture: By upgrading the 101-layer
ResNeXt to its 152-layer counterpart [19], we observe an
increase of 0.5 mask AP and 0.6 box AP. This shows a
deeper model can still improve results on COCO.
Using the recently proposed non-local (NL) model [43],
we achieve 40.3 mask AP and 45.0 box AP. This result is
without test-time augmentation, and the method runs at 3fps
on an Nvidia Tesla P100 GPU at test time.
Test-time augmentation: We combine the model results
evaluated using scales of [400, 1200] pixels with a step of
100 and on their horizontal flips. This gives us a singlemodel result of 41.8 mask AP and 47.3 box AP.
The above result is the foundation of our submission to
the COCO 2017 competition (which also used an ensemble,
not discussed here). The first three winning teams for the
instance segmentation task were all reportedly based on an
extension of the Mask R-CNN framework.
Appendix B: Enhanced Results on COCO
As a general framework, Mask R-CNN is compatible with complementary techniques developed for detection/segmentation, including improvements made to
Fast/Faster R-CNN and FCNs. In this appendix we describe some techniques that improve over our original results. Thanks to its generality and flexibility, Mask R-CNN
was used as the framework by the three winning teams in
the COCO 2017 instance segmentation competition, which
all significantly outperformed the previous state of the art.
Instance Segmentation and Object Detection
We report some enhanced results of Mask R-CNN in Table 8. Overall, the improvements increase mask AP 5.1
points (from 36.7 to 41.8) and box AP 7.7 points (from 39.6
to 47.3). Each model improvement increases both mask AP
and box AP consistently, showing good generalization of
the Mask R-CNN framework. We detail the improvements
next. These results, along with future updates, can be reproduced by our released code at https://github.com/
facebookresearch/Detectron, and can serve as
higher baselines for future research.
Updated baseline: We start with an updated baseline
with a different set of hyper-parameters. We lengthen the
training to 180k iterations, in which the learning rate is reduced by 10 at 120k and 160k iterations. We also change
the NMS threshold to 0.5 (from a default value of 0.3). The
updated baseline has 37.0 mask AP and 40.5 box AP.
End-to-end training: All previous results used stagewise training, i.e., training RPN as the first stage and Mask
R-CNN as the second. Following [37], we evaluate endto-end (‘e2e’) training that jointly trains RPN and Mask RCNN. We adopt the ‘approximate’ version in [37] that only
computes partial gradients in the RoIAlign layer by ignoring the gradient w.r.t. RoI coordinates. Table 8 shows that
e2e training improves mask AP by 0.6 and box AP by 1.2.
ImageNet-5k pre-training: Following [45], we experiment with models pre-trained on a 5k-class subset of ImageNet (in contrast to the standard 1k-class subset). This 5×
increase in pre-training data improves both mask and box 1
AP. As a reference, [40] used ∼250× more images (300M)
and reported a 2-3 box AP improvement on their baselines.
Keypoint Detection
We report enhanced results of keypoint detection in Table 9. As an updated baseline, we extend the training schedule to 130k iterations in which the learning rate is reduced
by 10 at 100k and 120k iterations. This improves APkp by
about 1 point. Replacing ResNet-50 with ResNet-101 and
ResNeXt-101 increases APkp to 66.1 and 67.3, respectively.
With a recent method called data distillation [35], we are
able to exploit the additional 120k unlabeled images provided by COCO. In brief, data distillation is a self-training
strategy that uses a model trained on labeled data to predict annotations on unlabeled images, and in turn updates
the model with these new annotations. Mask R-CNN provides an effective framework for such a self-training strategy. With data distillation, Mask R-CNN APkp improve by
1.8 points to 69.1. We observe that Mask R-CNN can benefit from extra data, even if that data is unlabeled.
By using the same test-time augmentation as used for
instance segmentation, we further boost APkp to 70.4.
10
Acknowledgements: We would like to acknowledge Ilija
Radosavovic for contributions to code release and enhanced
results, and the Caffe2 team for engineering support.
[21] J. Huang, V. Rathod, C. Sun, M. Zhu, A. Korattikara,
A. Fathi, I. Fischer, Z. Wojna, Y. Song, S. Guadarrama, et al.
Speed/accuracy trade-offs for modern convolutional object
detectors. In CVPR, 2017. 2, 3, 4, 6, 7
[22] M. Jaderberg, K. Simonyan, A. Zisserman, and
K. Kavukcuoglu.
Spatial transformer networks.
In
NIPS, 2015. 4
[23] A. Kirillov, E. Levinkov, B. Andres, B. Savchynskyy, and
C. Rother. Instancecut: from edges to instances with multicut. In CVPR, 2017. 3, 9
[24] A. Krizhevsky, I. Sutskever, and G. Hinton. ImageNet classification with deep convolutional neural networks. In NIPS,
2012. 2
[25] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E.
Howard, W. Hubbard, and L. D. Jackel. Backpropagation
applied to handwritten zip code recognition. Neural computation, 1989. 2
[26] Y. Li, H. Qi, J. Dai, X. Ji, and Y. Wei. Fully convolutional
instance-aware semantic segmentation. In CVPR, 2017. 2,
3, 5, 6
[27] T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and
S. Belongie. Feature pyramid networks for object detection.
In CVPR, 2017. 2, 4, 5, 7
[28] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 2, 5
[29] S. Liu, J. Jia, S. Fidler, and R. Urtasun. SGN: Sequential grouping networks for instance segmentation. In ICCV,
2017. 3, 9
[30] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional
networks for semantic segmentation. In CVPR, 2015. 1, 3, 6
[31] V. Nair and G. E. Hinton. Rectified linear units improve restricted boltzmann machines. In ICML, 2010. 4
[32] G. Papandreou, T. Zhu, N. Kanazawa, A. Toshev, J. Tompson, C. Bregler, and K. Murphy. Towards accurate multiperson pose estimation in the wild. In CVPR, 2017. 8
[33] P. O. Pinheiro, R. Collobert, and P. Dollar. Learning to segment object candidates. In NIPS, 2015. 2, 3
[34] P. O. Pinheiro, T.-Y. Lin, R. Collobert, and P. Dollár. Learning to refine object segments. In ECCV, 2016. 2, 3
[35] I. Radosavovic, P. Dollár, R. Girshick, G. Gkioxari, and
K. He. Data distillation: Towards omni-supervised learning.
arXiv:1712.04440, 2017. 10
[36] S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NIPS, 2015. 1, 2, 3, 4, 7
[37] S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In TPAMI, 2017. 10
[38] A. Shrivastava, A. Gupta, and R. Girshick. Training regionbased object detectors with online hard example mining. In
CVPR, 2016. 2, 5
[39] A. Shrivastava, R. Sukthankar, J. Malik, and A. Gupta. Beyond skip connections: Top-down modulation for object detection. arXiv:1612.06851, 2016. 4, 7
[40] C. Sun, A. Shrivastava, S. Singh, and A. Gupta. Revisiting
unreasonable effectiveness of data in deep learning era. In
ICCV, 2017. 10
References
[1] M. Andriluka, L. Pishchulin, P. Gehler, and B. Schiele. 2D
human pose estimation: New benchmark and state of the art
analysis. In CVPR, 2014. 8
[2] P. Arbeláez, J. Pont-Tuset, J. T. Barron, F. Marques, and
J. Malik. Multiscale combinatorial grouping. In CVPR,
2014. 2
[3] A. Arnab and P. H. Torr. Pixelwise instance segmentation
with a dynamically instantiated network. In CVPR, 2017. 3,
9
[4] M. Bai and R. Urtasun. Deep watershed transform for instance segmentation. In CVPR, 2017. 3, 9
[5] S. Bell, C. L. Zitnick, K. Bala, and R. Girshick. Insideoutside net: Detecting objects in context with skip pooling
and recurrent neural networks. In CVPR, 2016. 5
[6] Z. Cao, T. Simon, S.-E. Wei, and Y. Sheikh. Realtime multiperson 2d pose estimation using part affinity fields. In CVPR,
2017. 7, 8
[7] M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler,
R. Benenson, U. Franke, S. Roth, and B. Schiele. The
Cityscapes dataset for semantic urban scene understanding.
In CVPR, 2016. 9
[8] J. Dai, K. He, Y. Li, S. Ren, and J. Sun. Instance-sensitive
fully convolutional networks. In ECCV, 2016. 2
[9] J. Dai, K. He, and J. Sun. Convolutional feature masking for
joint object and stuff segmentation. In CVPR, 2015. 2
[10] J. Dai, K. He, and J. Sun. Instance-aware semantic segmentation via multi-task network cascades. In CVPR, 2016. 2, 3,
4, 5, 6
[11] J. Dai, Y. Li, K. He, and J. Sun. R-FCN: Object detection via
region-based fully convolutional networks. In NIPS, 2016. 2
[12] R. Girshick. Fast R-CNN. In ICCV, 2015. 1, 2, 3, 4, 6
[13] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic
segmentation. In CVPR, 2014. 2, 3
[14] R. Girshick, F. Iandola, T. Darrell, and J. Malik. Deformable
part models are convolutional neural networks. In CVPR,
2015. 4
[15] B. Hariharan, P. Arbeláez, R. Girshick, and J. Malik. Simultaneous detection and segmentation. In ECCV. 2014. 2
[16] B. Hariharan, P. Arbeláez, R. Girshick, and J. Malik. Hypercolumns for object segmentation and fine-grained localization. In CVPR, 2015. 2
[17] Z. Hayder, X. He, and M. Salzmann. Shape-aware instance
segmentation. In CVPR, 2017. 9
[18] K. He, X. Zhang, S. Ren, and J. Sun. Spatial pyramid pooling
in deep convolutional networks for visual recognition. In
ECCV. 2014. 1, 2
[19] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In CVPR, 2016. 2, 4, 7, 10
[20] J. Hosang, R. Benenson, P. Dollár, and B. Schiele. What
makes for effective detection proposals? PAMI, 2015. 2
11
[41] C. Szegedy, S. Ioffe, and V. Vanhoucke. Inception-v4,
inception-resnet and the impact of residual connections on
learning. In ICLR Workshop, 2016. 7
[42] J. R. Uijlings, K. E. van de Sande, T. Gevers, and A. W.
Smeulders. Selective search for object recognition. IJCV,
2013. 2
[43] X. Wang, R. Girshick, A. Gupta, and K. He. Non-local neural
networks. arXiv:1711.07971, 2017. 10
[44] S.-E. Wei, V. Ramakrishna, T. Kanade, and Y. Sheikh. Convolutional pose machines. In CVPR, 2016. 8
[45] S. Xie, R. Girshick, P. Dollár, Z. Tu, and K. He. Aggregated
residual transformations for deep neural networks. In CVPR,
2017. 4, 10
12
| 1cs.CV
|
A Simple Parallel and Distributed Sampling Technique:
Local Glauber Dynamics
arXiv:1802.06676v1 [cs.DS] 19 Feb 2018
Manuela Fischer
ETH Zurich
[email protected]
Mohsen Ghaffari
ETH Zurich
[email protected]
Abstract
Sampling constitutes an important tool in a variety of areas: from machine learning and combinatorial optimization to computational physics and biology. A central class of sampling algorithms
is the Markov Chain Monte Carlo method, based on the construction of a Markov chain with
the desired sampling distribution as its stationary distribution. Many of the traditional Markov
chains, such as the Glauber dynamics, do not scale well with increasing dimension. To address this
shortcoming, we propose a simple local update rule based on the Glauber dynamics that leads to
efficient parallel and distributed algorithms for sampling from Gibbs distributions.
Concretely, we present a Markov chain that mixes in O(log n) rounds when Dobrushin’s condition for the Gibbs distribution is satisfied. This improves over the LubyGlauber algorithm by
Feng, Sun, and Yin [PODC’17], which needs O(∆ log n) rounds, and their LocalMetropolis algorithm, which converges in O(log n) rounds but requires a considerably stronger mixing condition.
Here, n denotes the number of nodes in the graphical model inducing the Gibbs distribution, and
∆ its maximum degree. In particular, our method can sample a uniform proper coloring with
(2 + ε)∆ colors in O(log n) rounds, for any constant ε > 0, which
almost matches the threshold of
√
the sequential Glauber dynamics and improves on the (2 + 2 + ε)∆ threshold of Feng et al.
1
Introduction
Markov Chain Monte Carlo Method: The Markov Chain Monte Carlo (MCMC) method is a
central class of algorithms for sampling, that is, for randomly drawing an element from a ground set according to a certain probability distribution. It works by constructing a Markov chain with the targeted
sampling distribution as its stationary distribution. Within a number of steps, known as the mixing
time, the Markov chain converges; its state then (approximately) follows this distribution. Besides
the intrinsic interest of such a general sampling method, in particular for complex distributions where
simple sampling techniques fail, the MCMC method gives rise to efficient approximation algorithms in
a variety of areas: enumerative combinatorics (due to the fundamental connection between sampling
and counting established by Jerrum, Valiant, and Vazirani [JVV86]), simulated annealing [NSS86] in
combinatorial optimization, Monte Carlo simulations [MRR+ 53] in statistical physics, computation of
intractable integrals for, among many others, Bayesian inference [ADFDJ03] in machine learning.
Parallel and Distributed Sampling: The employment of MCMC methods is particularly important
when confronted with high-dimensional data, where traditional (exact) approaches quickly become
intractable. Such data sets are not only increasingly frequent, but also critical for the success of many
applications. For instance in machine learning, higher-dimensional models help expressability and
hence predictability. It is thus central that MCMC algorithms scale well with increasing dimensions.
This is not the case, however, for most sequential methods, as they process and update the variables
one by one, that is, a single site per step.
To speed up the sampling process, Markov chain updates can be parallelized by spreading the
variables across several processors. In other settings, such as distributed machine learning, the (data
associated to) variables might already be naturally distributed among several machines, and the overhead of aggregating them into one machine, if they fit there in the first place, would be untenable.
Local Sampling: In either case, to avoid overhead in communication and coordination, local update
rules for Markov chains are needed: a machine must be able to change the value of its variables without
knowing all the values of the variables on other machines. Yet, the joint distribution, over all variables
in the system, must converge to a certain distribution. This local sampling problem was introduced
in a recent work by Feng, Sun, and Yin [FSY17], whose title asks “What can be sampled locally?”.
We address this question by providing a simple and generic sampling technique—the Local Glauber
Dynamics, informally introduced in Section 1.2 and formally described in Section 2—which is applicable for a wide range of distributions, as stated in Section 1.1. This moves us a step closer towards an
answer of this question, thus towards the goal of generally understanding what can be sampled locally.
Besides its many practical ramifications, especially on the area of distributed machine learning, this
gives us a theoretical insight about the locality of problems, whose systematic study has been initiated
by the seminal works of Linial [Lin87] and Naor and Stockmeyer [NS95] with the pithy title “What
can be computed locally?”.
1.1
Our Result, and Related Work
For the sake of succinctness and comprehensibility of the presentation, we state and prove our main
result in terms of the special case that gets most attention for sequential sampling: sampling proper
colorings of a graph. We refer to [FV07] for a survey on sequential sampling of proper colorings. Our
result applies to a more general set of distributions, however, as explained in Remark 1.2.
Theorem 1.1. A uniform proper q-coloring of an n-node
graph with maximum degree ∆ can be
sampled within total variation distance ε > 0 in O log nε rounds, where q = α∆ for any α > 2.
Our parallel and distributed sampling algorithm
improves over the LubyGlauber algorithm by
Feng, Sun, and Yin [FSY17], which needs O ∆ log nε rounds, and their LocalMetropolis algorithm,
1
which converges in O log nε rounds but requires a considerably stronger mixing condition of α >
√
√
2 + 2. They state that “We also believe that the 2 + 2 threshold is of certain significance to this
[LocalMetropolis] chain as the Dobrushin’s condition to the Glauber dynamics.”, thus implying that
this value is a barrier for their approach. This is also justified by the supposedly easiest special case
of
√ a tree that leads to the same threshold for their algorithm. Our result gets rid of the additional
2 while not incurring any loss in the round complexity, with a considerably easier and more natural
update rule. Not least also our proof
is simpler and shorter. Moreover, our algorithm is asymptotically
best possible, as there is an Ω log nε lower bound [GJL17,FSY17] due to the exponential correlation
between variables.
The threshold of α > 2 corresponds to Dobrushin’s condition, thus almost matches the threshold of
the sequential Glauber dynamics [Jer95, SS97] at 2∆ + 1. In other words, we present a technique that
fully parallelizes the Glauber dynamics, speeding up the mixing time from poly n steps to O(log n)
rounds. In terms of number of colors needed, Dobrushin’s condition can be undercut: Vigoda [Vig00]
showed that α = 11
6 are enough, when resorting to a different highly non-local Markov chain. This
gives rise to the question whether efficient distributed algorithms intrinsically need to be stuck at
Dobrushin’s condition, which would imply that this bound is inherent to the locality of the sampling
process, or whether our threshold is an artifact of our possibly suboptimal dynamics.
Remark 1.2. In fact, our technique directly applies for sampling from the Gibbs distribution induced
by a Markov random field1 if Dobrushin’s condition [Dob68] is satisfied. More generally, it can used for
sampling from any local (that is, constant-radius) constraint satisfaction problems, which is universal
for conditional independent joint distributions, due to Hammersley-Clifford’s fundamental theorem
[HC71]. Moreover, our proof presented here captures all the difficulties that arise in these more general
cases, thus can be adapted in a straight-forward manner. We defer this generalization to the full version
of the paper.
1.2
Our Sampling Technique, and Related Approaches
Over the past few years, several methods to parallelize sequential Markov chains have been proposed.
Most of them rely on a heavy coordination machinery, are special purpose, and/or do not provide any
theoretical guarantees. In the following, we briefly outline two of the most promising and more generic
parallel and distributed sampling techniques, in the context of colorings.
The most natural one follows a standard decentralization approach, also implemented in the
LubyGlauber algorithm by [FSY17]: an independent set of nodes (e.g., a color class of a proper
coloring) simultaneously updates their color [FSY17], ensuring that no two neighboring nodes change
their color at the same time. This approach mainly suffers from the limitation that the number of
independent sets needed to cover all nodes might be large, which slows down mixing. In particular,
a multiplicative ∆-term in the mixing time seems inevitable [GLGG11, FSY17]. In the worst case of
a clique, this approach falls back to sequential sampling, updating one node after the other. Moreover, this method requires an independent set to be computed, which incurs a significant amount of
additional communication and coordination.
An orthogonal direction was pursued by [NSWA08,YXQ09,FSY17], where methods are introduced
to update the colors of all nodes simultaneously. One example is the LocalMetropolis algorithm by
[FSY17]. This extreme parallelism, however, comes at a cost of either introducing a bias in the
stationary distribution, resulting in a non-uniform coloring [NSWA08, YXQ09], or having to demand
stronger mixing conditions [FSY17].
1
This captures many graph problems—such as independent set, vertex cover, graph homomorphism—and physical
models—such as Ising model, Potts model, general spin systems, and hardcore gas model.
2
Our Local Sampling Technique: We aim for the middle ground between these two approaches,
motivated by the following observation: we do not need to prevent simultaneous updates of adjacent
nodes, only simultaneous conflicting updates of adjacent nodes. Preventing two adjacent nodes in the
first place from picking a new color in the same round seems to be way too restrictive, in particular
because it is unlikely that both nodes aim for the same new color. On the other hand, if all nodes
update their colors simultaneously, a node is expected to have a conflict with at least one of its
neighbors, which prevents any progress.
We interpolate between the two extreme cases by introducing a marking probability, so that only
a small fraction of a node’s neighbors is expected to update the color, and hence also, in worst
case, only these can conflict with its update. Concretely, we propose the following generic sampling
method, which we call Local Glauber Dynamics: In every step, every variable independently marks
itself at random with a certain (low) probability. If it is marked, it samples a proposal at random
and checks with its neighbors whether the proposal leads to a conflict with their current state or their
new proposals (if any). If there is a conflict, the variable rolls back and stays with its current state,
otherwise the state is updated. As opposed to sequential sampling, where only one variable per step
updates its value, here the expected number of variables simultaneously updating their value is Ω(n),
resulting in the desired speed-up from O(n log n), say, to O(log n). Of course, the main technical
aspect lies in showing that this simple update rule converges to the uniform distribution in O(log n)
rounds, which we prove in Section 2.
1.3
Notation and Preliminaries
Model: We work with the standard distributed message-passing model for the study of locality: the
LOCAL model introduced by Linial [Lin87], defined as follows. Given a graph G = (V, E) on n nodes
with maximum degree ∆, the computation proceeds in rounds. In every round, every node can send
a message to each of its neighbors. We do not limit the message sizes, but for the algorithm that we
present, O(log n)-bit messages suffice. In the end of the computation, every node v outputs a color.
The quantity of main interest is the round complexity, i.e., the number of rounds until the joint output
of all nodes satisfies a certain condition. We assume that all nodes have knowledge of log n and ∆.
(t)
Markov Chain: We will consider a Markov chain X = X (t) t≥0 , where X (t) = Xv
∈ [q]V is
v∈V
the coloring of the graph in round t. We will omit the round index, and use X = (Xv )v∈V ∈ [q]V for
the coloring at time t and X 0 = (Xv0 )v∈V ∈ [q]V for the coloring at time t + 1, for a t ≥ 0, instead.
(t)
Mixing Time: For a Markov chain X (t) t≥0 with stationary distribution µ, let πσ denote the
distribution of the random coloring nX (t) of the
chain atotime t ≥ 0, conditioned on X (0) = σ. The
(t)
mixing time τmix (ε) = maxσ∈Ω min t ≥ 0 : dTV πσ , µ
is defined to be the minimum number of
rounds needed so that the Markov chain is ε-close (in terms of total variation distance) to its stationary
distribution µ, regardless P
of X (0) . The total variation distance between two distributions µ, ν over Ω
is defined as dTV (µ, ν) = σ∈Ω 21 |µ(σ) − ν(σ)|.
Path Coupling: The Path Coupling Lemma by Bubley and Dyer [BD97, Theorem 1] (also see [FSY17,
Lemma 4.3]) gives rise to a particularly easy way of designing couplings. In a simplified version, it says
that it is enough to define the coupling of a Markov chain only for pairs of colorings that are adjacent,
that is, differ at exactly one node. The expected number of differing nodes after one coupling step
then can be used to bound the mixing time of the Markov chain.
Lemma 1.3 (Path Coupling [BD97], simplified). For σ, σ 0 ∈ [q]V , let φ(σ, σ 0 ) := |{v ∈ V : σv 6= σv0 }|.
If there is a coupling (X, Y ) → (X 0 , Y 0 ) of the Markov chain, defined only for (X, Y ) with φ(X,
Y ) = 1,
that satisfies E[φ(X 0 , Y 0 ) | X, Y ] ≤ 1 − δ for some 0 < δ < 1, then τmix (ε) = O 1δ · log nε .
3
2
Local Glauber Dynamics
Local Glauber Dynamics: We define a transition from X = (Xv )v∈V to X 0 = (Xv0 )v∈V in one round
as follows. Every node v ∈ V marks itself independently with probability 0 < γ < 1. If it is marked,
it proposes a new color cv ∈ [q] uniformly at random, independently from all the other nodes. If this
proposed color
proposed colors of any neighbor,
S does not lead to a conflict with the current and the
that is, cv ∈
/ u∈N (v) {Xu , cu } and cu ∈
/ {Xv , cv } for any u ∈ N (v)2 , then v accepts color cv , thus sets
Xv0 = cv . Otherwise, v keeps its current color, thus sets Xv0 = Xv .
Stationary Distribution: The local Glauber dynamics is ergodic: it is aperiodic, as there is always a
positive probability of not changing any of the colors, and irreducible, since any (proper) coloring can
be reached from any coloring. Moreover, the chain might possibly start from an improper coloring, but
it will never move from a proper to an improper coloring, that is, it is absorbing to proper colorings.
It is easy to verify that this local Glauber dynamics, due to its symmetric update rule, satisfies the
detailed balance equation for the uniform distribution, meaning that the transition from X to X 0 has
the same probability as a transition from X 0 to X for proper colorings. The chain thus is reversible
and has the uniform distribution over all proper colorings as unique stationary distribution.
Mixing Time: Informally speaking, the Path Coupling Lemma says that if for all X and Y which
differ in one node, we can define a coupling (X, Y ) → (X 0 , Y 0 ) in such a way that the expected number
of nodes at which X 0 and Y 0 differ is bounded away from 1 from above, then the chain converges quickly.
In Section 2.1, we formally describe such a path coupling, in Section 2.2, we list necessary (but not
necessarily sufficient) conditions for a node to have two different colors after one coupling step, which
is then used in Section 2.3 to bound the expected number of differing nodes by 1 − δ for some constant
0 < δ < 1, depending on α. Application of Lemma 1.3 then concludes the proof of Theorem 1.1.
2.1
Description of Path Coupling
We look at two colorings X and Y that differ at a node v0 ∈ V only. That is, r = Xv0 6= Yv0 = b, for
some r 6= g ∈ [q], which we will naturally refer to as red and blue, respectively, and Xv = Yv for all
Y
v 6= v0 ∈ V . In the following, we explain how every node v ∈ V comes up with a pair (cX
v , cv ) of new
proposals, which then will be accepted or rejected based on the local Glauber dynamics rules.
Marking: In both chains, every node v ∈ V is marked independently with probability γ, using the
same randomness in both chains. In the following, we restrict our attention to marked nodes only;
Y
non-marked nodes are thought of proposing their current color as new color, i.e., cX
v = Xv and cv = Yv .
Consistent, Mirrored, and Flipped Proposals: We introduce two possible ways of how proposals
for a node v can be sampled: consistently and mirroredly. For the consistent proposals, both chains
Y
propose the same randomly chosen color, that is, cX
v0 = cv0 = c for a u.a.r. c ∈ [q]. For the mirrored
proposals, both chains assign the same random proposal if it is neither red nor blue, and a flipped
Y
proposal (i.e., red to one and blue to the other chain) otherwise. More formally, cX
v0 = c and cv0 = c if
X
Y
c ∈ {r, g} and c the element in {r, g} \ {c}, and cv0 = cv0 = c if c ∈
/ {r, g}, for a u.a.r. c ∈ [q]. We say
Y . Note that we say mirrored proposal to refer to the process of
that v has flipped proposals if cX
=
6
c
v
v
sampling mirroredly, and we say flipped if, as a result of sampling mirroredly, a node proposes different
colors in the two chains.
Breadth-First Assignment of Proposals: Let B = {v ∈ V \ {v0 } : Xv ∈S{r, b}} ⊆ V \ {v0 } be
+
the set of nodes v 6= v0 with current color red or blue, as well as K =
v∈B N (v) \ {v0 } its
inclusive neighborhood, without v0 . We now ignore this set K for the moment, thus focus on the
set S ⊆ V \ K of marked nodes that are not adjacent to a node with color red or blue (except for
2
To simplify notation, we assume that cu = Xu in case u is not marked.
4
v0
v0
M0
M1
M1
u
u
M2
w
v
M0
M2
w
M3
v
M4
M3
M4
Figure 1: The breadth-first layers M d for d ≥ 0 of two chains that differ at v0 ∈ M 0 . The disk color corresponds
to the node’s current color, where black means any color except red and blue. The color of the box around a node
shows this node’s proposed color, where white stands for any color (possibly also red or blue, but consistent).
Dashed boxes indicate the sets F d of nodes with flipped proposals. Note that node v appears in layer 4 even
though it has distance 3 to v0 . This is because we perform the breadth-first assignment only on nodes with
flipped proposals. v’s neighbor u does not have flipped proposals, thus is in M 2 \ F 2 , which means that u’s
neighbors are not added to the next layer. Only v’s neighbor w ∈ F 3 leads to v being added to M 4 .
possibly v0 ). Informally speaking, we will go through these nodes in a breadth-first manner, with
increasing distance d ≥ 0 to node v0 , and fix their proposals layer by layer, but defer the assignment
of nodes not (yet) adjacent to a node with flipped proposals, as follows. We repeatedly add all (still
remaining) nodes that have a node in the last layer with flipped proposals to a new layer, and sample
their proposals mirroredly, thus perform a breadth-first assignment on nodes with flipped proposals
only. All remaining nodes sample their proposals consistently. Note that this in particular guarantees
that a node is sampled consistently only if it not adjacent to a node with flipped proposals.
More formally, this can be described as follows. We define M 0 = F 0 = {v0 }, even if v0 is not
marked. For node v0 , if marked, the proposals are sampled consistently. For d ≥ 1 and v ∈ M d ,
the proposals are sampled mirroredly. For the subsequent layer, we restrict the attention
Sd to (new)
d
d+1
d
neighbors of nodes in M with flipped proposals only, i.e., consider M
= N F \ i=0 M d for
d
d
X
Y
F = {v ∈ M : cv0 6= cv0 }.
For all remaining (marked) nodes, that is, nodes in S \ M and nodes in K, proposals are sampled
consistently. See Figure 1 for an illustration of this breadth-first-based approach.
Y
Accept Proposals: The proposals (cX
v )v∈V and (cv )v∈V in the chains X and Y are accepted or
rejected based on the local Glauber dynamics rules, leading to colorings X 0 , Y 0 ∈ [q]V .
2.2
Properties of the Coupling
The main observation is the following. If we ignore nodes with current colors red and blue for the
moment, one can argue that X 0 and Y 0 can only differ at a node different from v0 if its proposals are
flipped. Flipped proposals, however, can only arise when the proposals are sampled mirroredly, which
happens only if there is a node in the preceding layer with flipped proposals (due to the breadth-first
order in which we assign the proposals). A node thus can lead to an inconsistency only if there is path
in G from v0 to this node consisting of nodes with flipped proposals, called a flip path.
We will next make this intuition with the flip paths more precise, in two parts: for nodes in S
5
(that sample their proposals mirroredly if adjacent to a node with flipped proposals) in Lemma 2.1
and for nodes in K (that always sample their proposals consistently) in Lemma 2.2. See Figure 2 for
an illustration of these two cases.
Lemma 2.1. If X 0 and Y 0 differ at v 6= v0 ∈ S, there is a flip path (v0 , . . . , v` = v) ∈ F 0 × · · · × F `
of length ` ≥ 1 in G, with the additional property that the proposal of v is the opposite of the last
Y
color (in red and blue) seen on this path, in both chains. More formally, cY = cX
v 6= cv = cX , where
X
Y
cX = cv`−1 and cY = cv`−1 if ` > 1, and cX = Xv0 and cY = Yv0 if ` = 1.
Proof. We first argue that v’s proposals must be flipped and accepted in both chains. Trivially,
acceptance of a consistent proposal in both chains or rejection in both chains leads to Xv0 = Yv0 .
Moreover, observe that flipped proposals are, by construction, either accepted in both or rejected in
both chains, as flipping changes the role of red and blue, but not the overall behavior. Indeed, suppose,
without loss of generality, that cX
v = c ∈ {r, g} is rejected by X. Thus, in particular, v has a neighbor
u with current color or proposal c in X. As we are restricting our attention to the set S which does
not have any adjacent node with current color red or blue, except for v0 , either u = v0 or u proposes
c. So u either must have different current colors (if u = v0 ) or have mirrored proposals (if v ∈ F d ,
0
then u ∈ M d for some d0 ≤ d + 1, because at the latest v’s flipped proposal leads to u being added
to the subsequent layer, by how we assign the proposals in breadth-first manner) and hence flipped
proposals. Thus, v’s proposal c in Y will be rejected by Y , since either u = v0 ∈ N (v) has color c or
u ∈ N (v) proposes c.
It thus remains to rule out the case of consistent proposals that are accepted in one and rejected
in the other chain. Towards a contradiction, suppose that v proposes the same color cv in both chains,
Y
and that it is accepted in one and rejected in the other. Since Xv = Yv and cX
v = cv , this can
happen only if v is adjacent either to v0 or to at least one node with flipped proposals, as otherwise
all proposals and all current colors in v’s inclusive neighborhood would be the same, leading to the
same behavior in both chains. In both cases, v ∈ M d for some d ≥ 1, which means that its proposals
are sampled mirroredly. Hence, cv ∈
/ {r, g}, as otherwise the proposals would be flipped. Now, since
neither v’s current color nor v’s proposals is red or blue, and neighbors of v can differ in their colors or
proposals only if red or blue is involved, the proposals are either accepted or rejected in both chains.
It follows that indeed only nodes in S with flipped proposals that are accepted in both chains can
have different colors in X 0 and Y 0 .
By construction of the layers, and since v ∈ F ` for some ` ≥ 1, there must exist a sequence of nodes
v1 ∈ F 1 , . . . , v`−1 ∈ F `−1 connecting v0 to v in G: a flip path of length `. Moreover, the proposal is
accepted in a chain only if the proposed color is the opposite of the color (red or blue) that is seen on
the path (either as proposal if ` > 1, or as current color of v0 if ` = 1).
Lemma 2.2. If X 0 and Y 0 differ at v 6= v0 ∈ K, there is an almost flip path (v0 , . . . , v` = v) ∈
F 0 × · · · × F `−1 × K of length ` ≥ 1 in G, with the additional property that the proposal of v is either
Y
red or blue, that is, cv = cX
v = cv ∈ {r, b}.
Proof. Since, by definition of the coupling, v ∈ K samples its proposals consistently, X 0 and Y 0 can
only differ at v 6= v0 if the proposal is accepted in one and rejected in the other chain. This can
happen only if v is adjacent to either v0 or to at least one node with flipped proposals. Otherwise, all
proposals and all current colors in v’s inclusive neighborhood would be the same, leading to the same
behavior. Hence, v is adjacent to some u ∈ F d for some d ≥ 0. By construction of the layers, there
must exist a sequence of nodes v1 ∈ F 1 , . . . , v`−1 = u ∈ F `−1 connecting v0 to v in G: an almost flip
path of length ` = d + 1. Note that, in particular, because neighbors of nodes in B are by definition
sampled consistently (as they are in K), and a node at the end of an almost flip path has a neighbor
with flipped proposals, this last node on an almost flip path must be in K \ B.
6
The proposal cv is accepted in one and rejected in the other chain only if cv ∈ {r, g}. In that case,
the chain with the same color on the end of the path will reject, the other will (possibly) accept.
v0
F0
v0
F0
v0
v0
F0
F0
F1
F1
F1
F1
F2
F2
F2
F2
F3
F3
v
K nB
K nB
v
B
v
F4
X
v
B
F4
X
Y
Y
Figure 2: A flip path on the left: v’s flipped proposals are accepted in both chains, yielding Xv0 = r and Yv0 = b.
An almost flip path on the right: v ∈ K \ B samples its proposals consistently. In chain X, the proposal r will
be accepted, in chain Y , it will be rejected, leading to Xv0 = r 6= Yv0 = Yv . The disk color corresponds to the
node’s current color, where black means any color except red and blue. The color of the box around a node
indicates this node’s proposed color, where white means any color ( also red and blue, but consistent).
2.3
Bounding the Expected Number of Differing Nodes
P
We show that E[φ(X 0 , Y 0 ) | X, Y ] ≤ 1 − δ for some 0 < δ < 1, by bounding E[ v6=v0 ∈V 1 (Xv0 6= Yv0 ) |
X, Y ] and E[1 Xv0 0 6= Yv00 | X, Y ] separately. We will see that, as δ → 0, both terms can be bounded
by ≈ α1 , leading to an expected number of roughly α2 , which is strictly less than 1 for α > 2.
Nodes v 6= v0 : Section 2.2, or more precisely, Lemmas 2.1 and 2.2, show that the number of nodes
(different from v0 ) that have different colors in X 0 and Y 0 can be bounded by the number of (almost)
flip paths with an additional property. We will next see that the expected number of such (almost)
flip paths can be expressed as a geometric series summing over the depths of the layers.
There are at most ∆` paths (v0 , . . . , v` ) of length ` in G. Moreover, each such path has probability
(2γ/q)`−1 γ/q of being a flip or almost flip path with the mentioned additional property, since all
intermediate nodes v1 , . . . , v`−1 need to mark themselves and to propose one arbitrary color in {r, g},
and v` needs to mark itself and to propose the one color in {r, g} as specified in Lemma 2.1 and
Lemma 2.2, respectively. Note that a path in G can either be a flip path or an almost flip path, but
never both. Moreover, observe that node v0 does not need to be marked. We get
`−1
γ∆
∞
∞
X
X
γ
1 X 2γ∆ `
2γ
q
· =
≤
.
(1)
E
1(Xv0 6= Yv0 ) | X, Y ≤
∆` ·
q
q
2
q
1 − 2γ∆
v6=v0 ∈V
`=1
`=1
7
q
Node v0 : Chains X 0 and Y 0 can agree at node v0 only if at least one the proposals is accepted. For
X
Y
that, v0 needs to be marked and its proposal cvS
0 = cv0 = cv0 needs to be different from all the at most
∆ current colors of its neighbors, that is, cv ∈
/ v∈N (v0 ) {Xv }, which happens with probability at least
γ (1 − ∆/q). Moreover, the proposals of v0 ’s neighbors (if marked) need to avoid at most three colors
in {cv0 , r, g}, possibly less, which happens with probability at least 1 − 3γ/q. We thus get
∆
3γ ∆
0
0
E 1 Xv0 6= Yv0 ≤ 1 − γ 1 −
1−
(2)
q
q
Wrap-Up: Overall, combining Equations (1) and (2), we get
1
E[φ(X 0 , Y 0 ) | X, Y ] ≤ 1 − γ 1 −
α
− 6γ
α
e
+
γ
α
1−
2γ
α
1 − 1
= 1 − γe
| {z }
α
→1
− 6γ
α
as γ→0
!
6γ
eα
1+
2γ .
1− α
{z
}
|
→2 as γ→0
For α > 2 and γ := γ(α) small enough, this is strictly bounded away from 1 from above, where the
hidden constant depends on α (but not on ∆ or n).
References
[ADFDJ03] Christophe Andrieu, Nando De Freitas, Arnaud Doucet, and Michael I. Jordan. An
Introduction to MCMC for Machine Learning. Machine Learning, 50(1-2):5–43, 2003.
[BD97]
Russ Bubley and Martin Dyer. Path Coupling: A Technique for Proving Rapid Mixing
in Markov Chains. In the Proceedings of the Symposium on Foundations of Computer
Science (FOCS), pages 223–231, 1997.
[Dob68]
Roland L. Dobruschin. The Description of a Random Field by Means of Conditional
Probabilities and Conditions of its Regularity. Theory of Probability & Its Applications,
13(2):197–224, 1968.
[FSY17]
Weiming Feng, Yuxin Sun, and Yitong Yin. What Can Be Sampled Locally? In Proceedings of the International Symposium on Principles of Distributed Computing (PODC),
pages 121–130, 2017.
[FV07]
Alan Frieze and Eric Vigoda. A Survey on the Use of Markov Chains to Randomly Sample
Colourings. Oxford Lecture Series in Mathematics and its Applications, 34:53, 2007.
[GJL17]
Heng Guo, Mark Jerrum, and Jingcheng Liu. Uniform Sampling Through the Lovász
Local Lemma. Proceedings of the Symposium on Theory of Computing (STOC), pages
342–355, 2017.
[GLGG11] Joseph Gonzalez, Yucheng Low, Arthur Gretton, and Carlos Guestrin. Parallel Gibbs
Sampling: From Colored Fields to Thin Junction Trees. In the Proceedings of the International Conference on Artificial Intelligence and Statistics, pages 324–332, 2011.
[HC71]
John M. Hammersley and Peter Clifford. Markov Fields on Finite Graphs and Lattices.
1971.
[Jer95]
Mark Jerrum. A Very Simple Algorithm for Estimating the Number of k-Colorings of a
Low-Degree Graph. Random Structures & Algorithms, 7(2):157–165, 1995.
8
[JVV86]
Mark R. Jerrum, Leslie G. Valiant, and Vijay V. Vazirani. Random Generation of Combinatorial Structures from a Uniform Distribution. Theoretical Computer Science, 43:169–
188, 1986.
[Lin87]
Nathan Linial. Distributive Graph Algorithms - Global Solutions From Local Data. In
the Proceedings of the Symposium on Foundations of Computer Science (FOCS), pages
331–335. IEEE, 1987.
[MRR+ 53] Nicholas Metropolis, Arianna W. Rosenbluth, Marshall N. Rosenbluth, Augusta H. Teller,
and Edward Teller. Equation of State Calculations by Fast Computing Machines. The
Journal of Chemical Physics, 21(6):1087–1092, 1953.
[NS95]
Moni Naor and Larry Stockmeyer. What Can Be Computed Locally? SIAM Journal on
Computing, 24(6):1259–1277, 1995.
[NSS86]
Surendra Nahar, Sartaj Sahni, and Eugene Shragowitz. Simulated Annealing and Combinatorial Optimization. In Proceedings of the Design Automation Conference, pages
293–299. IEEE Press, 1986.
[NSWA08] David Newman, Padhraic Smyth, Max Welling, and Arthur U. Asuncion. Distributed
Inference for Latent Dirichlet Allocation. In Advances in Neural Information Processing
Systems, pages 1081–1088, 2008.
[SS97]
Jesús Salas and Alan D. Sokal. Absence of Phase Transition for Antiferromagnetic Potts
Models via the Dobrushin Uniqueness Theorem. Journal of Statistical Physics, 86(34):551–579, 1997.
[Vig00]
Eric Vigoda. Improved Bounds for Sampling Colorings. Journal of Mathematical Physics,
41(3):1555–1569, 2000.
[YXQ09]
Feng Yan, Ningyi Xu, and Yuan Qi. Parallel Inference for Latent Dirichlet Allocation on
Graphics Processing Units. In Advances in Neural Information Processing Systems, pages
2134–2142, 2009.
9
| 8cs.DS
|
Mining relevant interval rules
Thomas Guyet1 , René Quiniou2 , and Véronique Masson3
arXiv:1709.03267v1 [cs.AI] 11 Sep 2017
1
AGROCAMPUS-OUEST/IRISA-UMR 6074
2
Inria, Centre de Rennes
3
University Rennes-1/IRISA-UMR 6074
Abstract. This article extends the method of Garriga et al. for mining relevant rules to numerical attributes by extracting interval-based
pattern rules. We propose an algorithm that extracts such rules from
numerical datasets using the interval-pattern approach from Kaytoue et
al. This algorithm has been implemented and evaluated on real datasets.
Keywords: rule learning, interval patterns, relevant rules, closed patterns
1
Introduction
Garriga et al. [2] proposed a method to extract relevant association rules from
labeled itemsets. We extend the work of Garriga et al. to numerical attributes
using the pattern mining approach of Kaytoue et al. [3] which is based on FCA
(Formal Concept Analysis). Kaytoue et al. [3] proposed to extend the mining
of frequent closed interval pattern to numerical data. Our work bridges the gap
between these two approaches to extract relevant interval pattern rules.
2
Closed interval patterns
Let F = {f1 , . . . , fn } be a fixed set of n features. We represent a training example as a tuple of real values x = {x1 , . . . , xn }, where xi ∈ Dom(fi ), with an
associated class label. The tuple stores one value per feature of F . We consider
two-class learning problems where the set of examples E is divided in positives
(P ) and negatives (N ) such that E = P ∪N and P ∩N = ∅. Multi-class problems
can be transformed in two-class learning problems.
An n-dimensional interval pattern is a tuple of intervals h[li , ui ]ii∈[1,...n] ,
where li , ui ∈ Mi ⊂ R, li ≤ ui and Mi is an ordered finite set of modalities (i.e. each Mi is a set of feature values, i.e. Dom(fi ), or a subset of values Mi ⊂ Dom(fi )). An interval pattern P = h[li , ui ]ii∈[1,...n] covers a tuple
x = {x1 , . . . , xn }, denoted x ⊑ P , iff ∀i ∈ [1, ...n], li < xi ≤ ui .
Let X = h[li , ui ]ii∈[1,...n] and Y = h[li′ , u′i ]ii∈[1,...n] be two n-dimensional
interval patterns. We define X ⊔Y = h[min (li , li′ ) , max (ui , u′i )]ii∈[1,...n] . Further,
X ⊑ Y iff ∀i ∈ [1, ...n], [li , ui ] ⊆ [li′ , u′i ]. This definition extends the previous one
for tuple covering considering a value v as a singleton interval [v, v].
Let X → + be a positive rule where X is an interval pattern. True positives
are positive examples covered by the rule: T P (X) = {e|e ∈ P ∧ e ⊑ X}.
False positives are negative examples covered by the rule: F P (X) = {e|e ∈
N ∧ e ⊑ X}. True negatives are negative examples not covered by the rule:
T N (X) = {e|e ∈ N ∧ e 6⊑ X}. supp(X), the support of pattern X is defined
as supp(X) = |{e|e ∈ E ∧ e ⊑ X}|. We also define supp+ (X) = |T P (X)| and
supp− (X) = |F P (X)|. supp+ is antimonotone w.r.t the ⊑ relation and supp− is
monotone w.r.t ⊑. This means that ∀X, Y, X ⊑ Y, supp+ (X) ≤ supp+ (Y ) and
supp− (Y ) ≤ supp− (X).
The learning task consists in constructing all interval patterns X such that
supp+ (X) > minsup and supp− (X) < maxf p where minsup and maxf p are
given parameters.
From the practical point of view of data mining algorithms, closed patterns
are the largest patterns (w.r.t. a partial order ⊑ on the set of patterns, denoted
P) among patterns occurring in the exact same set of examples. Formally, a
set X ∈ P is closed when there is no other set Y ∈ P such that X ⊏ Y
(i.e. Y ⊑ X ∧ Y 6= X) and supp(X) = supp(Y ). Closed patterns are interesting
because they carry the same information as the total set of frequent patterns.
Kaytoue et al. [3] have investigated the problem of mining frequent closed
interval patterns with Formal Concept Analysis (FCA). They proposed the MinIntChange algorithm which enumerates all frequent closed frequent patterns.
It starts from the most generic interval pattern that covers all the examples:
IP = h[min (Mi ) , max (Mi )]ii∈1...n . Then, each interval pattern is specialized
applying minimal changes on the left or on the right of the interval.
3
Mining relevant interval-rules
The theory of relevancy, described in [4], aims mainly at reducing the hypothesis
space by eliminating irrelevant features. This theory has been used by Garriga
et al. [2] to extract relevant features in example database where an example
is a tuple of symbolic features. Here, we extend the definition of relevancy of
Garriga et al. [2] to the relevancy of interval patterns. First, we define two
closure operators, Γ + and Γ − , that respectively stand for the closure of interval
pattern on P (positive examples) and on N (negative examples).
Definition 1 (Relevancy of an interval pattern) Let X and Y be two interval patterns. X is more relevant than Y iff Γ + (Y ) = Γ + (X ⊔ Y ) and Γ − (X) =
Γ − (X ⊔ Y ).
Thus, similar results as those of Garriga et al. [2] can be deduced about the
characterization of the space of relevant interval patterns.
Theorem 1. Let X and Y be two interval patterns. If Γ + (Y ) = X and Y 6= X
then Y is less relevant than X.
Theorem 2. Let X and Y be two different closed interval patterns such that
X ⊏ Y . Then, we have that Y is less relevant than X iff Γ − (X) = Γ − (Y ).
Algorithm 1 Closed interval rule mining algorithm. P is the set of positive
examples, N is a set of negative examples and M is the set of modalities.
1: F CIP ←MinIntChange(P , M)
2: for (X, Y ) ∈ F CIP do
3:
if F P (X) = F P (Y ) and X ⊏ Y then
4:
F CIP ← F CIP \ {Y }
5:
end if
6: end for
The first theorem shows that the relevant rules X → + are those for which
the interval pattern X is closed over the positive examples. According to the
second theorem, in case of similar negative supports, the interval pattern with
largest intervals is preferred. Proofs for Theorems 1 and 2 may be deduced from
proofs on features sets [2].
Algorithm 1 is based on these theorems to extract the relevant interval patterns. The first step of the algorithm is to extract F CIP , the set of frequent
interval patterns closed over the positives. Then, line 3 prunes irrelevant patterns in accordance with Theorem 2. For any closed interval pattern Y ∈ F CIP ,
if there exists another closed interval pattern X such that both have the same
support in the negatives (i.e. same number false-positives) and such that X ⊏ Y
then Y is removed.
The size of the interval patterns search space is O m2×n where n is the
number of features and m is the number of modalities Mi of one attribute. Thus,
we are facing a memory usage constraint. Keeping all the frequent concept in
memory require a large memory. This memory issue is classically encountered in
formal concept analysis but it becomes harder when the number of modalities
increases.
To tackle the issue of memory usage, we reduce the modalities to a subset Mi
of a fixed maximal size, defined by parameter eqmod. The overall rule mining
algorithm has not to be modified. There are several methods to reduce the
number of modalities. We choose to extract the equi-probable intervals from the
positives examples.
4
Implementation and results
We evaluated our algorithm on three UCI datasets [1] (Haberman, Iris and Vertebral column). The algorithm is implemented in C++. Experiments are conducted
on an Intel Core-I5 with 8Go of RAM with Linux system.
For all experiments in this section, f pmax = 10% and eqmod = 10. Figure 1
illustrates the number of closed interval patterns in positive examples, the number of frequent and accurate rules; and the number of relevant interval pattern
rules. We can see that the computing times (see Figure 2) are strongly correlated
to the number of patterns.
Fig. 1. Number of closed interval patterns in positives, number of rules satisfying
minsup and maxf p; and number of relevant interval-rules w.r.t. minimal support.
Fig. 2. Computing time (in millisecond) w.r.t. minimal support.
Even for small data such as the Iris dataset, the number of patterns is high
for low thresholds (≈ 3000) but the number of relevant patterns is significantly
lower than the total number of closed rules. Moreover, the number of patterns
increases exponentially with the number of modalities.
5
Conclusions
We have presented a new algorithm for extracting relevant rules from a numerical
dataset. It offers a wider choice of possibly interesting rules for experts. The
number of extracted patterns is high but more representative of the input dataset
whereas standard algorithms such as CN2 or Ripper select a priori a very limited
set of rules simply based on covering and accuracy criteria. Future work will be
devoted to proposing additional selection criteria which enable the expert to
express his/her preferred set of relevant rules.
References
1. K. Bache and M. Lichman.
UCI machine learning repository.
http://archive.ics.uci.edu/ml, 2013.
2. Gemma C. Garriga, Petra Kralj, and Nada Lavrač. Closed sets for labeled data.
Journal of Machine Learning Research, 9:559–580, 2008.
3. Mehdi Kaytoue, Sergei O. Kuznetsov, and Amedeo Napoli. Revisiting numerical
pattern mining with formal concept analysis. In Proceedings of International Join
Conference on Artificial Intelligence (IJCAI), pages 1342–1347, 2011.
4. Nada Lavrač and Dragan Gamberger. Relevancy in constraint-based subgroup
discovery. In Jean-François Boulicaut, Luc Raedt, and Heikki Mannila, editors,
Constraint-Based Mining and Inductive Databases, volume 3848 of Lecture Notes in
Computer Science, pages 243–266. Springer Berlin Heidelberg, 2006.
| 2cs.AI
|
An Integrated and Scalable Platform for Proactive
Event-Driven Traffic Management
Alain Kibangou
Alexander Artikis
Evangelos Michelioudakis
Georgios Paliouras
Marius Schmitt
John Lygeros
Chris Baber
Natan Morar
Fabiana Fournier
Inna Skarbovsky
arXiv:1703.02810v1 [cs.AI] 8 Mar 2017
Abstract
Traffic on freeways can be managed by means of ramp meters from Road Traffic
Control rooms. Human operators cannot efficiently manage a network of ramp meters.
To support them, we present an intelligent platform for traffic management which
includes a new ramp metering coordination scheme in the decision making module,
an efficient dashboard for interacting with human operators, machine learning tools
for learning event definitions and Complex Event Processing tools able to deal with
uncertainties inherent to the traffic use case. Unlike the usual approach, the devised
event-driven platform is able to predict a congestion up to 4 minutes before it really
happens. Proactive decision making can then be established leading to significant
improvement of traffic conditions.
1
Introduction
Congestion can be defined as a situation when traffic is moving at speed below the designed
capacity of a roadway [9] or as a state of traffic flow on a transportation facility characterized
by high densities and low speeds, relative to some chosen reference state [3]. It results
of various root causes (e.g. traffic incidents, work zones, weather, special events, physical
bottlenecks), often interacting with one another [9] and induces excess delays, reduced safety,
and increased environmental pollution due to stop-and-go behaviour. One approach to tackle
congestion could be to increase the capacity of the traffic infrastructure by constructing new
roads. This approach is very costly and it is often not possible due to societal constraints
as citizens are more and more aware of environment protection. The solution is then to
control traffic in order to avoid, reduce or at least postpone congestion. To do so, most
cities in the world have taken important decisions to invest in road sensor capabilities to get
measurements of traffic parameters and to build modern Road Traffic Control rooms where
traffic operators monitor the traffic situation based on video images and measurements from
loop detectors and wireless magnetic sensors, for instance [18]. Actions to control traffic are
two-fold: manage ramp metering and/or change speed limits according to the current traffic
status [27]. We will focus on ramp metering which is the most common regulation policy.
As stated above, existing traffic management platforms are mainly based on human
operators and noisy data arriving from various sensors [18]. They are based on the paradigm
sense-respond. In contrast, here, we consider a sense-recognise-forecast-decide-act-explain
paradigm where decisions are triggered by forecasting events, whether they correspond to
problems or opportunities, instead of reacting to them once they happen. We present
SPEEDD (Scalable Proactive Event-Driven Decision Making), an integrated platform for
1
proactive event-driven decision-making and demonstrate its capabilities to be resilient to
the inherent uncertainty of the sensor readings, which include incomplete data streams,
erroneous data and imprecise definitions of the events that need to be detected and/or
forecasted. The following steps are to be considered. First, data are continuously acquired
from various types of sensor and fused in order to recognise, in real-time, events of special
significance. Second, the recognised events are correlated with historical information to
forecast congestion that may take place in the near future. Third, both forecasted and
recognised events are leveraged for real-time operational decision-making. Fourth, visual
analytics [34] prioritise and explain possible proactive actions, enabling human operators
to reach and execute the correct decision. The novelty of the proposed platform lies in
the difficult task of on-the-fly, low-latency processing of large, geographically distributed,
noisy event streams and historical data, for recognising and forecasting congestion, making
decisions to reduce the impact of the congestion, and explaining the decisions to human
operators in order to facilitate correct decision execution.
Operators in Road Traffic Control rooms have to monitor several on-ramps and actions
are in general restricted to the identified bottleneck. In absence of coordination of all the
ramp meters, the on-ramp immediately upstream of a bottleneck will solely attempt to
prevent a congestion forming at the bottleneck. This local control often results in a quick
growth of the queue length on the ramp. Then, to avoid an unacceptable spill of the
congestion in the nearby urban area, the metering action needs to be limited, resulting in a
congestion starting at the bottleneck and propagating upstream. By contrast, coordination
between the ramps allows to distribute the control burden onto multiple ramps, thereby
preventing ramp overflow without causing a congestion on the mainline. The main challenge
is to determine when it is necessary to use on-ramps to hold vehicles back in the queue
and reduce traffic demand from a downstream bottleneck. Such an action necessarily has
to happen in a proactive way, since any effects of ramp metering travel downstream at
most with the free-flow speed. SPEEDD supports a hierarchical coordination scheme with
predictions made by means of complex event processing tools. The hierarchical coordination
scheme decomposes the controller into distributed, local feedback loops, and a high-level
coordination scheme based on events. Unlike existing approaches [26, 25], it is based on the
optimality analysis of decentralized ramp metering carried out in [31].
The remainder of this paper is organized as follows: in Section 2 we describe the scenario
under study. Then, following the sense-recognise-forecast-decide-act-explain paradigm we
describe the methods for event recognition and forecasting in Section 3, decision making in
Section 4, and dashboard design in Section 5. Based on the developed method, we describe
the integrated prototype in Section 6 and evaluate its performance in Section 7. Finally in
Section 8 we propose directions for future work and conclude.
2
Scenario description
We consider the Grenoble South Ring road, in France, as the case study. This freeway links
the city of Grenoble from the north-east to the south-west. In addition to sustaining local
traffic, it has a major role since it connects two highways: the A480, which goes from Paris
and Lyon to Marseille, and the A41, which goes from Grenoble to Switzerland. Moreover,
the mountains surrounding Grenoble prevent the development of new roads, and also have a
negative impact on pollution dispersion, making the problem of traffic regulation on this road
even more crucial. From the Road traffic Control room of DIR CE1 , operators can monitor
1 Direction
Interdépartementale des Routes Centre-Est
2
traffic on the road network, including the South Ring road, though hundreds of CCTV
cameras, verbal reports (primarily from traffic operators on the road but also from police or
other emergency service personnel), emails and other text messages, and data from sensors
placed on the road. They can also effect traffic though variable message signs (VMS) and
soon they will be involved in management of ramp metering, which is still under deployment.
DIR CE is also a partner of the GTL (Grenoble Traffic Lab), which offers a dense network of
wireless magnetic sensors (see Fig. 1 and [6] for a full description of the sensing platform).
GTL also provides a microscopic calibrated simulator of the Grenoble south ring where the
dynamics of each each vehicle in the road are simulated. This simulator has been developed
using the AIMSUN platform2 . It gives the opportunity to test the entire system in closed
loop, from sensing to actuation; which is not possible with the actual freeway. In addition,
synthetic data produced by the simulator have annotations that can serve as baseline in
order to test the effectiveness and efficiency of the developed system.
Figure 1: Grenoble South Ring Network: the road is divided in 45 cells numbered from
east to west; Cells equipped with sensors have an S symbol; nodes, marked with N , are
constituted by on-ramps (blue arrow) or off-ramps (red arrow).
Our objective is to detect congestions a few minutes before they happen. So, proactive
suggestions to traffic operators can be provided, or automatic actions can be carried out, to
alleviate the forecasted congestion. The following sections describe the methods allowing us
to reach this objective.
3
Event-driven Congestion detection and forecasting
The detection of congestions is based on information received from the sensors. Special
behaviors of variables that describe the system such as speed, density, occupancy allow to
infer the existence of congestion. In what follows, we adopt a Complex Event Processing
2 http://www.aimsun.com/
3
(CEP) approach, sometimes called event stream processing, which is a method that combines
data from multiple sources for tracking and analyzing (processing) streams of information
(data) to infer events or patterns that suggest more complicated circumstances. The goal of
complex event processing is to identify meaningful events (such as opportunities or threats)
and respond to them as quickly as possible [21]. In general, there exist two methods to
define the rule patterns for a CEP application: machine learning and domain experts. In
the first case, the patterns are learnt automatically by a computer program, while in the
second, they are given by an external entity; usually a subject expert matter specialized
in the domain. It is also possible to combine these two methods. Historical data used at
design time contain raw events reported during the observed period along with annotations
provided by domain experts. These annotations mark important situations that have been
observed in the past and should be detected automatically in the future. Due to the dynamic
nature of the proactive traffic management application, the knowledge base of event pattern
definitions may require to be refined or enhanced with new ones.
3.1
Machine learning for event definitions
In order to effectively learn definitions for traffic congestion using sensor data, we have developed OSLα [22], an online structure learner for Markov Logic Networks (MLNs) [30]. OSLα
extends the procedure of OSL [16] by exploiting a given background knowledge to effectively
constrain the space of possible structures during learning. The space is constrained subject
to characteristics imposed by the rules governing a specific task, herein stated as axioms.
As a background knowledge we make use of MLN−EC [32], a probabilistic variant of the Event
Calculus [19, 23] for event recognition. Fig. 2 presents the components of OSLα.
OSLα
Learnt Hypothesis Ht :
0.4 HoldsAt(congestion(lid), t+1) ⇐
HappensAt(fast Slt20(lid), t)∧
HappensAt(fast Ogt45(lid), t)
Data Stream/Training Examples
...
Hypergraph
Inference
+
Paths to
Clauses
MLN−EC Axioms:
HoldsAt(f, t+1) ⇐
InitiatedAt(f, t)
HoldsAt(f, t+1) ⇐
HoldsAt(f, t) ∧
¬TerminatedAt(f, t)
¬HoldsAt(f, t+1) ⇐
TerminatedAt(f, t)
¬HoldsAt(f, t+1) ⇐
¬HoldsAt(f, t) ∧
¬InitiatedAt(f, t)
Weight
Learning
Clause
Evaluation
Micro-Batch Dt
HappensAt(fast Slt25(53708), 99)
HappensAt(fast Ogt55(53708), 99)
HappensAt(slow Slt15(53708), 99)
HappensAt(slow Ogt65(53708), 99)
Next(99, 100)
HoldsAt(congestion(53708), 100)
...
...
Micro-Batch Dt+1
HappensAt(fast Sgt70(53708), 200)
HappensAt(fast Olt25(53708), 200)
HappensAt(slow Sgt40(53708), 200)
HappensAt(slow Olt18(53708), 200)
Next(200, 201)
¬HoldsAt(congestion(53708), 201)
...
...
Figure 2: The procedure of OSLα.
The background knowledge consists of the MLN−EC axioms (i.e., domain-independent
rules) and an already known (possibly empty) hypothesis (i.e., set of clauses). Each axiom contains query predicates HoldsAt ∈ Q that consist of the supervision and template
predicates InitiatedAt, TerminatedAt ∈ P that specify the conditions under which a complex event starts and stops being recognized. The latter form the target complex event
definitions that we want to learn. OSLα exploits these axioms in order to create mappings
of supervision predicates into template predicates and search only for explanations of these
4
template predicates. Upon doing so, OSLα does not need to search over time sequences,
instead only needs to find appropriate bodies over the current time-point for the following
definite clauses:
InitiatedAt(f , t) ⇐ body
TerminatedAt(f , t) ⇐ body
At any step t of the online procedure a training example (micro-batch) Dt arrives containing sensor readings, e.g. a fast lane in a highway has average speed less than 25 km/hour
and sensor occupancy greater than 55%. Dt is used together with the already learnt hypothesis to predict the truth values ytP of the complex events of interest. This is achieved by
(maximum a posteriori) MAP inference based on LP-relaxed Integer Linear Programming
[15]. Given Dt OSLα constructs a hypergraph that represents the space of possible structures
as graph paths. Then for all incorrectly predicted complex events the hypergraph is searched,
guided by MLN−EC axioms and path mode declarations [16] using relational pathfinding [29]
up to a predefined length, for definite clauses explaining these complex events. The paths
discovered during the search correspond to conjunctions of true ground atoms and are generalized into first-order clauses by replacing constants in the conjunction with variables. Then,
these conjunctions are used as a body to form definite clauses using as head the template
predicate present in each path. The resulting set of formulas is converted into clausal normal
form and evaluated. The weights of the retained clauses are then optimized by the AdaGrad
online learner [10]. Finally, the weighted clauses are appended to the hypothesis Ht and the
procedure is repeated for the next training example Dt+1 .
3.2
Event-driven approach to forecast congestions in real-time
Event definitions learnt as described above are then used to forecast congestions in realtime by means of an event-driven application which can be defined by an event processing
network (EPN) [12]. An EPN, a conceptual model describing the event processing flow execution, comprises a collection of event processing agents (EPAs), event producers, events,
and consumers. The network describes the flow of events originating at event producers
and flowing through various event processing agents to eventually reach event consumers.
We resort to the IBM PROactive Technology ONline (PROTON3 ) as the CEP engine. In
our scenario, the CEP component receives events emitted from the sensors (producers) every 15 seconds and based on predefined event rules, it alerts in case of a detection of a
possible congestion. In this scenario, the input events are certain (a sensor reading event
happens) but the derived event is not certain (e.g., the fact that we have 15 sensor readings
in 5 minutes that show an increase in the density, doesn’t necessarily imply there will be
a traffic congestion for sure). In other words, the capability to forecast events requires the
inclusion of uncertainty aspects. Proactive event-driven computing deals with the inherent
uncertainty in the event inputs, in the output events, or in both ([1, 36, 11]).
The EPN for the proposed system includes the following EPAs:
• Congestion at a specific location: it exists if the density in a specific location is above
a certain given value (density threshold1) and the speed is below a certain given value
(speed threshold1) for at least 15 input events within the time period of 5 min or until
a ClearCongestion EPA occurs.
3 https://github.com/ishkin/Proton/
5
• ClearCongestion at a specific location: it occurs when a congestion is over, i.e. whenever the density is below a certain given value (density threshold2) and the speed is
above a certain given value (speed threshold2) for at least 15 input events within a
time window that is opened with either a Congestion or a PredictedCongestion events
and is closed after 5 min.
• PredictedCongestion: it occurs when a forecasted congestion is identified at a specific
location. This event pattern or rule is probabilistic in the sense that the output or
derived event has a certainty attribute value associated to it. It is of type TREND,
meaning that it derives an event whenever a specific change (increasing or decreasing)
over time of the density value in the input events is satisfied over a temporal window.
This EPA emits a derived event if at least 5 input events show an increase in the
density in a temporal window which is opened with the first input event that comes
and is closed when either a Congestion or ClearCongestion is detected for the same
location.
• Calculations: they concern calculations on sensor readings (such as averages). They
are emitted to be consumed by the decision making module.
To cope with uncertainties inherent in the event rules, a Sigmoid function is used to calculate
the confidence of the occurrence of a derived event. The idea is: whenever the number of
events in the matching set of the TREND pattern in the PredictedCongestion EPA is high
enough, the certainty of the derived event is close to one.
4
Decision making
In the previous section, we have described how to derive smart rules to recognize and/or
predict a congestion. Now, we will describe a smart way to manage or to avoid congestion. A
cause of congestion is related to an excess of demand of using the road infrastructure. Such
a demand can be managed by means of ramp metering. Given a set of equipped on-ramps,
the objective is to regulate the entering flow in a smart way while avoiding congestion to
spill back to the arterial network. We adopt a hierarchical approach, which decomposes
the controller into distributed, local feedback loops, and a high-level coordination scheme
based on events (see Fig. 3). Existing solutions are used for the local feedback laws but a
new coordination scheme is proposed to deal with non-monotonic effects in traffic dynamics.
In other words, for monotonic traffic, local feedback laws are enough but the coordination
scheme is necessary when non-monotonic effects occur. The non-monotonic behavior makes
the design of optimal controllers difficult, as non-convex problems arise. The main innovation
in our approach lies in the usage of a model-free, data- and event-driven solution to the
problem of ramp coordination.
4.1
Data-driven System Identification
To understand how non-monotonic effects affect the ramp metering problem, it is necessary
to briefly review road traffic dynamics. Freeway traffic conditions at some location can be
described by the traffic density ρ(t), measured in number of vehicles per kilometer, and the
(mainline) traffic flow φ(t). First-order traffic model Lighthill-Whitham-Richards (LWR)
[20] postulates a static flow-density relationship, which is called the fundamental diagram.
It is usual to associate a piecewise-affine fundamental diagram to the LWR model, a shape
confirmed with real data but with significant levels of variance (see Fig. 4). This spread is
6
Complex Event
Processing
all measurements
High-Level
Coordination
High-Level Coordination
Sets traffic light 4488
State Estimator, SysId
Metered Onramp
Metered Onramp
4132
4134
1675
4355
3810
1679
3811
3812
1683
4085
4489
4488
4061
1670
State Estimator,
SysId
Uses measurements from 1683, 1679, 3811,
3812, 4132
Metered Onramp
4381
Low-Level
Controller
Low-Level Controller
State Estimator,
SysId
1666
High-Level
Coordination
Coordinates with traffic lights 4489,
4487, ....
Low-Level
Controller
4487
User Interface
1687
3813
1703
4244
4804
1708
4087
1691
Figure 3: Hierarchical control approach for ramp metering.
most notable at and above the critical density (density value corresponding to the maximal
flow) and it might partially be caused by a dependency of the flow on further variables.
Figure 4: Fundamental diagram: theoretical (left) and reconstructed from real data using
GP regression with mean and 90% confidence interval.
Considering a subdivision of the road in cells, the density of cell k evolves as
ρk (t + 1) = ρk (t) +
T
(φk−1 (t) − φk (t))
Lk
(1)
where T and Lk denote the sampling time and the cell length respectively, whereas the
flow φk (t) = min {dk (ρk (t)), sk+1 (ρk+1 (t))} between to adjacent locations k and k + 1 is
computed as the minimum of the upstream demand dk (ρk (t)) of vehicles that seek to travel
downstream and the downstream supply sk+1 (ρk+1 (t)) of free space. This model is called
Cell Transmission Model (CTM) [8]. In the standard CTM, the flow is non-decreasing in the
upstream density and non-increasing in the downstream density. However, there is empirical
evidence of a capacity drop at a congested bottleneck, that is, the demand function dk (·)
slightly decreases as the upstream density exceeds the critical density. To deal with such
7
capacity drop at critical density, we resort to a 2D representation of the fundamental diagram
(see Fig. 5). We propose to estimate the capacity drop for bottleneck locations offline, using
model-free Gaussian Process (GP) regressions [28] to obtain a data driven estimate of the
two-dimensional fundamental diagram.
Figure 5: GP regression for two-dimensional fundamental diagram where ρds (resp. ρus )
stands for downstream density (resp. upstream density).
4.2
Low-level Control
The main objective of the low-level control is to maximize local traffic flows by shifting
the local traffic density towards the critical density, which is sufficient for close-to-optimal
performance for a monotonic freeway. It can be achieved with the successful ALINEA
algorithm [24], an integral feedback law in which the ideal metering rate is given as
r̃k (t) := r̂k (t − 1) + KI · (ρck − ρk (t)),
where, KI is the integral gain chosen as in [24] and r̂k (t − 1) is the on-ramp flow measured
during the last sampling period. The only road parameter used by the feedback law is
the critical density ρck , which is estimated online from data, as outlined before. However,
the actual metering rate is subject to certain constraints. Obviously, it is non-negative
and upper-bounded by some constant maximal on-ramp flow r̄k . We also allow for a userdefined (see Section 5) lower bound rk , which can be used to limit the maximal waiting time
of drivers on the on-ramp. In addition, the space on the on-ramp is finite and it is paramount
that the queue length qk (t) (in number of vehicles) does not exceed the maximal capacity
q̄k to avoid spill-back of the queue into adjacent arterial roads. Conversely, sometimes it
may be required to hold back a certain amount of vehicles on the on-ramp to ease the traffic
situation downstream, even if no congestion is imminent right at the on-ramp. To this end,
we define a desired queue length 0 < qk∗ (t) ≤ q̄k , which will be chosen by the coordination
algorithm as described in the following section. Thus, the actual metering rate rk (t) is
saturated to the interval
1
1
max rk ,
(qk (t) − q̄k ) + dˆk (t) ≤ rk (t) ≤ min r̄k ,
(qk (t) − qk∗ (t)) + dˆk (t) .
∆t
∆t
Here, dˆk (t) is the prediction of the traffic demand arriving at the on-ramp in the next
time interval and ∆t is the sampling time. The states ρk (t) and qk (t) are estimated from
measurement streams using a standard Kalman filter [17].
8
4.3
Ramp Coordination
The aim of coordinated ramp metering is to target inefficiencies that result from limited
space on the on-ramps in conjunction with the non-monotonic behavior of a congested
bottleneck. In the spirit of the proactive approach of the proposed platform, it is necessary
to predict congestion. As described in Section 3, we predict congestion by learning patterns
from historic, large data sets. Then, efficiency of the coordinated ramp metering scheme
hinges mainly on the accuracy of the predictions made by CEP, while the coordination
algorithm can be described as a simple finite state machine that reacts accordingly. As
control
and
control
active
inactive
inactive
(a) State diagram to determine the ac- (b) Activation of upstream
tive local control algorithm.
coordination.
Figure 6: State diagrams of the coordination algorithm. The symbols and abbreviations are
explained in the text.
depicted in Fig. 6a, the local feedback law described in Section 4.2 either controls the local
density ρ or both the local density and the on-ramp queue length q. In general, control
of the density is activated if a congestion is detected and the queue length is controlled
only if an upstream ramp requests coordination via the Ramp Coordination (RC) event. If
coordination is active between two ramps, we seek to balance the occupancies on both the
∗
ds
upstream (us) and downstream (ds) on-ramp, that is, qds
(t) = q̄q̄us
·qus (t) [26]. Fig. 6b shows
when a downstream ramp will request coordination from an upstream ramp. RC events
are periodically sent if upstream coordination is active. Here, condition (?) is shorthand
notation for (?) := ((TOPC ∨ TOPR ∨ C) ∧ CA) ∨ (q ≥ γ4 · q̄), where the relevant events
are Predicted Congestion (PC), Predicted Ramp Overflow (PR) and Congestion (C). The
Boolean variable CA is true if the local control algorithm is active, that is, it controls ρ, or
ρ and q and the Boolean variables TOP C/P R are used to describe a trade-off with respect to
the total expected travel time as described below. The remaining events Clear Congestion
(CC) and Clear Ramp Coordination 4 are used to determine when control can safely be
deactivated. In both state diagrams in Fig. 6, conditions in bold can be interpreted as the
“default” conditions. If these are comprised of events that might rarely not be predicted/
detected correctly, alternative conditions are also provided as a fall-back solution. The
parameters γ1 = 0.8, γ2 = 0.7, γ3 = 0.7 and γ4 = 0.8 are tuning parameters for defining
thresholds for transitions.
Note that using multiple ramps comes at a cost. More cars are held back on the onramps, resulting in time lost for the drivers if it turns out that congestions would have
been avoided without the use of coordination. Recall that the CEP engine also estimates
4 This
event is implemented as a Ramp Coordination event with particular attribute values.
9
the probability P[E] of a predicted event happening within some time horizon T . It is
used to perform a trade-off between the (potential) benefits of preventing a congestion
and the possibility of wasting driver’s time on the on-ramps. To do so, we compare the
additional waiting time ∆Tramp on the upstream on-ramp to the time potentially wasted
in congestion on the mainline ∆Tml
and define the Boolean trade-off variable TOE :=
0
0
P[E] · ∆Tml
> (1 − P[E]) · ∆Tramp
. Here, the event E is either PC or PR, as both affect
condition (?) in the same way. For a pair of ramps, the additional waiting time can be
bounded by
T /∆t
∆Tramp = ∆t ·
X
qus (t) ≤ T ·
t=0
q̄ds
0
q̄us =: ∆Tramp
,
q̄ds + q̄us
for a situation in which coordinated ramp metering was ultimately unnecessary and therefore, the total amount of cars stored on both ramps is less than the space available on
the downstream ramp qus (t) + qds (t) ≤ q̄ds , for all t ∈ [t, t + T ]. Conversely, if a congestion does arise within a time horizon T , the average, surplus demand is equal to at least
∆d ≥ (l · (ρcds − ρds (t)) + q̄ds − qds (t))/T . If coordination is used, the surplus demand can
us
. Therefore,
also be stored on the upstream on-ramp, delaying the congestion by ∆T = q̄∆d
the additional time spent in congestion can be bounded by
∆Tml ≥ ∆T · ∆φ · Tcon =
q̄us · T
0
· ∆φ · Tcon := ∆Tml
.
l · (ρcds − ρds (t)) + q̄ds − qds (t)
with Tcon the expected duration of the congestion and ∆φ the bottleneck capacity drop,
which have to be estimated from historic data. Note that the inequality used in the tradeoff provides a sufficient, but not a necessary condition for efficiency of coordination because
0
0
can be computed. This is less restrictive than it might
only the bound ∆Tml
and ∆Tramp
seem at first, since CEP will continue to produce updated predictions as traffic conditions
evolve. Therefore, adopting a conservative approach at worst delays the usage of coordination slightly, until congestion can be predicted with sufficient accuracy.
5
Dashboard design
While the automated event forecasting and decision making can address challenges relating
to congestion and ramp metering, operators in the Road Traffic Control room, such as
Grenoble’s DIR CE are required to monitor and manage many other aspects of the road
network. Consequently, it was essential to integrate the output from the SPEEDD system
into a visualisation which supported these other aspects of their work. The goals of the Road
Traffic Control room is to: maximize the available capacity of the road system, minimize the
impact of incidents, manage demand regulation and congestion, assist in emergency service
response, maintain public confidence in messages displayed on the VMS, and maintain a
record of actions and events on the road system.
Design of the dashboard for SPEEDD began with visits to DIR CE in which we spoke
with and observed operators at work. In addition to such observations, we were able to
conduct eye-tracking studies in order to explore which information sources were most useful
to the operators and what their information handling strategies were [33, 5]. Combining the
material collected during field studies, we developed a set of Cognitive Work Analysis [35]
descriptions which provided different views on the operators’ activity. A dashboard which
was implemented in the SPEEDD architecture and which resulted directly from our understanding of operators’ decision making, information use and communications was produced
10
as the first version of the dashboard. This was presented to operators and its usability
explored. From this evaluation, a revised dashboard, depicted in Fig. 7, was developed. A
detailed description of the design process and decisions made in the development process of
the design can be found in [2].
Figure 7: Developed traffic monitoring dashboard.
Fig. 7 consists of three main areas. In the top right of the screen, an event list provides
a time-ordered set of outputs from the SPEEDD congestion prediction algorithm. This
highlights to the operators event which need to be managed. Some of the events will be
automatically handled by the ramp-metering and, as such, need not be brought to the
operators’ attention. When ramp-metering is implemented, the status of the ramp is changed
in the bar charts associated with each ramp on the map (on the left of the screen). The map
shows the Grenoble South Ring and, through colour-coding of the discrete segments of the
road, indicating the current level of congestion. The bar charts at each ramp also indicate
the traffic flow through the ramp and the congestion adjacent to the ramp. In the centre of
the map, one can see a small camera icon. This indicates the site of the CCTV camera which
has produced the largest image in the CCTV panel (on the bottom right of the screen). The
CCTV panel shows a collection of images from the CCTVs which the operators can use to
diagnose the level and possible causes of congestion. The aim of the dashboard design was
to produce a clear and simple overview of the road system, in order to allow operators to
maintain a high level of situation awareness, and to allow them to see the decisions that
the automated system was enacting (both in terms of the changes to ramps and display of
congestion on the road, and in terms of those aspects of congestion which were not handled
directly by ramp metering). For the events in the event list, the operator will indicate what
action was taken, e.g., in terms of calling up VMS messages to alert drivers to congestion
or to advise them to decrease their speed. The operator actions can be combined with the
initial changes in ramp metering, e.g., ramp metering initiated at a given time, in order to
produce a detailed log of operations in the control room.
11
6
Proactive Event-driven system architecture
Fig. 8 shows the event-driven architecture run time represented as a group of loosely-coupled
components interacting through events. The event bus serves as the communication and
integration platform for the run time. In general, input from the operational systems (traffic
Figure 8: Run time architecture of the SPEEDD system.
sensor readings) are represented as events and injected into the system by posting a new event
message to the event bus. These events are consumed by complex event processing (CEP).
The derived events representing detected or forecasted situations that the CEP component
outputs are posted to the event bus as well. The decision making module listens to these
events so that the decision making procedure is triggered upon a new event representing a
situation that requires a decision. The output of the decision making represents the action
to be taken to mitigate or resolve the situation. These actions are posted as action events.
The visualization component (dashboard) consumes events coming from two sources: the
situations (detected as well as forecasted) and the corresponding actions suggested by the
automatic decision components. The user can accept the suggested action as it is, modify
the suggested action’s parameters, or reject it (and even decide upon a different action). In
the case where an action is to be performed, the resulting action will be sent as a new event
to the event bus so that the corresponding actuators are notified. The full description of
the proposed proactive event-driven architecture can be found in [13].
12
The proposed event-driven architecture can be run in an open, closed, or hybrid loop
mode. In the current scenario, operators interact with the outputs of the prototype through a
dashboard. The dashboard client communicates, via the dashboard server, with the modules
of the architecture. Operators can accept, respond to, or make suggestions and control
actions. Actions taken by operators via the dashboard are fed back into the run time as
events, thus allowing for the seamless integration of expert knowledge and the outputs of
complex algorithms.
It is worth noting that the Machine Learning for event definitions is not part of the run
time architecture described in Fig. 8. The automated construction of traffic congestion
patterns (see Section 3) is performed at design time.
7
Evaluation
In this section, we first evaluate our approach of learning traffic event definitions using
Machine Learning techniques and then the different components of the proposed run time
platform.
7.1
Learning Event Definitions
We applied OSLα (see Section 3.1) to traffic management using real data from the magnetic
sensors mounted on the Grenoble South Ring, consisting of approximately 3.3GiB of sensor
readings (one month data). Annotations of traffic congestion are provided by human traffic
controllers, but only very sparsely. To deal with this issue, we also used a synthetic dataset
generated by the traffic micro-simulator of GTL (see Section 2). The synthetic dataset
concerns the same location and consists of 6 simulations of one hour each (≈ 18.6MiB).
A set of first-order logic functions is used to discretize the numerical data (speed, occupancy) and produce input events such as, for instance, HappensAt(fast Slt55(53708), 100),
representing that the speed in the fast lane of location 53708 is less than 55 km/hour at
time 100. The total length of the training sequence in the real data case consists of 172, 799
time-points, while in the synthetic data it consists of 238 time-points. The evaluation results
were obtained using MAP inference [15] and are presented in terms of F1 score. In the real
dataset, all reported statistics are micro-averaged over the instances of recognized CEs using
10-fold cross validation over the entire dataset, using varying batch sizes. At each fold, an
interval of 17, 280 time-points was left out and used for testing. In the synthetic data, the
reported statistics are micro-averaged using 6-fold cross validation over 6 simulations by
leaving one out for testing.
Fig. 9 presents the experimental results on the real dataset. We compare OSLα against
the AdaGrad online weight learner [10] that optimizes the weights of a manually constructed
traffic congestion definition. The predictive accuracy of the learned models, both for OSLα
and AdaGrad, is low. This arises mainly from the largely incomplete supervision. In OSLα,
the predictive accuracy increases (almost) monotonically as the learning steps increase. On
the contrary, the accuracy of AdaGrad is more or less constant. OSLα outperforms AdaGrad
in terms of accuracy. (OSLα achieves a 0.64 F1 score, while the best score of AdaGrad is
0.59.) This is a notable result. The absence of proper supervision penalizes the hand-crafted
rules, compromising the accuracy of AdaGrad that uses them. OSLα is not penalized in this
way, and is able to construct rules with a better fit in the data, given enough learning steps.
For some locations of the highway, OSLα has constructed rules with different thresholds for
speed and occupancy than those of the hand-crafted rules. With respect to efficiency (see
the right diagrams of Fig. 9), unsurprisingly AdaGrad is faster and scales better to the
13
avg. batch processing (seconds)
0.65
0.6424
F 1 score
0.6
0.55
0.5281
0.5565
0.56
0.5
0.45
0.4
50
40
30
5
4
2.43
2
0
10
0.6714
1000
0.5986
0.5788
0.55
0.5
50
40
3000
30
10
1000
1000
batch size
(#sensor events)
10
8
6
4
2
0
10
0.23
0.241
0.421
30
40
batch size (minutes)
#batches
0.532
20
2000
20
2000
50
#batches
0.6
0.5844
4000
3000
batch size (minutes)
avg. batch processing (seconds)
10
5000
20
30
0.65
F 1 score
8.84
6
40
2000
20
batch size (minutes)
8
3000
batch size (minutes)
0.5714
10
50
1000
2000
3000
4000
5000
batch size
(#sensor events)
Figure 9: Real dataset: F1 score (left) and average batch processing time (right) for OSLα
(top), and AdaGrad operating on manually constructed traffic congestion rules (bottom).
In the left figures, the number of batches (see the Y axes) refers to number of learning steps.
14
increase in the batch size. At the same time, OSLα processes data batches efficiently, much
faster than their duration. For example, OSLα takes less than 9 sec to process a 50-minute
batch including 4, 220 sensor readings.
0.9
0.88
0.9
0.8957
0.8938
0.88
0.86
0.8785
0.84
F 1 score
F 1 score
0.886
0.8531
0.8269
0.82
0.8995
0.8979
0.8941
0.8903
0.8577
0.8747
0.86
0.84
0.82
0.8
30
20
10
batch size (minutes)
0
0
5
10
15
20
25
0.8
30
20
10
batch size (minutes)
#batches
0
0
5
10
15
20
25
#batches
Figure 10: Synthetic dataset: F1 score for OSLα (left) and AdaGrad operating on manually
constructed traffic congestion rules (right).
To test the behavior of OSLα under better supervision, we made use of a synthetic
dataset produced by the traffic micro-simulator of GTL. Fig. 10 presents the experimental
results. Not surprisingly, the predictive accuracy of the learned models in these experiments
is much higher as compared to real dataset. Moreover, the accuracy of OSLα and AdaGrad
is affected mostly by the batch size: accuracy increases as the batch size increases. The
synthetic dataset is smaller than the real dataset and thus, as the batch size decreases, the
number of learning steps is not large enough to improve accuracy. The best performance
of OSLα and AdaGrad is almost the same (approximately 0.89). In other words, OSLα can
match the performance of techniques taking advantage of rules crafted by human experts.
This is another notable result.
7.2
Event forecasting
In order to explore the quality of our CEP module, we ran a test comprising of 20 simulations
generated by the traffic micro-simualtor of GTL along with annotations of congestions. The
annotations of congestions include the location and the time the congestion is detected. First,
we evaluated the quality of our Congestion pattern against the annotated data. We checked
the proportion of detections by our EPA that were annotated in the data as congestions
(precision) and second, the proportion of congestions we were able to detect out of all
the annotated congestions (recall). In all our simulations our precision was 100%, while
the average recall over all the simulations was 72%. This can be easily explained: the
rule implemented has been given to us by the domain expert, who is the one to identify the
congestions in the simulations, thus giving a perfect precision. However, when implementing
the pattern we applied a “stricter” criterion for the rule than the one in the simulator: we
took into account not just the average speed critical thresholds, but also density thresholds,
therefore we have a less success rate in the recall of the results, i.e., there were annotations
of congestion in the data that we “missed”.
As a second step, we aimed at checking a more interesting question, that is, whether the
inclusion of uncertainty aspects enables us to predict a congestion in the highway before it
reaches critical thresholds, as opposed to detecting it once it happens. We addressed this
question by having two EPNs, once including uncertainty aspects and the other one without
uncertainty, i.e. deterministic; and running the tests twice, one time for each EPN (with and
15
without uncertainty). This is a common approach in CEP engines dealing with uncertainty
(see for example in [7]). The deterministic case served as baseline, as we knew at this stage
that all our congestions have been detected correctly. The precision of our results indicates
the proportion of congestions we were able to predict (in other words, PredictedCongestion pointed out correctly to a congestion), whereas the recall indicates the proportion of
congestions we were able to detect out of all the annotated congestions (in other words,
PredictedCongestion pointed out correctly out of all congestions). We used a threshold of
0.6 in the certainty attribute to determine whether to consider PredictedCongestion as a
congestion. In other words, only PredictedCongestion alerts with a certainty value larger
than 0.6 were considered in our calculations of precision and recall. In these tests, the average precision was 91% and the average recall was 75%. Furthermore, PredictedCongestion
event is emitted 3 to 4 minutes before a Congestion is detected, thus enabling the system to
take proactive actions in order to alleviate these congestions. The recall average indicates
that there are other situations that cause congestions which are not detected by our pattern.
Further analysis shows that these situations are characterized by “jumping data”, meaning,
the values of speed and density tend to jump thus not satisfying the increasing build-up
which is required in our pattern. We are currently investigating these “jumping” cases to
see if we can identify some common behavior/pattern.
7.3
Decision making
We evaluated the decision making module by considering that the ramps with indices k ∈
{2, 6, 7, 8, 9} as depicted in Figure 1 are used for ramp metering. We assume that onramp queues are extended to provide storage space for up to 50 cars each. The simulation
is conducted as described in [31] with non-monotonic demand functions. To quantify the
benefits of ramp metering, we use the Total Time Spent (TTS), a standard metric defined as
the sum of the travel times of all cars for a certain day. We perform three types of simulations.
First, we simulate traffic without ramp metering to obtain a baseline performance, TTSol .
Second, simulations using local ramp metering as described in Section 4.2 are performed,
but no coordination between ramps is used. The corresponding travel time is denoted
TTScl . Third, we employ coordinated ramp metering with the coordination along the lines
of Section 4.3 and denote the corresponding total time spent as TTSco . The parameters of
the coordination are chosen as γ3 = 0.1 and γ4 = 0.2. For the five-week period, we obtain
relative savings of
TTSol − TTScl
= 9.9%
TTSol
and
TTSol − TTSco
= 13.6%.
TTSol
Benefits of coordination tend to increase as traffic demand increases, while conversely, no
benefits are obtained on days with no or only light congestion for an uncontrolled freeway.
However, TTS does not only quantify time wasted in congestion and in on-ramp queues,
but vehicles traveling at free-flow velocity contribute significantly as well. Ramp metering
cannot provide any benefits during times at which the uncontrolled freeway is not congested.
Therefore, we define the Total-Free-flow-Time TFT as the travel time accumulated by all
vehicles on a hypothetical freeway, that is always uncongested, i.e. all vehicles travel at
free-flow velocity at all times. The relative savings in terms of time wasted in congestion
and in on-ramp queues for all days amount to
TTSol − TTScl
= 25.2% and
TTSol − TFT
16
TTSol − TTSco
= 34.6%.
TTSol − TFT
2
4
6
8
10
12
14
16
18
20
2
onramp index
cell index
The results are visualized in Figure 11 for one day that provides average savings5 . Note
that these savings are larger than the ones reported in [31] and [14]. This is no surprise,
since these papers use the monotonic CTM. In a monotonic setting, the only benefit of ramp
metering is to avoid blocking off-ramps with spilled-back mainline congestion. It shall be
noted however that both the time spent in congestion and the relative savings of coordination
seem to be sensitive to the traffic demands. In a non-monotonic setting, small changes in
demands may cause large differences in open- or closed-loop behavior. As stated earlier,
coordination tends to provide larger relative savings for more severe congestion.
11
14
16
19
6:00
9:00
12:00
15:00
18:00
6:00
9:00
12:00
15:00
18:00
time
time
(a) Density for local control.
(b) Queue occupancies for local control.
2
4
6
8
10
12
14
16
18
20
2
onramp index
cell index
7
7
11
14
16
19
6:00
9:00
12:00
15:00
18:00
6:00
9:00
12:00
15:00
18:00
time
time
(c) Density for coordinated control.
(d) Queue occupancies for coordinated control.
Figure 11: Simulation results for traffic demands of April 19th , 2014. Coordination distributes vehicles among on-ramps, thereby reducing traffic on the mainline and increasing
the bottleneck flows, in particular for cell 11 and 19.
7.4
Evaluation of dashboard design
As part of the usability evaluation of the dashboard design, participants (N = 24) completed
a series of tasks using the original dashboard and the revised version which was described
in Section 5. Participants were asked to complete the Software Usability Scale (SUS) questionnaire [4] for each dashboard. The SUS questionnaire consists of 10 simple questions
concerning the potential usefulness and benefit that users feel that the dashboard might
provide them with. Each statement is rated on a scale of 0 to 4. The scoring of responses
then involves subtracting 1 from odd-numbered questions and subtracting scores of evennumbered questions from 5. This is because the questions alternate between positive and
negative connotations. Scores are then summed and multiplied by 2.5, to give a final score
5 April
19th , 2014:
TTSol −TTScl
TTSol
TFT
= 21.2%,
TTSol −TTSco
TTSol −TFT
17
= 33.8%
out of 100. As a rule of thumb, scores in excess of 65 are deemed “acceptable”. Fig. 12
compares the evaluation of the dashboard described in section 5 with the original version:
the median scores were 49 for the original version (indicating that the design was of lower
than acceptable usability) and 69 for the revised version (indicating that participants felt
the design to be acceptable).
Figure 12: Comparison of the original dashboard (one) with the revised dashboard (two) in
terms of subjective rating of usability.
In addition to collecting subjective opinion of the usability of the two dashboard designs,
an experiment was conducted in which 24 participants each completed 60 ramp metering and
traffic congestion tasks (e.g. task involving decisions on whether to alter the ramp metering
rate and whether level of congestion has changed on different sections of the road). The tasks
were completed with each dashboard and under different levels of automation reliability (low
= 20%; medium = 50% and high = 80%). This latter condition was introduced to explore
how users might respond to recommendations which were based on noisy or incomplete
data (hence resulting in erroneous advice). Analysis of decision time showed that responses
were significantly faster with dashboard 2 (mean decision time approximately 14.25s for
dashboard 1 and 10.5s for dashboard 2) and also varied with reliability (mean decision time
approximately 13.0s for low; 12.75s for medium; 11.5s for high). In terms of decisions,
users were able to match the reliability of the automation, i.e. when the reliability levels
were low and medium, users would only respond to the “correct” answers and were able
to compensate for the errors to some extent. However, decision accuracy for the low and
medium reliability automation was ≤ 85%, and for the high reliability was around 92%.
Thus, when the automation performed poorly, human decision making could be affected.
18
8
Conclusions
We have presented an intelligent platform for traffic management which includes a new
ramp metering coordination scheme in the decision making module, an efficient dashboard
for interacting with human operators, machine learning tools for learning event definitions
and complex event processing tools able to deal with uncertainties inherent to the traffic
use case. It has been shown that the developed machine learning tool can match the performance of techniques taking advantage of rules crafted by human experts while complex
event processing tools are able to predict congestion 3 to 4 minutes before Congestion happens even with uncertain and noisy data. The decision making module using coordinated
ramp metering improves Total Spent Time compared to ramp metering without coordination using the current standard local feedback algorithm ALINEA. It is worth noting that
even though the system is able to take proactive actions in order to alleviate congestions,
the recall average indicates that there are other situations that cause congestions which are
not detected by our patterns. Hence, there remains a need to ensure an integrated humanautomation decision system as we have implemented in SPEEDD. Future works also include
analysis and definition of patterns exhibiting jumps rather than trends.
Acknowledgments
This work is part of the EU-funded SPEEDD project (FP7-ICT 619435). We would also
thank Philippe Mansuy of DIR CE for giving us access to the Grenoble South Ring Control
Room and valuable comments of traffic operators.
References
[1] A. Artikis, O. Etzion, Z. Feldman, and F. Fournier. Event Processing under Uncertainty. In DEBS 2012, 2012.
[2] C. Baber, S. Starke, X. Chen, N. Morar, and A. Howes. The design of user interfaces for
the speedd prototype,3rd report. Technical Report FP7-619435/SPEEDD-D5.3, EU,
Scalable Data Analytics, Scalable Algorithms, Software Frameworks and Visualization
ICT-2013 4.2.a, 2016.
[3] P. Bovy and I. Salomon. Congestion in Europe: Measurements, patterns and policies.
Travel behaviour: Spatial patterns, congestion and modelling, pages 143–179, 2002.
[4] J. Brooke. SUS: a quick and dirty usability scale. In P.W. Jordan, B. Weerdmeester,
B.A. Thomas, and I.L. McLelland, editors, Usability Evaluation in Industry, pages
189–194. Taylor and Francis, London, 1996.
[5] C. Canudas de Wit, I. Bellicot, F. Garin, P. Grandinetti, A.Y. Kibangou, F. Morbidi,
M. Schmitt, A. Hempel, C. Baber, and N Cooke. User requirements and scenario
definition. Technical Report FP7-619435/SPEEDD-D8.1, EU, Scalable Data Analytics,
Scalable Algorithms, Software Frameworks and Visualization ICT-2013 4.2.a, 2014.
[6] C. Canudas de Wit, F. Morbidi, L. Leon Ojeda, A.Y. Kibangou, I. Bellicot, and P. Bellemain. Grenoble Traffic Lab: An experimental platform for advanced traffic monitoring
and forecasting. IEEE Control Systems, 35(3):23–39, 2015.
19
[7] G. Cugola, A. Margara, M. Matteucci, and G. Tamburrelli. Introducing uncertainty in
complex event processing: model, implementation, and validation. Computing, pages
1–42, 2014.
[8] C.F. Daganzo. The cell transmission model: A dynamic representation of highway traffic
consistent with the hydrodynamic theory. Transportation Research Part B: Methodological, 28(4):269–287, 1994.
[9] A. Downs. Why Traffic Congestion is Here To Stay ... and Will Get Worse. ACCESS
Magazine, 25(1), 2004.
[10] J. Duchi, E. Hazan, and Y. Singer. Adaptive Subgradient Methods for Online Learning
and Stochastic Optimization. Journal of Machine Learning Research, 12:2121–2159,
July 2011.
[11] Y. Engel, O. Etzion, and Z. Feldman. A Basic Model for Proactive Event-Driven
Computing. In DEBS 2012, pages 107–118, 2012.
[12] O. Etzion and P. Niblett. Event Processing in Action. Manning Publication, 2010.
[13] F. Fournier, A. Kofman, I. Skarbovsky, and Skarlatidis A. Extending Event-Driven
Architecture for Proactive Systems. In Proc. of Event Processing, Forecasting and
Decision-Making in the Big Data Era (EPForDM), EDBT/ICDT Workshops., pages
104–110, 2015.
[14] G. Gomes and R. Horowitz. Optimal freeway ramp metering using the asymmetric cell transmission model. Transportation Research Part C: Emerging Technologies,
14(4):244–262, 2006.
[15] T.N. Huynh and R.J. Mooney. Max-Margin Weight Learning for Markov Logic Networks. In Proceedings of the European Conference on Machine Learning and Principles
and Practice of Knowledge Discovery in Databases (ECML PKDD), volume 5781 of
Lecture Notes in Computer Science, pages 564–579. Springer, 2009.
[16] T.N. Huynh and R.J. Mooney. Online Structure Learning for Markov Logic Networks.
In Proceedings of the European Conference on Machine Learning and Principles and
Practice of Knowledge Discovery in Databases (ECML-PKDD 2011), volume 2, pages
81–96, September 2011.
[17] R.E. Kalman. A new approach to linear filtering and prediction problems. Journal of
Basic Engineering, 82(35), 1960.
[18] M. Kojima, C. Nowakowski, and P. Green. Organization and structure of traffic management centers: Two case studies in Michigan. Technical report, University of Michigan,
Ann Arbor, MI, USA, 1999.
[19] R. Kowalski and M. Sergot. A Logic-based Calculus of Events. New Generation Computing, 4(1):67–95, 1986.
[20] M.J. Lighthill and G.B. Whitham. On kinematic waves ii. a theory of traffic flow on long
crowded roads. Proceedings of the Royal Society of London A: Mathematical, Physical
and Engineering Sciences, 229(1178):317–345, 1955.
[21] D.C. Luckham. Event Processing for Business: Organizing the Real-Time Enterprise.
John Wiley and sons, Inc, New jersey, USA, 1st edition, 2012.
20
[22] E. Michelioudakis, A. Skarlatidis, G. Paliouras, and A. Artikis. Online Structure Learning using Background Knowledge Axiomatization. In Proceedings of the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in
Databases (ECML-PKDD 2016), volume 1, pages 242–237, September 2016.
[23] E.T. Mueller. Event Calculus. In Handbook of Knowledge Representation, volume 3 of
Foundations of Artificial Intelligence, pages 671–708. Elsevier, 2008.
[24] M. Papageorgiou, H. Hadj-Salem, and J.-M. Blosseville. Alinea: A local feedback
control law for on-ramp metering. Transportation Research Record, (1320):58–64, 1991.
[25] I. Papamichail, A. Kotsialos, I. Margonis, and M. Papageorgiou. Coordinated ramp
metering for freeway networks–a model-predictive hierarchical control approach. Transportation Research Part C: Emerging Technologies, 18(3):311–331, 2010.
[26] I. Papamichail, M. Papageorgiou, V. Vong, and J. Gaffney. Heuristic ramp-metering coordination strategy implemented at monash freeway, australia. Transportation Research
Record: Journal of the Transportation Research Board, 2178(1):10–20, 2010.
[27] D. Pisarski and C. Canudas de Wit. Optimal Balancing of Road Traffic Density Distributions for the Cell Transmission Model. In 51st IEEE Conference on Decision and
Control (CDC 2012), Maui, Hawaii, United States, December 2012.
[28] C.E. Rasmussen. Gaussian processes in machine learning. In Advanced lectures on
machine learning, pages 63–71. Springer, 2004.
[29] B.L. Richards and R.J. Mooney. Learning relations by pathfinding. In Proceedings of
the Tenth National Conference on Artificial Intelligence, AAAI’92, pages 50–55. AAAI
Press, 1992.
[30] M. Richardson and P.M. Domingos. Markov logic networks. Machine Learning, 62(12):107–136, 2006.
[31] M. Schmitt, C. Ramesh, and J. Lygeros. Sufficient optimality conditions for distributed,
non-predictive ramp metering in the monotonic cell transmission model. submitted to
Journal of Transportation Research, Part B: Methodological., 2016.
[32] A. Skarlatidis, G. Paliouras, A. Artikis, and G.A. Vouros. Probabilistic Event Calculus
for Event Recognition. ACM Transactions on Computational Logic, 16(2):11:1–11:37,
February 2015.
[33] S. Starke, N. Cooke, A. Howes, N. Morar, and C. Baber. Visual sampling in a road
traffic control management control room. In International Conference on Ergonomics
and Human Factors, pages 503–511. Taylor and Francis, 2015.
[34] J.J. Thomas and K.A. Cook. Illuminiating the Path: the research and development
agenda for visual analytics,. National Visualization and Analytics Center, Pacific Northwest National Laboratory, Richland, WA, 2005.
[35] K.J. Vicente. Cognitive work analysis: Toward safe, productive, and healthy computerbased work. Lawrence Erlbaum Associates, New York, 1999.
[36] S. Wasserkrug, A. Gal, O. Etzion, and Y. Turchin. Efficient processing of uncertain
events in rule-based systems. IEEE Transactions on Knowledge and Data Engineering,
24(1):45–58, 2012.
21
| 3cs.SY
|
A framework for small but rich vehicle routing problems.
Vladimir G. Deineko1
arXiv:1610.01876v1 [cs.DS] 5 Oct 2016
Warwick Business School, Coventry, United Kingdom
Abstract. In this paper we consider a 2-vehicle routing problem which can be viewed as a building block for the varieties of the vehicle routing problems (VRPs). To approach this problem, we
suggest a framework based on the Held and Karp dynamic programming algorithm for the classical travelling salesman problem. An algorithm based on this framework shows an exceptionally
good performance on published test data. Our approach can be easily extended to a variety of
constraints/attributes in the VRP, hence the wording “small but rich” in the title of our paper.
Keywords. Combinatorial optimization; vehicle routing problem; dynamic programming; 2-period
travelling salesman problem.
1
Introduction
In the vehicle routing problem (VRP) a set of customers with certain requests are to be visited by
vehicles. The vehicles are to be chosen from a fleet of heterogeneous vehicles, with various fixed and
variable costs of usage. The objective is to find a minimum cost schedule for customer visits to deliver
required services in the specified time and manner. The scheduling of visits may need to satisfy some
additional constraints, e.g. with some visits demanded on particular days, or by particular vehicles, etc.
It would not be an exaggeration to say that thousands of research papers devoted to the VRPs
are published every year. As rightly mentioned by Michael Drexl [16], this “is certainly due to the
intellectual challenge VRPs pose as well as to their relevance in logistics and transport”.
The VRP can be viewed as a combination of the two well known combinatorial optimization problems
- the travelling salesman problem (see e.g. [1]) and the bin packing/knapsack problem (see e.g. [22]).
It is not surprising that the combination of these two problems creates new computational challenges
for researchers and practitioners. For instance, the sizes of the TSP instances which are tractable by
recently developed computational algorithms [1] are much bigger than the sizes of easy tractable VRP
instances.
In many practical applications the number of customers visited by a single vehicle are not very
large. We mention here just a few contextual examples: teachers visiting special needs pupils, nurses
attending patients at home, food delivery in rural areas (long distances - hence few customers to
visit), bulk deliveries of industrial goods (large items - hence few customers). The task of allocation
of customers to vehicles adds a lot of complexity to the VRP, especially due to the various additional
constraints arising in practice. The VRPs with multiple practice related constraints are referred to in
recent publications as rich VRPs (see surveys [5,16,27]), or multi-attribute VRPs [38,39].
In our study we decided to concentrate on the simplest possible version of the rich VRP - the
VRP with many constraints/attributes but with only two vehicles, which we call the 2VRP. Further
on, keeping in mind the practical applications mentioned above, we studied small (with not many
customers) but rich VRPs (hence the title of the paper). We hope that better understanding of the 2VRP
would allow researchers to develop new algorithms for the VRPs with many vehicles. A straightforward
approach for the VRP with many vehicles could for instance be a heuristic which enumerates all possible
pairs of vehicles and iteratively solves the 2VRPs.
The approach described in this paper is based on the well known Held and Karp dynamic programming approach [20]. We are not the first who suggest using dynamic programming techniques for
2
Vladimir G. Deineko
the VRPs (see e.g. [13,18,24,32,33,39]). The way we use dynamic programming is different though to
what has been suggested so far. We suggest a framework which permits to easily incorporate additional
constraints and vary the objective function to optimise (see [18,21,35,36,40] for frameworks suggested
for various types of the VRPs). To address the curse of dimensionality in dynamic programming, an
aggregation scheme is included in the framework. We tested our framework on the 2-period balanced
travelling salesman problem [3]. The results of computational experiments are impressive: for 60 benchmark instances suggested in [3], better solutions have been found for 57 instances.
2
2.1
A framework for the 2VRP
Notations and definitions
In this section we introduce a model for a vehicle routing problem (VRP) with only two vehicles (2VRP).
We describe objects involved in the model and attributes associated with each object. In what follows
we use an approach which finds optimal solutions for small size problems. To reduce the size of an
initial problem we suggest making use of various aggregation techniques. As an example of a possible
aggregation, one can think of considering all customers on a street as one customer with the demand
defined as the sum of demands for the customers on that street.
There are two not homogeneous vehicles, vehicle 1 and vehicle 2, in the 2VRP. They have different
capacities W1 and W2 and different costs of travelling. The vehicles will be used for delivering demanded
goods to customers.
We assume that each customer is located in an estate with only two entry points. A network of
one-way roads within the estate connects these two points. So the travel costs within the estate are
asymmetric costs. To distinguish between two entry points to the estate, we refer to one of the points
as the left node and denote it by L(i). The other of the two points is referred to as the right node, and
is denoted by R(i).
We assume that customer i in our problem has a set of seven attributes
2
2
1
1
(i), w(i)} with the following meaning:
(i), lR
(i), lL
(i), lR
{L(i), R(i), lL
–
–
–
–
–
the left node L(i);
the right node R(i);
m
cost of travelling from the left node to the right node lL
(i), if travelled by vehicle m, m = 1, 2;
m
(i), if travelled by vehicle m, m = 1, 2;
cost of travelling from the right node to the left node lR
the demand w(i).
So, given a set of n customers {1, 2, . . . , n}, with the demands w(i), they have to be visited by one of
the two vehicles which delivers the demanded goods from depots to the customers. The total demand of
all customers is less than the total capacity of the two vehicles, so only one route per vehicle is needed.
The vehicles may travel from different depots and return to different depots. Assume that vehicle 1
travels from depot d1R and returns to depot d1L , and vehicle 2 travels from depot d2R and returns to
depot d2L .
The costs of travelling between the nodes are given by the two 2(n + 4) × 2(n + 4) cost matrices
C 1 = (c1ij ) and C 2 = (c2ij ), where cm (i, j) is the cost of travelling by vehicle m from node i to node j.
m
m
A visit of customer i adds to the travel costs either cost lL
(i), or cost lR
(i) depending on the way
the customer is visited - from the left entry node, or from the right entry node.
Our approach will utilise the well known dynamic programming algorithm for the travelling salesman problem. So we would like to view routes for the two vehicles as one route. To simplify further
considerations, we introduce an auxiliary customer 0 with the set of attributes {d1L , d2R , 0, ∞, 0, ∞, 0}.
The reason for introduction of the auxiliary customer 0 and placing it into the two-vehicle route is to
A framework for the 2-VRP
3
separate points visited by the two different vehicles. Cost of travelling from d1L to d2R costs nothing,
while travelling in opposite direction is forbidden (by the infinitely large cost).
The two-vehicle route starts from the node d1R , visits all customers from the set U = {0} ∪
{1, 2, . . . , n}, and ends in the node d2L . Vehicle 1 is the first to start visiting customers in the route. Visiting customer 0 means that the mode of travelling is changed from travelling by vehicle 1 to travelling
by vehicle 2. Vehicle 2 will travel from customer 0, i.e. from depots d2R , to depot d2L .
The objective is to find a minimum cost route of delivering the requested demand to all customers.
The total demand of customers visited by each vehicle should not exceed the corresponding vehicle’s
capacity.
2.2
Dynamic programming recursions
In this section, the well known Held & Karp [20] dynamic programming algorithm for the TSP is
adapted for the case of the 2VRP formulated above.
Let J be a subset of customers not containing i, so J ⊂ U , i ∈
/ J. Denote as VL[i, J] the minimum
cost of an optimal 2-vehicle route among all routes which start from a visit of customer i from the
corresponding left node, then visit all the customers in set J, and stop in depot d2R . Similarly, define
VR[i, J] to be the cost of the optimal route that starts visiting customer i from the corresponding right
node, and visits then the customers in J. The optimal cost of a 2-vehicle tour can be calculated as
V = mini∈U \{0} {c1d1 ,L(i) + VL[i, U \ {i}], c1d1 ,R(i) + VR[i, U \ {i}]}.
R
d1L
+V L[j, J \ {j}]
1
lL
(i)
c1R(i)L(j)
(1)
R
L(j)
d2R
customer 0
R(j)
L(i)
R(j)
L(j)
customer i
c1R(i)R(j)
Set of customers J \ {j}
customer j
+V R[j, J \ {j}]
Set of customers J
Fig. 1: Illustration of the calculation of length VL[i, J], the shortest route from customer i through all
customers in set J (the case of 0 ∈ J.)
4
Vladimir G. Deineko
We assume here that the total demand from customers is bigger than the capacity of each of the
vehicles, so in the formula above 0 cannot be the first customer. Values VL[i, J] and VR[i, J] for all
customers i and subsets J, J ⊂ {0} ∪ {1, 2, . . . , n}, are calculated as shown in the recursions below:
(
)
1
lL
(i) + c1R(i),L(j) + VL[j, J \ {j}]
minj∈J 1
if 0 ∈ J,
lL (i) + c1R(i),R(j) + VR[j, J \ {j}]
(
)
VL[i, J] =
2
2
l
(i)
+
c
+
VL[j,
J
\
{j}]
L
R(i),L(j)
minj∈J
if 0 ∈
/ J, w({i} ∪ J) ≤ W2
2
lL
(i) + c2R(i),R(j) + VR[j, J \ {j}]
∞
otherwise
(2)
(
)
1
1
l
(i)
+
c
+
VL[j,
J
\
{j}]
L(i),L(j)
minj∈J R
if 0 ∈ J,
1
1
l
(i)
+
c
+ VR[j, J \ {j}]
R
L(i),R(j)
)
(
VR[i, J] =
2
(i) + c2L(i),L(j) + VL[j, J \ {j}]
lR
if 0 ∈
/ J, w({i} ∪ J) ≤ W2
minj∈J 2
lR (i) + c2L(i),R(j) + VR[j, J \ {j}]
∞
otherwise
(3)
VR[0, J] =
min
(
j∈J
c2R(0),L(j) + VL[j, J \ {j}
c2R(0),R(j) + VR[j, J \ {j}
∞
)
(
w(U \ ({i} ∪ J)) ≤ W1 ,
if
and w({i} ∪ J) ≤ W2
(4)
otherwise
The boundary conditions are:
2
VL[i, ∅] = lR
(i) + c2L(i),L(0) ,
2
VR[i, ∅] = lL
(i) + c2R(i),L(0) .
(5)
Recursions (1)-(5) extend Held & Karp recursions to the case of 2VRP: since there are only two
vehicles, the capacity constraints are easily verified without any extra dimensions or complicated calculations. Notice that we use notation w(J) for the sum of demands of all items in set J. For instances
with relatively small sizes, these days computers perform calculations (1)-(5) within seconds. However
when the number of customers approaches 20, the computations on a standard laptop become more and
more problematic. To make the dynamic programming approach practical, we suggest an aggregation
strategy described in the next section.
2.3
Aggregation strategies and local search
As was mentioned above, the dynamic programming approach can be used for small size problems
but may become impractical in real life applications. In this section we describe an approach for data
aggregation and reducing the size of the initial problem.
We start with a feasible 2VRP solution with n customers. We “cut” this solution into a small number
of sub-paths. One can think of various possible ways of cutting a 2-VRP solution into sub-paths. Each
subpath is replaced then by a new customer. The way we describe the customers in the 2VRP in the
previous section permits an easy replacement of any subpath by a new customer. A left node for the new
customer is the first node in the subpath (i.e. either left or right node of the first customer); the right
A framework for the 2-VRP
5
Fig. 2: Illustration of the “sliding subsets” heuristic: first step of disassembling; each subset contains 2
customers.
node of the new customer is the last node in the subpath; the demand of the new customer is the sum
of the demands of customers in the sub-path; the left/right lengths are calculated as the corresponding
lengths of the subpath.
Obviously, an exact solution obtained for a new small-size problem can be viewed only as an approximate solution for the initial 2-VRP. This approximate solution can be again “disassembled” into
a small number of sub-paths, and the process of solving small size problems is repeated again, until
all possible ways of disassembling and aggregating a solution have been enumerated and no further
improvement was achieved.
We suggest the following straightforward approach, which we call the sliding sub-sets method. Assume that we have an initial solution to the 2VRP: τ = hd1 , t11 , t12 , . . . , t1k1 , 0, t21 , t22 , . . . , t2k2 , d2 i. Here
vehicle one’s route is τ = hd1 , t11 , t12 , . . . , t1k1 , 0i and vehicle two’s route is h0, t21 , t22 , . . . , t2k2 , d2 i.
We disassemble this solution into a new set of customers for the new 2VRP according to the following
procedure. First, define 0 as a customer in the new 2VRP and delete it from τ . Let s now be a small
constant, a parameter of the algorithm. Choose two subsets of customers containing s items each so that
the customers in each subset are picked from consecutive positions in τ . Subset S1 will always contain
at least one customer from vehicle 1’s route, and S2 contains at least one customer from vehicle 2’s
route. On the first disassembling step define the first subset as S1 = {t11 , . . . , t1s }, and the second subset
as S2 = {t1k1 −s+2 , . . . , t1k1 , t21 }. Notice that subset S2 is chosen to ensure that at least one customer
from vehicle 2’s route is included in the subset. Delete S1 and S2 from τ and add them to the set of
customers in the new 2VRP. Consider d2 as a new customer, delete it from τ . Sub-paths which are left
in τ are considered as aggregated customers and added to the new 2VRP. On the first step these are:
depot d1 , sub-path ht1s+1 , . . . , t1k1 −s+1 i, and the sub-path ht22 , . . . , t2k2 i. So the new 2VRP contains 2s + 5
customers (depots are counted here as customers).
Figure 2 illustrates the concept of sliding subsets. To simplify drawings the customers are depicted
as points, however we treat them rather as segments, as described in the previous section. On the other
6
Vladimir G. Deineko
hand, the nodes in the tour are not connected by straight lines, as one would expect. Recall that in our
model we view customers as “paths”, therefore curves are chosen to illustrate the connections.
If no better solution is found, we redefine subset S2 by deleting, say, l first elements and adding l
new elements (l is a parameter, to which we refer as the step). So subset S2 slides along the tour. We
repeat the process until we reach the end of the tour. Then we change (slide) set S1 (with the step l)
and redefine set S2 to follow set S1 similar to the first step settings described above.
If the solution of the 2VRP was improved, the process of disassembling is applied to the new
solution. The process is stopped when all possible positions for subsets S1 and S2 are considered and
no improvements found.
Fig. 3: Illustrations of disassembling: (a) first step: S1 and S2 are separated by one sub-path only; (b) an
example when S1 and S2 are separated by two sub-paths and a depot; (c) depot and the first subpath
are considered as one customer; (d) modified first step: a subpath between S1 and S2 (case (a)) is
partitioned into a sub-path and a single customer to keep the size of the new 2VRP fixed.
Details of implementation. In the follow up computational experiments we used a modification of the
approach described above. The parameters of the algorithm are the size of the subsets s and step l.
To keep the size of the small 2VRP fixed and defined by only these two parameters, we implement the
disassembling procedure as described below. We refer to Figure 3 in our explanations of various steps
(and possibilities) for disassembling. Figure 3(a) illustrates the first step of the disassembling process
as was described above (with parameter s = 2). Sets S1 and S2 are separated by one sub-path in this
case. The size of the new small 2VRP is 2s + 5.
In the illustrations we set step l = 2. Fig. 3(b)) illustrates the outcome of disassembling the tour on
the second step. Notice that depot 0 was already removed from the tour, therefore setting l = 2 yields
A framework for the 2-VRP
7
the position of S2 as shown in the figure. There are two sub-paths between sets S1 and S2 , and the size
of the new 2VRP problem is 2s + 6.
Consider the step when S1 and S2 are chosen as shown in Fig. 3(c). If the first sub-path did not
contain the depot, the size of the problem would have been 2s + 7. On the implementation step, it was
convenient to keep the size of the problem fixed at 2s + 6. Therefore it was decided to “glue” the first
sub-path with the depot and have it as the depot in the new problem as shown in the figure.
If subsets S1 and S2 are separated by a single path (for example, on the first step of disassembling),
it was decided to consider the last node in the sub-path as a second sub-path: in this case the problem
with 2s + 5 customers becomes the problem with 2s + 6 customers (compare Fig. 3(a) and Fig. 3(d)).
We will use notation H(s, l) for a heuristic from the family of heuristics described above, where the
size of the subsets is s: |S1 | = |S2 | = s, and the step for moving the subsets is l.
2.4
Varieties of the VRP covered by the framework.
Below we illustrate the advantages of the suggested dynamic programming approach by listing some
types of the VRPs that can be tackled by the approach proposed.
– Arc routing. Our model made no difference between the classic capacitated VRP, where the
customers are represented by one node in a road network, and the arc routing VRP, where the
customers are the streets/arcs in the network (see [41,10]).
– Heterogeneous fleet. The dynamic programming recursions take into account individual characteristics of the vehicles, so both homogeneous and heterogeneous fleets (see [2] and recent survey
[23]) can be managed.
– Multi-depot and open VRPs. Incorporation of the multi-depot feature into the model is straightforward. The reader is referred to recent papers [31,25] for the specifics of the multi-depot VRP. For
the open VRP (see references in recent papers [28,29]), it is enough to introduce a dummy depot
with zero distances to this depot from all customers.
– Tight capacity constraint. For some instances of the VRP, the main difficulties lie in packing
all goods into a bounded number of vehicles. Since the dynamic programming approach enumerates
all possible subsets, the “bin packing”/loading part of the VRP is resolved at the same time as
the routing part. See arguments in [11] for the benefits of integrating loading and routing. It is
easy to see that the framework can incorporate more complicated packing constraints, e.g. twodimensional loading constraints [26], by solving the corresponding packing sub-problems on each
step of calculations.
– Fixed items in a vehicle. In some VRPs it is important to allocate customers to particular vehicle
visits in advance. We refer to these customers as fixed items in a vehicle. This feature can easily be
added to the dynamic programming recursions. Fixed items can be useful, e.g. with multiple visits
of customers (see Section 3.1 in this paper). Another example is the so-called site-dependent VRP
[2] - some customers can be served only by a specific type of vehicles (so-called docking constraints
[7]). An interesting case study with fixed customers was described in [15]. An Austrian red cross
considered introducing two tiers for a blood delivery service. Urgent delivery (with a higher price)
is delivery within one day, and standard delivery (at a lower price) is delivery on the second day.
Hospital customers for the current day are known, while the next day’s customers are unknown and
only the probabilities of requests can be evaluated. So, on each day a dispatcher knows undelivered
requests from yesterday, requests which arrived today, and probable requests for tomorrow. The
requests from yesterday are urgent and have to be delivered today - hence a fixed allocation of these
customers to today’s route (it is assumed here that the delivery is done by one vehicle). A sample
of (probable) requests for tomorrow should be fixed for tomorrow’s delivery. Today’s requests are
flexible and can be allocated to either vehicle, however they will be charged different prices.
8
Vladimir G. Deineko
– Penalties for wrong day deliveries. In the previous paragraph we mentioned charges/costs for
deliveries in different vehicles/days. Another example is given in [12], where penalties for wrong day
deliveries were introduced. These can easily be incorporated into our model by changing left/right
lengths of intervals (assuming that costs of travelling and penalties are measured in the same
monetary units).
– Cumulative VRPs. In the cumulative VRP the objective is to minimise the sum of arrival times
to all customers; this problem is also known as the latency problem (see references in recent papers
[8,30,34]). It is easy to see that the dynamic programming approach can be adapted for this type
of objective function.
– Weight or time dependent travel costs. There are some practical situations when the travel
costs depend on the load of the vehicle (see [42,43] and references there) or on the time when the
vehicle travels ([17,37]). It can easily be seen that the dynamic programming recursions above allow
this type of calculation to be incorporated into the recursions.
Fig. 4: Illustration of the definition of the balanced 2TSP; customers 1, 3, 5, and 8 are visited in two
periods.
3
An application of the framework
In this section we provide evidence of computational efficiency of the proposed framework. As a test
problem we have chosen a 2-period balanced travelling salesman problem and a set of the benchmark
test problems from Bassetto & Mason [3]. In their algorithms, Bassetto & Mason used powerful exact
techniques and even used a visualisation and human interventions for improving solutions. We believe
that we have chosen a strong competitor to test the potential of the suggested framework.
3.1
A 2-period travelling salesman problem
Given m customers to be visited in each of the two periods and n customers to be visited once in
either of the periods, and the distance matrix between the customers, the 2-period travelling salesman
problem (2TSP) asks for tours for the two periods with the minimal total distance travelled.
A framework for the 2-VRP
9
Butler, Williams, and Yarrow [4] considered a practical application of the 2TSP in milk collection in
Ireland. They applied what they called a “man-machine method”, combining an integer programming
technique with a human being intervention for identifying violated constraints.
The 2TSP can be modelled as a special case of the 2VRP as follows. For each of the customers to be
visited twice, an identical copy of the customer is created. The identical customers are allocated then to
the different vehicles. This allocation to the vehicles is fixed. The 2VRP now has n + 2 × m customers:
each vehicle has to serve m fixed customers, n customers to be served by only one of the two vehicles.
In the balanced 2TSP (2VRP) an additional constraint demands that the number of customers
visited in each period (each vehicle) differs by no more than 1. Bassetto and Mason [3] considered
a balanced 2TSP in the Euclidean plane. In this variant of the problem the customers’ locations are
points in the Euclidean plane, and the costs of travelling between the customers are standard Euclidean
distances. Figure 4 illustrates a solution to the Euclidean balanced 2TSP with 10 customers. Four out
of ten customers in the example are visited in 2 periods (the depot is considered as a customer here).
A short summary of the approaches used in [3] is as follows. First a TSP tour on the set of all customers, called a general tour (GT), is constructed. The GT is used to obtain a partition of customers
into two subsets visited in two periods. The initial partition is improved by applying decision rules motivated by geometry (e.g. removal of crossing edges). For each period an optimal TSP tour is constructed
by applying an exact TSP algorithm (see [1], Chapter 16, and [9]). Three algorithms A1, A2, and A3 are
suggested and tested on the benchmark instances. Since the test instances are defined by the Euclidean
coordinates, a visualisation of solutions is an option. The authors used the visualisation and human
intervention to improve some of the solutions. Columns labelled as “PC” in the tables in the Appendix
refer to the best solutions found by algorithms A1 and A3 from [3]. Columns “PC+Manual” refer to
the best solutions obtained after manual improvements (the exceptions are the results for instances I5 ,
I6 , and I10 the best solution for which were found by algorithm A2).
3.2
Computational experiments
The set of benchmark instances from [3] contains 60 randomly generated instances with 48 customers.
The set of these instances is divided into three subsets with a different number of customers to be visited
in the two periods (8, 12, and 24 customers). We used the same set of instances in our computational
experiments. However, we used only distance matrices, without using in our algorithms any knowledge
on the sets of coordinates in the instances.
We used the framework discussed in the previous sections in the following algorithmic setting:
Algorithm for the 2VRP {
repeat n(= 48) times:
Generate a random allocation of n customers into two TSP tours;
Insert into each tour the missing fixed customers to be visited by the two vehicles;
Apply a tour improvement TSP heuristic of Carlier & Villon [6] to improve the tours;
repeat
Apply dynamic programming heuristic H(s, l) for improving the partition into the tours;
Apply a tour improvement TSP heuristic of Carlier & Villon [6] to improve the tours;
while (improvement found);
}
In our computational experiments we used three heuristics: H(3, 1), H(5, 2), and H(6, 3). So the
dynamic programming recursions were used for sub-problems with no more than 18 customers. We
found that the results are much better if a multi-start strategy is used, therefore we repeated our search
starting with randomly generated initial solutions. The number of repetitions in our experiments was
chosen as n = 48, however for some of the instances we tried 2n repetitions to improve the solutions (see
10
Vladimir G. Deineko
20
15
10
5
0
-8%
-7%
-6%
-5%
-4%
-3%
-2%
-1%
0%
Fig. 5: Histogram for the distribution of improvements achieved with H(6, 3) on all 60 instances.
comments in the appendix). Surprisingly random partitions into tours in our experiments yield better
solutions than partitions obtained from a “good” (general) TSP solution for all customers. We did not
use an exact algorithm for finding TSP solutions. We used Carlier & Villon’s [6] TSP heuristic which is
an O(n3 ) heuristic that finds an optimal TSP tour in an exponential neighbourhood of tours. (The use
of this particular heuristic is rather a matter of author’s preferences. Notice that the neighbourhood
searched by this heuristic contains at least 75% of tours that can be obtained from a tour by applying
the well-known 3opt heuristic, as proved in Theorem 5 in [14]. For the experiments we used a desktop
computer with Intel i7-3770 3.40 GHz CPU, 16 GB of RAM, and GNU C++ compiler. The length of
the solutions found are represented in the tables in the Appendix. In this section we provide a summary
of the results obtained.
Figure 5 shows the improvement (measured in percent of achieved cost) that our technique yields
using heuristic H(6, 3) for the dynamic programming search. Our results here are compared to the best
solution found by algorithms A1 and A3 in [3]. It can be seen that all solutions have been improved, with
the majority of improvements being 2% or better. By varying the size of the neighbourhoods searched
in the dynamic programming heuristic, one can influence the computation time (and the accuracy).
Tables 1-3 below summarise the outcomes of computational experiments for each of the settings used
and for each type of instance. We compared our computational results with both the computer results
(labelled as “PC”) and the results improved by manual intervention (labelled as “PC+Manual”) from
[3].
Settings in [3]
Mean %
Best %
Worst %
Improved #
H(3, 1), tmean = 27 sec
PC
PC+manual
-2.22%
-0.51%
-7.19%
-1.75%
+0.46%
+0.86%
19/20
13/20
H(5, 2), tmean = 172 sec
PC
PC+manual
-2.54%
-0.83%
-6.92%
-2.65%
+0.28%
+0.93%
19/20
17/20
H(6, 3), tmean = 415 sec
PC
PC+manual
-2.63%
-0.93%
-7.31%
-2.77%
-0.12%
-0.00%
20/20
20/20
Table 1: Summary of results for instances with 8 (out of 48) nodes visited in two periods.
A framework for the 2-VRP
Settings in [3]
Mean %
Best %
Worst %
Improved #
H(3, 1), tmean = 23 sec
PC
PC+manual
-1.76%
-0.63%
-3.81%
-2.46%
+0.30%
+2.09%
19/20
17/20
H(5, 2), tmean = 172 sec
PC
PC+manual
-2.14%
-1.02%
-3.93%
-2.59%
-0.53%
+0.42%
20/20
18/20
11
H(6, 3), tmean = 415 sec
PC
PC+manual
-2.17%
-1.04%
-3.93%
-2.59%
-0.53%
+0.46%
20/20
19/20
Table 2: Summary of results for instances with 16 (out of 48) nodes visited in two periods.
Settings in [3]
Mean %
Best %
Worst %
Improved #
H(3, 1), tmean = 17 sec
PC
PC+manual
-1.20%
-0.58%
-4.44%
-1.50%
-0.00%
+0.24%
20/20
19/20
H(5, 2), tmean = 131 sec
PC
PC+manual
-1.25%
-0.64%
-3.85%
-1.52%
-0.17%
+0.24%
20/20
19/20
H(6, 3), tmean = 299 sec
PC
PC+manual
-1.36%
-0.75%
-4.44%
-2.17%
-0.17%
+0.16%
20/20
18/20
Table 3: Summary of results for instances with 24 (out of 48) nodes visited in two periods.
4
Summary
In this paper we have presented a framework for small but rich VRPs. The framework can be used
for a variety of VRP settings. The computational scheme which implements the framework permits an
easy scaling, hence varying computational time and the accuracy of the solutions found. Computational
experiments for the balanced 2-period TSP have shown an impressive performance from the framework.
Our next step would be performing extensive computational experiments to test the potential of the
framework for other VRP settings.
Acknowledgements
Vladimir Deineko acknowledges support by the Center for Discrete Mathematics and Its Applications,
University of Warwick. The author also thanks Tatiana Bassetto and Francesco Mason for providing
the benchmark problems, Douglas Miranda and Alex Taylor for useful comments on an early version of
the paper.
12
A
Vladimir G. Deineko
Appendix
Results for instances with 8 out of 48 nodes visited in two periods
Solutions from [3]
H(3, 1)
H(5, 2)
H(6, 3)
Instance PC PC+manual time (sec) length time (sec) length time (sec) length
I21
25217
24937
26
24517
185
24493
395
24517
I22
26996
26549
19
26413
141
26347
352
26347
I23
26476
26192
23
26222
168
26131
412
26050
I24
26802
26038
30
26091
169
26057
369
26002
I25
27728
27408
23
27053
150
26914
391
26914
I26
24348
24268
29
23894
186
23837
463
23837
I27
27335
26857
26
26435
181
26417
422
26417
I28
24679
24232
26
24006
162
23985
408
24143
I29
26890
26466
28
26466
180
26466
405
26466
I30
24978
24200
32
24407
212
23915
549
23915
I31
26266
26130
23
26130
147
26130
369
26130
I32
26360
26054
32
26067
196
26296
429
26032
I33
26418
26418
28
26540
174
26493
419
26387
I34
28733
27074
28
26666
166
26744
380
26633
I35
25043
24587
27
24587
172
24526
434
24517
I36
27103
26790
26
26321
179
26079
409
26079
I37
25662
25123
24
25072
153
25072
432
25101
I38
26459
25709
26
25816
161
25588
372
25588
I39
27209
26994
27
26553
186
26314
474
26247
I40
25416
24964
28
25081
170
24850
419
24824
Results for instances with 16 out of 48 nodes visited in two periods
Solutions from [3]
H(3, 1)
H(5, 2)
H(6, 3)
Instance PC PC+manual time (sec) length time (sec) length time (sec) length
I1
33804
32556
25
32566
154
32566
392
32556
I2
30929
30929
23
30716
150
30716
379
30716
I3
30596
30382
24
30280
157
29967
385
29954
I4
28563
28441
23
28274
139
28260
337
28223
I5
27323
27206
29
27177
184
27177
452
27177
I6
33065
32396
22
31971
152
31890
346
31890
I7
32854
31861
24
31615
153
31606
370
31637
I8
30850
30571
22
30215
120
30266
328
30266
I9
34709
34024
19
33951
124
33911
331
33911
I10
31451
30867
25
30844
154
30660
368
30660
I61
27158
26934
22
26799
146
26644
348
26644
I62
27774
27619
24
27325
168
27248
360
27248
I63
25308
24960
21
24345
142
24313
324
24313
I64
27875 27285*
19
27856
138
27399
377
27410
I65
27060
26888
20
27141
140
26872
373
26806
I66
27677
27624
26
27135
169
27188
429
27131
I67
30268
30203
21
29928
156
29841
340
29841
I68
28033
27923
24
27478
163
27425
393
27447
I69
27958
27638
22
27621
154
27143
383
27143
I70
28483
28427
22
27865
157
27799
350
27790
A framework for the 2-VRP
13
Solution for instance I64 was improved to 27236 by H(5, 2) with 2 × 48 random starts (in 280 seconds).
Results for instances with 24 out of 48 nodes visited in two periods
Solutions from [3]
H(3, 1)
H(5, 2)
H(6, 3)
Instance PC PC+manual time (sec) length time (sec) length time (sec) length
I41
30253
30147
16
30100
117
30064
258
30064
I42
33008
32020
17
31544
124
31738
284
31544
I43
31500
31500
18
31200
140
31166
311
31194
I44
30313
30170
21
29812
156
29757
335
29757
I45
27986
27857
17
27780
128
27780
310
27780
I46
30073
30013
20
30013
148
30013
328
30013
I47
32106
32106
17
31704
133
31735
299
31728
I48
31004
30942
19
30478
149
30471
337
30471
I49
33663
33185
17
33173
130
33173
312
33197
I50
31266
31266
16
31266
117
31213
259
31213
I51
33722
33627
19
33358
146
33344
322
33358
I52
32353
32280
15
32251
113
32251
274
32200
I53
33287
33200
17
32943
118
32871
268
32726
I54
31973
31600
18
31368
149
31373
345
31368
I55
33837 33507*
16
33587
115
33587
292
33560
I56
29696
29476
16
29176
115
29156
266
28835
I57
31954
31640
18
31427
133
31427
299
31427
I58
30705
30246
18
30165
140
30178
345
30165
I59
31549
31549
19
31471
137
31223
296
31223
I60
32384
32317
15
32193
115
32137
244
32140
The best known value for instance I55 , which is 33507, can be achieved within 620 seconds by our
algorithm as well, if we set parameters as s = 6 and l = 2.
References
1. D. L. Applegate, R. E. Bixby, V. Chvatal, W. J. Cook, 2006. The traveling salesman problem. A
computational study. Princeton University Press.
2. R.Baldacci, M. Battara, and D. Vigo, 2008. Routing a heterogeneous fleet of vehicles, in [19], 3–27.
3. T. Bassetto, F. Mason, 2011. Heuristic algorithms for the 2-period balanced travelling salesman problem
in euclidean graphs. European Journal of Operational Research, 208(3): 253–262.
4. M. Butler, H.P. Williams, 1997. The two-period travelling salesman problem applied to milk collection
in Ireland, Computational Optimization and Applications 7, 291–306.
5. J. Caceres-Cruz, P. Arias, D. Guimarans, D. Riera, and A. A. Juan, 2014. Rich vehicle routing
problem: Survey. ACM Computing Surveys, 47, 2, Article 32.
6. J. Carlier, P. Villon, 1990. A new heuristic for the travelling salesman problem, RAIRO – Operations
Research 24, 245–253.
7. V. N. Coelho, A. Grasas, H. Ramalhinho, I. M. Coelho, M. J. F. Souza, R. C. Cruz, 2016. An ILSbased algorithm to solve a large-scale real heterogeneous fleet VRP with multi-trip and docking constraints,
European Journal of Operational Research, 250, 367–376.
8. S. Coene, F. C. R. Spieksma, G. J. Woeginger, 2011. Charlemagne’s challenge: the periodic latency
problem, Operations Research, 59 (3), 674–683.
9. CONCORDE TSP SOLVER. http://www.tsp.gatech.edu/concorde.
10. A. Corberan, C. Prins, 2010. Recent results on arc routing problems: An annotated bibliography, Networks, 56(1), 50–69.
14
Vladimir G. Deineko
11. J-F. Côté, G. Guastaroba, M. G. Speranza, 2015. The value of integrating loading and routing, Working
paper, CIRRELT-2015-31.
12. Y. Crama, M. Rezaei, T. Van Woensel, 2015. A branch-and-price algorithm for 2-period vehicle routing
problems, Working paper,University of Liege, ORBi, http://hdl.handle.net/2268/175663.
13. J. Desrosiers, Y. Dumas, F. Soumis, 1986. A dynamic programming solution of the large-scale singlevehicle dial-a-ride problem with time windows, American Journal of Mathematical and Management Sciences, 6, 301–325.
14. V. G. Deineko, G. J. Woeginger, 2000. A study of exponential neighborhoods for the travelling salesman
problem and for the quadratic assignment problem. Math. Program., 87, 519–542.
15. K.F. Doerner, W. J. Gutjahr, R. F. Hartl, and G. Lulli, 2008. Stochastic Local Search Procedures
for the Probabilistic Two-Day Vehicle Routing Problem. A. Fink and F. Rothlauf (Eds.): Advances in
Computational Intelligence, SCI 144, 153168.
16. M. Drexl, 2012. Rich vehicle routing in theory and practice, Logist. Res., 5, 47–63.
17. J. F. Ehmke, A. M. Campbell, B. W. Thomas, 2016. Vehicle routing to minimize time-dependent
emission in urban areas, European Journal of Operational Research, 251, 478–494.
18. J. Gromicho, J. J. van Hoorn, A. L. Kok, J. M. J. Shutten, 2012. Restricted dynamic programming:
A flexible framework for solving realistic VRPs. Computers & Operations Research, 39, 902–909.
19. B. Golden, S. Raghavan, E. Wasil (eds), 2008. The vehicle routing problem. Latest Advances and New
Challenges, Springer Science + Business Media, LLC.
20. M. Held, R.M. Karp, 1962. A dynamic programming approach to sequencing problems. SIAM Journal
of Applied Mathematics, 10, 196–210.
21. S. Irnich, 2008. A unified modeling and solution framework for vehicle routing and local search-based
metaheuristics. SIAM Journal on Computing, 20, 2, 270–287.
22. H. Kellerer, U. Pferchy, D. Pisinger, 2004. Knapsack problems, Springer Verlag, Berlin.
23. Ç. Koç, T. Bektaş, O. Jabali, G. Laporte, 2016. Thirty years of heterogeneous vehicle routing, European Journal of Operational Research, 249, 1–21.
24. A. L. Kok, C. M. Meyer, H. Kopfer, J. M. J. Shutten, 2010. A dynamic programming heuristic for
the vehicle routing problem with time windows and European community social legislation, Transportation
Science, 44.4, 442–454.
25. D. Krushinsky, T. van Woensel, 2015. An approach to the asymmetric multi-depot capacitated arc
routing problem, European Journal of Operational Research, 244, 100–109.
26. R. Lahyani, M. Khemakhem, F. Semet, 2015. Rich vehicle routing problems: From a taxonomy to a
definition, European Journal of Operational Research, 241, 1–14.
27. S. C. H. Leung, Z. Zhang, D. Zhang, X. Hua, M. K. Lim, 2013. A meta-heuristic algorithm for
heterogeneous fleet vehicle routing problems with two-dimensional loading constraints, European Journal of
Operational Research, 225, 199–210.
28. R. Liu, Z. Jiang, 2012. The close-open mixed vehicle routing problem, European Journal of Operational
Research, 220, 349–360.
29. A. D. López-Sánches, A. G. Hernández-Dı́az, D. Vigo, R. Caballero, J. Molina, 2014. A multistart algorithm for a balanced real-world open vehicle routing problem, European Journal of Operational
Research, 238, 104–113.
30. J. Lysgaard, S. Wøhlk , 2014. A branch-and-cut-and-price algorithm for the cumulative capacitated
vehicle routing problem, European Journal of Operational Research, 236, 800–810.
31. J. R. Montoya-Torres, J. L. Franco, S. N. Isaza, H. F. Jiménez, H. Herazo-Padilla, 2015. A
literature review on the vehicle routing problem with multiple depots, Computers & Industrial Engineering,
79, 115–129.
32. H. N. Psaraftis, 1980. An exact algorithm for the single-vehicle, many-to-many, immediate request diala-ride problem, Transportation Science, 14, 130–154.
33. H. N. Psaraftis, 1983. A dynamic programming solution to the single-vehicle, many-to-many dial-a-ride
problem with time windows, Transportation Science, 17, 351–357.
34. J. C. Rivera, H. M. Afsar, C. Prins, 2016. Mathematical formulations and exact algorithm for the
multitrip cumulative capacitated single-vehicle routing problem, European Journal of Operational Research,
249, 893–104.
A framework for the 2-VRP
15
35. K. Soonpracha, A. Mungwattana, G. K. Janssens, T. Manisri , 2014. Heterogeneous VRP review
and conceptual framework, Proceedings of the International Multiconference of Engineers and Computer
Scientists, IMECS 2014, Vol II, Hong Kong.
36. A. Subramanyam, C. E. Gounaris, 2016. A branch-and-cut framework for the consistent traveling salesman problem, European Journal of Operational Research, 248, 384–395.
37. D. Taş, M. Gendreau, O. Jabali, G. Laporte, 2016. The traveling salesman problem with timedependent service tomes, European Journal of Operational Research, 248, 372–383.
38. T. Vidal, T. G. Crainic, M. Gendreau, C. Prins, 2013. Heuristics for multi-attribute vehicle routing
problems: A survey and synthesis, European Journal of Operational Research, 231, 1–21.
39. T. Vidal, T. G. Crainic, M. Gendreau, C. Prins, 2014. Implicit depot assignments and rotations in
vehicle routing heuristics, European Journal of Operational Research, 237, 15–28.
40. T. Vidal, T. G. Crainic, M. Gendreau, C. Prins, 2014. A unified solution framework for multi-attribute
vehicle routing problems, European Journal of Operational Research, 234, 658–673.
41. S. Wøhlk, 2008. A decade of capacitated arc routing, in [19], 29–48.
42. E. E. Zachariadis, C. D. Tarantilis, C. T. Kiranoudis, 2015. The load-dependent vehicle routing
problem and its pick-up and delivery extension, Transportation Research Part B, 71, 158–181.
43. Z. Zhang, H. Qin, W. Zhu, A. Lim, 2012. The single vehicle routing problem with toll-by-weight scheme:
A branch-and-bound approach, European Journal of Operational Research, 220, 295–304.
| 8cs.DS
|
arXiv:1511.04478v1 [cs.DS] 13 Nov 2015
A Backward/Forward Recovery Approach for the
Preconditioned Conjugate Gradient Method
Massimiliano Fasi, Julien Langou, Yves Robert, and Bora Uçar
Abstract
Several recent papers have introduced a periodic verification mechanism to detect silent errors in iterative solvers. Chen [PPoPP’13, pp.
167–176] has shown how to combine such a verification mechanism (a stability test checking the orthogonality of two vectors and recomputing the
residual) with checkpointing: the idea is to verify every d iterations, and
to checkpoint every c × d iterations. When a silent error is detected by
the verification mechanism, one can rollback to and re-execute from the
last checkpoint. In this paper, we also propose to combine checkpointing and verification, but we use algorithm-based fault tolerance (ABFT)
rather than stability tests. ABFT can be used for error detection, but
also for error detection and correction, allowing a forward recovery (and
no rollback nor re-execution) when a single error is detected. We introduce an abstract performance model to compute the performance of all
schemes, and we instantiate it using the preconditioned conjugate gradient algorithm. Finally, we validate our new approach through a set of
simulations.
1
Introduction
Silent errors (or silent data corruptions) have become a significant concern in
HPC environments [1]. There are many sources of silent errors, from bit flips in
cache caused by cosmic radiations, to wrong results produced by the arithmetic
logic unit. The latter source becomes relevant when the computation is performed in the low voltage modeto reduce the energy consumption in large-scale
computations. But the low levels of voltage dramatically reduces the dependability of the system.
The key problem with silent errors is the detection latency: when a silent
error strikes, the corrupted data is not identified immediately, but instead only
when some numerical anomaly is detected in the application behavior. It is clear
that this detection can occur with an arbitrary delay. As a consequence, the de
facto standard method for resilience, namely checkpointing and recovery, cannot
be used directly. Indeed, the method of checkpointing and recovery applies to
fail-stop errors (e.g., hardware crashes): such errors are detected immediately,
and one can safely recover from the last saved snapshot of the application state.
1
On the contrary, because of the detection latency induced by silent errors, it is
often impossible to know when the error struck, and hence to determine which
checkpoint (if any) is valid to safely restore the application state. Even if an
unlimited number of checkpoints could be kept in memory, there would remain
the problem of identifying a valid one.
In the absence of a resilience method, the only known remedy to silent errors
is to re-execute the application from scratch as soon as a silent error is detected.
On large-scale systems, the silent error rate grows linearly with the number of
components, and several silent errors are expected to strike during the execution
of a typical large-scale HPC application [2, 3, 4, 5]. The cost of re-executing the
application one or more times becomes prohibitive, and other approaches need
to be considered.
Several recent papers have proposed to introduce a verification mechanism
to be applied periodically in order to detect silent errors. These papers mostly
target iterative methods to solve sparse linear systems, which are natural candidates to periodic detection. If we apply the verification mechanism every, say,
d iterations, then we have the opportunity to detect the error earlier, namely
at most d − 1 iterations after the actual faulty iteration, thereby stopping the
progress of a flawed execution much earlier than without detection. However,
the cost of the verification may be non-negligible in front of the cost of one
iteration of the application, hence the need to trade off for an adequate value of
d. Verification can consist in testing the orthogonality of two vectors (cheap) or
recomputing the residual (cost of a sparse matrix-vector product, more expensive). We survey several verification mechanisms in Section 2. Note that in all
these approaches a selective reliability model is enforced, where the parts of the
application that are not protected are assumed to execute in a reliable mode.
While verification mechanisms speed up the detection of silent errors, they
cannot provide correction, and thus they cannot avoid the re-execution of the application from scratch. A solution is to combine checkpointing with verification.
If we apply the verification mechanism every d iterations, we can checkpoint
every c × d iterations, thereby limiting the amount of re-execution considerably.
A checkpoint is always valid because it is being preceded by a verification. If an
error occurs, it will be detected by one of the c verifications performed before
the next checkpoint. This is exactly the approach proposed by Chen [6] for a
variety of methods based on Krylov subspaces, including the widely used conjugate Gradient (CG) algorithm. Chen [6] gives an equation for the overhead
incurred by checkpointing and verification, and determines the best values of c
and d by finding a numerical solution of the equation. In fact, computing the
optimal verification and checkpoint intervals is a hard problem. In the case of
pure periodic checkpointing, closed-form approximations of the optimal period
have been given by Young [7] and Daly [8]. However, when combining checkpointing and verification, the complexity of the problem grows. To the best
of our knowledge, there is no known closed-form formula, although a dynamic
programming algorithm to compute the optimal repartition of checkpoints and
verifications is available [9].
For linear algebra kernels, another widely used technique for silent error
2
detection is algorithm-based fault tolerance (ABFT). The pioneering paper of
Huang and Abraham [10] describes an algorithm capable of detecting and correcting a single silent error striking a dense matrix-matrix multiplication by
means of row and column checksums. ABFT protection has been successfully applied to dense LU [11], LU with partial pivoting [12], Cholesky [13]
and QR [14] factorizations, and more recently to sparse kernels like SpMxV
(matrix-vector product) and triangular solve [15]. The overhead induced by
ABFT is usually small, which makes it a good candidate for error detection at
each iteration of the CG algorithm.
The beauty of ABFT is that it can correct errors in addition to detecting
them. This comes at the price of an increased overhead, because several checksums are needed to detect and correct, while a single checksum is enough when
just detection is required. Still, being able to correct a silent error on the fly
allows for forward recovery. No rollback, recovery nor re-execution are needed
when a single silent error is detected at some iteration, because ABFT can correct it, and the execution can be safely resumed from that very same iteration.
Only when two or more silent errors strike within an iteration we do need to
rollback to the last checkpoint. In many practical situations, it is reasonable to
expect no more than one error per iteration, which means that most roll-back
operations can be avoided. In turn, this leads to less frequent checkpoints, and
hence less overhead.
The major contributions of this paper are an ABFT framework to detect
multiple errors striking the computation and a performance model that allows
to compare methods that combine verification and checkpointing. The verification mechanism is capable of error detection, or of both error detection and
correction. The model tries to determine the optimal intervals for verification
and checkpointing, given the cost of an iteration, the overhead associated to
verification, checkpoint and recovery, and the rate of silent errors. Our abstract
model provides the optimal answer to this question, as a function of the cost of
all application and resilience parameters.
We instantiate the model using a CG kernel, preconditioned with a sparse
approximate inverse [16], and compare the performance of two ABFT-based verification mechanisms. We call the first scheme, capable of error detection only,
ABFT-Detection and the second scheme, which enhances the first one by providing single error correction as well, ABFT-Correction. Through numerical
simulations, we compare the performance of both schemes with Online-Detection, the approach of Chen [6] (which we extend to recover from memory errors
by checkpointing the sparse matrix in addition to the current iteration vectors).
These simulations show that ABFT-Correction outperforms both OnlineDetection and ABFT-Detection for a wide range of fault rates, thereby
demonstrating that combining checkpointing with ABFT correcting techniques
is more efficient than pure checkpointing for most practical situations.
Our discussion focuses on the sequential execution of iterative methods. Yet,
all our techniques extend to parallel implementation based on the message passing paradigm (with using, e.g., MPI). In an implementation of SpMxV in such
a setting, the processing elements (or processors) hold a part of the matrix and
3
the input vector, and hold a part of the output vector at the end. A recent
exposition of different algorithms can be found elsewhere [17]. Typically, the
processors perform scalar multiply operations on the local matrix and the input vector elements, when all required vector elements have been received from
other processors. The implementations of the MPI standard guarantees correct message delivery, i.e., checksums are incorporated into the message so as
to prevent transmission errors (the receives can be done in-place and hence are
protected). However, the receiver will obviously get corrupted data if the sender
sends corrupted data. Silent error can indeed strike at a given processor during
local scalar multiply operations. Performing error detection and correction locally implies global error detection and correction for the SpMxV. Note that, in
this case, the local matrix elements can form a matrix which cannot be assumed
to be square in general (for some iterative solvers they can be). Furthermore,
the mean time between failures (MTBF) reduces linearly with the number of
processors. This is well-known for memoryless distributions of fault inter-arrival
times and remains true for arbitrary continuous distributions of finite mean [18].
Therefore, resilient local matrix vector multiplies are required for resiliency in
a parallel setting.
The rest of the paper is organized as follows. Section 2 provides an overview
of related work. Section 3 provides background on ABFT techniques for the
PCG algorithm, and presents both the ABFT-Detection and ABFT-Correction approaches. Section 5 is devoted to the abstract performance model.
Section 6 reports numerical simulations comparing the performance of ABFTDetection, ABFT-Correction and Online-Detection. Finally, we outline main conclusions and directions for future work in Section 7.
2
Related work
We classify related work along the following topics: silent errors in general,
verification mechanisms for iterative methods, and ABFT techniques.
2.1
Silent errors
Considerable efforts have been directed at error-checking to reveal silent errors. Error detection is usually very costly. Hardware mechanisms, such as
ECC memory, can detect and even correct a fraction of errors, but in practice
they are complemented with software techniques. The simplest technique is
triple modular redundancy and voting [19], which induces a costly verification.
For high-performance scientific applications, process replication (each process
is equipped with a replica, and messages are quadruplicated) is proposed in the
RedMPI library [20]. Elliot et al. [21] combine partial redundancy and checkpointing, and confirm the benefit of dual and triple redundancy. The drawback
is that twice the number of processing resources is required (for dual redundancy). A comprehensive list of general-purpose techniques and references is
provided by Lu et al. [22].
4
Application-specific information can be very useful to enable ad-hoc solutions, which dramatically decrease the cost of detection. Many techniques have
been advocated. They include memory scrubbing [23] and ABFT techniques
(see below).
As already stated, most papers assume on a selective reliability setting [24,
25, 26, 27]. It essentially means that one can choose to perform any operation
in reliable or unreliable mode, assuming the former to be error-free but energy consuming and the latter to be error-prone but preferable from an energy
consumption point of view.
2.2
Iterative methods
Iterative methods offer a wide range of ad-hoc approaches. For instance, instead of duplicating the computation, Benson et al. [28] suggest coupling a
higher-order with a lower-order scheme for PDEs. Their method only detects
an error but does not correct it. Self-stabilizing corrections after error detection in the CG method are investigated by Sao and Vuduc [27]. Heroux and
Hoemmen [29] design a fault-tolerant GMRES capable of converging despite
silent errors. Bronevetsky and de Supinski [30] provide a comparative study of
detection costs for iterative methods.
As already mentioned, a nice instantiation of the checkpoint and verification
mechanism that we study in this paper is provided by Chen [6], who deals with
sparse iterative solvers. For PCG, the verification amounts to checking the
orthogonality of two vectors and to recomputing and checking the residual (see
Section 3 for further details).
As already mentioned, our abstract performance model is agnostic of the
underlying error-detection technique and takes the cost of verification as an
input parameter to the model.
2.3
ABFT
The very first idea of algorithm-based fault tolerance for linear algebra kernels
is given by Huang and Abraham [10]. They describe an algorithm capable of
detecting and correcting a single silent error striking a matrix-matrix multiplication by means of row and column checksums. This germinal idea is then elaborated by Anfinson and Luk [31], who propose a method to detect and correct
up to two errors in a matrix representation using just four column checksums.
Despite its theoretical merit, the idea presented in their paper is actually applicable only to relatively small matrices, and is hence out of our scope. Bosilca
et al. [32] and Du et al. [11] present two relatively recent survey.
The problem of algorithm-based fault-tolerance for sparse matrices is investigated by Shantharam et al. [15], who suggest a way to detect a single error in
an SpMxV at the cost of a few additional dot products. Sloan et al. [33] suggest
that this approach can be relaxed using randomization schemes, and propose
several checksumming techniques for sparse matrices. These techniques are less
5
effective than the previous ones, not being able to protect the computation from
faults striking the memory, but provide an interesting theoretical insight.
3
CG-ABFT
We streamline our discussion on the CG method, however, the techniques that
we describe are applicable to any iterative solver that use sparse matrix vector
multiplies and vector operations. This list includes many of the non-stationary
iterative solvers such as CGNE, BiCG, BiCGstab where sparse matrix transpose
vector multiply operations also take place. In particular, we consider a PCG
variant where the application of the preconditioner reduces to the computation
of two SpMxV with triangular matrices [16], which are a sparse factorization of
an approximate inverse of the coefficient matrix. In fact, the model discussed
in this paper can be profitably employed for any sparse inverse preconditioner
that can be applied by means of one or more SpMxV.
We first provide a background on the CG method and overview both Chen’s
stability tests [6] and our ABFT protection schemes.
Algorithm 1 The PCG algorithm.
Input: A, M ∈ Rn×n , b, x0 ∈ Rn , ε ∈ R
Output: x ∈ Rn : kAx − bk ≤ ε
1: r0 ← b − Ax0 ;
2: z0 ← M⊺ Mr0 ;
3: p0 ← z0 ;
4: i ← 0;
5: while kri k > ε (kAk · kr0 k + kbk)
6:
q ← Api ;
7:
αi ← kri k2 /p⊺i q;
8:
xi+1 ← xi + α pi ;
9:
ri+1 ← ri − α q;
10:
zi+1 ← M⊺ Mri+1 ;
11:
β ← kri+1 k2 / kri k2 ;
12:
pi+1 ← zi+1 + β pi ;
13:
i ← i + 1;
14: end while
15: return xi ;
The code for the variant of the PCG method we use is shown in Algorithm 1.
The main loop features three sparse matrix-vector multiply, two inner products
(for p⊺i q and kri+1 k2 ), and three vector operations of the form axpy.
Chen’s stability tests [6] amount to checking the orthogonality of vectors
p⊺
q
i+1
pi+1 and q, at the price of computing kpi+1
kkqi k , and to checking the residual
at the price of an additional SpMxV operation Axi − b. The dominant cost of
these verifications is the additional SpMxV operation.
Our only modification to Chen’s original approach is that we also save the
sparse matrix A in addition to the current iteration vectors. This is needed
6
when a silent error is detected: if this error comes for a corruption in data
memory, we need to recover with a valid copy of the data matrix A. This holds
for the three methods under study, Online-Detection, ABFT-Detection
and ABFT-Correction, which have exactly the same checkpoint cost.
We now give an overview of our own protection and verification mechanisms.
We use ABFT techniques to protect the SpMxV, its computations (hence the
vector q), the matrix A and the input vector pi . Since ABFT protection for
vector operations is as costly as repeated computation, we use triple modular
redundancy (TMR) for them for simplicity.
Although theoretically possible, constructing ABFT mechanism to detect up
to k errors is practically not feasible for k > 2. The same mechanism can be
used to correct up to ⌊k/2⌋. Therefore, we focus on detecting up to two errors
and correcting the error if there was only one. That is, we detect up to two
errors in the computation q ← Api (two entries in q are faulty), or in pi , or
in the sparse representation of the matrix A. With TMR, we assume that the
errors in the computation are not overly frequent so that two out of three are
correct (we assume errors do not strike the vector data here). Our fault-tolerant
PCG versions thus have the following ingredients: ABFT to detect up to two
errors in the SpMxV and correct the error, if there was only one; TMR for vector
operations; and checkpoint and roll-back in case errors are not correctable.
We assume the selective reliability model in which all checksums and checksum related operations are non-faulty, also the tests for the orthogonality checks
are non-faulty.
4
ABFT-SpMxV
Here, we discuss the proposed ABFT method for the SpMxV (combining ABFT
with checkpointing is described later in Section 5.2). The proposed ABFT mechanisms are described for detecting single errors (Section 4.1), multiple errors
(Section 4.2), and correcting a single error (Section 4.3).
4.1
Single error detection
The overhead of the standard single error correcting ABFT technique is too
high for the sparse matrix-vector product case. Shantharam et al. [15] propose
a cheaper ABFT-SpMxV algorithm that guarantees the detection of a single
error striking either the computation or the memory representation of the two
input operands (matrix and vector). Because their results depend on the sparse
storage format adopted, throughout the paper we will assume that sparse matrices are stored in the compressed storage format by rows (CSR), that is by
means of three distinct arrays, namely Colid ∈ Nnnz(A) , Val ∈ Rnnz(A) and
Rowidx ∈ Nn+1 [34, Sec. 3.4]). Here nnz(A) is the number of non-zero entries
in A.
Shantharam et al. can protect y ← Ax, where A ∈ Rn×n and x, y ∈ Rn .
7
To perform error detection, they rely on a column checksum vector c defined by
cj =
n
X
ai,j
(1)
i=0
and an auxiliary copy x′ of the x vector. After having
the actual
Pperformed
n
SpMxV, to validate the result, it suffices to compute i=1 yi , c⊺ x and c⊺ x′ ,
and to compare their values. It can be shown [15] that in case of no errors,
these three quantities carry the same value, whereas if a single error strikes
either the memory or the computation, one of them must differ from the other
two. Nevertheless, this method requires A to be strictly diagonally dominant,
a condition that seems to restrict too much the practical applicability of their
method. Shantharam et al. need this condition to ensure detection of errors
striking an entry of x corresponding to a zero checksum column of A. We
further analyze that case and show how to overcome the issue without imposing
any restriction on A.
A nice way to characterize the problem is expressing it in geometrical terms.
Consider the computation of a single entry of the checksum as
(w A)j =
⊺
n
X
wi ai,j = w⊺ Aj ,
i=1
where w ∈ Rn denotes the weight vector and Aj the j-th column of A. Let
us now interpret such an operation as the result of the scalar product h·, ·i :
Rn × Rn → R defined by hu, vi 7→ u⊺ v. It is clear that a checksum entry is
zero if and only if the corresponding column of the matrix is orthogonal to the
weight vector. In (1), we have chosen w to be such that wi = 1 for 1 ≤ i ≤ n,
in order to make the computation easier. Let us see now what happens without
this restriction.
The problem reduces to finding a vector w ∈ Rn that is not orthogonal to
any vector out of a basis B = {b1 , . . . , bn } of Rn – the rows of the input matrix.
Each of these n vectors is perpendicular to a hyperplane hi of Rn , and w does
not verify the condition
hw, bi i 6= 0,
(2)
for any i, if and only if it lies on hi . Since the Lebesgue measure in Rn of an
hyperplane of Rn itself is zero, the union of these hyperplanes is measurable
and has measure 0. Therefore, the probability that a vector w randomly picked
in Rn does not satisfy condition (2) for any i is zero.
Nevertheless, there are many reasons to consider zero checksum columns.
First of all, when working with finite precision, the number of elements in Rn
one can have is finite, and the probability of randomly picking a vector that
is orthogonal to a given one could be larger than zero. Moreover, a coefficient
matrix usually comes from the discretization of a physical problem, and the
distribution of its columns cannot be considered as random. Finally, using a
randomly chosen vector instead of (1, . . . , 1)⊺ increases the number of required
8
floating point operations, causing a growth of both the execution time and
the number of rounding errors (see Section 6). Therefore, we would like to
keep w = (1, . . . , 1)⊺ as the vector of choice, in which case we need to protect
SpMxV with matrices having zero column sums. There are many matrices with
this property, for example the Laplacian matrices of graphs [35, Ch. 1].
Algorithm 2 Shifting checksum algorithm.
Input: A ∈ Rn×n , x ∈ Rn
Output: y ∈ Rn such that y = Ax or the detection of a single error
1: x′ ← x;
2: [w, c, k, cr ] = computeChecksum(A);
3: return SpMxV(A, x, x′ , w, c, k, cr );
4: function computeChecksum(A)
5:
Generate w ∈ Rn+1 ;
6:
w ← w1:n ;
7:
c ← w⊺ A;
8:
if min(| c |) = 0 ;
9:
Find k that verifies (4);
10:
c ← c + k;
11:
12:
cr ← w⊺ Rowindex ;
return w, c, k, cr ;
13: function SpMxV(A, x, x′ , w, c, k, cr )
14:
w ← w1:n ;
15:
sr ← 0;
16:
for i ← 1 to n
17:
yi ← 0;
18:
sr ← sr + Rowindex i ;
19:
for j ← Rowindex i to Rowindex i+1 − 1
20:
ind ← Colid j ;
21:
yi ← yi + Val j · xind ;
22:
23:
24:
25:
26:
27:
28:
29:
30:
yn+1 ← k w⊺ x′ ;
cy ← w⊺ y;
dx ← c⊺ x;
dx′ ← c⊺ x′ ;
dr ← cr − sr ;
if dx = 0 ∧ dx′ = 0 ∧ dr = 0
return y1:n ;
else
error (“Soft error is detected”);
In Algorithm 2, we propose an ABFT SpMxV method that uses weighted
checksums and does not require the matrix to be strictly diagonally dominant.
The idea is to compute the checksum vector and then shift it by adding to
all entries a constant value chosen so that all elements of the new vector are
different from zero. We give the generalized result in Theorem 1.
9
Theorem 1 (Correctness of Algorithm 2). Let A ∈ Rn×n be a square matrix,
let x, y ∈ Rn be the input and output vector respectively, and let x′ = x. Let us
assume that the algorithm performs the computation
e x,
e ← Ae
y
(3)
e ∈ Rn×n and x
e ∈ Rn are the possibly faulty representations of A and x
where A
e ∈ Rn is the possibly erroneous result of the sparse matrixrespectively, while y
vector product. Let us also assume that the encoding scheme relies on
1. an auxiliary checksum vector
#
" n
n
X
X
ai,n + k ,
ai,1 + k, . . . ,
c=
i=1
i=1
where k is such that
cj =
n
X
ai,j + k 6= 0,
(4)
i=1
for 1 ≤ j ≤ n,
2. an auxiliary checksum yn+1 = k
Pn
i=i
x
ei ,
3. an auxiliary counter sr initialized to 0 and updated at runtime by adding
the value of the hit element each time the Rowindex array is accessed (line
20 of Algorithm 2),
P
4. an auxiliary checksum cr = ni=1 Rowindex i ∈ N.
Then, a single error in the computation of the SpMxV causes one of the following
conditions to fail:
P
e = n+1
i. c⊺ x
ei ,
i=1 y
P
n+1
ii. c⊺ x′ = i=1 yei ,
iii. sr = cr .
Proof. We will consider three possible cases, namely
a. a faulty arithmetic operation during the computation of y,
b. a bit flip in the sparse representation of A,
c. a bit flip in an element of of x.
Case a. Let us assume, without loss of generality, that the error has struck
at the pth position of y, which implies yei = yi for 1 ≤ i ≤ n with i 6= p and
10
yep = yp +ε, where ε ∈ R \{0} represents the value of the error that has occurred.
e gives
Summing up the elements of y
n+1
X
i=1
yei
=
=
=
n X
n
X
i=1 j=1
n
X
j=1
⊺
ai,j x
ej + k
n
X
j=1
cj x
ej + ε
x
ej + ε
e + ε,
c x
that violates condition (i).
Case b. A single error in the A matrix can strike one of the three vectors
that constitute its sparse representation:
• a fault in Val that alters the value of an element ai,j implies an error in the
computation of yei , which leads to the violation of the safety condition (i)
because of (a),
• a variation in Colid can zero out an element in position ai,j shifting its
value in position ai,j ′ , leading again to an erroneous computation of yei ,
• a transient fault in Rowindex entails an incorrect value of sr and hence a
violation of condition (iii).
Case c. Let us assume, without loss of generality, an error in position p of
x. Hence we have that x
ei = xi for 1 ≤ i ≤ n with i 6= p and x
ep = xp + ε, for
e gives
some ε ∈ R \ {0}. Noting that x = x′ , the sum of the elements of y
n+1
X
i=1
yei
=
=
n X
n
X
i=1 j=1
n X
n
X
ai,j x
ej + k
ai,j xj + k
=
cj xj + ε
= c x +ε
n
X
i=1
j=1
⊺ ′
n
X
Pn
i=1
x
ej
xj + ε
ai,p + k
!
!
n
X
ai,p + εk
i=1
ai,p + k ,
i=1
that violates (ii) since
j=1
n
X
j=1
i=1 j=1
n
X
n
X
ai,p + k 6= 0 by definition of k.
Let us remark that computeChecksum in Algorithm 2 does not require the
input vector x of SpMxV as an argument. Therefore, in the common scenario
of many SpMxV with the same matrix, it is enough to invoke it once to protect
several matrix-vector multiplications. This observation will be crucial when
discussing the performance of these checksumming techniques.
11
Table 1: Overhead comparison for Algorithm 2 and Algorithm 3. Here n denotes
the size of the matrix and n′ the number of null sum columns.
Algorithm 2
Algorithm 3
n
n
2n
n
-
n
n′
2n + 2n′
n + n′
n′
5n
4n + 5n′
initialization of y
computation of yn+1
SpMxV overhead
checksum check
computation of cy and cyb
b
computation of y + y
Total SpMxV overhead
Shifting the sum checksum vector by an amount is probably the simplest
deterministic approach to relax the strictly diagonal dominance hypothesis, but
it is not the only one. An alternative solution is described in Algorithm 3, which
basically exploits the distributive property of matrix multiplication over matrix
addition. The idea is to split the original matrix A into two matrices of the
b such that no column of either matrix has a zero checksum.
same size, A and A,
b ← Ab
Two standard ABFT multiplications, namely y ← Ax and y
x, are then
performed. If no error occurs neither in the first nor in the second computation,
b is computed in reliable mode and then returned. Let us
the sum of y and y
b to be much smaller than
note that, as we expect the number of non-zeros of A
b
b vector.
n, we store sparsely both the checksum vector of A and the y
We do not write down an extended proof of the correctness of this algorithm,
and limit ourselves to a short sketch. We consider the same three cases as
in the proof of Theorem 1, without introducing any new idea. An error in
b can be detected using the dot product between the
the computation of y or y
corresponding column checksum and the x error. An error in A can be detected
b , as the matrix loop structure of the
by either cr or an erroneous entry in y or y
sparse multiplication algorithm has not been changed. Finally, an error in the
b differ from
pth component of x would make the sum of the entries of y and y
c⊺ x′ and b
c⊺ x′ , respectively.
The evaluation of the performance of the two algorithms, though straightforward from the point of view of the computational cost, has to be carefully
assessed in order to devise a valid and practical trade-off between Algorithm 2
and Algorithm 3.
In both cases computeChecksum introduces an overhead of O (nnz(A)),
but the shift version should in general be faster containing less assignments than
its counterpart, and this changes the constant factor hidden by the asymptotic
notation. Nevertheless, as we are interested in performing many SpMxV with a
same matrix, this pre-processing overhead becomes negligible.
The function SpMxV has to be invoked once for each multiplication, and
hence more care is needed. Copying x and initializing y both require n operations, and the multiplication is performed in time O (nnz(A)), but the split
version pays an n′ more to read the values of the sparse vector b. The cost
12
Algorithm 3 Splitting checksum algorithm.
Input: A ∈ Rn×n , x ∈ Rn
Output: y ∈ Rn such that y = Ax or the detection of a single error
1: [c, k, cr ] = computeChecksum(A);
2: return SpMxV(A, x, c, b
c, b, cr );
3: function computeChecksum(A)
4:
c, m ← 0;
5:
for i ← 1 to nnz(A)
6:
ind ← Colid i ;
7:
cind ← cind + Val i ;
8:
mind ← i;
9:
10:
11:
12:
13:
14:
15:
16:
k ← 0;
for i ← 1 to n
if ci = 0 ∧ mi 6= 0
bmi ← true;
ci ← Val mi ;
b
ci ← ci − b
ci ;
P
cr ← n
Rowindex
i;
i=1
return c, b
c, b, cr ;
17: function SpMxV(A, x, c, b
c, b, cr )
18:
x′ ← x;
19:
for i ← 1 to n
20:
yi ← 0;
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
sr ← 0;
for i ← 1 to n
sr ← sr + Rowindex i ;
for j ← Rowindex i to Rowindex i+1 − 1
ind ← Colid j ;
if bj
ybi ← ybi + Val j · xind ;
else
yi ← yi + Val j · xind ;
P
P
cy ← n
y
;
cyb ← n
bi ;
i
i=1
i=1 y
dx ← c⊺ x − cy ; dxb ← b
c⊺ x − cyb;
c⊺ x′ − cyb;
dx′ ← c⊺ x′ − cy ; dxb′ ← e
dr ← cr − sr ;
if dx = 0 ∧ dx′ = 0 ∧ dxe = 0 ∧ dxe′ = 0 ∧ dr = 0
b;
return y + y
else
error (“Soft error is detected”);
13
of the verification step depends instead on the number of zeroes of the original
checksum vector, that is also the number of non-zero elements of the sparse
vector b
c. Let us call this quantity n′ . Then the overhead is 4n for the shifting
and 3n + 3n′ for the splitting, that requires also the sum of two sparse vectors
to return the result. Hence, as summarized in Table 1, the two methods bring
different overhead into the computation. Comparing them, it is immediate to
see that the shifting method is cheaper when
n′ >
n
,
5
while it has more operations to do when the opposite inequality holds. For
the equality case, we can just choose to use the first method because of the
cheaper preprocessing phase. In view of this observation, it is possible to devise
a simple algorithm that exploits this trade-off to achieve better performance. It
suffices to compute the checksum vector of the input matrix, count the number
of non-zeros and choose which detection method to use accordingly.
We also note that by splitting the matrix A into say ℓ pieces and checksumming each piece separately we can possibly protect A from up to ℓ errors, by
protecting each piece against a single error (obviously the multiple errors should
hit different pieces).
4.2
Multiple error detection
With some effort, the shifting idea in Algorithm 2 can be extended to detect
errors striking a single SpMxV. Let us consider the problem of detecting up to
k errors in the computation of y ← Ax introducing an overhead of O (kn). Let
k weight vectors w(1) , . . . , w(k) ∈ Rn be such that any sub-matrix of
i
h
W = w(1) w(2) . . . w(k)
has full rank. To build our ABFT scheme let us note that, if no error occurs,
for each weight vector w(ℓ) it holds that
" n
#
n
X
X (ℓ)
(ℓ)
(ℓ)⊺
w A=
wi ai,n ,
wi ai,1 , . . . ,
i=1
i=1
and hence that
w(ℓ) Ax
⊺
=
n
X
(ℓ)
wi ai,1 x1 + · · · +
i=1
i=1
=
n X
n
X
n
X
(ℓ)
wi ai,j xj .
i=1 j=1
14
(ℓ)
wi ai,n xn
Similarly, the sum of the entries of y weighted with the same w(ℓ) is
n
X
(ℓ)
wi yi
(ℓ)
=
w1 y1 + · · · + wn(ℓ) yn
=
w1
=
j=1
n X
n
X
i=1
(ℓ)
n
X
a1,j xj + · · · + wn(ℓ)
n
X
an,j xj
j=1
(ℓ)
wi ai,j xj ,
i=1 j=1
and we can conclude that
n
X
i=1
⊺
(ℓ)
wi yi = w(ℓ) A x,
(ℓ)
for any w with 1 ≤ ℓ ≤ k.
To convince ourself that with these checksums it is actually possible to detect
up to k errors, let us suppose that k ′ errors, with k ′ ≤ k, occur in positions
e the faulty vector where yepi = ypi + εpi for
p1 , . . . , pk′ , and let us denote by y
εpi ∈ R \ {0} and 1 ≤ i ≤ k ′ and yei = yi otherwise. Then for each weight vector
we have
k′
n
n
X
X
X
(ℓ)
(ℓ)
εpj .
wp(ℓ)
wi yi =
wi yei −
j
j=1
i=1
i=1
′
Said otherwise, the occurrence of the k errors is not detected if and only if, for
1 ≤ ℓ ≤ k, all the εpi respect
′
k
X
εpj = 0 .
wp(ℓ)
j
(5)
j=1
′
We claim that there cannot exist a vector (εp1 , . . . , εp′k )⊺ ∈ Rk \ {0} such that
all the conditions in (5) are simultaneously verified. These conditions can be
expressed in a more compact way as a linear system
(1)
(1)
0
εp1
wp1 · · · wpk′
. .
.
..
..
.
.
. .. = .. .
.
(k)
wp1
···
(k)
wpk′
εpk′
0
Denoting by W∗ the coefficient matrix of this system, it is clear that the
errors cannot be detected if only if (εp1 , . . . εpk′ )⊺ ∈ ker(W∗ ) \ {0}. Because of
the properties of W, we have that rk(W∗ ) = k. Moreover, it is clear that the
rank of the augmented matrix
(1)
(1)
wp1 · · · wpk′ 0
.
..
..
..
.
.
.
.
.
(k)
(k)
wp1 · · · wpk′ 0
15
is k as well. Hence, by means of the Rouché–Capelli theorem, the solution
of the system is unique and the null space of W∗ is trivial. Therefore, this
construction can detect the occurrence of k ′ errors during the computation of
y by comparing the values of the weighted sums y⊺ w(ℓ) with the result of the
⊺
dot product (w(ℓ) A)x, for 1 ≤ ℓ ≤ k.
However, to get a true extension of the algorithm described in the previous
section, we also need to make it able to detect errors that strike the sparse
representation of A and that of x. The first case is simple, as the k errors can
e , or in Rowindex ,
strike the Val or Colid arrays, leading to at most k errors in y
where they can be caught using k weighted checksums of the Rowindex vector.
Detection in x is much trickier, since neither the algorithm just described
nor a direct generalization of Algorithm 2 can manage this case. Nevertheless,
a proper extension of the shifting technique is still possible. Let us note that
there exists a matrix M ∈ Rk×n such that
W⊺ A + M = W.
The elements of such an M can be easily computed, once that the checksum
e ∈ Rn be the faulty vector, defined by
rows are known. Let x
1 ≤ i ≤ k′ ,
xi + εpi ,
x
ei =
x
otherwise.
i
e = Ae
for some k ′ ≤ k, and let us define y
x. Now, let us consider a checksum
vector x′ ∈ Rn such that x′ = x and let assume that it cannot be modified by
a transient error. For 1 ≤ ℓ ≤ k, it holds that
n
X
i=1
(ℓ)
wi yei
+
n
X
j=1
mℓ,j x
ej =
=
n X
n
X
+
i=1 j=1
n
n X
X
(ℓ)
wi ai,j xj +
=
n
X
n
X
n
X
(ℓ)
wi ai,j
i=1
j=1
n
X
ε pi
n
X
j=1
(ℓ)
wi ai,j
!
xj +
(ℓ)
wj aj,pi
k
X
i=1
n
X
+ mℓ,j
i=1
ε pi
′
mℓ,j xj +
j=1
!
′
mℓ,j xj +
j=1
n
X
j=1
k
X
i=1
j=1 i=1
=
′
(ℓ)
wi ai,j xj
k
X
+
n
X
′
mℓ,j xj +
j=1
n
X
j=1
(ℓ)
ε pi w pi
′
xj +
(ℓ)
ε pi w pi
i=1
′
=w
(ℓ)⊺
x+
k
X
(ℓ)
ε pi w pi
i=1
′
⊺
= w(ℓ) x′ +
k
X
(ℓ)
ε pi w pi .
i=1
Therefore, an error is not detected if and only if the linear system
(1)
(1)
0
εp1
wp1 · · · wpk′
. .
.
..
..
.
.
.
. . = ..
.
(k)
wp1
···
(k)
wpk′
16
εpk′
0
i=1
εpi ml,p
(ℓ)
wj aj,pi + ml,p
i=1
k
X
k
X
i
i
has a non-trivial solution. But we have already seen that such a situation can
never happen, and we can thus conclude that our method, whose pseudocode
we give in Algorithm 4, can also detect up to k errors occurring in x. Therefore,
we have proven the following theorem.
Theorem 2 (Correctness of Algorithm 4). Let us consider the same notation
as in Theorem 1. Let W ∈ Rn+1×n be a matrix such that any square submatrix
is full rank, and let us denote by W ∈ Rn×n the matrix of its first n rows. Let
us assume an encoding scheme that relies on
⊺
1. an auxiliary checksum matrix C = (W⊺ A) ,
2. an auxiliary checksum matrix M = W − C,
3. a vector of auxiliary counters sRowindex initialized to the null vector and
updated at runtime as in lines 16 – 17 of Algorithm 4),
4. an auxiliary checksum vector cRowindex = W⊺ Rowindex .
Then, up to k errors striking the computation of y or the memory locations that
store A or x, cause one of the following conditions to fail:
i. W⊺ y = C⊺ x,
ii. W⊺ (x′ − y),
iii. sRowindex = cRowindex .
Let us note that we have just shown that our algorithm can detect up to k
errors striking only A, or only x or only the computation. Nevertheless, this
result holds even when the errors are distributed among the possible cases, as
long as at most k errors rely on the same checkpoint.
It is clear that the execution time of the algorithm depends on both nnz(A)
and k. For the computeChecksum function, the cost is, assuming that the
weight matrix W is already known, O (k nnz(A)) for the computation of C,
and O (kn) for the computation of M and cRowindex . Hence the number of performed operations is O (k nnz(A)). The overhead added to the SpMxV depends
instead on the computation of four checksum matrices that lead to a number of
operations that grows asymptotically as kn.
4.3
Single error correction
We now discuss single error correction, using Algorithm 4 as a reference. We describe how a single error striking either memory or computation can be not only
detected but also corrected at line 27. We use only two checksum vectors, that
is, we describe correction of single errors assuming that two errors cannot strike
the same SpMxV. By the end of the section, we will generalize this approach
and discuss how single error correction and double error detection can be performed concurrently by exploiting three linearly independent checksum vectors.
17
Algorithm 4 Shifting checksum algorithm for k errors detection.
Input: A ∈ Rn×n , x ∈ Rn
Output: y ∈ Rn such that y = Ax or the detection of up to k errors
1: x′ ← x
c] = computeChecksums(A, k);
2: [W, C, M, e
3: return SpMxV(A, x, x′ , W, C, M, k, e
c);
4: function computeChecksums(A,
h
i k)
5:
6:
7:
8:
9:
10:
Generate W = w(1) . . . w(k) ∈ Rn+1×n ;
W ← W1:n,∗ ∈ Rn×k
C⊺ ← W⊺ A;
M ← W − C;
cRowindex ← W⊺ Rowindex ;
return W, C, M, cRowindex ;
c)
11: function SpMxV(A, x, x′ , W, C, M, k, e
12:
W ← W1:n,∗ ∈ Rn×k
e
13:
s ← [0, . . . , 0];
14:
for i ← 1 to n
15:
yi ← 0;
16:
for j = 1 to k
17:
sej ← sej + wij Rowindex i ;
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
for j ← Rowindex i to Rowindex i+1 − 1
ind ← Colid j ;
yi ← yi + Val j · xind ;
⊺
dx ← W y − C⊺ x;
dx′ ← W⊺ (x′ − y) − M⊺ x;
dr ← e
c−e
s;
if dx = 0 ∧ dx′ = 0 ∧ dr = 0
return y;
else
error (“Soft errors are detected”);
18
Whenever a single error is detected, regardless of its location (computation or
memory), it is corrected by means of a succession of various steps. When one
or more errors are detected, the correction mechanism tries to determine their
multiplicity and, in case of a single error, what memory locations have been
corrupted or what computation has been miscarried. Errors are then corrected
using the values of the checksums and, if need be, partial recomputations of the
result are performed.
As we did for multiple error detection, we require that any 2 × 2 submatrix
of W ∈ Rn×2 has full rank. The simplest example of weight matrix having this
property is probably
1 1
1 2
W = 1 3 .
.. ..
. .
1 n
To detect errors striking Rowidx , we compute the ratio ρ of the second
component of dr to the first one, and check whether its distance from an integer
is smaller than a certain threshold parameter ε. If this distance is smaller, the
algorithm concludes that the σth element, where σ = Round(ρ) is the nearest
integer to ρ, of Rowidx is faulty, performs the correction by subtracting the first
component of dr from Rowidx σ , and recomputes yσ and yσ−1 , if the error in
Rowindexσ is a decrement; or yσ+1 if it was an increment. Otherwise, it just
emits an error.
The correction of errors striking Val, Colid and the computation of y are
performed together. Let now ρ be the ratio of the second component of dx to
the first one. If ρ is near enough to an integer σ, the algorithm computes the
checksum matrix C′ = W⊺ A and considers the number zC
e of non-zero columns
e = |C − C′ |. At this stage, three cases are possible:
of the difference matrix C
• If zC
e = 0, then the error is in the computation of yσ , and can be corrected
by simply recomputing this value.
• If zC
e = 1, then the error has struck an element of Val . Let us call f the
e The algorithm finds the element of
index of the non-zero column of C.
Val corresponding to the entry at row σ and column f of A and corrects
it by using the column checksums much like as described for Rowidx .
Afterwards, yd is recomputed to fix the result.
• If zC
e = 2, then the error concerns an element of Colid . Let us call f1 and
f2 the index of the two non-zero columns and m1 , m2 the first and last
elements of Colid corresponding to non-zeros in row σ. It is clear that
there exists exactly one index m∗ between m1 and m2 such that either
Colid m∗ = f1 or Colid m∗ = f2 . To correct the error it suffices to switch
the current value of Colid m∗ , i.e., putting Colid m∗ = f2 in the former case
and Colid m∗ = f1 in the latter. Again, yσ has to be recomputed.
19
• if zC
e > 2, then errors can be detected but not corrected, and an error is
emitted.
To correct errors striking x, the algorithm computes ρ, that is the ratio of the
second component of dx′ to the first one, and checks that the distance between
d and the nearest integer σ is smaller than ε. Provided that
Pnthis condition is
verified, the algorithm computes the value of the error τ = i=1 xi − cxσ and
corrects xσ = xσ − τ . The result is updated by subtracting from y the vector
yτ = Axτ , where xτ ∈ Rn×n is such that xτσ = τ and xτi = 0 otherwise.
Let us now investigate how detection and correction can be combined and
let us give some details about the implementation of ABFT-Correction as
defined in Section 3. Indeed, note that double errors could be shadowed when
using Algorithm 2, although the probability of such an event is negligible.
Let us restrict ourselves to an easy case, without considering errors in x. As
usual, we compute the column checksums matrix
⊺
C = (W⊺ A) ,
and then compare the two entries of C⊺ x ∈ R2 with the weighted sums
ye1c =
and
ye2c =
n
X
i=1
n
X
yei
ie
yi
i=1
e is the possibly faulty vector computed by the algorithm. It is clear
where y
e c − c = 0.
that if no error occurs, the computation verifies the condition δ = y
δ2
Furthermore, if exactly one error occurs, we have δ1 , δ2 6= 0 and δ1 ∈ N, and if
two errors strike the vectors protected by the checksum c, the algorithm is able
to detect them by verifying that δ 6= 0.
At this point it is natural to ask whether this information is enough to build
a working algorithm or some border cases can bias its behavior. In particular,
when δδ12 = p ∈ N, it is not clear how to discern between single and double
errors. Let ε1 , ε2 ∈ R \ {0} be the value of two errors occurring at position p1
e ∈ Rn be such that
and p2 respectively, and let y
1 ≤ i ≤ n, i 6= p1 , p2
yi ,
yi + ε1 ,
i = p1
yei =
.
yi + ε2 ,
i = p2
Then the conditions
δ1
=
ε1 + ε2 ,
(6)
δ2
=
p1 ε 1 + p2 ε 2 ,
(7)
hold. Therefore, if ε1 and ε2 are such that
p (ε1 + ε2 ) = p1 ε1 + p2 ε2 ,
20
(8)
it is not possible to distinguish these two errors from a single error of value
ε1 + ε2 occurring in position p. This issue can be solved by introducing a new
set of weights and hence a new row of column checksums. Let us consider a
c ∈ Rn×3 that includes a quadratic weight vector
weight matrix W
1 1 1
1 2 4
c =
W
1 3 9 ,
.. ..
..
. .
.
1 n
n2
and the tridimensional vector
c ⊺A x − W
c ⊺y
e,
δ̂ = W
whose components can be expressed as
δ1
δ2
= ε1 + ε2
= p1 ε 1 + p2 ε 2
δ2
= p21 ε1 + p22 ε2 .
To be confused with a single error in position p, ε1 and ε2 have to be such that
p (ε1 + ε2 ) = p1 ε1 + p2 ε2
and
p2 (ε1 + ε2 ) = p21 ε1 + p22 ε2
hold simultaneously for some p ∈ N. In other words, possible values of the
errors are the solution of the linear system
0
(p − p1 ) (p − p2 )
ε1
=
.
0
(p2 − p21 ) (p2 − p22 )
ε2
It is easy to see that the determinant of the coefficient matrix is
(p − p1 ) (p − p2 ) (p2 − p1 ) ,
which always differs from zero, as long as p, p1 and ps differ pairwise. Thus,
the matrix is invertible, and the solution space of the linear system is the trivial
c as weight matrix guarantees that it is
kernel (ε1 , ε2 ) = (0, 0). Thus using W
always possible to distinguish a single error from double errors.
5
Performance model
In Section 5.1, we introduce the general performance model. Then in Section 5.2
we instantiate it for the three methods that we are considering, namely OnlineDetection, ABFT-Detection and ABFT-Correction.
21
5.1
General approach
We introduce an abstract performance model to compute the best combination
of checkpoints and verifications for iterative methods. We execute T time-units
of work followed by a verification, which we call a chunk, and we repeat this
scheme s times, i.e., we compute s chunks, before taking a checkpoint. We say
that the s chunks constitute a frame. The whole execution is then partitioned
into frames. We assume that checkpoint, recovery and verification are error-free
operations. Let Tcp , Trec and Tverif be the respective cost of these operations.
Finally, assume an exponential distribution of errors and let q be the probability
of successful execution for each chunk: q = e−λT , where λ is the fault rate.
The goal of this section is to compute the expected time E (s, T ) needed to
execute a frame composed of s chunks of size T . We derive the best value of s as a
function of T and of the resilience parameters Tcp , Trec , Tverif , and q, the success
probability of a chunk. Each frame is preceded by a checkpoint, except maybe
the first one (for which we recover by reading initial data again). Following
earlier work [36], we derive the following recursive equation to compute the
expected completion time of a single frame:
E (s, T ) = q s (s(T + Tverif )) + Tcp )
+ (1 − q s ) (E (Tlost ) + Trec + E (s, T )) .
(9)
Indeed, the execution is successful if all chunks are successful, which happens
with probability q s , and in this case the execution time simply is the sum of
the execution times of each chunk plus the final checkpoint. Otherwise, with
probability 1 − q s , we have an error, which we detect after some time E (Tlost ),
and that forces us to recover (in time Trec ) and restart the frame anew, hence
in time E (s, T ). The difficult part is to compute E (Tlost ).
For 1 ≤ i ≤ s, let fi be the following conditional probability:
fi = P(error strikes at chunk i|there is
an error in the frame) . (10)
Given the success probability q of a chunk, we obtain that
fi =
q i−1 (1 − q)
.
1 − qs
Indeed, the first i − 1 chunks were successful (probability q i−1 ), the ith one had
an error (probability 1 − q), and we condition by the probability of an error
within the frame, namely 1 − q s . With probability fi , we detect the error at
the end of the ith chunk, and we have lost the time spent executing the first i
chunks. We derive that
E (Tlost ) =
s
X
fi (i(T + Tverif )) .
i=1
22
Ps
(1−q)h(q)
where h(q) = 1 + 2q + 3q 2 + · · · + sq s−1 . If
1−qs
s+1
′
m(q) = q +q 2 +· · ·+q s = 1−q
1−q −1, we get by differentiation that m (q) = h(q),
s
s+1
+ 1−q
hence h(q) = −(s+1)q
1−q
(1−q)2 and finally
We have
i=1
fi =
E (Tlost ) = (T + Tverif )
sq s+1 − (s + 1)q s + 1
.
(1 − q s )(1 − q)
Plugging the expression of E (Tlost ) back into (9), we obtain
E (s, T ) = s(T + Tverif )) + Tcp + (q −s − 1)Trec
+T
sq s+1 − (s + 1)q s + 1
,
q s (1 − q)
which simplifies into
E (s, T ) = Tcp + (q −s − 1)Trec + (T + Tverif )
1 − qs
.
q s (1 − q)
We have to determine the value of s that minimizes the overhead of a frame:
E (s, T )
.
(11)
s = argmin
sT
s≥1
The minimization is complicated and should be conducted numerically (because
T , the size of a chunk, is still unknown). Luckily, a dynamic programming
algorithm to compute the optimal value of T and s is available [9].
5.2
Instantiation to PCG
For each of the three methods, Online-Detection, ABFT-Detection and
ABFT-Correction, we instantiate the previous model and discuss how to
solve (11).
5.2.1
Online-Detection
For Chen’s method [6], we have chunks of d iterations, hence T = dTiter , where
Titer is the raw cost of a PCG iteration without any resilience method. The verification time Tverif is the cost of the orthogonality check operations performed
as described in Section 3. As for silent errors, the application is protected from
arithmetic errors in the ALU, as in Chen’s original method, but also for corruption in data memory (because we also checkpoint the matrix A). Let λa be
the rate of arithmetic errors, and λm be the rate of memory errors. For the
latter, we have λm = M λword if the data memory consists of M words, each
susceptible to be corrupted with rate λword . Altogether, since the two error
sources are independent, they have a cumulative rate of λ = λa + λm , and the
success probability for a chunk is q = e−λT .
Plugging these values in (11) gives an optimization formula very similar to
that of Chen [6, Sec. 5.2], the only difference being that we assume that the
verification is error-free, which is needed for the correctness of the approach.
23
5.2.2
ABFT-Detection
When using ABFT techniques, we detect possible errors every iteration, so a
chunk is a single iteration, and T = Titer . For ABFT-Detection, Tverif is
the overhead due to the checksums and redundant operations to detect a single
error in the method.
ABFT-Detection can protect the application from the same silent errors
as Online-Detection, and just as before the success probability for a chunk
(a single iteration here) is q = e−λT .
5.2.3
ABFT-Correction
In addition to detection, we now correct single errors at every chunk. Just as
for ABFT-Detection, a chunk is a single iteration, and T = Titer , but Tverif
corresponds to a larger overhead, mainly due to the extra checksums needed to
detect two errors and correct a single one.
The main difference lies in the error rate. An iteration with ABFT-Correction is successful if zero or one error has struck during that iteration, so that the
success probability is much higher than for Online-Detection and ABFTDetection. We compute that value of the success probability as follows. We
have a Poisson process of rate λ, where λ = λa + λm as for Online-Detection and ABFT-Detection. The probability of exactly k errors in time T is
(λT )k −λT
[37], hence the probability of no error is e−λT and the probability of
k! e
exactly one error is λT e−λT , so that q = e−λT + λT e−λT .
6
6.1
Experiments
Setup
There are two different sources of advantages in combining ABFT and checkpointing. First, the error detection capability lets us perform a cheap validation
of the partial result of each PCG step, recovering as soon as an error strikes.
Second, being able to correct single errors makes each step more resilient and
increases the expected number of consecutive valid iterations. We say an iteration is valid if it is non-faulty, or if it suffers from a single error that is corrected
via ABFT.
For our experiments, we use a set of positive definite matrices from the UFL
Sparse Matrix Collection [38], with size between 17456 and 74752 and density
lower than 10−2 . We perform the experiments under Matlab and use the factored approximate inverse preconditioners [16, 39] in the PCG. The application
of these preconditioners requires two SpMxV, which are protected against error
using the methods proposed in Section 4 (in all methods Online-Detection,
ABFT-Detection, and ABFT-Correction).
At each iteration of PCG, faults are injected during vector and matrix-vector
operations but, since we are assuming selective reliability, all the checksums and
checksum operations are considered non-faulty. Faults are modeled as bit flips
24
Table 2: Test matrices used in the experiments. Name and id are from the
University of Florida Sparse Matrix Collection.
name
Boeing/bcsstk36
Mulvey/finan512
Andrews/Andrews
GHS psdef/wathen100
GHS psdef/wathen120
GHS psdef/gridgena
GHS psdef/jnlbrng1
UTEP/Dubcova2
JGD Trefethen/Trefethen 20000
id
size
density
steps
residual
341
752
924
1288
1289
1311
1312
1848
2213
23052
74752
60000
30401
36441
48962
40000
65025
20000
2.15e-03
1.07e-04
2.11e-04
5.10e-04
4.26e-04
2.14e-04
1.24e-04
2.44e-04
1.39e-03
50
25
20
50
50
50
50
50
10
6.41e-04
2.19e-14
1.59e-04
2.55e-13
9.16e-14
5.61e-05
5.83e-13
1.16e-05
6.00e-16
occurring independently at each step, under an exponential distribution of parameter λ, as detailed in Section 5.2. These bit flips can strike either the matrix
(the elements of Val , Colid and Rowidx ), or any entry of the PCG vectors ri ,
zi , q, pi or xi . We chose not to inject errors during the computation explicitly, as they are just a special case of error we are considering. Moreover, to
simplify the injection mechanism, Titer is normalized to be one, meaning that
each memory location or operation is given the chance to fail just once per iteration [27]. Finally, to get data that are homogeneous among the test matrices,
the fault rate λ is chosen to be inversely proportional to M (memory size) with
a proportionality constant α ∈ (0, 1); this makes sense as larger the memory
used by an application, larger is the chance to have an error. It follows that the
expected number of PCG iterations between two distinct fault occurrences does
not depend either on the size or on the sparsity ratio of the matrix.
We compare the performance of three algorithms, namely Online-Detection, ABFT-Detection (single detection and rolling back as soon as an error
is detected), and ABFT-Correction (correcting single errors during a given
step and rolling back only if two errors strike a single operation). We instantiate
them by limiting the maximum number of PCG steps to 50 (20 for #924, whose
convergence is sublinear) and setting the tolerance parameter ǫ at line 5 of
Algorithm 1 to 10−14 . The number of iterations for a non-faulty execution and
the achieved accuracy are detailed in Table 2.
Implementing the null checks in Algorithm 2, Algorithm 3 and Algorithm 4
poses a challenge. The comparison dr = 0 is between two integers, and can
be correctly evaluated by any programming language using the equality check.
However, the other two, having floating point operands, are problematic. Since
the floating point operations are not associative and the distributive property
does not hold, we need a tolerance parameter that takes into account the rounding operations that are performed by each floating point operation. Here, we
give an upper bound on the difference between the two floating point checksums,
25
using the standard model [40, Sec. 2.2] to make sure that errors caught by our
algorithms really are errors and not merely inaccuracies due to floating point
operations (which is tolerable, as non-faulty executions can give rise to the same
inaccuracy).
Theorem 3 (Accuracy of the floating point weighted checksums). Let A ∈
Rn×n , x ∈ Rn , c ∈ Rn . If all of the sums involved into the matrix operations
are performed using some flavor of recursive summation [40, Ch. 4], it holds
that
|f l ((c⊺ A) x) − f l (c⊺ (Ax)) | ≤ 2 γ2n |c⊺ | | A | | x | .
(12)
We refer the reader to the technical report for the proof [41, Theorem 2].
Let us note that if all of the entries of c are positive, as it is often the case in
our setting, the absolute value of c in (12) can be safely replaced with c itself.
It is also clear that these bounds are not computable, since c⊺ | A | | x | is
not, in general, a floating point number. This problem can be alleviated by
overestimating the bound by means of matrix and vector norms.
Since we are interested in actually computing the bound at runtime, we
consider a weaker bound. Recalling that [42, Sec. B.7]
kAk1 = max
1≤j≤n
n
X
|ai,j |.
(13)
i=1
we can upper bound the right hand side in so that
|f l ((c⊺ A) x) − f l (c⊺ (Ax)) | ≤ 2 γ2n n kc⊺ k∞ kAk1 kxk∞ .
(14)
Though the right hand side of (14) is not exactly computable in floating point
arithmetic, it requires an amount of operations dramatically smaller than (12);
just a few sums for the norm of A. As this norm is usually computed using
the identity in (13), any kind of summation yields a relative error of at most
n′ u [40, Sec. 4.6], where n′ is the maximum number of nonzeros in a column
of A, and u is the machine epsilon. Since we are dealing with sparse matrices,
we expect n′ to be very small, and hence the computation of the norm to be
accurate. Moreover, since the right hand side in (14) does not depend on x, it
can be computed just once for a given matrix and weight vector.
Clearly, using (14) as tolerance parameter guarantees no false positive (a
computation without any error that is considered as faulty), but allows false
negatives (an iteration during which an error occurs without being detected)
when the perturbations of the result are small. Nonetheless, this solution works
almost perfectly in practice, meaning that though the convergence rate can
be slowed down, the algorithms still converges towards the “correct” answer.
Though such an outcome could be surprising at first, Elliott et al. [43, 44]
showed that bit flips that strike the less significant digits of the floating point
representation of vector elements during a dot product create small perturbations of the results, and that the magnitude of this perturbation gets smaller as
26
Table 3: Experimental validation of the model. Here sei and s∗i represent the best
checkpointing interval according to our model and to our simulations respectively, whereas Et (sei ) and Et (s∗i ) stand for the execution time of the algorithm
using these checkpointing intervals.
id se1 Et (se1 ) s∗1 Et (s∗1 )
l1 se2 Et (se2 ) s∗2 Et (s∗2 )
l2
341
752
924
1288
1289
1311
1312
1848
2213
4
30
23
22
16
4
25
4
19
305.42
13.81
49.82
11.12
16.56
216.70
14.41
321.70
2.31
1
24
30
19
13
1
22
1
12
293.22
13.34
47.53
10.82
16.38
207.97
13.86
309.28
2.19
4.16
3.57
4.82
2.72
1.07
4.19
3.97
4.01
5.58
4
24
23
22
23
4
23
4
24
305.45
12.17
44.52
11.32
13.49
220.20
12.30
366.03
2.33
1
23
26
19
23
1
22
1
23
293.16
11.93
42.42
11.03
13.49
208.19
12.06
314.20
2.20
4.19
2.01
4.96
2.58
0.00
5.77
1.96
16.49
5.94
the size of the vectors increases. Hence, we expect errors that are not detected
by our tolerance threshold to be too small to impact the solution of the linear
solver.
6.2
Simulations
To validate the model, we perform the simulation whose results are illustrated
in Table 3. For each matrix, we set λ = 161M and consider the average execution
time of 100 repetitions of both ABFT-Detection (columns 5-8) and ABFTCorrection (columns 6-9). In the table we record the checkpointing interval
s∗i which achieves the shortest execution time Et (s∗1 ), and the checkpointing
interval sei which is the best stepsize according to our method, along with its
execution time Et (sei ). Finally, we evaluate the performance of our guess by
means of the quantity
li =
Et (sei ) − Et (s∗i )
· 100 ,
Et (s∗i )
that expresses the loss, in terms of execution time, of executing with the checkpointing interval given by our model with respect to the best possible choice.
From the table, we clearly see that the values of sei and s∗i are close, since the
time loss reaches just above 5% for l1 and just below 15% for l2 . This sometimes
poor result depends just on the small number of repetitions we are considering,
that leads to the presence of outliers, lucky runs in which a small number of
errors occur and the computation is carried on in a much quicker way. Similar
results hold for other values of λ.
We also compare the execution time of the three algorithms to empirically
asses how much their relative performance depend on the fault rate. The results
on our test matrices are shown in Fig. 1, where the y-axis is the execution time
27
200
4
40
100
2
20
0
102
104
0
(a) Matrix #341
104
0
(b) Matrix #752
10
6
102
104
(c) Matrix #924
200
8
4
6
100
4
2
0
102
2
102
104
0
(d) Matrix #1288
102
104
0
(e) Matrix #1289
102
104
(f) Matrix #1311
8
1
6
100
4
2
0
0.5
50
102
104
(g) Matrix #1312
0
102
104
(h) Matrix #1848
0
102
104
(i) Matrix #2213
Figure 1: Execution time in seconds (y axis) of Online-Detection (dotted),
ABFT-Detection (solid line) and ABFT-Correction (dashed) with respect
to the normalized MTBF (x-axis). The matrix number is in the subcaption.
28
(in seconds), and the x-axis is the normalized mean time between failure (the
reciprocal of α). Here, the larger x = α1 , the smaller the corresponding value
α
, hence the smaller the expected number of errors. For each value of
of λ = M
λ, we draw the average execution time of 50 runs of the three algorithms, using
the best checkpointing interval predicted in Section 5.1 for ABFT-Detection
and ABFT-Correction, and by Chen [6, Eq. 10] for Online-Detection.
In terms of execution time, Chen’s method is comparable to ours for middle to
high fault rates, since it clearly outperforms ABFT-Detection in five out of
nine cases, being slightly faster than ABFT-Correction for lower fault rates.
Intuitively, this behavior is not surprising. When λ is large, many errors
occur but, since α is between zero and one, we always have, in expectation, less
than one error per iteration. Thus ABFT-Correction requires fewer checkpoints than ABFT-Detection and almost no rollback, and this compensates
for the slightly longer execution time of a single step. When the fault rate is very
low, instead, the algorithms perform almost the same number of iterations, but
ABFT-Correction takes slightly longer due to the additional dot products
at each step.
Altogether, the results show that ABFT-Correction outperforms both
Online-Detection and ABFT-Detection for a wide range of fault rates,
thereby demonstrating that combining checkpointing with ABFT correcting
techniques is more efficient than pure checkpointing for most practical situations.
7
Conclusion
We consider the problem of silent errors in iterative linear systems solvers. At
first, we focus our attention on ABFT methods for SpMxV, developing algorithms able to detect and correct errors in both memory and computation using
various checksumming techniques. Then, we combine ABFT with replication,
in order to develop a resilient PCG kernel that can protect axpy’s and dot products as well. We also discuss how to take numerical issues into account when
dealing with actual implementations. These methods are a worthy choice for
a selective reliability model, since most of the operations can be performed in
unreliable mode, whereas only checksum computations need to be performed
reliably.
In addition, we examine checkpointing techniques as a tool to improve the
resilience of our ABFT PCG and develop a model to trade-off the checkpointing
interval so to achieve the shortest execution time in expectation. We implement
two of the possible combinations, namely an algorithm that relies on roll back
as soon as an error is detected, and one that is able to correct a single error and recovers from a checkpoint just when two errors strike. We validate
the model by means of simulations and finally compare our algorithms with
Chen’s approach, empirically showing that ABFT overhead is usually smaller
than Chen’s verification cost.
We expect this combined approach to be interesting for other variants of
29
the preconditioned conjugate gradient algorithm [34]. Triangular preconditioners seem to be particularly attracting, in that it looks possible to treat them
by adapting the techniques described in this paper (Shantharam et al. [15] addressed the triangular case).
Acknowledgements
Y. Robert and B. Uçar were partly supported by the French Research Agency
(ANR) through the Rescue and SOLHAR (ANR MONU-13-0007) projects.
Y. Robert is with Institut Universitaire de France. J. Langou was fully supported by NSF award CCF 1054864.
References
References
[1] A. Moody, G. Bronevetsky, K. Mohror, B. de Supinski, Design, modeling, and evaluation of a scalable multi-level checkpointing system, in: High
Performance Computing, Networking, Storage and Analysis (SC), 2010 International Conference for, 2010, pp. 1–11.
[2] F. Cappello, Fault tolerance in petascale/ exascale systems: Current knowledge, challenges and research opportunities, International Journal of High
Performance Computing Applications 23 (3) (2009) 212–226.
[3] F. Cappello, A. Geist, B. Gropp, L. Kale, B. Kramer, M. Snir, Toward
exascale resilience, International Journal of High Performance Computing
Applications 23 (4) (2009) 374–388.
[4] F. Cappello, A. Geist, W. Gropp, S. Kale, B. Kramer, M. Snir, Toward
exascale resilience: 2014 update, Supercomputing frontiers and innovations
1 (1)
[5] B. Schroeder, G. A. Gibson, Understanding failures in petascale computers,
Journal of Physics: Conference Series 78 (012022).
[6] Z. Chen, Online-ABFT: An Online Algorithm Based Fault Tolerance
Scheme for Soft Error Detection in Iterative Methods, in: Proc. 18th ACM
SIGPLAN Symposium on Principles and Practice of Parallel Programming,
PPoPP ’13, ACM, 2013, pp. 167–176.
[7] J. W. Young, A first order approximation to the optimum checkpoint interval, Comm. of the ACM 17 (9) (1974) 530–531.
[8] J. T. Daly, A higher order estimate of the optimum checkpoint interval for
restart dumps, FGCS 22 (3) (2004) 303–312.
30
[9] A. Benoit, A. Cavelan, Y. Robert, H. Sun, Assessing general-purpose algorithms to cope with fail-stop and silent errors, in: Workshop on Performance Modeling, Benchmarking and Simulation (PMBS), 2014, extended
version available as INRIA Research Report RR-8599.
[10] K.-H. Huang, J. A. Abraham, Algorithm-Based Fault Tolerance for Matrix
Operations, Computers, IEEE Transactions on C–33 (6) (1984) 518–528.
[11] P. Du, A. Bouteiller, G. Bosilca, T. Herault, J. Dongarra, Algorithm-based
fault tolerance for dense matrix factorizations, in: PPoPP, ACM, 2012, pp.
225–234.
[12] E. Yao, J. Zhang, M. Chen, G. Tan, N. Sun, Detection of soft errors in LU
decomposition with partial pivoting using algorithm-based fault tolerance,
International Journal of High Performance Computing Applications 29 (4)
(2015) 422–436.
[13] D. Hakkarinen, P. Wu, Z. Chen, Fail-stop failure algorithm-based fault
tolerance for cholesky decomposition, Parallel and Distributed Systems,
IEEE Transactions on 26 (5) (2015) 1323–1335.
[14] P. Du, P. Luszczek, S. Tomov, J. Dongarra, Soft error resilient QR factorization for hybrid system with GPGPU, Journal of Computational Science
4 (6) (2013) 457 – 464, scalable Algorithms for Large-Scale Systems Workshop (ScalA2011), Supercomputing 2011.
[15] M. Shantharam, S. Srinivasmurthy, P. Raghavan, Fault Tolerant Preconditioned Conjugate Gradient for Sparse Linear System Solution, in: Proceedings of the 26th ACM International Conference on Supercomputing,
ICS ’12, ACM, 2012, pp. 69–78.
[16] M. Benzi, M. Tuma, A sparse approximate inverse preconditioner for nonsymmetric linear systems, SIAM Journal on Scientific Computing 19 (3)
(1998) 968–994.
[17] K. Kaya, B. Uçar, U. V. Çatalyürek, Analysis of partitioning models and
metrics in parallel sparse matrix-vector multiplication, in: Parallel Processing and Applied Mathematics (PPAM2014), Springer LNCS, Warsaw,
Poland, 2014, pp. 174–184.
[18] G. Aupy, Y. Robert, F. Vivien, D. Zaidouni, Checkpointing algorithms
and fault prediction, Journal of Parallel and Distributed Computing 74 (2)
(2014) 2048–2064.
[19] R. E. Lyons, W. Vanderkulk, The use of triple-modular redundancy to
improve computer reliability, IBM J. Res. Dev. 6 (2) (1962) 200–209.
[20] D. Fiala, F. Mueller, C. Engelmann, R. Riesen, K. Ferreira, R. Brightwell,
Detection and correction of silent data corruption for large-scale highperformance computing, in: Proc. of the ACM/IEEE SC Int. Conf., SC
’12, IEEE Computer Society Press, 2012.
31
[21] J. Elliott, K. Kharbas, D. Fiala, F. Mueller, K. Ferreira, C. Engelmann,
Combining partial redundancy and checkpointing for HPC, in: Proc.
ICDCS ’12, IEEE Computer Society, 2012.
[22] G. Lu, Z. Zheng, A. A. Chien, When is multi-version checkpointing needed,
in: 3rd Workshop for Fault-tolerance at Extreme Scale (FTXS), ACM
Press, 2013.
[23] A. A. Hwang, I. A. Stefanovici, B. Schroeder, Cosmic rays don’t strike
twice: understanding the nature of dram errors and the implications for
system design, SIGARCH Comput. Archit. News 40 (1) (2012) 111–122.
[24] M. Hoemmen, M. A. Heroux, Fault-tolerant iterative methods via selective
reliability, Tech. rep., Sandia Corporation (2011).
[25] M. Hoemmen, M. A. Heroux, Fault-Tolerant Iterative Methods via Selective
Reliability, in: Proceedings of the 2011 International Conference for High
Performance Computing, Networking, Storage and Analysis (SC). IEEE
Computer Society, Vol. 3, 2011, p. 9.
[26] P. G. Bridges, K. B. Ferreira, M. A. Heroux, M. Hoemmen, Fault-tolerant
linear solvers via selective reliability, preprint (2012).
[27] P. Sao, R. Vuduc, Self-stabilizing iterative solvers, in: Proc. ScalA ’13,
ACM, 2013.
[28] A. R. Benson, S. Schmit, R. Schreiber, Silent error detection in numerical
time-stepping schemes., CoRR abs/1312.2674.
[29] M. Heroux, M. Hoemmen, Fault-tolerant iterative methods via selective reliability, Research report SAND2011-3915 C, Sandia National Laboratories
(2011).
[30] G. Bronevetsky, B. de Supinski, Soft error vulnerability of iterative linear
algebra methods, in: Proc. 22nd Int. Conf. on Supercomputing, ICS ’08,
ACM, 2008, pp. 155–164.
[31] C. Anfinson, F. Luk, A Linear Algebraic Model of Algorithm-Based Fault
Tolerance, IEEE Trans. Computers 37 (12) (1988) 1599–1604.
[32] G. Bosilca, R. Delmas, J. Dongarra, J. Langou, Algorithm-based fault tolerance applied to high performance computing, J. Parallel and Distributed
Computing 69 (4) (2009) 410 –416.
[33] J. Sloan, R. Kumar, G. Bronevetsky, Algorithmic Approaches to Low Overhead Fault Detection for Sparse Linear Algebra, in: Dependable Systems
and Networks (DSN), 2012 42nd Annual IEEE/IFIP International Conference on, 2012, pp. 1–12.
[34] Y. Saad, Iterative Methods for Sparse Linear Systems, 2nd Edition, SIAM
Press, 2003.
32
[35] F. R. K. Chung, Spectral Graph Theory, American Mathematical Society,
1997.
[36] M. Bougeret, H. Casanova, M. Rabie, Y. Robert, F. Vivien, Checkpointing
strategies for parallel jobs, in: SC’2011, IEEE, 2011, pp. 1–11.
[37] M. Mitzenmacher, E. Upfal, Probability and Computing: Randomized Algorithms and Probabilistic Analysis, Cambridge University Press, 2005.
[38] T. A. Davis, Y. Hu, The University of Florida Sparse Matrix Collection,
ACM Trans. Math. Softw. 38 (1) (2011) 1:1–1:25.
[39] M. Benzi, R. Kouhia, M. Tůma, Stabilized and block approximate inverse
preconditioners for problems in solid and structural mechanics, Computer
Methods in Applied Mechanics and Engineering 190 (49–50) (2001) 6533 –
6554.
[40] N. J. Higham, Accuracy and Stability of Numerical Algorithms, 2nd Edition, SIAM Press, 2002.
[41] M. Fasi, Y. Robert, B. Uçar, Combining Algorithm-based Fault Tolerance
and Checkpointing for Iterative Solvers, Research Report RR-8675, INRIA
(2015).
[42] N. J. Higham, Functions of Matrices: Theory and Computation, SIAM
Press, 2008.
[43] J. Elliott, F. Mueller, M. Stoyanov, C. Webster, Quantifying the impact of
single bit flips on floating point arithmetic, preprint (2013).
[44] M. Stoyanov, C. Webster, Quantifying the impact of single bit flips on floating point arithmetic, Tech. rep., Oak Ridge National Laboratory (2013).
33
| 8cs.DS
|
LARGEST ACYLINDRICAL ACTIONS AND STABILITY
IN HIERARCHICALLY HYPERBOLIC GROUPS
arXiv:1705.06219v1 [math.GR] 17 May 2017
CAROLYN ABBOTT, JASON BEHRSTOCK, AND MATTHEW GENTRY DURHAM
Abstract. We consider two manifestations of non-positive curvature: acylindrical actions
(on hyperbolic spaces) and quasigeodesic stability. We study these properties for the class of
hierarchically hyperbolic groups, which is a general framework for simultaneously studying
many important families of groups, including mapping class groups, right-angled Coxeter
groups, most 3–manifold groups, right-angled Artin groups, and many others.
A group that admits an acylindrical action on a hyperbolic space may admit many such
actions on different hyperbolic spaces. It is natural to try to develop an understanding
of all such actions and to search for a “best” one. The set of all cobounded acylindrical
actions on hyperbolic spaces admits a natural poset structure, and in this paper we prove
that all hierarchically hyperbolic groups admit a unique action which is the largest in this
poset. The action we construct is also universal in the sense that every element which acts
loxodromically in some acylindrical action on a hyperbolic space does so in this one. Special
cases of this result are themselves new and interesting. For instance, this is the first proof
that right-angled Coxeter groups admit universal acylindrical actions.
The notion of quasigeodesic stability of subgroups provides a natural analogue of quasiconvexity which can be considered outside the context of hyperbolic groups. In this paper,
we provide a complete classification of stable subgroups of hierarchically hyperbolic groups,
generalizing and extending results that are known in the context of mapping class groups
and right-angled Artin groups. Along the way, we provide a characterization of contracting
quasigeodesics; interestingly, in this generality the proof is much simpler than in the special
cases where it was already known.
Contents
1. Introduction
2. Background
3. Altering the hierarchically hyperbolic structure
4. Characterization of contracting geodesics
5. Universal and largest acylindrical actions
6. Characterizing stability
7. Almost hierarchically hyperbolic spaces
References
1
5
9
13
15
18
21
24
1. Introduction
Hierarchically hyperbolic groups were recently introduced by Behrstock, Hagen, and Sisto
[BHS17b] to provide a uniform framework in which to study many important families of
groups, including mapping class groups, right-angled Coxeter groups, most 3–manifold groups,
right-angled Artin groups and many others. A hierarchically hyperbolic space consists of: a
quasigeodesic space, X ; a set of domains, S, which index a collection of δ–hyperbolic spaces
Date: May 18, 2017.
1
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
2
to which X projects; and, some additional information about these projections, including,
for instance, a partial order on the domains and a unique maximal element in that order.
Largest acylindrical actions. The study of acylindrical actions on hyperbolic spaces, as
initiated in its modern form by Osin [Osi16] following earlier work of [Sel97] and [Bow08],
has proven to be a powerful tool for studying groups with some aspects of non-positive curvature. As established in [BHS17b], hierarchically hyperbolic groups admit non-elementary
acylindrical actions when the δ–hyperbolic space associated to the maximal element in S has
infinite diameter, a property which holds in all the above examples except for those that are
direct products.
Any given group with an acylindrical action may actually admit many acylindrical actions
on many different spaces. A natural question is to try and find a “best” acylindrical action.
There are different ways that one might try to optimize the acylindrical action. For instance,
the notion of a universal acylindrical action, for a given group G, is an acylindrical action
on a hyperbolic space X such that every element of G which acts loxodromically in some
acylindrical action on some hyperbolic space, must act loxodromically in this action. As
established by Abbott, there exist finitely generated groups which admit acylindrical actions,
but no universal acylindrical action [Abb16]; we also note that universal actions need not be
unique [ABO16].
In forthcoming work, Abbott, Balasubramanya, and Osin [ABO16], introduce a partial
order on cobounded acylindrical actions; when there exist an element in this partial ordering
which is comparable to and larger than all other elements it is called a largest action. By
construction, any largest action is necessarily a universal action and unique.
In this paper we prove that all hierarchically hyperbolic groups have largest actions. Special
cases of this theorem recover some recent results of [ABO16], as well as a number of new cases.
For instance, in the case of right-angled Coxeter groups (and more generally for cubulated
groups), even the existence of a universal acylindrical action was unknown. Further, outside of
the relatively hyperbolic setting, our result provides a single construction that simultaneously
covers these new cases as well as all previously known largest and universal acylindrical
actions of finitely presented groups. The following summarizes the main results of Section 5
(where there are also further details on the background and comparison with known results).
Theorem A (HHG have actions that are largest and universal). Every hierarchically hyperbolic group admits a largest acylindrical action. In particular, the following admit acylindrical
actions which are largest and universal:
(1) Hyperbolic groups and their subgroups.
(2) Mapping class groups.
(3) Fundamental groups of three manifolds with no Nil or Sol in their prime decomposition.
(4) Groups that act properly and cocompactly on a proper CAT(0) cube complex. This includes right angled-Artin groups and right-angled Coxeter groups.
Stability in hierarchically hyperbolic groups. One of the key features of a Gromov
hyperbolic space is that every geodesic is uniformly Morse, a property also known as (quasigeodesically) stable; that is, any uniform quasigeodesic beginning and ending on a geodesic
must lie uniformly close to it. In fact, any geodesic metric space in which each geodesic is
uniformly Morse is hyperbolic.
In the context of geodesic metric spaces, the presence of Morse geodesics has important
structural consequences for the space; for instance, any asymptotic cone of such a space has
global cut points [DMS10]. Moreover, quasigeodesic stability in groups is quite prevalent,
since any finitely generated acylindrically hyperbolic group contains Morse geodesics [Osi16,
Sis16].
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
3
There has been much interest in developing alternative characterizations [DMS10, CS15,
ACGH16, ADT16] and understanding this phenomenon in various important contexts [Min96,
Beh06, DMS10, DT15, ADT16]. There is also a nascent theory of Morse boundaries, which
encode all Morse geodesics of a group [CS15, Cor15, CH16, CD16, CM17]. In [DT15], Durham
and Taylor generalized the notion of stability to subspaces and subgroups.
A subset Y Ă X is said to have D–bounded projections when diampπU pYqq ă D for all nonmaximal U P S; when the constant doesn’t matter we simply say the subset has uniformly
bounded projections.
We prove a complete characterization of stability in hierarchically hyperbolic groups.
Theorem B (Equivalent conditions for subgroup stability). Any hierarchically hyperbolic
group G admits a hierarchically hyperbolic group structure pG, Sq such that for any finitely
generated H ă G, the following are equivalent:
(1) H is stable in G;
(2) H is undistorted in G and has uniformly bounded projections;
(3) Any orbit map H Ñ CS is a quasi-isometric embedding, where S is Ď–maximal in S.
Theorem B generalizes some previously known results. In the case of mapping class groups:
[DT15] proved equivalence of (1) and (3); equivalence of (2) and (3) follows from the distance
formula; moreover, [KL08, Ham] yield that these conditions are also equivalent to convex
cocompactness in the sense of [FM02]. The case of right-angled Artin groups was studied in
[KMT14], where they prove equivalence of (1) and (3).
Section 6 contains a more general version of Theorem B, as well as further applications, including Theorem 6.6 which concerns the Morse boundary of hierarchically hyperbolic groups,
and proves that all hierarchically hyperbolic groups have finite stable asymptotic dimension.
As a sample application of Theorem B and using work of Taylor–Tiozzo [TT16], we prove
the following in Section 6.4.
Theorem C (Random subgroups are stable). Let pX , Sq be a HHS for which CS has infinite diameter, where S is the Ď–maximal element, and consider G ă AutpX , Sq which acts
properly and cocompactly on X . Then any k–generated random subgroup of G stably embeds
in X .
We note that one immediate consequence of this result is a new proof of a theorem of
Maher–Sisto: any random subgroup of a hierarchically hyperbolic group which is not the
direct product of two infinite groups is stable [MS17]. The mapping class group and rightangled Artin groups cases of this result were first established in [TT16].
On purely loxodromic subgroups. In the mapping class group setting [BBKL16] proved
that the conditions in Theorem B are also equivalent to being undistorted and purely pseudoAnosov. Similarly, in the right-angled Artin group setting, it was proven in [KMT14] that
(1) and (3) are each equivalent to being purely loxodromic.
Subgroups of right-angled Coxeter groups all of whose elements act loxodromically on
the contact graph were studied in the recent preprint [Tra, Theorem 1.4], who proved that
property is equivalent to (3). Since there often exist elements in a right-angled Coxeter group
which do not act loxodromically on the contact graph, his condition is not equivalent to (1);
it is the ability to change the hierarchically hyperbolic structure as we do in Theorem 3.11,
discussed below, which allows us to prove our more general result which characterizes all
stable subgroups, not just the ones acting loxodromically on the contact graph.
Mapping class groups and right-angled Artin groups have the property that in their standard hierarchically hyperbolic structure they admit a universal acylindrical action on CS
where S is the Ď–maximal domain. On the other hand, right-angled Coxeter groups often
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
4
don’t admit universal acylindrical actions on CS in their standard structure. Accordingly, we
believe the following questions are interesting. The first item would generalize the situation
in the mapping class group as established in [BBKL16], and the second item would generalize what is known in right-angled Artin groups as proven in [KMT14], and partial results
about right-angled Coxeter groups as in [Tra]. In the case of the mapping class group, if the
second item was true this would resolve a question of Farb–Mosher [FM02]. See also [ADT16,
Question 1].
Question D. Let pG, Sq be a hierarchically hyperbolic group which admits a universal acylindrical action on CS, where S is Ď–maximal in S. Let H be a finitely generated subgroup of
G. Are the conditions in Theorem B also equivalent to:
(1) . . . H is undistorted and acts purely loxodromically on CS?
(2) . . . H acts purely loxodromically on CS?
Note that in the context of Question D, an element acts loxodromically on CS if and only
if it has positive translation length. This holds since the action is acylindrical and thus each
element either acts elliptically or loxodromically.
New hierarchically hyperbolic structures. In order to establish the above results, we
provide some new structural theorems about hierarchically hyperbolic spaces.
One of the key technical innovations in this paper is provided in Section 3. There we
prove Theorem 3.11 which allows us to modify a given hierarchically hyperbolic structure
pX , Sq by removing CU for some U P S and, in their place, enlarging the space CS where S
is the Ď–maximal element of S. For instance, this is how we construct the space on which
a hierarchically hyperbolic group has its largest acylindrical action.
Another important tool is Theorem 4.4 which provides a simple characterization of contracting geodesics in a hierarchically hyperbolic space The following is a restatement of that
result in the case of groups:
Theorem E (Characterization of contracting quasigeodesics). Let G be a hierarchically hyperbolic group. For any D ą 0 and K ě 1 there exists a D 1 ą 0 depending only on D and G
such that the following holds for every pK, Kq–quasigeodesic γ Ă X : the quasigeodesic γ has
D–bounded projections if and only if γ is D 1 –contracting.
Since the presence of a contracting geodesic implies the group has at least quadratic
divergence, an immediate consequence of Theorem E is that any hierarchically hyperbolic
group has quadratic divergence whenever X projects to an infinite diameter subset of CS.
Finally, through much of this paper we impose a technical hypothesis on our hierarchically
hyperbolic structures, called having clean containers. Although, in Proposition 3.5 this
hypothesis is shown to hold for many groups, it does not hold in all cases. In Section 7
we introduce a technical trick which allows us to prove a weakened version of the results of
Section 3 without this hypothesis. In turn, this allows us to remove this hypothesis from
the remaining results in the paper, for instance allowing us in the case of groups to upgrade
Theorem 5.1 to Theorem A and Theorem 4.4 to Theorem E.
Acknowledgments. The authors thank Mark Hagen and Alessandro Sisto for lively conversations about hierarchical hyperbolicity. The authors were supported in part by NSF
grant DMS-1440140 while at the Mathematical Sciences Research Institute in Berkeley during Fall 2016 program in Geometric Group Theory. Abbott was supported by the NSF RTG
awards DMS-1502553 and Durham was supported by DMS-1045119. Behrstock thanks Chris
Leininger for an interesting conversation which led to the formulation of Conjecture D. We
thank Ivan Levcovitz and Jacob Russell for helpful feedback on an early version of this article.
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
5
2. Background
2.1. Hierarchically hyperbolic spaces. In this section we recall the basic definitions and
properties of hierarchically hyperbolic spaces as introduced in [BHS17b, BHS15].
Definition 2.1 (Hierarchically hyperbolic space). A q–quasigeodesic space pX , dX q is said
to be hierarchically hyperbolic if there exists δ ě 0, an index set S, and a set tCW | W P Su
of δ–hyperbolic spaces pCU, dU q, such that the following conditions are satisfied:
(1) (Projections.) There is a set tπW : X Ñ 2CW | W P Su of projections sending
points in X to sets of diameter bounded by some ξ ě 0 in the various CW P S.
Moreover, there exists K so that each πW is pK, Kq–coarsely Lipschitz.
(2) (Nesting.) S is equipped with a partial order Ď, and either S “ H or S contains
a unique Ď–maximal element; when V Ď W , we say V is nested in W . We require
that W Ď W for all W P S. For each W P S, we denote by SW the set of V P S
such that V Ď W . Moreover, for all V, W P S with V Ĺ W there is a specified subset
CV .
ρVW Ă CW with diamCW pρVW q ď ξ. There is also a projection ρW
V : CW Ñ 2
(3) (Orthogonality.) S has a symmetric and anti-reflexive relation called orthogonality:
we write V KW when V, W are orthogonal. Also, whenever V Ď W and W KU , we
require that V KU . Finally, we require that for each T P S and each U P ST for
which tV P ST | V KU u ‰ H, there exists W P ST ´ tT u, so that whenever V KU
and V Ď T , we have V Ď W ; we say W is a container associated with T P S and
U P ST . Finally, if V KW , then V, W are not Ď–comparable.
(4) (Transversality and consistency.) If V, W P S are not orthogonal and neither is
nested in the other, then we say V, W are transverse, denoted V &W . There exists
κ0 ě 0 such that if V &W , then there are sets ρVW Ď CW and ρW
V Ď CV each of
diameter at most ξ and satisfying:
(
min dW pπW pxq, ρVW q, dV pπV pxq, ρW
V q ď κ0
for all x P X .
For V, W P S satisfying V Ď W and for all x P X , we have:
(
min dW pπW pxq, ρVW q, diamCV pπV pxq Y ρW
V pπW pxqqq ď κ0 .
&
V
Finally, if U Ď V , then dW pρU
W , ρW q ď κ0 whenever W P S satisfies either V Ĺ W
or V &W and W U .
(Finite complexity.) There exists n ě 0, the complexity of X (with respect to S),
so that any set of pairwise–Ď–comparable elements has cardinality at most n.
(Large links.) There exist λ ě 1 and E ě maxtξ, κ0 u such that the following holds.
Let W P S and let x, x1 P X . Let N “ λdW pπW pxq, πW px1 qq ` λ. Then there exists
tTi ui“1,...,tN u Ď SW ´ tW u such that for all T P SW ´ tW u, either T P STi for some
i, or dT pπT pxq, πT px1 qq ă E. Also, dW pπW pxq, ρTWi q ď N for each i.
(Bounded geodesic image.) For all W P S, all V P SW ´ tW u, and all geodesics
V
γ of CW , either diamCV pρW
V pγqq ď E or γ X NE pρW q ‰ H.
(Partial Realization.) There exists a constant α with the following property. Let
tVj u be a family of pairwise orthogonal elements of S, and let pj P πVj pX q Ď CVj .
Then there exists x P X so that:
‚ dVj px, pj q ď α for all j,
(5)
(6)
(7)
(8)
V
‚ for each j and each V P S with Vj Ď V , we have dV px, ρVj q ď α, and
V
‚ if W &Vj for some j, then dW px, ρWj q ď α.
(9) (Uniqueness.) For each κ ě 0, there exists θu “ θu pκq such that if x, y P X and
dpx, yq ě θu , then there exists V P S such that dV px, yq ě κ.
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
6
Notation 2.2. Note that below we will often abuse notation by simply writing pX , Sq or
S to refer to the entire package of an hierarchically hyperbolic structure, including all the
associated spaces, projections, and relations given by the above definition.
Notation 2.3. When writing distances in CU for some U P S, we often simplify the notation
slightly by suppressing the projection map πU , i.e., given x, y P X and p P CU we write
dU px, yq for dU pπU pxq, πU pyqq and dU px, pq for dU pπU pxq, pq. Note that when we measure
distance between a pair of sets (typically both of bounded diameter) we are taking the
minimum distance between the two sets. Given A Ă X and U P S we let πU pAq denote
YaPA πU paq.
It is often convenient to work with equivariant hierarchically hyperbolic structures, we
now recall the relevant structures for doing so. For details see [BHS15].
Definition 2.4 (Hierarchically hyperbolic groups). Let pX , Sq be a hierarchically hyperbolic
space. The automorphism group of pX , Sq is denoted AutpX , Sq and is defined as follows. An
element of AutpX , Sq consists of a map g : X Ñ X , together with a bijection g♦ : S Ñ S and,
for each U P S, an isometry g˚ pU q : CU Ñ Cpg♦ pU qq so that the following diagrams coarsely
commute whenever the maps in question are defined (i.e., when U, V are not orthogonal):
g
X
πU
CU
and
CU
ρU
V
CV
// X 1
g ˚ pU q
g ˚ pV
q
πg♦ pU q
g ˚ pU q
// Cpg ♦ pU qq
// Cpg ♦ pU qq
g ♦ pU q
g ♦ pV q
ρ
// Cpg ♦ pV qq
A finitely generated group G is said to be a hierarchically hyperbolic group (HHG) if there is
a hierarchically hyperbolic space pX , Sq and an action G Ñ AutpX , Sq so that the induced
uniform quasi-action of G on X is metrically proper, cobounded, and S contains finitely
many G–orbits. Note that when G is a hyperbolic group then with respect to any word
metric it inherits a hierarchically hyperbolic structure.
An important consequence of being a hierarchically hyperbolic space is the following distance formula, which relates distances in X to distances in the hyperbolic spaces CU for
U P S. The notation ttxuus means include x in the sum if and only if x ą s.
Theorem 2.5 (Distance formula for HHS; [BHS15]). Let pX , Sq be a hierarchically hyperbolic
space. Then there exists s0 such that for all s ě s0 , there exist C, K so that for all x, y P X ,
ÿ
dpx, yq —K,C
ttdU px, yquus .
U PS
We now recall an important construction of subspaces in a hierarchically hyperbolic space
called standard product regions introduced in [BHS17b, Section 13] and studied further in
[BHS15]. First we define the two factors in the product space.
Definition 2.6 (Nested partial tuple (FU )). Recall SU “ tV P S | V Ď U u. Fix κ ě
κ0 and let FU be the
κ–consistent tuples (i.e., tuples satisfying the conditions of
ś set of CV
Definition 2.1.(4)) in V PSU 2 .
Definition 2.7 (Orthogonal partial tuple (EU ) ). Let SK
U “ tV P S | V KU u Y tAu, where
A is a Ď–minimal elementśW such that V Ď W for all V KU . Fix κ ě κ0 , let EU be the set
of κ–consistent tuples in V PSK ´tAu 2CV .
U
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
7
Definition 2.8 (Product regions in X ). Given X and U P S, there are coarsely well-defined
maps φĎ , φK : FU , EU Ñ X which extend to a coarsely well-defined map φU : FU ˆ EU Ñ X .
Indeed, for each p~a, ~bq P FU ˆ EU , and each V P S, the coordinate pφU p~a, ~bqqV is defined as
follows. If V Ď U , then pφU p~a, ~bqqV “ aV . If V KU , then pφU p~a, ~bqqV “ bV . If V &U , then
pφU p~a, ~bqqV “ ρU
a, ~bqqV “ ρU
V . Finally, if U Ď V , and U ‰ V , let pφU p~
V . We refer to FU ˆ EU
as a product region, which we denote PU .
We often abuse notation slightly and use the notation EU , FU , and PU to refer to the
image in X of the associated set. In [BHS15, Lemma 5.9] it is proven that these standard
product regions have the property that they are “hierarchically quasiconvex subsets” of X .
Here we leave out the definition of hierarchically quasiconvexity, because its only use here is
that product regions have “gate maps,” as given by the following in [BHS15, Lemma 5.4]:
Lemma 2.9 (Existence of coarse gates; [BHS15]). If Y Ď X is k–hierarchically quasiconvex
and non-empty, then there exists a gate map for Y, i.e., for each x P X there exists y P Y
such that for all V P S, the set πV pyq (uniformly) coarsely coincides with the projection of
πV pxq to the kp0q–quasiconvex set πV pYq.
Remark 2.10 (Surjectivity of projections). As one can always change the hierarchical structure so that the projection maps are coarsely surjective [BHS15, Remark 1.3], we will assume
that S is such a structure. That is, for each U P S, if πU is not surjective, then we identify
CU with πU pX q.
We also need the notion of a hierarchy path, whose existence was proven in [BHS15,
Theorem 5.4]:
Definition 2.11. For R ě 1, a path γ in X is a R–hierarchy path if
(1) γ is a pR, Rq–quasigeodesic,
(2) for each W P S, πW ˝ γ is an unparametrized pR, Rq–quasigeodesic. An unbounded
hierarchy path r0, 8q Ñ X is a hierarchy ray.
2.2. Acylindrical actions. We recall the basic definitions related to acylindrical actions;
the canonical references are [Bow08] and [Osi16]. We also discuss a partial order on these
actions which was recently introduced in [ABO16].
Definition 2.12 (Acylindrical). The action of a group G on a metric space X is acylindrical
if for any ε ą 0 there exist R, N ě 0 such that for all x, y P X with dpx, yq ě R,
|tg P G | dpx, gxq ď ε and dpy, gyq ď εu| ď N.
Recall that given a group G acting on a hyperbolic metric space X, an element g P G is
loxodromic if the map Z Ñ X defined by n ÞÑ gn x is a quasi-isometric embedding for some
(equivalently any) x P X. However, an element of G may be loxodromic for some actions and
not for others. Consider, for example, the free group on two generators acting on its Cayley
graph and acting on the Bass-Serre tree associated to the splitting F2 » xxy ˚ xyy. In the
former action, every non-trivial element is loxodromic, while in the latter action, no powers
of x and y are loxodromic.
Definition 2.13 (Generalized loxodromic). An element of a group G is called generalized
loxodromic if it is loxodromic for some acylindrical action of G on a hyperbolic space.
Remark 2.14. By [Osi16, Theorem 1.1], every acylindrical action of a group on a hyperbolic space either has bounded orbits or contains a loxodromic element. By [Osi16, Theorem
1.2.(L4)] and Sisto [Sis16, Theorem 1], every generalized loxodromic element is Morse. Therefore, if a group H does not contain any Morse elements, it does not contain any generalized
loxodromics, and thus H must have bounded orbits in every acylindrical action on a hyperbolic space. This is the case when, for example, H is a non-trivial direct product.
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
8
Definition 2.15 (Universal acylindrical action). An acylindrical action of a group on a
hyperbolic space is a universal acylindrical action if every generalized loxodromic element is
loxodromic.
Notice that if every acylindrical action of a group G on a hyperbolic space has bounded
orbits, then G does not contain any generalized loxodromic elements, and the action of G on
a point (which is acylindrical) is a universal acylindrical action.
The following notions are discussed in detail in [ABO16]. We give a brief overview here.
Fix a group G. Given a (possibly infinite) generating set X of G, let | ¨ |X denote the word
metric with respect to X. Given two generating sets X and Y , we say X is dominated by Y
and write X ĺ Y if
sup |y|X ă 8.
yPY
Note that when X ĺ Y , then the action G ñ ΓpG, Y q provides more information about the
group than G ñ ΓpG, Xq, and so, in some sense, is a “larger" action. The two generating
sets X and Y are equivalent if X ĺ Y and Y ĺ X; when this happens we write X „ Y .
Let AHpGq be the set of equivalence classes of generating sets X of G such that ΓpG, Xq
is hyperbolic and the action G ñ ΓpG, Xq is acylindrical, where ΓpG, Xq is the Cayley graph
of Γ with respect to the generating set X. We denote the equivalence class of X by rXs. The
preorder ĺ induces an order on AHpGq, which we also denote ĺ.
Definition 2.16 (Largest). We say an equivalence class of generating sets is largest if it is
the largest element in AHpGq under this ordering.
Given a cobounded acylindrical action of G on a hyperbolic space S, a Milnor–Schwartz
argument gives a (possibly infinite) generating set Y of G such that there is a G–equivariant
quasi-isometry between G ñ S and G ñ ΓpG, Y q. By a slight abuse of language, we will
say that a particular cobounded acylindrical action G ñ S on a hyperbolic space is largest,
when, more precisely, it is the equivalence class of the generating set associated to this action
through the above correspondence, rY s, that is the largest element in AHpGq.
Remark 2.17. Notice that by definition, every largest acylindrical action is a universal
acylindrical action.
2.3. Stability. Stability is strong coarse convexity property which generalizes quasiconvexity
in hyperbolic spaces and convex cocompactness in mapping class groups [DT15]. In the
general context of metric spaces, it is essentially the familiar Morse property generalized to
subspaces, so we begin there.
Definition 2.18 (Morse/stable quasigeodesic). Let X be a metric space. A quasigeodesic
γ Ă X is called Morse (or stable) if there exists a function N : R2ě0 Ñ Rě0 such that if q is
a pK, Cq–quasigeodesic in X with endpoints on γ, then
q Ă NN pK,Cq pγq.
We call N the stability gauge for γ and say γ is N –stable if we want to record the constants.
We can now define a notion of stable embedding of one metric space in another which is
equivalent to the one introduced by Durham and Taylor [DT15]:
Definition 2.19 (Stable embedding). We say a quasi-isometric embedding f : X Ñ Y between quasigeodesic metric spaces is a stable embedding if there exists a stability gauge N
such that for any quasigeodesic constants K, C and any pK, Cq–quasigeodesic γ Ă X, then
f pγq is an N –stable quasigeodesic in Y .
The following generalizes the notion of a Morse quasigeodesic to subgroups:
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
9
Definition 2.20 (Subgroup stability). Let H be subgroup of a finitely generated group G.
We say H is a stable subgroup of G if some (equivalently, any) orbit map of H into some
(any) Cayley graph (with respect to a finite generating set) of G is a stable embedding.
If for some h P G, H “ xhy is stable, then we call h stable. Such elements are often called
Morse elements.
Stability of a subset is preserved under quasi-isometries. Note that stable subgroups are
undistorted in their ambient groups and, moreover, they are quasiconvex with respect to any
choice of finite generating set for the ambient group.
3. Altering the hierarchically hyperbolic structure
The goal of this section is to prove that any hierarchically hyperbolic space satisfying two
technical assumptions—the bounded domain dichotomy and the clean container property—
admits a hierarchically hyperbolic structure with unbounded products, i.e., every non-trivial
product region in the ambient space has unbounded factors; see Theorem 3.11 below.
In particular, this establishes that most of the standard examples of hierarchically hyperbolic groups admit a hierarchically hyperbolic group structure with unbounded products.
This is a key ingredient in our complete characterization of the contracting property in such
spaces, which we establish in Section 4.
3.1. Unbounded products. Fix a hierarchically hyperbolic space pX , Sq.
Let M ą 0 and let SM Ă S be the set of domains U P S such that there exists V P S
and W P SK
V satisfying: U Ď V , diampCV q ą M , and diampCW q ą M .
Recall that a set of domains U Ă S is closed under nesting if whenever U P U and V Ď U ,
then V P U.
Lemma 3.1. For any M ą 0, the set SM is closed under nesting.
Proof. Let U P SM and V Ď U . By definition of U P SM , there exists Z P SM with U Ď Z
and satisfying: diampCZq ą M and there exists W P SK
Z such that diampCZq ą M . Since
M
V Ď Z, it follows that V P S , as desired.
l
Definition 3.2 (Bounded domain dichotomy). We say pX , Sq has the M –bounded domain
dichotomy if there exists M ą 0 such that any U P S with diampCU q ą M satisfies
diampCU q “ 8. If the value of M is not important, we simply refer to the bounded domain dichotomy.
We note that any hierarchically hyperbolic group has the bounded domain dichotomy.
(Also, note that this property implies the space is “asymphoric” as defined in [BHS17c].)
Definition 3.3 (Unbounded products). We say that a hierarchically hyperbolic space pX , Sq
has unbounded products if it has the bounded domain dichotomy and the property that if
U P S has diampFU q “ 8, then diampEU q “ 8.
3.2. Clean containers. The clean container property is a technical assumption related to
the orthogonality axiom.
Definition 3.4 (Clean containers). A hierarchically hyperbolic space pX , Sq has clean containers if for each T P S and each U P ST with tV P ST | V K U u ‰ H, the associated
container provided by the orthogonality axiom is orthogonal to U .
We first describe some interesting examples with clean containers. Then we show that this
property is preserved under some combination theorems for hierarchically hyperbolic spaces.
We refer the reader to [BHS15, Sections 8 & 9] and [BHS17a, Section 6] for details on the
structure in the new spaces.
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
10
Proposition 3.5. The following spaces admit hierarchically hyperbolic structures with clean
containers.
(1) Hyperbolic groups
(2) Mapping class groups
(3) Cubical groups
(4) π1 pM q, for M a 3–manifold with no Nil or Sol in its prime decomposition.
Proof. Hierarchically hyperbolic structures for these spaces were constructed in [BHS17b]
and [BHS15].
(1) The statement is immediate for hyperbolic groups, as they all admit hierarchically
hyperbolic structure with no orthogonality, and thus the containers axiom is vacuous.
(2) For mapping class groups, in the standard structure, a container for domains orthogonal to a given subsurface U is the complementary subsurface, which is orthogonal
to U .
(3) The statement follows immediately from [HS16, Corollary 3.7].
(4) Given a geometric 3–manifold M of the above form, π1 pM q is quasi-isometric to a
(possibly degenerate) product of hyperbolic spaces, and so has clean containers by
Lemma 3.6. Given an irreducible non-geometric graph manifold M , the hierarchically hyperbolic structure comes from considering π1 pM q as a tree of hierarchically
hyperbolic spaces with clean containers and hence has clean containers by Lemma 3.8.
Finally, the general case of a non-geometric 3–manifold M follows immediately from
Lemma 3.7 and the fact that π1 pM q is hyperbolic relative to its maximal graph
manifold subgroups.
l
Lemma 3.6. The product of two hierarchically hyperbolic spaces which both have clean containers has clean containers.
Proof. Let pX0 , S0 q and pX1 , S1 q be hierarchically hyperbolic spaces with clean containers.
In the hierarchically hyperbolic structure pX0 ˆ X1 , Sq given by [BHS15, Theorem 8.25] there
are two types of containers, those that come from one of the original structures and those
that do not. Containers of the first type are clean, as both original structures have clean
containers.
The second type of domain consists of new domains obtained as follows. Given a domain
U P Si , a new domain VU is defined with the property that it contains under nesting any
domain in Si which is orthogonal to U and also any domain in Si`1 . Thus, by construction
VU is a container for everything orthogonal to U . As VU K U , the result follows.
l
Lemma 3.7. If G is hyperbolic relative to a collection of hierarchically hyperbolic spaces which
all have clean containers, then G is a hierarchically hyperbolic space with clean containers.
Proof. That G is a hierarchically hyperbolic space follows from [BHS15, Theorem 9.1]. In
the hierarchically hyperbolic structure on G, no new orthogonality relations are introduced,
and thus all containers are containers in the hierarchically hyperbolic structure of one of the
peripheral subgroups. As each of these structures have clean containers, it follows that G
does, as well.
l
Lemma 3.8. Let T be a tree of hierarchically hyperbolic spaces such that XpT q is hyperbolic.
If for each v P T , the hierarchically hyperbolic space pXv , Sv q has clean containers, then so
does XpT q.
Proof. This follows immediately from the proof of [BHS15, Lemma 8.10] and the fact that
edge-hieromorphisms are full and preserve orthogonality. In the notation from that result,
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
11
we note that, if Sv has clean containers for each v P T , then the domain Av P Sv described
in the proof also has the property that Av K Uv . Therefore, as edge-hieromorphisms are full
and preserve orthogonality, rAv s K rU s.
l
The following uses the notion of hyperbolically embedded subgroups introduced in [DGO17].
Lemma 3.9. Let pG, Sq be a hierarchically hyperbolic group with clean containers, and let
H ãÑhh pG, Sq. Then there exists a finite set F Ă H ´ t1u such that for all N Ÿ H with
F X N “ H and H{N is hyperbolic, then the group G{N̂ , obtained by quotienting by the
normal closure, is a hierarchically hyperbolic group with clean containers.
Proof. Recall that in the hierarchically hyperbolic structure pG{N̂ , SN q obtained in [BHS17a,
Theorem 6.2] (and in the notation used there), two domains U, V P SN satisfy U Ď V
(respectively U K V) if there exists a linked pair tU, V u with U P U and V P V such that
U Ď V (respectively U K V ). Let T P SN and U P pSN qT with V “ tV P ST | V K Uu ‰
H. To prove the container axiom, we consider domains T, U, V P S such that T P T, U P U
and V P V for all V P V, and such that any pair is a linked pair. Then the orthogonality
axiom for pG, Sq provides a domain W such that W Ě V and W Ď T . As pG, Sq has clean
W
containers, we also have that W K U . This implies that ρU
S and ρS are coarsely equal by
[DHS15, Lemma 1.5], and so tU, W u is a linked pair. Therefore, W K U.
l
3.3. A new hierarchically hyperbolic structure. In this section we describe a new hierarchically hyperbolic structure on hierarchically hyperbolic spaces with the bounded domain
dichotomy and clean containers. We first describe the hyperbolic spaces that will be part of
the new structure.
Let pX , Sq be a hierarchically hyperbolic space with the M –bounded domain dichotomy.
Let SM Ă S be the set of U P S such that there exists a V P S and W P SK
V satisfying
M
diampCV q ą M and diampCW q ą M . For each U P S, define SU Ă SU similarly.
Remark 3.10 (Factored spaces). As defined in [BHS17a], given pX , Sq and T Ă S the
p T is the space obtained from X by coning-off all FV for V P T. Sometimes we
factored space F
abuse language slightly and refer to this as the factored space obtained from X by collapsing
T. In particular, when S is the Ď–maximal element of S, then CS is identified with the space
p SztSu , which is obtained from X by coning-off FU for all U P SztSu.
F
We often consider the case of a fixed pX , Sq and U P S and then applying this construction
to the hierarchy hyperbolic structure pFU , SU q. For this application, note that πU pX q is
p S ztU u , by [BHS17a, Corollary 2.9], and thus so is CU , by Remark 2.10.
quasi-isometric to F
U
p SM . By Lemma 3.1, SM is closed under nesting and hence
Consider the factored space F
XpSM is a hierarchically hyperbolic space. Moreover, since this hierarchically hyperbolic space
has the property that no pair of orthogonal domains both have diameter larger than M , by
[BHS17c, Corollary 2.16] it is hyperbolic for some constant depending only on pX , Sq and
M.
For each U P SM , we similarly define TU to be the factored space obtained from FU
by collapsing SM
U . This setup again satisfies the assumptions in Lemma 3.1 and [BHS17c,
Corollary 2.16], so we obtain that TU is δ–hyperbolic for some δ which we can take to be
uniform by finite complexity of pX , Sq.
The next result uses the above spaces to obtain a hierarchically hyperbolic structure with
particularly nice properties from a given hierarchically hyperbolic structure.
Theorem 3.11. Every hierarchically hyperbolic space with the bounded domain dichotomy
and clean containers admits a hierarchically hyperbolic structure with unbounded products.
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
12
Proof. Let pX , Sq be a hierarchically hyperbolic space. Let T Ă S which we define to include
S as well as the set of domains U with both FU and EU unbounded.
We begin to define our new hierarchically hyperbolic structure on X by taking T as the
index set. We let TS “ XpSM be the hyperbolic space associated to the top level domain S,
and let TU be the hyperbolic spaces associated to each U P T, as defined in the discussion
before the theorem.
Notice that for each U P T, the space TU is identical to the space CU if and only if for
every W P S for which W Ď U satisfies W P T.
Now, suppose TU ‰ CU (and hence that there exists at least one W P S with W Ď U
satisfying W R T). As noted above, we can take FU , TU , and CU to have the same underlying
space with different metrics. Even though the identity map on FU no longer induces an
isometry ψU : TU Ñ CU , this map is still 1–Lipschitz. In fact, we now show that ψU is
actually bi-Lipschitz for all U P TztSu. By definition of T, any W P S for which W R T
either FW or EW is bounded. Since U ‰ S, EU is unbounded, and it follows that EW is
unbounded. Since W R T, it follows that FW must be bounded; moreover, pX , Sq has the
M –bounded domain dichotomy, so FW is uniformly bounded. Thus, by the distance formula
in FU , for any pair of points in in FU , their distance in TU can be at most a uniformly
bounded multiple of their distance in CU , and so ψU is a bi-Lipschitz map, with constant M .
To avoid confusion, if U P T, we use the notation dU for distance in TU and the notation
dCU for distance in CU .
We take the associated projections πU to be the composition of the nearest point projection
X Ñ FU and the factor map FU Ñ TU . If U, V P T, we take the relative projections ρVU to be
the preimage under ψU of the corresponding relative projections in pX , Sq whenever U ‰ S.
In the case that U “ S, we take ρVS to be the image of FV under the factor map X Ñ TS .
We now check the axioms for pX , Tq.
Projections: Since the metrics on our new spaces TU for U P T are not the same, we
need to check that these new projections are still coarsely Lipschitz. This, however, is clear,
as πU is the composition of a nearest point projection and a factor map, both of which are
coarsely Lipschitz.
Nesting: The partial order and projections are given by construction. The diameter
bound in the case of nesting projections is immediate from the bound from pX , Sq and the
fact that the maps ψU are bi-Lipschitz for all U P T.
Orthogonality: This is essentially inherited from the hierarchically hyperbolic structure
pX , Sq. The first and last statements are immediate. To prove the second statement, let
T P T and let U P TT be such that tV P TT | V K U u ‰ H. Then as T Ď S, by the
orthogonality axiom for pX , Sq there is a domain W P S such that W Ĺ T and whenever
V Ď T and V K U , we have V Ď W . We will show that W P T.
Indeed, FW is unbounded as there exists some V P T with V Ď W . Furthermore, as the
hierarchically hyperbolic structure pX , Sq has clean containers, W K U . As U P T, it follows
that EW is unbounded, as well.
Transversality and consistency: For U, V P T Ď S with U &V , the relative projections
V
ρU
V , ρU are well-defined and satisfy the required bounds using the constant M ¨ ξ, where ξ is
the original constant from the hierarchically hyperbolic structure on pX , Sq.
We now check the consistency conditions. For all U P T distances in TU can only increase
from distances in CU by factor of M , so the consistency inequalities clearly hold by this fact
and the consistency axiom from pX , Sq. Moreover, we may take the constant to be M ¨ κ0 ,
where κ0 is the original constant from pX , Sq.
Partial realization: If tVj u is a family of pairwise orthogonal domains of T, then tVj u
is a family of pairwise orthogonal domains of S. By the partial realization axiom for pX , Sq,
there is a constant α and a point x P X such that the conclusion holds for all W P S. By
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
13
increasing the constant to M ¨ α, we also have a bound on the appropriate distances in TW ,
and the axiom holds.
Finite complexity: This clearly holds by construction.
Large link axiom: Let λ and E be the constants from the large link axiom for pX , Sq, let
W P T, and let x, x1 P X . Consider the set tTi u Ă SW ´tW u provided by the large link axiom
for pX , Sq. Since Ti Ď W , it follows that ETi is unbounded for each i. Let T P TW ´ tW u.
If dT px, x1 q ą E ¨ M , it follows that FT is unbounded. Furthermore, dCT px, x1 q ą E, whence
T Ď Ti for some i by the large link axiom for pX , Sq. Therefore FTi is unbounded, and so
Ti P T. The result follows.
Bounded geodesic image: For all domains in TztSu, we have increased distances in
the corresponding hyperbolic spaces by no more than M . Hence this axiom holds with the
original constants multiplied by M .
Uniqueness: Let κ ą 0. We can take θu1 ą maxtθu pκq, M u, where θu pκq is the original
constant from the uniqueness axiom for pX , Sq. Then if x, y P X with dpx, yq ą θu1 , then
uniqueness for pX , Sq implies there exists U P S with dCU px, yq ą M . Either U P T or
diampCU q “ 8 and EU is bounded. We are done in the first case. In the second case, by
construction TU is uniformly quasi-isometrically embedded in TS , and hence dS px, yq is at
least a uniform constant depending only on M and the quasi-isometry constants.
l
Corollary 3.12. If pG, Sq is a hierarchically hyperbolic group with clean containers, then
pG, Tq is a hierarchically hyperbolic group with unbounded products.
Proof. Recall that every hierarchically hyperbolic group has the bounded domain dichotomy.
Thus by Theorem 3.11, pG, Tq is a hierarchically hyperbolic space with unbounded products.
It remains only to show that is a hierarchically hyperbolic group structure. The action of
G on itself is clearly metrically proper and cobounded, so it only remains to show that T
contains finitely many G–orbits. If U P S but U R T, then either FU or EU must be bounded.
Then for each g P G, the same will be true for FgU or EgU , which shows that gU R T. Thus
G ¨ U Ć T. Since S has finitely many G–orbits, the result follows.
l
4. Characterization of contracting geodesics
For this section, fix a hierarchically hyperbolic space pX , Sq with the bounded domain
dichotomy; denote the Ď–maximal element S P S.
Definition 4.1 (Bounded projections). Let Y Ă X and D ą 0. We say that Y has D–
bounded projections if for every U P SztSu, we have dU pYq ă D.
Definition 4.2 (Contracting). A quasigeodesic γ in a metric space X is said to be D–
contracting if there exist a map πγ : X Ñ γ and constants A, D ą 0 satisfying:
(1) For any x P γ, we have diamX px, πγ pxqq ă D;
(2) If x, y P X with dX px, yq ă 1, then diamX pπγ pxq, πγ pyqq ă D;
(3) For all x P X, if we set R “ A ¨ dpx, γq, then diampπγ pBR pxqqq ď D.
Sometimes authors refer to any quasigeodesic satisfying (3) as contracting. Nonetheless, for
applications one also needs to assume the coarse idempotence and coarse Lipschitz properties
given by (1) and (2), so for convenience we combine them all in one property.
A useful well-known fact is stability of contracting quasigeodesics. Two different proofs of
the following occur as special cases of the results [MM99, Lemma 6.1] and [Beh06, Lemma 6.2];
this explicit statement can also be found in [DT15, Section 4].
Lemma 4.3. If γ is a D–contracting pK, Kq–quasigeodesic in a metric space X, then γ is
D 1 –stable for some D 1 depending only on D and K.
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
14
The following result and argument both generalize and simplify the analogous result for
mapping class groups in [Beh06].
Theorem 4.4. Let pX , Sq be a hierarchically hyperbolic space. For any D ą 0 and K ě 1
there exists a D 1 ą 0 depending only on D and pX , Sq such that the following holds for every
pK, Kq–quasigeodesic γ Ă X . If γ has D–bounded projections, then γ is D 1 –contracting.
Moreover, if pX , Sq has the bounded domain dichotomy and clean containers, then X admits
a hierarchically hyperbolic structure pX , Tq with unbounded products where, additionally, we
have that if γ is D–contracting, then γ has D 1 –bounded projections.
Remark 4.5. In Section 7, we introduce a technical trick that allows us to remove the
assumption that pX , Sq has clean containers. Thereby Theorem E follows from Theorem 4.4.
Proof. First suppose that γ has D–bounded projections. It follows immediately from the
definition that γ is a hierarchically quasiconvex subset of X . Hierarchical quasiconvexity is
the hypothesis necessary to apply [BHS17a, Lemma 5.4], which then yields existence of a
coarsely Lipschitz gate map g : X Ñ γ, i.e., for each x P X , the image gpxq P γ has the
property that for all U P S the set πU pgpxqq is a uniformly bounded distance from the
projection of πU pxq to πU pγq.
We will use g as the map to prove γ is contracting. Gate maps satisfy condition (1) of
Definition 4.2 by definition and condition (2) since they are coarsely Lipschitz. Hence it
remains to prove that condition (3) of Lemma 4.3 holds.
Fix a point x P X with dX px, γq ě B0 and let y P X be any point with dX px, yq ă
B1 dX px, γq for constants B0 and B1 as determined below.
Since g is a gate map and γ has D–bounded projections, for all U P S ´ tSu we have
dU pgpxq, gpyqq ă D. Thus, by taking a threshold L for the distance formula (Theorem 2.5)
larger than D, we have
dX pgpxq, gpyqq —pK,Cq dS pgpxq, gpyqq,
for uniform constants K, C. Thus it suffices to prove that dS pgpxq, gpyqq is bounded by some
uniform constant B2 . We also choose L to be larger than the constants in Definition 2.1.(4).
By Definition 2.1.(1), the maps πU are Lipschitz with a uniform constant. Taking B0
sufficiently large, it follows that there exists U P S so that dU px, gpxqq ą L. By choosing B1
to be sufficiently small, and applying the distance formula to the pairs px, yq and px, gpxqq, the
fact that the projections πU are Lipschitz implies that the sum of the terms in the distance
formula associated to px, gpxqq ř
is much greater than
of those associated to px, yq.
ř the sum ř
Having chosen B1 ă 21 , we have dU px, gpxqq ą 2 dU px, yq ą pdU px, yq`Lq. Thus, there
exists W P S for which dW px, gpxqq ą dW px, yq ` L.
If W “ S, then having dS px, gpxqq ą dS px, yq ` L (where we enlarge L if necessary)
would already show that the CS–geodesic between x and y was disjoint from πS pγq and then
hyperbolicity of CS would yield a uniform bound on the dS pgpxq, gpyqq.
Otherwise, we may assume W ‰ S. By the triangle inequality, we have dW py, gpxqq ą L.
Further, since, as noted above, the CW projections between gpxq and gpyq are uniformly
bounded, by choosing B0 large enough and B1 small enough, we also have dW py, gpyqq ą L.
By the bounded geodesic image axiom (Definition 2.1.(7)), any geodesic in CS either has
bounded projection to CU or satisfies πS pγq X NE pρU
S q ‰ H for any U P S ´ tSu. For any
geodesic from πS pxq to πS pgpxqq (or from πS pyq to πS pgpyq), the above argument implies that
the first condition doesn’t hold for W . Thus, in both cases, we know that any such geodesic
must pass uniformly close to ρW
S . Hence the hyperbolicity of CS implies γ is contracting, and
the first implication holds.
We prove the second implication by contradiction. By Theorem 3.11, we obtain a new
structure pX , Tq which has unbounded products. For every U P TztSu we have that both
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
15
FU and EU are unbounded, hence every U P TztSu yields a non-trivial product region
PU “ EU ˆ FU which is uniformly quasi-isometrically embedded in X .
Suppose γ is contracting but doesn’t have D–bounded projections. Then we obtain a
sequence tUi u P TztSu with diampπCUi pγqq Ñ 8. Thus there is a sequence of pairs of points
xi , yi P γ, so that dUi pxi , yi q — Ki , with Ki Ñ 8. For each i, let qi be a R–hierarchy path
between xi , yi . By [BHS15, Proposition 8.24], there exists ν ą 0 depending only on R and
pX , Sq, such that
diamUi pqi X Nν pPUi qq — Ki .
Since γ is contracting, it is uniformly stable by Lemma 4.3. Since γ is uniformly stable
and the qi are uniform quasigeodesics, it follows that each qi is contained in a uniform
neighborhood of γ. Hence arbitrarily long segments of γ are uniformly close to the product
regions PUi . This contradicts the assumption that γ is contracting and completes the proof.
l
5. Universal and largest acylindrical actions
The goal of this section is to show that for every hierarchically hyperbolic group pG, Sq
with clean containers AHpGq has a largest element. Recall that the action associated to such
an element is necessarily a universal acylindrical action.
We prove the following stronger result which, in addition to providing new largest and
universal acylindrical actions for cubulated groups, gives a single construction that recovers
all previously known largest and universal acylindrical actions of finitely presented groups
that are not relatively hyperbolic.
Theorem 5.1. Every hierarchically hyperbolic group with clean containers admits a largest
acylindrical action.
Remark 5.2. In Section 7, we introduce a technical trick that allows us to remove the
assumption that pX , Sq has clean containers. Whence Theorem 5.1 implies Theorem A.
Before giving the proof, we record the following result which gives a sufficient condition
for an action to be largest. This result follows directly from the proof of Theorem 6.3 in
[ABO16]; we give a sketch of the argument here. Recall that an action H ñ S is elliptic if
H has bounded orbits.
Proposition 5.3 ([ABO16]). Let G be a group, tHŤ
1 , . . . , Hn u a finite collection of subgroups
of G, and F be a finite subset of G such that F Y p ni“1 Hi q generates G. Assume that:
Ť
(1) ΓpG, F Y p ni“1 Hi qq is hyperbolic and acylindrical.
(2) Each Hi is elliptic in every acylindrical action of G on a hyperbolic space.
Ť
Then rF Y p ni“1 Hi qs is the largest element in AHpGq.
Ť
Proof. First notice that by assumption (1), ΓpG, F Y p ni“1 Hi q is an element of AHpGq.
Let G ñ S be a cobounded acylindrical action of G on a hyperbolic space, S,Ťand fix a
basepoint s P S. Then there exists a bounded subspace B Ă S such that S Ď gPG g ¨ B.
By assumption (2), the orbit Hi ¨ s is bounded for all i “ 1, . . . , n. Since |F | ă 8, we know
diampF ¨ sq ă 8 and thus
K “ maxtdiampBq, diampH1 ¨ sq, . . . , diampHn ¨ sq, diampF ¨ squ
is finite. Let C “ ts1 P S | dps1 , sq ď 3Ku, and let
Z “ tg P G | g ¨ C X C ‰ Hu.
The standard Milnor-Schwartz Lemma argument shows that Z is an infinite generating
set of G and there exists a G–equivariant quasi-isometry S Ñ ΓpG, Zq. It is clear that Z
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
16
Ť
contains F , as well as Hi for all i “ 1, . . . , n and thus rZs ĺ rF Y p ni“1 Hi s. The result
follows.
l
Proof of Theorem 5.1. Let pG, Sq be a hierarchically hyperbolic group with finite generating
set F . By Corollary 3.12, there is a hierarchically hyperbolic group structure pG, Tq with
unbounded products. Recall that S is the Ď–maximal element of T with associated hyperbolic
space TS . The action on TS is acylindrical by [BHS17b, Theorem K].
Moreover, the action of G on TS is cobounded, so let B be a fundamental domain for
G ñ TS and
U “ tU P T | πS pFU q Ă B and U is Ď–maximal in TztSuu.
Notice that U will contain exactly one representative from each G–orbit of domains, and so
must be a finite set. Indeed, for a hierarchically hyperbolic group, this follows from the fact
that the action of G on T is cofinite.
Let Hi ď G be the stabilizer of FUi for each Ui P U . By a standard Milnor-Schwartz
argument
(see [ABO16] for details) there is a G–equivariant quasi-isometry between ΓpG, F Y
Ť
p ni“1 Hi qq and TS , where n “ |U |. Therefore condition (1) of Proposition 5.3 is satisfied.
By definition, each Hi sits inside a non-trivial direct product in G, the product region PUi
associated to each Ui P U . It follows that Hi must be elliptic in every acylindrical action of
G on a hyperbolic space (see Remark 2.14), satisfying condition (2).
Therefore, by Proposition 5.3, the action is largest.
l
Remark 5.4. The proof of Theorem 5.1 can be extended to treat a number of groups which
are hierarchically hyperbolic spaces, but not hierarchically hyperbolic groups. For example,
it was shown in [BHS15, Theorem 10.1] that every fundamental group of a 3–manifold with
no Nil or Sol in its prime decomposition admits a hierarchically hyperbolic space structure,
but as explained in [BHS15, Theorem 10.2] it is likely that these don’t all admit hierarchically
hyperbolic group structures. Nonetheless, the proof of the above theorem works in this case
by replacing the use of the fact that the action of G on T is cofinite, with the fact that for
π1 pM q, U is precisely the set of Ď–maximal domains in the hierarchically hyperbolic structure
on each of the Seifert-fibered components of the prime decomposition of M , and so is finite.
Remark 5.5. There is an instructive direct proof of the universality of the above action
using the characterization of contracting elements in Section 4, which we now give. Let
g P G be an infinite order element and consider the geodesic xgy in ΓpG, F q.
If xgy is contracting in ΓpG, F q, then by Theorem 4.4 all proper projections are bounded,
and thus by the distance formula, g is loxodromic for the action on TS .
If xgy is not contracting in ΓpG, F q, then there exists some U P T such that πU pxgyq is
unbounded. Thus for any increasing sequence of constants pKi q with Ki Ñ 8, there are
sequences of pairs of points xi , yi P xgy such that dpxi , yi q Ñ 8 as i Ñ 8 and dU pxi , yi q ě Ki .
For each i, let γi be an R–hierarchy path between xi and yi . By definition, γi is a uniform
quasigeodesic. Then by [BHS15, Proposition 8.24], there exists ν ą 0 depending only on R
and pX , Tq such that diamU pγi X Nν pPU qq ě Ki . If g is a generalized loxodromic, then xgy
is stable, by [Sis16], and so the subgeodesic rxi , yi s stays within a uniform bounded distance
of γi . Thus arbitrarily long subgeodesics of xgy stay within a uniformly bounded distance of
a product region, PU . This contradicts xgy being Morse, and therefore g is not a generalized
loxodromic element.
This remark directly implies that the action on TS is a universal acylindrical action. (The
universality of the action can also be proven using the classification of elements of AutpSq
described in [DHS15].)
Another immediate consequence of the above remark is the following, which for hierarchically hyperbolic groups strengthens a result obtained by combining Osin [Osi16, Theorem
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
17
1.2.(L4)] and Sisto [Sis16, Theorem 1], which together prove that a generalized loxodromic
element in an acylindrically hyperbolic group is quasi-geodesically stable.
Corollary 5.6. Let pG, Sq be a hierarchically hyperbolic group. An element g P G is generalized loxodromic if and only if xgy is contracting in Γ.
The next result provides information about the partial ordering of acylindrical actions. Of
the groups listed below, the largest and universal acylindrical action of the class of CAT(0)
cubical groups is new; the other cases were recently established to be largest in [ABO16].
Corollary 5.7. The following groups admit acylindrical actions that are largest (and therefore universal):
(1) Hyperbolic groups and their subgroups
(2) Mapping class groups
(3) Fundamental groups of three manifolds with no Nil or Sol in their prime decomposition
(4) Groups that act properly and cocompactly on a proper CAT(0) cube complex. This includes right angled-Artin groups and right-angled Coxeter groups.
Proof. With the exception of p3q the above are all hierarchically hyperbolic groups [BHS17b,
BHS15, HS16] and therefore have the bounded domain dichotomy. If G is the fundamental
group of a three-manifold with no Nil or Sol in its prime decomposition, then while G is not
always a hierarchically hyperbolic group, it has a hierarchically hyperbolic structure pX , Sq
such that X is the Cayley graph of G and G ă AutpSq. Additionally, all of the associated
hyperbolic spaces are infinite, and therefore pX , Sq has the bounded domain dichotomy. By
Theorem 3.5, the above groups each have clean containers, so the result follows.
l
We give an explicit description of these actions for each hierarchically hyperbolic group in
the corollary, in the sense that we describe the set W of domains which are removed from the
standard hierarchical structure of the group. Recall that the space TS is constructed from X
by coning off all elements of T “ SzW.
(1) Hyperbolic groups (and their subgroups) have a canonical simplest hierarchically hyperbolic group structure given by taking S “ tSu, where CS is the Cayley graph of the
group with respect to a finite generating set. For this structure, W “ H, and the action
on the Cayley graph is clearly a universal acylindrical action.
(2) For mapping class groups, the natural hierarchically hyperbolic group structure is S is
the set of stabilizers of simple closed curves on the surface and a maximal element S,
where CS is the curve complex. For this structure, W “ H, and the action on the
curve complex is universal. Universality of this action was shown by Osin in [Osi16], and
follows from results of Masur-Minsky and Bowditch [Bow08, MM99].
(3) If M is a 3–manifold with no Nil or Sol in its prime decomposition and G “ π1 M , then W
is exactly the set of vertex groups in the prime decomposition that are fundamental groups
of hyperbolic 3–manifolds (each of which has exactly one domain in its hierarchically
hyperbolic structure).
(4) If G is a group that acts properly and cocompactly on a CAT(0) cube complex X, then
by [HS16], X has a G–equivariant factor system. This factor system gives a hierarchically
hyperbolic group structure in which S is the closure under projection of the set of
hyperplanes along with a maximal element S, where CS is the contact graph as defined
in [Hag14]. In this structure, W is the set of indices whose stabilizer in G contains a
power of a rank one element.
In the particular case of right-angled Artin groups, no power of a rank one element will
stabilize a hyperplane, so W “ H. In this case, the contact graph CS is quasi-isometric
to the extension graph defined by [KK14]. That the action on the extension graph is a
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
18
universal acylindrical action follows from the work of [KK14] and the centralizer theorem
for right-angled Artin groups. This action is also shown to be largest in [ABO16].
We give a concrete example of the situation in the case of a right-angled Coxeter group.
Example 5.8. Let G be the right-angled Coxeter group whose defining graph is a pentagon. Then G “ xa, b, c, d, e | ra, bs, rb, cs, rc, ds, rd, es, ra, esy, and the Cayley graph of G
is the tiling of the hyperbolic plane by pentagons. We consider the dual square complex
to this tiling. To form the contact graph CS, we start with the square complex and cone
off each hyperplane carrier, which is equivalent to coning off the hyperplane stabilizers in
the Cayley graph. The result is a quasi-tree. Thus a fundamental domain for the hierarchically hyperbolic group structure of G is tUa , Ub , Uc , Ud , Ue , Su where Uv is associated
to the stabilizer of the hyperplane labeled by v and S is associated to the contact graph
described above.
Consider the hyperplane Jb that is labeled by b. Then the stabilizer of Jb is the link
of the vertex b, which contains the infinite order element ac. As G is a hyperbolic group,
all infinite order elements are generalized loxodromic, but ac is not loxodromic for the
action on the contact graph since its axis lies in a hyperplane stabilizer that has been
coned-off. Thus the action on the contact graph is not universal.
Let Ub P S be the element associated to StabpJb q. Then StabpJb q “ xa, b, c |
ra, bs, rb, csy » D8 ˆ Z{2Z » FUb ˆ EUb is a product region, and the maximal orthogonal
component EUb is bounded. Thus Ub P W, as is Uv , for each vertex v of the defining graph.
The contact graph associated to pFUb , SUb q is a line, and the element ab is loxodromic
for the action on this space.
Note that once W has been removed from S, the resulting hierarchically hyperbolic
structure is pG, tSuq, the canonical hierarchically hyperbolic structure for a hyperbolic
group, in which CS “ ΓpG, ta, b, c, d, euq.
6. Characterizing stability
In this section, we will give several characterizations of stability which hold in any hierarchically hyperbolic group. In fact, we will characterize stable embeddings of geodesic metric
spaces into hierarchically hyperbolic spaces with unbounded products. One consequence of
this will be a description of points in the Morse boundary of a proper geodesic hierarchically
hyperbolic space with unbounded products as the subset of the hierarchically hyperbolic
boundary consisting of points with bounded projections.
6.1. Stability. While it is well-known that contracting implies stability [Beh06, DMS10,
MM99], the converse is not true in general. Nonetheless, in several important classes of
spaces the converse holds, including in hyperbolic spaces, CAT(0) spaces, the mapping class
group, and Teichmüller space [Sul14, Beh06, DT15, Min96]. We record the following relation
between stability and contracting subsets which holds in a fairly general context:
Corollary 6.1. If pX , Sq has unbounded products and Y Ă X , then Y is N –stable if and
only if Y is D–contracting, where N and D determine each other.
Proof. Lemma 4.3 shows that contracting implies stable (the assumption on unbounded products is not necessary for this implication). For the other direction, the fact that X has unbounded products implies that Y has bounded projections, since otherwise one could find
large segments of quasigeodesics contained inside direct products, contradicting stability. The
result now follows from Theorem 4.4.
l
The following provides a general characterization of stability:
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
19
Corollary 6.2. Let i : Y Ñ X be map from a metric space into a hierarchically hyperbolic
space pX , Sq with unbounded products. The following are equivalent:
(1) i is a stable embedding;
(2) ipYq is undistorted and has uniformly bounded projections;
(3) πS ˝ i : Y Ñ CS is a quasi-isometric embedding.
Proof. Items (1) and (2) are equivalent via Corollary 6.1. Equivalence of (2) and (3) follows
from the distance formula and the assumption that i is a quasi-isometric embedding.
l
6.2. The Morse boundary. In the rest of this section, we turn to studying the Morse
boundary and use this to give a bound on the stable asymptotic dimension of a hierarchically
hyperbolic space. We begin by describing two notions of boundary.
In [DHS15], Durham, Hagen, and Sisto introduced a boundary for any hierarchically hyperbolic space. We collect the relevant properties we need in the following theorem:
Theorem 6.3 (Theorem 3.4 and Proposition 5.8 in [DHS15]). If pX , Sq is a proper hierarchically hyperbolic space, then there exists a topological space BX such that BX Y X “ X
compactifies X , and the action of AutpX , Sq on X extends continuously to an action on X .
Moreover, if Y is a hierarchically quasiconvex subspace of X , then, with respect to the induced hierarchically hyperbolic structure on Y, the limit set of ΛY of Y in BX is homeomorphic
to BY and the inclusion map i : Y Ñ X extends continuously an embedding Bi : BY Ñ BX .
Building on ideas in [CS15], Cordes introduced the Morse boundary of a proper geodesic
metric space [Cor15], which was then refined further by Cordes–Hume in [CH16]. The Morse
boundary is a stratified boundary which encodes the asymptotic classes of Morse geodesic
rays based at a common point. Importantly, it is a quasi-isometry invariant and generalizes
the Gromov boundary of a hyperbolic space [Cor15].
We briefly discuss the construction of the Morse boundary and refer the reader to [Cor15,
CH16] for details.
Consider a a proper geodesic metric space X with a basepoint e P X. Given a stability
pN q
gauge N : R2ě0 Ñ Rě0 , define a subset Xe
Ă X to be the collection of points y P X
pN q
such that e and y can be connected by an N –stable geodesic in X. Each such Xe is δN –
hyperbolic for some δN ą 0 depending on N and X [CH16, Proposition 3.2]; here, we use
pN q
the Gromov product definition of hyperbolicity, as Xe need not be connected. Moreover,
pN q
any stable subset of X embeds in Xe for some N [CH16, Theorem A.V].
The set of stability gauges admits a partial order: N1 ă N2 if and only if N1 pK, Cq ă
pN q
pN q
N2 pK, Cq for all constants K, C. In particular, if N1 ă N2 , then Xe 1 Ă Xe 2 .
pN q
pN q
Since each Xe is Gromov hyperbolic, each admits a Gromov boundary BXe . Take the
direct limit with respect to this partial order to obtain a topological space Bs X called the
Morse boundary of X.
We fix pX , Sq, a hierarchically hyperbolic structure with unbounded products.
Definition 6.4. We say λ P BX has bounded projections if for any e P X , there exists D ą 0
such that any R–hierarchy path re, λs has D–bounded projections. Let Bc X denote the set
of points λ P BX with bounded projections.
The boundary BX contains BCU for each U P S, by construction. The next lemma shows
that the boundary points with bounded projections are contained in BCS, as a subset of BX ,
where S is the Ď–maximal element. In general, the set of cobounded boundary points may
be a very small subset of BCS. For instance, in the boundary of the Teichmüller metric, these
points are a proper subset of the uniquely ergodic ending laminations and have measure zero
with respect to any hitting measure of a random walk on the mapping class group.
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
20
Lemma 6.5. The inclusion Bc X Ă BCS holds for any pX , Sq with unbounded products where
S is the Ď–maximal element of S. Moreover, if X is also proper, then for any D ą 0 there
exists D 1 ą 0 depending only on D and pX , Sq such that if pxn q Ă X is a sequence with
xn Ñ λ P BX such that re, xn s has D–bounded projections for some e P X and each n, then
re, λs has D 1 –bounded projections.
Proof. Let λ P Bc X . If re, λs is an R–hierarchy path, then re, λs has an infinite diameter
projection to some CU , see, e.g., [DHS15, Lemma 3.3]. As λ has bounded projections, we
must have U “ S. Since πS pre, λsq Ă CS is a quasigeodesic ray, the first statement follows.
Now suppose that X is also proper. For each n, let γn “ re, xn s be any R–hierarchy
path between e and xn in X . The Arzela-Ascoli theorem implies that after passing to a
subsequence, γn converges uniformly on compact sets to some R1 –hierarchy path γ with
R1 depending only on R and pX , Sq. Hence γ has D 1 –bounded projections for some D 1
depending only on D and pX , Sq. Moreover, since xn Ñ λ in CS, it follows that πS pγq is
asymptotic to λ in CS.
If re, λs is any other R1 –hierarchy path, it follows from uniform hyperbolicity of the CU
and the definition of hierarchy paths that dHaus
pγ, re, λsq is uniformly bounded for all U P S.
U
Since γ has D 1 –bounded projections, the distance formula implies that re, λs has D 2 –bounded
projections for some D 2 depending only on D and pX , Sq, as required.
l
6.3. Bounds on stable asymptotic dimension. The asymptotic dimension of a metric
space is a coarse notion of topological dimension which is invariant under quasi-isometry.
Introduced by Cordes–Hume [CH16], the stable asymptotic dimension of a metric space X
is the maximal asymptotic dimension a stable subspace of X.
The stable asymptotic dimension of a metric space X is always bounded above by its
asymptotic dimension. Behrstock, Hagen, and Sisto, [BHS17a] proved that all proper hierarchically hyperbolic spaces have finite asymptotic dimension (and thus have finite stable
asymptotic dimension, as well). The bounds on asymptotic dimension obtained in [BHS17a]
are functions of the asymptotic dimension of the top level curve graph.
In the following theorem, we prove that a hierarchically hyperbolic space pX , Sq has finite
stable asymptotic dimension under the assumption that asdimpCSq ă 8. Recall that asymptotic dimension is monotonic under taking subsets. Thus, if X is assumed to be proper,
so that asdimpCSq ă 8, then X (and therefore the stable subsets) have finite asymptotic
dimension by [BHS17a]. Here, using some geometry of stable subsets we obtain a sharper
bound on asdims pX q than asdimpX q.
Theorem 6.6. Let pX , Sq be a hierarchically hyperbolic space with unbounded products such
that CS has finite asymptotic dimension. Then asdims pX q ď asdimpCSq. Moreover, if X is
also proper and geodesic, then there exists a continuous bijection pi : Bs X Ñ Bc X .
pN q
Proof. By [CH16, Lemma 3.6], for any stability gauge N there exists N 1 such that Xe is
pN q
N 1 –stable. Hence, there exists D 1 ą 0 depending only on N 1 and pX , Sq such that Xe is
pN q
D 1 –cobounded. By Corollary 6.2, it follows that the projection πS : Xe Ñ CS is a quasiisometric embedding with constants depending only on D 1 and pX , Sq. Since every stable
pN q
subset of X embeds into some Xe [CH16, Theorem A.V], the first conclusion then follows
from the definition of stable asymptotic dimension.
Now suppose that X is proper.
pN q
Since each Xe is stable in X , these sets have bounded projections by Corollary 6.2; from
pN q
this it follows that Xe is hierarchically quasiconvex for each N . Hence by [DHS15, ProposipN q
pN q
tion 5.8], the canonical embedding ipN q : Xe ãÑ X extends to an embedding pipN q : BXe ãÑ
BX .
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
21
¯
´
pN q
Ă Bc X Ă BCS. Let pi : Bs X Ñ Bc X
By Corollary 6.2 and Lemma 6.5, we have pipN q BXe
be the direct limit of the pipN q . Since it is injective on each stratum, pi is injective.
To prove surjectivity, let λ P Bc X . Let e P X and fix a hierarchy path re, λs. Since
λ P Bc X , re, λs has D–bounded projections for some D ą 0. Let xn P re, λs be such that
xn Ñ λ in X . If re, xn s is a sequence of geodesics between e and xn , then, by properness, the
Arzela–Ascoli theorem, and passing to a subsequence if necessary, there exists a geodesic ray
γ : r0, 8q Ñ X with γp0q “ e such that re, xn s converges on compact sets to γ. Since each
re, xn s has D–bounded projections, it follows that γ has D 1 –bounded projections for some D 1
depending only on D and pX , Sq. Moreover, by hyperbolicity of CS and the construction of
γ we have that dHaus
CS pπS pγq, re, λsq is uniformly bounded and thus, by the distance formula,
Haus
so is dX pγ, re, λsq. Since rpxn qs “ rγs by construction, it follows that pipγq “ λ, as required.
Continuity of pipN q for each N follows from [DHS15, Proposition 5.8], as above. This and
the definition of the direct limit topology implies continuity of pi.
l
The following corollary is immediate:
Corollary 6.7. If G is a hierarchically hyperbolic group with clean containers, then G has
finite stable asymptotic dimension.
6.4. Random subgroups. Let G be any countable group and µ a probability measure on
G whose support generates a non-elementary semigroup. A k–generated random subgroup of
G, denoted Γpnq is defined to be the subgroup xwn1 , wn2 , . . . , wnk y Ă G generated by the nth
step of k independent random walks on G, where k P N. For other recent results on the
geometry of random subgroups of acylindrically hyperbolic groups, see [MS17].
Following Taylor-Tiozzo [TT16], we say a k–generated random of G has a property P if
PrΓpnq
has
Ps Ñ 1
as
n Ñ 8.
Theorem 6.8. Let pX , Sq be a HHS for which the Ď–maximal element, S, has CS infinite
diameter, and consider G ă AutpX , Sq which acts properly and cocompactly on X . Then any
k–generated random subgroup of G stably embeds in X .
Proof. By [BHS17b, Theorem K], G acts acylindrically on CS. Let Γpnq be generated by
k–random independent walks as above. Now, [TT16, Theorem 1.2] implies that Γpnq a.a.s.
quasi-isometrically embeds into CS, and hence any orbit of Γpnq in X has bounded projections.
By Theorem 4.4, having bounded projections implies contracting; thus any orbit of Γpnq in
X is a.a.s. contracting, which gives the conclusion.
l
In particular, one consequence is a new proof of the following result of Maher–Sisto. This
result follows from the above, together with Rank Rigidity for HHG (i.e, [DHS15, Theorem 9.14]) which implies that a hierarchically hyperbolic group which is not a direct product
of two infinite groups has CS infinite diameter.
Corollary 6.9 (Maher–Sisto; [MS17]). If G is a hierarchically hyperbolic group which is not
the direct product of two infinite groups, then any k–generated random subgroup of G is stable.
7. Almost hierarchically hyperbolic spaces
While many hierarchically hyperbolic groups of interest have clean containers (see Proposition 3.5), groups exist that do not. In this section, we explain a trick which allows us to
generalize the results of the previous sections to remove the clean containers hypothesis.
In the case of a hierarchically hyperbolic group pG, Sq without clean containers, the construction in Theorem 3.11 yields a structure which is not a hierarchically hyperbolic space.
However, the only aspect of the definition of a hierarchically hyperbolic space that may fail
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
22
to hold is part of the orthogonality axiom, Definition 2.1.(3). Indeed, given T P T and some
U P TT for which tV P TT | V K U u ‰ H, the container W P ST ´ tT u provided by
the structure pX , Sq may not be an element of T. In particular, it is possible that EW is
bounded.
We introduce the notion of an almost hierarchically hyperbolic space, and use it to show
that such spaces satisfy many of the same properties as hierarchically hyperbolic spaces.
The following is a weaker version of the orthogonality axiom:
p31 q (Bounded pairwise orthogonality) T has a symmetric and anti-reflexive relation
called orthogonality: we write V K W when V, W are orthogonal. Also, whenever
V Ď W and W K U , we require that V K W . Moreover, if V K W , then V, W are
not Ď–comparable. Finally, the cardinality of any collection of pairwise orthogonal
domains is uniformly bounded by ξ.
By [BHS15, Lemma 2.1], the orthogonality axiom (Definition 2.1, (3)) for an hierarchically
hyperbolic structure implies axiom p31 q. However, the converse does not hold; that is, the
last condition of p31 q does not imply the container statement in (3), and thus this is a strictly
weaker assumption. However, it suffices for the applications in this paper.
Definition 7.1 (Almost HHS). If pX , Sq satisfies all axioms of a hierarchically hyperbolic
space except (3) and additionally satisfies axiom p31 q, then pX , Sq is an almost hierarchically
hyperbolic space.
Theorem 7.2. Given a hierarchically hyperbolic space pX , Sq with the bounded domain
dichotomy, there exists an almost hierarchically hyperbolic structure pX , Tq on X with unbounded products satisfying the following properties:
(1) (Distance formula) There exists s0 such that for all x ě x0 there exist constants
K, C such that for all x, y P X ,
ÿ
dX px, yq —pK,Cq
ttdW px, yquus .
W PT
(2) (Acylindricity) Let G ď AutpX , Tq act properly and cocompactly on X and let S
be the unique Ď–maximal element of T. Then G acts acylindrically on the hyperbolic
space TS associated with S.
(3) (Realization) For each κ ě 1 there exist θe , θu ě 0 such that the following holds.
ś
Let ~b P W PT 2TW be κ–consistent; for each W , let bW denote the TW –coordinate of
~b. Then there exists x P X such that dW pbw , πW pxqq ď θe for all W P T. Moreover, x
is coarsely unique in the sense that the set of all x which satisfy dW pbW , πW pxqq ď θe
in each W P T has diameter at most θu .
(4) (Hierarchy paths) There exists D0 such that any x, y P X are joined by a D0 –
hierarchy path.
(5) (Gate maps) If Y Ď X is k–hierarchically quasiconvex and non-empty, then there
exists a gate map for Y, i.e., for each x P X there exists y P Y such that for all
W P T, the set πV pyq (uniformly) coarsely coincides with the projection of πV pxq to
the kp0q–quasiconvex set πV pYq.
(6) (Coarse median structure) The space X is coarse median of rank at most the
complexity of pX , Tq.
Remark 7.3. In fact, many of the above results hold for all almost hierarchically hyperbolic
spaces and do not require the existence of an original hierarchically hyperbolic structure.
However, to avoid complicated notation and proofs, we restrict ourselves to this situation.
Proof. The new structure pX , Tq is as described in section 3.4.2. All of the axioms of Definition 2.1 except (3) hold as in the proof of Theorem 3.14. Finally, we now show axiom p31 q is
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
23
satisfied by this new structure. Indeed, the first three conditions are clear, since T Ď S. For
the last condition, any collection of pairwise orthogonal domains in T is also a collections of
pairwise orthogonal domains in S, and thus by [BHS15, Lemma 2.2] has uniformly bounded
size. Therefore pX , Tq is an almost hierarchically hyperbolic space. That it has unbounded
products is clear from the construction.
We now prove the properties listed above.
(1) The proof of the distance formula from [BHS15, Theorem 4.5] goes through almost
verbatim. The only use of the container part of the orthogonality axiom throughout
that entire paper is in that proof of [BHS15, Lemma 2.2] which proves that the
cardinality of any collection of pairwise orthogonal domains is uniformly bounded by
ξ. As we have adopted the conclusion of that result as part of p31 q, the result follows.
(2) The original proof of acylindricity in [BHS17b, Theorem 14.3] does not use the orthogonality axiom, only the distance formula, and therefore goes through as written.
(3) The proof of realization is more involved, and so we provide it in detail.
First, let R “ tU P S | FU is boundedu, and let S1 “ SzR. As R is closed under
nesting, by [BHS17a, Proposition 2.4] pX , S1 q is a hierarchically hyperbolic space.
Notice that T Ď S1 .
We follow the original proof of realization in [BHS15, Theorem 3.1].
Let tVj u be a family of pairwise orthogonal elements of T, all of level at most l.
By the last clause of the new orthogonality axiom (Definition 7.1 (31 )), we have
|tVj u| ď ξ. Thus, there exists some l1 such that tVj u is a family of pairwise orthogonal
elements of S1 , all of level at most l1 . Then Claim 1 of the proof of [BHS15, Theorem
3.1] provides a constant θe “ θe pl1 , κq ą 100Eκα and a collection tUi u of pairwise
orthogonal elements of S1 so that:
(a) Each Ui is nested into some Vj ,
(b) For each Vj there exists some Ui nested into it, and
(c) Any E–partial realization point x for tUi u satisfies dW pbW , xq ď θe for each
W P S1 for which there exists j with W Ď Vj .
As Ui P S1 for all i, FUi is unbounded. By (1), it follows that EUi is unbounded
for all i, as well, and therefore Ui P T for all i. After possibly increasing the constant,
condition (c) still holds for W P T. Therefore, Claim 1 holds for pX , Tq, as well.
Now, applying Claim 1 when l “ lS , where S P T is the unique Ď–maximal
element, along with the partial realization axiom, completes the proof of existence.
If x, y both have the desired property, then dV px, yq ď 2θe ` κ for all V P T, whence
the uniqueness axiom ensures that dpx, yq ď θu , for an appropriate θu .
(4) The hierarchy paths in the new structure pX , Tq are the same as those in pX , Sq. As
T Ă S, all the required properties hold.
(5) The definition of a hierarchically quasiconvex subset passes through to the almost
hierarchically hyperbolic setting without issue. The proof of the existence of gate
maps to hierarchically quasiconvex subsets [BHS15, Lemma 5.4] uses only the existence of hierarchy paths, realization, and consistent centers [BHS15, Lemma 2.6].
The first two are shown to hold for almost hierarchically hyperbolic spaces above,
and the proof of consistent centers does not use the orthogonality axiom. Therefore,
the proof of the existence of gate maps goes through as written.
(6) The proof of [BHS15, Theorem 7.3] relies only on consistent centers and realization,
both of which hold in the present setting, as discussed above.
l
As all hierarchically hyperbolic groups satisfy the bounded domain dichotomy, the following is immediate:
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
24
Corollary 7.4. Every hierarchically hyperbolic group admits an almost hierarchically hyperbolic group structure with unbounded products.
The results in this paper hold in the more general setting of almost hierarchically hyperbolic
spaces, as the proofs only rely on axioms (1)-(2),(31 ),(4)-(9) of a hierarchically hyperbolic
space and the consequences listed in Theorem 7.2. However, to avoid complicated statements
and notation, we stated the results only in the more restricted context of hierarchically
hyperbolic spaces.
References
[Abb16]
[ABO16]
[ACGH16]
[ADT16]
[BBKL16]
[Beh06]
[BHS15]
[BHS17a]
[BHS17b]
[BHS17c]
[Bow08]
[CD16]
[CH16]
[CM17]
[Cor15]
[CS15]
[DGO17]
[DHS15]
[DMS10]
[DT15]
[FM02]
[Hag14]
[Ham]
[HS16]
[KK14]
Carolyn R. Abbott. Not all finitely generated groups have universal acylindrical actions. Proc.
Amer. Math. Soc., 144(10):4151–4155, 2016.
C. Abbott, S. Balasubramanya, and D. Osin. Hyperbolic structures on groups. 2016.
Goulnara N Arzhantseva, Christopher H Cashen, Dominik Gruber, and David Hume. Characterizations of Morse quasi-geodesics via superlinear divergence and sublinear contraction.
arXiv:1601.01897, 2016.
Tarik Aougab, Matthew Gentry Durham, and Samuel J Taylor. Middle recurrence and pulling
back stability. arXiv:1609.06698, 2016.
M. Bestvina, K. Bromberg, R.P. Kent, and C.J. Leininger. Undistorted purely pseudo-anosov
groups. arXiv:1608.01583, 2016.
J. Behrstock. Asymptotic geometry of the mapping class group and Teichmüller space. Geometry
& Topology, 10:2001–2056, 2006.
Jason Behrstock, Mark F Hagen, and Alessandro Sisto. Hierarchically hyperbolic spaces II: combination theorems and the distance formula. arXiv:1509.00632v2, pages 1–53, 2015.
J. Behrstock, M.F. Hagen, and A. Sisto. Asymptotic dimension and small-cancellation for hierarchically hyperbolic spaces and groups. Proc. London Math. Soc., 114(5):890–926, 2017.
Jason Behrstock, Mark F Hagen, and Alessandro Sisto. Hierarchically hyperbolic spaces I: curve
complexes for cubical groups. Geometry & Topology, 21:1731–1804, 2017.
Jason Behrstock, Mark F Hagen, and Alessandro Sisto. Quasiflats in hierarchically hyperbolic
spaces. arXiv:1704.04271, 2017.
Brian H. Bowditch. Tight geodesics in the curve complex. Inventiones mathematicae, 171(2):281–
300, 2008.
Matthew Cordes and Matthew Gentry Durham. Boundary convex cocompactness and stability of
subgroups of finitely generated groups. arXiv:1607.08899, 2016.
Matthew Cordes and David Hume. Stability and the Morse boundary. To appear in Groups,
Geometry, and Dynamics, 2016.
Christopher H Cashen and John M Mackay. A metrizable topology on the contracting boundary
of a group. arXiv:1703.01482, 2017.
Matthew Cordes. Morse boundaries of proper geodesic metric spaces. arXiv:1502.04376, 2015.
Ruth Charney and Harold Sultan. Contracting boundaries of CATp0q spaces. J. Topol., 8(1):93–
117, 2015.
F. Dahmani, V. Guirardel, and D. Osin. Hyperbolically embedded subgroups and rotating families
in groups acting on hyperbolic spaces. Mem. Amer. Math. Soc., 245(1156):v+152, 2017.
Matthew Gentry Durham, Mark F. Hagen, and Alessandro Sisto. Boundaries of hierarchically
hyperbolic spaces. Geometry & Topology, 2015. To appear.
Cornelia Druţu, Shahar Mozes, and Mark Sapir. Divergence in lattices in semisimple Lie groups
and graphs of groups. Trans. Amer. Math. Soc., 362(5):2451–2505, 2010.
Matthew Gentry Durham and Samuel J. Taylor. Convex cocompactness and stability in mapping
class groups. Algebr. Geom. Topol., 15(5):2839–2859, 2015.
Benson Farb and Lee Mosher. Convex cocompact subgroups of mapping class groups. Geom.
Topol., 6:91–152 (electronic), 2002.
Mark F. Hagen. Weak hyperbolicity of cube complexes and quasi-arboreal groups. J. Topol.,
7(2):385–418, 2014.
U. Hamenstädt. Word hyperbolic extensions of surface groups. Preprint, arXiv:0505244.
M.F. Hagen and T. Susse. Hierarchical hyperbolicity of all cubical groups. arXiv:1609.01313, 2016.
Sang-Hyun Kim and Thomas Koberda. The geometry of the curve graph of a right-angled Artin
group. International Journal of Algebra and Computation, 24(02):121–169, 2014.
ACYLINDRICAL ACTIONS AND STABILITY IN HHG
[KL08]
[KMT14]
[Min96]
[MM99]
[MS17]
[Osi16]
[Sel97]
[Sis16]
[Sul14]
[Tra]
[TT16]
25
Richard P. Kent, IV and Christopher J. Leininger. Shadows of mapping class groups: capturing
convex cocompactness. Geom. Funct. Anal., 18(4):1270–1325, 2008.
Thomas Koberda, Johanna Mangahas, and Samuel J Taylor. The geometry of purely loxodromic
subgroups of right-angled Artin groups. arXiv:1412.3663, 2014.
Yair N. Minsky. Quasi-projections in Teichmüller space. J. Reine Angew. Math., 473:121–136,
1996.
Howard A Masur and Yair N Minsky. Geometry of the complex of curves I: Hyperbolicity. Inventiones mathematicae, 138(1):103–149, 1999.
Joseph Maher and Alessandro Sisto. Random subgroups of acylindrically hyperbolic groups and
hyperbolic embeddings. arXiv:1701.00253, 2017.
D. Osin. Acylindrically hyperbolic groups. Trans. Amer. Math. Soc., 368(2):851–888, 2016.
Z. Sela. Acylindrical accessibility for groups. Invent. Math., 129(3):527–565, 1997.
Alessandro Sisto. Quasiconvexity of hyperbolically embedded subgroups. Mathematische
Zeitschrift, 283:649–658, 2016.
Harold Sultan. Hyperbolic quasi-geodesics in CAT(0) spaces. Geom. Dedicata, 169:209–224, 2014.
Hung Cong Tran. Purely loxodromic subgroups in right-angled Coxeter groups. arXiv:1703.09032.
Samuel J. Taylor and Giulio Tiozzo. Random extensions of free groups and surface groups are
hyperbolic. Int. Math. Res. Not. IMRN, (1):294–310, 2016.
University of Wisconsin-Madison, Madison, Wisconsin, USA
E-mail address: [email protected]
Lehman College and The Graduate Center, CUNY, New York, New York, USA
E-mail address: [email protected]
University of Michigan, Ann Arbor, Michigan, USA
E-mail address: [email protected]
| 4math.GR
|
CERTAIN CLASSES OF COHEN-MACAULAY MULTIPARTITE GRAPHS
arXiv:1803.07880v1 [math.AC] 21 Mar 2018
RAJIV KUMAR AND AJAY KUMAR
Abstract. The Cohen-Macaulay property of a graph arising from a poset has been studied
by various authors. In this article, we study the Cohen-Macaulay property of a graph arising
from a family of reflexive and antisymmetric relations on a set. We use this result to find
classes of multipartite graphs which are Cohen-Macaulay.
1. Introduction
Graphs and simplicial complexes play an important role in combinatorial commutative
algebra. In order to see the relationship between commutative algebra and combinatorics,
one can associate monomial ideals to graphs or simplicial complexes. Many authors have
studied the connection between the algebraic properties of these ideals and the combinatorial
properties of the corresponding combinatorial objects, see [4, Chapter 9]. In this article, our
main focus is to study the edge ideal of a graph. A graph is called Cohen-Macaulay if the
corresponding edge ideal is Cohen-Macaulay.
The Cohen-Macaulay property of graphs has been well studied for various classes. HerzogHibi ([3]) prove that a bipartite graph is Cohen-Macaulay if and only if it is arising from a
poset. For a finite poset and r, s ∈ N, Ene-Herzog-Mohammadi ([2]) associated a monomial
ideal generated in degree s, to the set of all multichains of length r in a poset, and proved
that this ideal is Cohen-Macaulay. Note that if s = 2, then these ideals are edge ideals of
some r-partite graphs. Motivated by these results, we associate a monomial ideal to a family
of posets, and find a class of Cohen-Macaulay r-partite graphs.
Our main tool in this article is the following well known relationship between the StanleyReisner ideal and its Alexander dual: The Stanley-Reisner ideal is Cohen-Macaulay if and
only if its Alexander dual has a linear resolution. For more details see [1, Theorem 3].
This paper has been organized in the following manner. In Section 2, we introduce the
basic notions which are used throughout the article, more details can be found in [4]. In
Section 3, we associate a monomial ideal Hr (P) to a family P of partial order relations on a
finite set. In Lemma 3.3, we prove that monomial ideal Hr (P) has a linear resolution. This
forces the Alexander dual of Hr (P) to be Cohen-Macaulay.
Section 4 is devoted to finding the classes of Cohen-Macaulay r-partite graphs. In Theorem
4.4, we see that the Alexander dual of Hr (P) is an edge ideal of an r-partite graph associated
to a family of reflexive and antisymmetric relations on a given set. Using this we find classes
of Cohen-Macaulay graphs which are recorded in Theorems 4.7 and 4.10.
2. Preliminaries
2.1. Notation. The following notation is used throughout the article.
Date: March 22, 2018.
2010 Mathematics Subject Classification. 05E40, 13C14, 13D02.
Key words and phrases. Posets, Multipartite Graphs, Cohen-Macaulay Graphs, Linear Resolutions.
1
2
R. KUMAR AND A. KUMAR
i) For n ∈ N, we denote [n] = {1, . . . , n}.
ii) By Pa , we mean that the set P with partial relation ≤a .
iii) Let S = k[X1 , . . . , Xn ] be a polynomial ring with deg(Xi ) = 1, where k is a field. Then
by S(−j), we mean a graded free S-module of rank 1 with S(−j)n = Sn−j .
iv) Let M and N be graded S-modules. Then a homomorphism φ : M −→ N is called a
graded homomorphism if φ(Mn ) ⊂ Nn .
2.2. Graphs and Edge Ideals.
Definition 2.1.
i) A graph G = (V, E) is an ordered pair, where V is the set of vertices of G and E is a
collection of subsets of V of cardinality 2.
ii) An element of E is called an edge of G. For all i, j ∈ V , we say that i is adjacent to j
if and only if {i, j} ∈ E.
iii) For an integer r ≥ 2, a graph G is called an r-partite if there exists a partition of
V = V1 ∪ · · · ∪ Vr such that for all 1 ≤ k ≤ r and i, j ∈ Vk implies that i is not adjacent
to j. If r = 2, we say that G is a bipartite graph. A bipartite graph on vertex set
V = V1 ∪ V2 is called a complete bipartite graph if i and j are adjacent for all i ∈ V1 and
j ∈ V2 .
iv) Let G be a graph on a vertex set V and W ⊂ V . Then a graph H is called a induced
subgraph of G on W if for i, j ∈ W , i and j are adjacent in H if and only if so in G.
v) Let S = k[X1 , . . . , Xn ] be a polynomial ring over k and G be a graph on a vertex set
V = [n]. Then the monomial ideal I(G) = hXi Xj : {i, j} ∈ Ei is called the edge ideal
of G. A graph G is called Cohen-Macaulay if S/I(G) is Cohen-Macaulay.
2.3. Simplicial Complexes.
Definition 2.2. For fixed n ∈ N, let V = [n].
i) A simplicial complex on V , denoted by ∆ or ∆V , is a collection of subsets of V with the
following properties:
a) φ ∈ ∆ and {i} ∈ ∆ for all i ∈ V .
b) If F ∈ ∆ and G ⊂ F , then G ∈ ∆.
ii) An element of ∆ is called a face of ∆, and a maximal face with respect to inclusion is
called a facet.
iii) A subset F ⊂ V is called a nonface of ∆ if F ∈
/ ∆, and it is called a minimal nonface if
it is minimal with respect to inclusion.
iv) The Alexander dual of ∆, denoted by ∆∨ , is defined as
∆∨ = {V \ F : F is a nonface of ∆}.
v) Let ∆ be a simplicial complex on [n] and S = k[X1 , . . . , Xn ]. The Stanley-Reisner ideal
of ∆ is the squarefree monomial ideal, denote as I∆ , is defined as follows:
I∆ = hXi1 · · · Xir : {i1 , . . . , ir } is a minimal nonface of ∆i .
Further, let ∆∨ be the Alexander dual of ∆. Then we say that I∆∨ is the Alexander dual
∨
of I∆ , and denote it by I∆
.
CERTAIN CLASSES OF COHEN-MACAULAY MULTIPARTITE GRAPHS
3
2.4. Free Resolution. Let S = k[X1 , . . . , Xn ] and I be a homogeneous ideal of S. Then a
free resolution of I over S is an exact sequence
φn
F• : · · · −→ Fn −→ Fn−1 −→ · · · −→ F0 −→ I −→ 0
such that for each i ≥ 0, Fi is a graded free S-module and φi is a graded homomorphism.
Further, if I is generated in degree d and Fi ' S(−d − i)βi for some βi ∈ N and for all i, then
we say that I has a linear resolution.
3. Linear Resolution
In this section, we associate a monomial ideal Hr (P) to a family of posets P, and we show
that this ideal has a linear resolution.
Definition 3.1. Let (P, ≤) be a finite partial ordered set. A subset I ⊂ P is called a poset
ideal if for all p ∈ I and q ∈ P with q ≤ p implies that q ∈ I.
For a given set P = {p1 , . . . , pn } and an integer r ≥ 2, we consider a family of partial
ordered relations P = {≤a : a ∈ [r − 1]} on P . For the sake of simplicity, we denote (P, ≤a )
by Pa . Let J(Pa ) denotes the set of poset ideals of Pa . Now corresponding to a family P, we
define a set
Kr (P) = {I = (I1 , I2 , . . . , Ir−1 ) : Ia ∈ J(Pa ) ∀ a ∈ [r − 1] and Ir−1 ⊂ · · · ⊂ I1 },
with a partial order ≺ given by J ≺ I if and only if Ja ⊂ Ia ∀ a.
Let S = k[X1 , X2 , . . . , Xr ], where Xa = {Xa,1 , . . . , Xa,n } for all a ∈ [r]. For I ∈ Kr (P), we
associate a squarefree monomial
Y
Y
Xa,i
Xa+1,i for all 1 ≤ a ≤ r − 1.
uI = uI1 uI2 · · · uIr−1 , where uIa =
pi ∈Ia
pi ∈Pa \Ia
Let Hr (P) = {uI }I∈Kr (P) be the squarefree monomial ideal of the polynomial ring S,
generated by monomials uI , where I ∈ Kr (P).
Example 3.2. Let P = {p1 , p2 , p3 } and r = 3. Suppose we have the following partial order
relations on P :
p3
p2
p1
p2
p3
p1
(P, ≤1 )
(P, ≤2 )
Then the collection of poset ideals of P1 is J(P1 ) = {∅, {p1 }, {p2 }, {p1 , p2 }, {p2 , p3 }, P } and,
of P2 is J(P2 ) = {∅, {p1 }, {p3 }, {p1 , p2 }, {p1 , p3 }, P }. Also, We see that
(φ, φ), ({p1 }, φ), ({p1 }, {p1 }), ({p2 }, φ), ({p1 , p2 }, φ),
({p1 , p2 }, {p1 }), ({p1 , p2 }, {p1 , p2 }), ({p2 , p3 }, φ), ({p2 , p3 }, {p3 }), ,
K3 (P) =
(P, φ), (P, {p }), (P, {p }), (P, {p , p }), (P, {p , p }), (P, P )}
1
3
1 2
1 3
4
R. KUMAR AND A. KUMAR
and hence H3 (P) is generated by the following set of squarefree monomials
X2,1 X2,2 X2,3 X3,1 X3,2 X3,3 , X1,1 X2,2 X2,3 X3,1 X3,2 X3,3 , X1,1 X2,1 X2,2 X2,3 X3,2 X3,3 ,
X1,2 X2,1 X2,3 X3,1 X3,2 , X3,3 , X1,1 X1,2 X2,3 X3,1 X3,2 X3,3 , X1,1 X1,2 X2,1 X2,3 X3,2 X3,3 ,
X1,1 X1,2 X2,1 X2,2 X2,3 X3,3 , X1,2 X1,3 X2,1 X3,1 X3,2 X3,3 , , X1,2 X1,3 X2,1 X2,3 X3,1 X3,2
X1,1 X1,2 X1,3 X3,1 X3,2 X3,3 , X1,1 X1,2 X1,3 X2,1 X3,2 X3,3 , X1,1 X1,2 X1,3 X2,3 X3,1 X3,2 ,
X X X X X X ,X X X X X X ,X X X X X X
1,1 1,2 1,3 2,1 2,2 3,3
1,1 1,2 1,3 2,1 2,3 3,2
1,1 1,2 1,3 2,1 2,2 2,3
in the polynomial ring k[Xa,i , 1 ≤ a, i ≤ 3].
Note that For r = 2, Herzog-Hibi ([3]) proved that the ideal H2 (P) has a linear resolution.
In fact, if all partial order relations ≤a are same for a ∈ [r −1], then Ene-Herzog-Mohammadi
([2]) studied the ideal Hr (P), and proved that it has a linear resolution. More generally, in
Proposition 3.4, we prove that Hr (P) has a linear resolution.
Let G(Hr (P)) denotes the minimal set of monomial generators of the monomial ideal
Hr (P). Define a partial order on G(Hr (P)) by uJ ≺ uI if J ≺ I. We fix a total order ≺0 on
G(Hr (P)), which extends the partial order ≺. For more details see [5, Theorem 1.1].
In order to prove that Hr (P) has a linear resolution, we use the following remark.
Remark 3.3. Let I be a monomial ideal with monomial generators u1 , . . . , um of I. Suppose
for all j < i, there exists an integer k < i and an integer l such that
uk
uj
= Xl and Xl divides
.
gcd(uk , ui )
gcd(uj , ui )
Then, by [4, Theorem 8.2.1 and Lemma 8.2.3] I has a linear resolution.
Proposition 3.4. The squarefree monomial ideal Hr (P) has a linear resolution.
Proof. Let J ≺0 I with J = (J1 , J2 , . . . , Jr−1 ) , I = (I1 , I2 , . . . , Ir−1 ) . For a ∈ [r − 1], define
0
). Since Ja0 ⊂ Ia for a ∈ [r − 1], we have J0 ≺ I.
Ja0 = Ja ∩ Ia and J0 = (J10 , . . . , Jr−1
Then take a = max{q : Jq0 ( Iq }. Since Ja0 ( Ia are poset ideals of Pa , there exists a
pi ∈ Ia \ Ja0 such that pi is a maximal element in Ia . This forces that δa = Ia \ {pi } is a poset
uI Xa+1,i
u δa
ideal of Pa . This gives us uδa = a
.
which implies that Xa+1,i =
Xa,i
gcd(uδa , uIa )
Our claim is K = (I1 , I2 , . . . , Ia−1 , δa , Ia+1 , . . . , Ir−1 ) ∈ Kr (P). Since δa ( Ia , to prove the
0
claim it is enough to prove that Ia+1 ⊂ δa . The fact that Ja+1
⊂ Ja0 and pi ∈
/ Ja0 implies that
0
0
pi ∈
/ Ja+1
, and by choice of a, we know that Ja+1
= Ia+1 . The claim follows from the fact
that δa = Ia \ pi .
0
Since we know that pi ∈ Ia and pi ∈
/ Ja+1
= Ia+1 , we get Xa+1,i does not divide uI . Now,
0
pi ∈
/ Ja and pi ∈ Ia and hence pi ∈
/ Ja . This forces that Xa+1,i divides uJa , and hence we get
uJ
Xa+1,i divides
.
gcd(uJ , uI )
uK
u δa
Finally, from the definition of K, we get
=
= Xa+1,i , and hence
gcd(uI , uK )
gcd(uIa , uδa )
the proof follows from Remark 3.3.
4. Cohen Macaulay Multipartite Graphs
For a ∈ [r − 1] and P = {p1 , p2 , . . . , pn }, let ≤a be a partial order relation on P such that
if pi ≤a pj , then we have i ≤ j. In this case, we prove that the Alexander dual of Hr (P) is
the edge ideal of some r-partite graph. Using this we identify two classes of Cohen-Macaulay
CERTAIN CLASSES OF COHEN-MACAULAY MULTIPARTITE GRAPHS
5
graphs. In order to find Hr (P)∨ , we define the following relation on P : For 1 ≤ a ≤ b ≤ r −1,
we define a relation ≤[a,b] on P as follows: pi ≤[a,b] pj if there exists a non-decreasing sequence
(t1 , t2 , . . . , tk ) such that
pi ≤a pt1 ≤a+1 · · · ≤b−1 ptk ≤b pj .
Example 4.1. Let ≤1 and ≤2 be partial order relations on a set P = {p1 , p2 , p3 } as defined
in the Example 3.2. Then the relation ≤[1,2] on P is shown as in the following diagram.
p2
p3
p1
p2
(P, ≤[1,2] )
For a simplicity, we denote P[a,b] = (P, ≤[a,b] ). Since ≤a is a reflexive relation on P for all
a ∈ [r − 1], by definition of ≤[a,b] it follows that so is ≤[a,b] . Further, the relation ≤[a,b] on P
is antisymmetric follows from the fact that if pi ≤a pj , then we have i ≤ j. In Example 4.1
observe that ≤[1,2] is not a transitive relation on P , and hence P[a,b] need not be a poset.
Definition 4.2.
i) Let u be a monomial in S. Then support of u, denoted by supp(u), is defined as
supp(u) = {Xa,i : Xa,i divides u}.
ii) Let V =
r
S
Va , where Va = {Xa,1 , . . . , Xa,n }. For F ⊂ V , define FXa = {X1,i : Xa,i ∈ F }
a=1
for a ≤ r.
F
,
iii) For F ⊂ V , we set γrF = φ. For 2 ≤ a ≤ r, we define a poset ideal, denoted as γa−1
F
F
F
generated by {pi : X1,i ∈ FXa } ∪ γa in Pa−1 . Note that γa+1 ⊂ γa for all a ∈ [r − 1], so
F
) ∈ Kr (P).
we denote γ F = (γ1F , . . . , γr−1
Lemma 4.3. Let pi ∈ γaF for some a ∈ [r − 1]. Then there exists pj such that X1,j ∈ FXb for
some a + 1 ≤ b ≤ r with pi ≤[a,b−1] pj .
Proof. Since pi ∈ γaF , there exists a maximal element pj ∈ γaF with pi ≤a pj . Now by
F
definition of γaF , it follows that either X1,j ∈ FXa+1 or pj ∈ γa+1
. If X1,j ∈ FXa+1 , then we
F
are through. Otherwise there exists a maximal element pk ∈ γa+1 with pj ≤a+1 pk . Again we
F
have either X1,k ∈ FXa+2 or pk ∈ γa+2
. Thus, we repeat this procedure till we get the desired
result.
Theorem 4.4. The monomial ideal Hr (P)∨ is minimally generated by the squarefree monomials of type Xs,i Xt,j , 1 ≤ s < t ≤ r if and only if pi ≤[s,t−1] pj .
Proof. Let V =
r
S
Va be a vertex set and ∆P be a simplicial complex on V such that
a=1
I∆P = Hr (P). Set w to be the product of all variables. By the definition of ∆∨P , facets of
∆∨P are given by supp(w/uI ), where I ∈ Kr (P). Now, by definition of uI , it follows that Xa,i
divides uI if and only if pi ∈
/ Ia−1 or pi ∈ Ia . Also, we know that
(supp(w/uI ))Xa = {X1,i : Xa,i does not divide uI },
6
R. KUMAR AND A. KUMAR
and hence, we get (supp(w/uI ))Xa = {X1,i : pi ∈ Ia−1 \ Ia }.
Let F ⊂ V be a face of ∆∨P . Since facets of ∆∨P are given by supp(w/uI ) for some
I ∈ Kr (P), there exists I ∈ Kr (P) such that F ⊂ supp(w/uI ). In particular, for a ∈ [r], we
have FXa ⊂ (supp(w/uI ))Xa .
First, assume that pi ≤[a,b−1] pj , where 1 ≤ a < b ≤ r. Then our claim is that the set
F = {Xa,i , Xb,j } is a minimal non-face of ∆∨P . Note that since |F | = 2, it is enough to prove
that F is a non-face of ∆∨P .
Let I ∈ Kr (P). If pj ∈
/ Ib−1 \ Ib , then Xb,j divides uI , and hence Xb,j ∈
/ supp(w/uI ). This
forces that FXb 6⊂ (supp(w/uI ))Xb . Otherwise, we claim that FXa 6⊂ (supp(w/uI ))Xa , and
hence by the claim F is a non-face. In order to prove the claim, it is enough to prove that
pi ∈ Ia .
Proof of claim: Since pi ≤[a,b−1] pj , there exist a non-decreasing sequence (t1 , t2 , . . . , tk )
such that pi ≤a pt1 ≤a+1 · · · ≤b−2 ptk ≤b−1 pj . The fact ptk ≤b−1 pj and Ib−1 is a poset ideal
in Pb−1 implies that ptk ∈ Ib−1 . Using Ib−1 ⊂ Ib−2 , we get ptk ∈ Ib−2 . Now, again repeat the
process, we get pi ∈ Ia . This proves the claim.
Conversely, let F be a minimal nonface of ∆∨P . This gives us the following:
for any I ∈ Kr (P ) ∃ a ∈ [r] such that FXa 6⊂ (supp(w/uI ))Xa .
(4.1)
Case 1 : Suppose FXa 6⊂ (supp(w/uγ F ))Xa for some 2 ≤ a ≤ r. Since we know that
F
\ γaF }, there exists some X1,i ∈ FXa such that either
(supp(w/uγ F ))Xa = {X1,i : pi ∈ γa−1
F
F
F
F
,
, we see that pi ∈ γa−1
pi ∈
/ γa−1 or pi ∈ γa . Also, note that X1,i ∈ FXa , by definition of γa−1
F
and hence pi ∈ γa . By Lemma 4.3, there exists a pj such that pi ≤[a,b−1] pj with X1,j ∈ FXb
for some a + 1 ≤ b ≤ r. Since {Xa,i , Xb,j } is a minimal nonface of ∆∨P . By assumption F is a
minimal nonface, and hence we get F = {Xa,i , Xb,j } with pi ≤[a,b−1] pj .
Case 2 : FXa ⊂ (supp(w/uγ F ))Xa for 2 ≤ a ≤ r. Since γ F ∈ Kr (P), by Equation (4.1), we
get FX1 6⊂ (supp(w/uγ F ))X1 . Thus, there exists pi ∈ γ1F such that X1,i ∈ FX1 .
Since pi ∈
/ γrF = φ, we can choose 2 ≤ b ≤ r such that b = min{j : pi ∈
/ γjF }. Then, we
F
F
F
see that pi ∈ γb−1
\ γbF . Let pj ∈ γb−1
\ γbF such that pi ≤b−1 pj . Since pj ∈ γb−1
\ γbF , by
Lemma 4.3, there exists some b ≤ c ≤ r such that pj ≤[b,c−1] pk with Xc,k ∈ F . Note that
pi ≤[1,c−1] pk and {X1,i , Xc,k } ⊂ F . By assumption F is a minimal nonface of ∆∨P and we
know {X1,i , Xc,k } ⊂ F is a nonface, and hence F = {X1,i , Xc,k } with pi ≤[1,c−1] pk .
Let F = {≤[a,b] : 1 ≤ a ≤ b ≤ r − 1} be a family of reflexive and antisymmetric relations on
a given set P = {p1 , p2 , . . . , pn }. Now corresponding to F, we associate an r-partite graph
r
S
on a vertex set V =
Va such that Xa,i is adjacent to Xb,j if and only if pi ≤[a,b−1] pj .
a=1
Example 4.5. Let P = {p1 , p2 , p3 } and F = {≤1 , ≤2 , ≤[1,2] }, where ≤1 , ≤2 , ≤[1,2] are given as
in Examples 3.2 and 4.1. We associate a following graph on vertices set V = {Xa,i : a, i ∈ [3]}:
CERTAIN CLASSES OF COHEN-MACAULAY MULTIPARTITE GRAPHS
X1,1
X1,2
X1,3
X2,1
X2,2
X2,3
7
X3,1
X3,2
X3,3
G
Note that by the following corollary G is Cohen-Macaulay.
Corollary 4.6. Assume that the above family F have the following properties:
i) For a ∈ [r − 1], P[a,a] = Pa is a poset.
ii) For 1 ≤ a ≤ b ≤ r − 1, if pi ≤[a,b] pj , then i ≤ j.
iii) For 1 ≤ a ≤ b ≤ r − 1, if pi ≤[a,b] pj , then there exists a non-decreasing sequence
(t1 , t2 , . . . , tk ) such that pi ≤a pt1 ≤a+1 · · · ≤b−1 ptk ≤b pj .
Then an r-partite graph associated to F is Cohen-Macaulay.
Proof. We notice that the edge ideal of an r-partite graph associated to a family F is equals
to Hr (P)∨ . Since, by Proposition 3.4, the ideal Hr (P) has a linear resolution, the result
follows.
Theorem 4.7. Let G be an r-partite graph with partitions Va = {Xa,1 , . . . , Xa,n }, for all
a ∈ [r] satisfying the following conditions:
i) {Xa,i , Xb,i } is an edge for all i ∈ [n] and 1 ≤ a < b ≤ r;
ii) if {Xa,i , Xb,j } is an edge with 1 ≤ a < b ≤ r, then i ≤ j;
iii) {Xa,i , Xb,j } is an edge if and only if there exists a non-decreasing sequence (t1 , t2 , . . . , tk )
such that {Xa,i , Xa+1,t1 }, . . . , {Xb−1,tk , Xb,j } are edges of G for 1 ≤ a ≤ b ≤ r and
i, j ∈ [n].
iv) if {Xa,i , Xa+1,j } and {Xa,j , Xa+1,k } are edges for a ∈ [r − 1], then {Xa,i , Xa+1,k } is an
edge.
Then G is Cohen-Macaulay.
Proof. Let P = {p1 , p2 , . . . , pn }. Then for each 1 ≤ a < b ≤ r, we define a relation ≤[a,b−1] on
P by pi ≤[a,b−1] pj if and only if {Xa,i , Xb,j } is an edge of G. Observe that for all 1 ≤ a < b ≤ r,
P[a,b−1] satisfies the hypothesis of Corollary 4.6, and hence G is Cohen-Macaulay.
Definition 4.8. Let G be a bipartite graph on the vertex set V with partition of V = V1 ∪V2 ,
where V1 = {X1,1 , . . . , X1,m } and V2 = {X2,1 , . . . , X2,n }. Then we say that G satisfies HerzogHibi conditions if it satisfy the following:
i) m = n.
ii) For all i, we have {X1,i , X2,i } ∈ E.
iii) If {X1,i , X2,j } ∈ E, then i ≤ j.
iv) If {X1,i , X2,j } ∈ E and {X1,j , X2,k } ∈ E, then {X1,i , X2,k } ∈ E.
Further, if for i ≤ j, we have {X1,i , X2,j } ∈ E, then we say G is a Herzog-Hibi complete
graph.
8
R. KUMAR AND A. KUMAR
Remark 4.9. Let V1 = {X1,1 , . . . , X1,m } and V2 = {X2,1 , . . . , X2,n }, and G be a bipartite
graph on a vertex set V = V1 ∪ V2 . If G satisfy Herzog-Hibi conditions, then, by Theorem
4.7, we know that G is Cohen-Macaulay. In fact, Herzog-Hibi ([3], 2005) prove that G is
Cohen-Macaulay if and only if G satisfy Herzog-Hibi conditions. In this case, further we have
the following:
i) Let G0 be induced subgraph on vertex set V \ {X1,1 , X2,1 }. Then note that G0 satisfy
Herzog-Hibi conditions, and hence G0 is Cohen-Macaulay.
ii) Let S = k[X1,1 , . . . , X1,n , X2,1 , . . . , X2,n ]. Since X2,1 is adjacent
to
only X1,1 , we get
S
S
hI(G), X1,1 i = hI(G0 ), X1,1 i and dim
= dim
= n, and hence by
hI(G), X1,1 i
I(G)
S
= n.
(i), we know that depth
hI(G), X1,1 i
S(−1) X1,1 S
S
iii) Consider the sequence 0 −→
−→
−→
−→ 0. This forces
I(G) : X1,1
I(G) hI(G), X1,1 i
S(−1)
S(−1)
S(−1)
that dim
is
≤ n and depth
≥ n, and hence
I(G) : X1,1
I(G) : X1,1
I(G) : X1,1
Cohen-Macaulay of dimension n.
Theorem 4.10. Let G be an r-partite graph with partition V1 , V2 , . . . , Vr which satisfy the
following:
i) |Va | = n for all 1 ≤ a ≤ r
ii) for all 1 ≤ a, b ≤ r − 1, the induced graph on the vertices Va ∪ Vb is a complete bipartite
graph,
iii) for all 1 ≤ a ≤ r − 1, the induced graph on vertices Va ∪ Vr satisfies the Herzog-Hibi
conditions.
Then G is Cohen-Macaulay.
Remark 4.11. Let G be a graph as in Theorem 4.10 and S = k[X1 , X2 , . . . , Xn ]. Suppose
Va = {Xa,1 , . . . , Xa,n } for all a ∈ [r].
i) If n = 1, then G is a complete graph on r vertices, and hence G is Cohen-Macaulay.
ii) For 1 ≤ a ≤ r, let Va0 = V \ {Xa,1 }, and let G0 be the induced subgraph of G on the
vertex set V 0 = V10 ∪ · · · ∪ Vr0 . Then G0 also satisfies the hypothesis of Theorem 4.10 with
|Va0 | = n − 1 for all a.
iii) Let Ga be the induced subgraph on vertex set Va ∪ Vr for all a ∈ [r − 1]. Since, for all
a, Ga is bipartite graph which satisfies Herzog-Hibi conditions, by Remark 4.9(i), Ga is
S
Cohen-Macaulay, and hence by 4.9(iii), we get
is Cohen-Macaulay.
I(Ga ) : Xa,1
iv) For a ∈ [r] and i ∈ [n], define st(Xa,i ) = {Xb,j ∈ V : Xb,j adjacent to Xa,i }. Since the
induced graph on vertices Va ∪ Vb is a complete bipartite graph for all a, b ≤ r − 1, we
get V1 ∪ · · · ∪ Va−1 ∪ Va+1 ∪ · · · ∪ Vr−1 ⊂ st(Xa,1 ) for all a ∈ [r − 1].
v) By (iv), we get
I(G) : Xa,1 = hV1 ∪ · · · ∪ Va−1 ∪ Va+1 ∪ · · · ∪ Vr−1 , I(Ga ) : Xa,1 i.
Since, by Remark 4.9(iii),
S
is Cohen-Macaulay,
I(Ga ) : Xa,1
CERTAIN CLASSES OF COHEN-MACAULAY MULTIPARTITE GRAPHS
9
S
is Cohen-Macaulay, and hence so is
I(G) : Xa,1
S
S
k[Va , Vr ]
. Note that
'
,
hI(G), X1,1 , . . . , Xa−1,1 i : Xa,1
hI(G),
X1,1 , . . . , Xa−1,1 i : Xa,1 I(Ga ) : Xa,1
S
and hence Remark 4.9(iii) forces that dim
= n.
hI(G), X1,1 , . . . , Xa−1,1 i : Xa,1
By Remark 4.9(iii), we know that
Proof of Theorem 4.10. In order to prove the result, we use the induction on n. If n = 1, by
Remark 4.11(i), we know G is Cohen-Macaulay. If n > 1, then take Va = {Xa,1 , Xa,2 , . . . , Xa,n }.
Take the following short exact sequence
0 −→
S(−1)
S
S
Xr−1,1
−→
−→
−→ 0.
hI(G), X1,1 , . . . , Xr−2,1 i : Xr−1,1
hI(G), X1,1 , . . . , Xr−2,1 i
hI(G), X1,1 , . . . , Xr−1,1 i
For 1 ≤ a ≤ r, let Va0 = V \ {Xa,1 }, and let G0 be the induced subgraph of G on the vertex
set V = V10 ∪· · ·∪Vr0 . Note that induce subgraph G0 also satisfies the hypothesis of the theorem
hypothesis, G0 is Cohen-Macaulay,
and
with |Va0 | = n−1 < |Va | for all a. Hence, by induction
S
S
, and dim
= n. By Remark
hence so is
0
0
hI(G ), X1,1 , . . . , Xr−1,1 i
hI(G ), X1,1 , . . . , Xr−1,1 i
S
is Cohen-Macaulay of dimension n, and
4.11(v), we know
hI(G), X1,1 , . . . , Xr−2,1 i : Xr−1,1
S
hence so
.
hI(G), X1,1 , . . . , Xr−2,1 i
Similarly, using the following short exact sequence
0→
S(−1)
hI(G), X1,1 , . . . , Xr−3,1 i : Xr−2,1
we get
get
Xr−2,1
→
S
S
→
→ 0,
hI(G), X1,1 , . . . , Xr−3,1 i
hI(G), X1,1 , . . . , Xr−2,1 i
S
is Cohen-Macaulay. By repeating the above process, we
hI(G), X1,1 , . . . , Xr−3,1 i
S
is Cohen-Macaulay.
I(G)
Note that if G is Cohen-Macaulay, then G need not be one of the graph described in
Theorems 4.7 and 4.10. For example, if G is cycle on 5 vertices, then G is Cohen-Macaulay.
But G does not satisfy the hypothesis of Theorems 4.7 and 4.10.
Remark 4.12. Let G be as in the above theorem such that for 1 ≤ a ≤ r − 1, the induced
graph on vertices Va ∪Vr is a Herzog-Hibi complete graph. Since r−1
partitions are complete
2
bipartite and r − 1 partitions are Herzog-Hibi complete, we have the number of edges is
n+1
(r − 1)n + 1
2 r−1
n
+ (r − 1)
=
.
2
2
2
Since the height of edge ideal of a given graph in the above remark is (r − 1)n, by [6,
Theorem 4.3.7], we have the following:
Corollary 4.13. Let G be given as in the above remark and I be the edge ideal of G. Then
S/I has a linear resolution.
10
R. KUMAR AND A. KUMAR
References
[1] Eagon, A., Reiner, V. (1998). Resolutions of Stanley-Reisner Rings and Alexander Duality. J. pure &
Applied Algebra. 130, 265 - 275.
[2] Ene, V., Herzog, J., Mohammadi F. (2011). Monomial Ideals and Toric Rings of Hibi Type Arising from
a Finite Poset. European Journal of Combinatorics. 32, 404 - 421.
[3] Herzog, J., Hibi, T. (2005). Distributive lattices, bipartite graphs and Alexander duality. J. Algebraic
Combin. 22, no. 3, 289 – 302.
[4] Herzog, J., Hibi, T. (2011). Monomial Ideals, Graduate Texts in Mathematics, Springer-Verlag London
Ltd., London.
[5] Szpilrajn, E. (1930). Sur lextension de lordre partiel. Fund. Math. 16, 386 - 389.
[6] Villarreal, R. H. (2001). Monomial Algebras, Marcel Dekker.
Indian Institute of Technology Bombay, India.
E-mail address: [email protected]
DAV University Jalandhar, India.
E-mail address: [email protected]
| 0math.AC
|
Sparse covariance matrix estimation in
high-dimensional deconvolution
DENIS BELOMESTNY1,2 , MATHIAS TRABS3 and ALEXANDRE B. TSYBAKOV4
arXiv:1710.10870v2 [math.ST] 26 Mar 2018
1
Duisburg-Essen University, Faculty of Mathematics
Thea-Leymann-Str. 9 D-45127 Essen, Germany
2
National Research University Higher School of Economics
Shabolovka, 26, 119049 Moscow, Russia,
E-mail: [email protected]
3
Universität Hamburg, Faculty of Mathematics
Bundesstraße 55, 20146 Hamburg, Germany
E-mail: [email protected]
4
CREST, ENSAE, Université Paris-Saclay
5, avenue Henry Le Chatelier, 91120 Palaiseau, France
E-mail: [email protected]
We study the estimation of the covariance matrix Σ of a p-dimensional normal random vector based
on n independent observations corrupted by additive noise. Only a general nonparametric assumption
is imposed on the distribution of the noise without any sparsity constraint on its covariance matrix. In
this high-dimensional semiparametric deconvolution problem, we propose spectral thresholding estimators
that are adaptive to the sparsity of Σ. We establish an oracle inequality for these estimators under model
miss-specification and derive non-asymptotic minimax convergence rates that are shown to be logarithmic
in n/ log p. We also discuss the estimation of low-rank matrices based on indirect observations as well as
the generalization to elliptical distributions. The finite sample performance of the threshold estimators is
illustrated in a numerical example.
MSC 2010 subject classifications: Primary 62H12; secondary 62F12, 62G05.
Keywords: Thresholding, minimax convergence rates, Fourier methods, severely ill-posed inverse problem.
1. Introduction
One of the fundamental problems of multivariate data analysis is to estimate the covariance matrix
Σ ∈ Rp×p of a random vector X ∈ Rp based on independent and identically distributed (i.i.d.)
realizations X1 , . . . , Xn of X. An important feature of data sets in modern applications is high
dimensionality. Since it is well known that classical procedures fail if the dimension p is large,
various novel methods of high-dimensional matrix estimation have been developed in the last
decade. However, an important question has not yet been settled: How can Σ be estimated in a
high-dimensional regime if the observations are corrupted by noise?
Let X1 , . . . , Xn be i.i.d. random variables with multivariate normal distribution N (0, Σ). The
maximum likelihood estimator of Σ is the sample covariance estimator
n
Σ∗X :=
1X
Xj Xj> .
n j=1
The estimation error of Σ∗X explodes for large p. To overcome this problem, sparsity assumptions
can be imposed on Σ, reducing the effective number of parameters. The first rigorous studies of
this idea go back to Bickel and Levina [3, 4] and El Karoui [21] who have assumed that most
1
2
D. Belomestny, M. Trabs and A.B. Tsybakov
entries of Σ are zero or very small. This allows for the construction of banding, tapering and
thresholding estimators based on Σ∗X , for which the dimension p can grow exponentially in n.
Subsequently, a rich theory has been developed in this direction including Lam and Fan [34] who
proposed a penalized pseudo-likelihood approach, Cai et al. [11] who studied minimax optimal
rates, Cai and Zhou [12] studying the `1 loss as well as Rothman et al. [46] and Cai and Liu [8]
for more general threshold procedures and adaptation, to mention only the papers most related
to the present contribution. For current reviews on the theory of large covariance estimation, we
refer to [9, 24]. Heading in a similar direction as noisy observations, covariance estimation in the
presence of missing data has been recently investigated by Lounici [37] as well as Cai and Zhang
[10].
Almost all estimators in the afore mentioned results build on the sample covariance estimator
Σ∗X . In this paper, we assume that only the noisy observations
Yj = Xj + εj ,
j = 1, . . . , n,
are available, where the errors ε1 , . . . , εn are i.i.d. random vectors in Rp independent of X1 , . . . , Xn .
Then the sample covariance estimator Σ∗Y is biased:
n
i
h1 X
Yi Yi> = Σ + Γ
E[Σ∗Y ] = E
n i=1
where Γ = E[ε1 ε>
1 ] is the covariance matrix of the errors. Assuming Γ known to correct the bias is
not very realistic. Moreover, for heavy tailed the errors εj that do not have finite second moments,
Γ is not defined and the argument based on Σ∗Y makes no sense. Several questions arising in this
context will be addressed below:
(i) How much information on the distribution of εj do we need to consistently estimate Σ?
(ii) Do we need finite second moments of εj and/or sparsity restrictions on Γ to estimate Σ?
(iii) What is the minimax optimal rate of estimating Σ based on noisy observations?
If the covariance matrix Γ of the errors exists and is known, the problem does not differ from
the direct observation case, since Γ can be simply subtracted from Σ∗Y . If Γ can be estimated, for
instance from a separate sample of the error distribution or from repeated measurements, we can
proceed similarly. However, in the latter case, we need to assume that Γ is sparse, since otherwise
we cannot find a good estimator for large dimensions. Reducing our knowledge about εj further,
we may only assume that the distribution of εj belongs to a given nonparametric class. This leads
to a high-dimensional deconvolution model. The difference from standard deconvolution problems
is that the density of Xj ’s is a parametric object known up to a high-dimensional matrix parameter
Σ. A related model in the context of stochastic processes has been recently studied by Belomestny
and Trabs [2]. Obviously, we need some assumption on the distribution of errors since otherwise
Σ is not identifiable as, for example, in the case of normally distributed εj . It turns out that we
do not need a sparse covariance structure for the error distribution and we can allow for heavy
tailed errors without any moments.
From the deconvolution point of view, it might seem surprising that Σ and thus the distribution
of Xj can be estimated consistently without knowing or estimating the distribution of errors εj ,
but as we will show it is possible. The price to pay for this lack of information is in the convergence
rates that turn out to be very slow - logarithmic in the sample size. In the pioneering works in
one-dimensional case, Matias [40], Butucea and Matias [5] have constructed a variance estimator
in deconvolution model with logarithmic convergence rate and a corresponding lower bound. In
this paper, we provide a general multidimensional analysis of the minimax rates on the class of
sparse covariance matrices.
To replace the sample covariance matrix Σ∗Y by a deconvolution counterpart, we use some ideas
from the literature on density deconvolution. Starting with Carroll and Hall [13] and Fan [22], the
deconvolution problem have been extensively studied. In particular, unknown (but inferable) error
distributions have been analysed by Neumann [42], Delaigle et al. [18], Johannes [31] and Delaigle
Covariance estimation in high-dimensional deconvolution
3
and Hall [17] among others. For adaptive estimation with unknown error distribution we refer to
Comte and Lacour [14], Kappus and Mabon [32], Dattner et al. [16] and references therein. Almost
all contributions to the deconvolution literature are restricted to a univariate model. Hence, our
study contributes to the deconvolution theory by treating the multivariate case; in particular, our
techniques for the lower bounds might be of interest. To our knowledge, only Masry [39], Eckle
et al. [20], and Lepski and Willer [35, 36] have studied the setting of multivariate deconvolution.
They deal with a different problem, namely that of nonparametric estimation of the density of Xj
or its geometric features when the distribution of εj is known.
Applying a spectral approach, we construct an estimator for the covariance matrix assuming
that Xj are normally distributed and that the characteristic function ψ of the distribution of
εj decays slower than the Gaussian characteristic function. A similar idea in a one-dimensional
deconvolution problem has been developed by Butucea et al. [6]. The assumption log |ψ(u)| =
o(|u|2 ) as |u| → ∞ implies identifiability of Σ and allows us to construct an estimator Σ̂, which is
consistent in the maximal entry norm. Based on Σ̂, we then construct hard and soft thresholding
S
estimators Σ̂H
τ and Σ̂τ , respectively, for sparse matrices. The sparsity is described by an upper
bound S on the `q -norm, q ∈ [0, 2), of entries of Σ. We establish sparsity oracle inequalities for
S
Σ̂H
τ and Σ̂τ when the estimation error is measured in the Frobenius norm. This choice of the norm
is naturally related to the distance between two multivariate normal distributions. The oracle
bounds reveal that the thresholding estimators adapt to the unknown sparsity S. For the soft
thresholding estimator we present an oracle inequality, which shows that the estimator adapts
also to approximate sparsity.
Assuming that the characteristic function ψ of εj satisfies log |ψ(u)| = O(|u|β ) for large u ∈ Rp
and some β ∈ [0, 2), we prove the following upper bound on the estimation error in the Frobenius
norm:
n −(1−β/2)(1−q/2)
1/2
(1)
kΣ̂H
log
τ − Σk 6 CS
log p
for some constant C > 0 and with high probability. The dependence of this bound on the sparsity
S is the same as found by Bickel and Levina [3] for the case direct observations; furthermore the
well-known quotient n/ log p drives the rate. However, the severely ill-posed nature of the inverse
problem causes the logarithmic dependence of the rate on n/ log p. We also see that the estimation
problem is getting harder if β gets closer to 2 where it is more difficult to distinguish the signal
from the noise. Furthermore, we establish a lower bound showing that the rate in (1) cannot be
improved in a minimax sense for q = 0. Let us emphasise that our observations Yj are by definition
not normally distributed. Therefore, the proof of the lower bound differs considerably from the
usual lower bounds in high-dimensional statistics, which rely on Gaussian models.
Covariance estimation is crucial in many applications where also observation errors appear.
For instance, many portfolio optimization approaches rely on the covariance matrix of a possibly
high number of assets where the financial data are typically perturbed due to bid-ask spreads,
micro-structure noise etc. [23, 49]. While in a high-frequency regime the observation noise can be
handled by local averages, in a low-frequency situation, as daily closing prices, the denoising is
more difficult and our deconvolution approach can be applied, cf. [2]. Note that the dimension
dependence in [2] can be improved with our analysis for low-rank matrices. As another application
the spatial empirical covariance matrices of climate data and their eigenvectors, called empirical
orthogonal functions, are important spatio-temporal statistics. Naturally recordings of climate
data, e.g. sea surface temperatures, may suffer from measurement errors [15] and should be taken
into account. Especially, sparse covariance structures appear in the problem of spatio-temporal
wind speed forecasting taking into account the time series data of a target station and data of
surrounding stations, see [47].
This paper is organized as follows. In Section 2 we construct and analyze the spectral covariance
matrix estimator. In Section 3 the resulting thresholding procedures are defined and analyzed. In
Section 4 we investigate upper and lower bounds on the estimation error. In Section 5 some
extensions of our approach are discussed including the estimation of low-rank matrices based
4
D. Belomestny, M. Trabs and A.B. Tsybakov
on indirect observations as well as the generalization to elliptical distributions. The numerical
performance of the procedure is illustrated in Section 6. Longer and more technical proofs are
postponed to Section 7 and to the appendix.
Notation: For any x ∈ Rp and q ∈ (0, ∞], the `q -norm of x is denoted by |x|q and we write
for brevity |x| := |x|2 . For x, y ∈ Rp the Euclidean scalar product is written as hx, yi. We denote
by Ip the p × p identity matrix, and by 1{·} the indicator function. For two matrices A, B ∈
Rp×p thepFrobenius scalar product is given by hA, Bi := tr(A>√
B) inducing the Frobeninus norm
Ai.
The
nuclear
norm
is
denoted
by
kAk
:=
tr(
A> A) and the spectral norm by
kAk := hA,
1
p
kAk∞ := λmax (A> A), where λmax (·) stands for the maximal eigenvalue. For A ∈ Rp×p and
q ∈ [0, ∞] we denote by |A|q the `q -norm of the entries of the matrix if q > 0 and the number of
non-zero entries for q = 0. We write A > 0 or A > 0 if the matrix A ∈ Rp×p is positive definite or
semi-definite. We denote by PΣ,ψ the joint distribution of Y1 , . . . , Yn when the covariance matrix
of Xj is Σ and the characteristic function of the noise εj is ψ. We will write for brevity PΣ,ψ = P
if there is no ambiguity.
2. Spectral covariance estimators
Let ψ denote the characteristic function of error distribution:
ψ(u) = E eihu,ε1 i , u ∈ Rp .
Then the characteristic function of Yj is given by
1
ϕ(u) := E eihu,Yj i = exp − hu, Σui + log ψ(u) ,
2
u ∈ Rp .
Here and throughout we assume that ψ(u) 6= 0 and we use the distinguished logarithm, cf. [48,
Lemma 7.6]. This assumption is standard in the literature on deconvolution. Allowing for some
zeros of ψ has been studied in [41, 19]. Note that our estimation procedure defined below does not
rely on all u in Rd , but uses only u with a certain radius |u|.
The canonical estimator for the characteristic function ϕ is the empirical characteristic function
n
ϕn (u) :=
1 X ihu,Yj i
e
,
n j=1
u ∈ Rp .
√
Since ϕn (u) concentrates around ϕ(u) with rate n, we have p
ϕn (u) 6= 0 with overwhelming
probability for sufficiently large frequencies u ensuring |ϕ(u)| > C (log(ep))/n for some constant
C > 1 (see Lemma 13 and Corollary 14). In this case log ϕn (u) is well defined. On the unlikely
event {ϕn (u) = 0}, we may set log ϕn (u) := 0.
Arguing similarly to Belomestny and Trabs [2], we consider the identity
log ϕn (u)
hu, Σui log ψ(u) log ϕn (u) − log ϕ(u)
=−
+
+
,
|u|2
|u|2
|u|2
|u|2
u ∈ Rp \ {0}.
(2)
Both sides are normalized by |u|2 being the order of the leading term hu, Σui. While the lefthand side of (2) is a statistic based on the observations Y1 , . . . , Yn , the first term on the righthand side encodes the parameter of interest, namely the covariance matrix Σ. The second term
is a deterministic error due to the unknown distribution of εj . If | log ψ(u)| = o(|u|2 ), i.e., the
error distribution is less smooth than the normal distribution, the deterministic error vanishes as
|u| → ∞. The third term in (2) is a stochastic error term. Using the first order approximation we
get
ϕ (u) − ϕ(u)
ϕ (u) − ϕ(u)
n
n
+1 ≈
.
(3)
log ϕn (u) − log ϕ(u) = log
ϕ(u)
ϕ(u)
Covariance estimation in high-dimensional deconvolution
5
The latter expression resembles the estimation error in classical deconvolution problems. However,
there is a difference since here in the denominator we have ϕ(u) rather than the characteristic
function of the distribution of errors. A similar structure was detected in the statistical analysis of
low-frequently observed Lévy processes by Belomestny and Reiß [1]. Following [1], one can call this
type of problems auto-deconvolution problems. Since |ϕ(u)| = e−hu,Σui/2 |ψ(u)|, and we assume that
| log ψ(u)| = o(|u|2 ), the stochastic error grows exponentially in |u|. Thus, the estimation problem
is severely ill-posed even in one-dimensional case.
These remarks lead us to the conclusion that Σ can be estimated consistently without any
particular knowledge of the error distribution as soon as | log ψ(u)| = o(|u|2 ), and the spectral
radius |u| in (2) is chosen to achieve a trade-off between the stochastic and deterministic errors.
To specify more precisely the condition | log ψ(u)| = o(|u|2 ), it is convenient to consider, for any
β ∈ (0, 2) and T > 0, the following nonparametric class of functions ψ:
Hβ (T ) := ψ characteristic function on Rp : log |ψ(u)| 6 T 1 + |u|ββ , u ∈ Rp .
Note that log |ψ(u)| = log(1/|ψ(u)|) since |ψ(u)| 6 1. Therefore, the condition that determines
the class Hβ (T ) can be written as the lower bound |ψ(u)| > exp −T 1+|u|ββ . If the characteristic
function of εj belongs to Hβ (T ), the decay |u|ββ for some β < 2 of the characteristic exponent
allows for separating the normal distribution of Xj from error distribution for large |u|. The decay
rate β determines the ill-posedness of the estimation problem. Noteworthy, we require neither
sparsity restrictions on the joint distribution of (ε1 , . . . , εn ) nor moment conditions of these random
variables.
A typical representative in the class Hβ is a characteristic function of a vector of independent
β-stable random variables. In the case of identically distributed marginals, it has the form ψ(u) =
exp(−σ|u|ββ ), u ∈ Rp , for some parameter σ > 0. A related example with correlated coefficients
is a p-dimensional stable distribution with characteristic function ψ(u) = exp(−σ|u|β2 ) (note that
|u|β2 6 |u|ββ ). Recalling that stable distributions can be characterized as limit distributions of
normalized sums of independent random variables and interpreting εj as accumulation of many
small measurement errors, suggests that these examples are indeed quite natural.
If ψ ∈ Hβ (T ), the deterministic error term in (2) is small for large values of |u|. We will choose
u in (2) in the form U u(i,j) where U > 0 is large, and u(i,j) are p-dimensional unit vectors defined
by
1
(4)
u(i,i) := u(i) := (1{i=k} )k=1,...,p and u(i,j) := √ u(i) + u(j) for i 6= j.
2
Using the symmetry of Σ = (σi,j )i,j=1,...,p , we obtain
hu(i) , Σu(i) i = σi,i
and hu(i,j) , Σu(i,j) i = σi,j +
σi,i + σj,j
2
for any i, j ∈ {1, . . . , p} with i 6= j. Motivated by (2) applied to U u(i,j) for some spectral radius
U > 0, we introduce the spectral covariance estimator:
(
− U12 Re log ϕn (U u(i) ) ,
if i = j,
Σ̂ = (σ̂i,j )i,j=1,...p with σ̂i,j :=
(5)
− U12 Re log ϕn (U u(i,j) ) − 12 (σ̂i,i + σ̂j,j ),
if i 6= j.
Equivalently, we can write Re(log ϕn (u)) = log |ϕn (u)| for any u ∈ Rp with |ϕn (u)| 6= 0. Since
ϕn (u) concentrates around ϕ(u), cf. Lemma 13, we have ϕn (u) 6= 0 with high probability if
ϕ(u) 6= 0.
The spectral covariance estimator Σ̂ can be viewed as a counterpart of the classical sample
covariance matrix for the case of indirect observations. The entries σ̂i,j of Σ̂ enjoy the following
concentration property.
6
D. Belomestny, M. Trabs and A.B. Tsybakov
Theorem 1. Assume that |Σ|∞ 6 R, and ψ ∈ Hβ (T ) for some β, R, T > 0. Let γ >
p
2
β
U > 1 satisfy 8γ (log(ep))/n < e−RU −3T U . Set
τ (U ) = 6γ
eRU
2
+3T U β
log(ep) 1/2
U2
n
+ 3T U −2+β .
√
2 and
(6)
Then, for any τ > τ (U ),
2
PΣ,ψ |σ̂i,j − σi,j | < τ > 1 − 12(ep)−γ
and
PΣ,ψ
2
max |σ̂i,j − σi,j | < τ > 1 − c∗ p2−γ
i,j=1,...,p
2
where c∗ = 12e−γ .
Proof. Set S(u) = Re(log ϕn (u) − log ϕ(u)). Using (2) we obtain, for all i, j = 1, . . . , p,
|σ̂i,i − σi,j | 6 U −2 |S(U u(i,j) )| + U −2 log |ψ(U u(i,j) )|
6 U −2 |S(U u(i,j) )| + U −2
max
log |ψ(U u(i,j) )| .
i∈{1,...,p}
For U > 1 the last summand in this display is bounded uniformly by 3T U −2+β on the class Hβ (T ).
This remark and Corollary 14 in Section 7.1 imply that
p
2
6γ log(ep)
−2+β
+
3T
U
6 12(ep)−γ
P |σ̂i,j − σi,j | > √ 2
(i,j)
nU mini,j∈{1,...,p} |ϕ(U u
)|
p
if the condition γ (log(ep))/n < |ϕ(U u(i,j) )|/8 is satisfied for all i, j. Note that for any i, j =
1, . . . , p, and any ψ ∈ Hβ (T ),
U 2 hu(i,j) , Σu(i,j) i
|ϕ(U u(i,j) )| = exp −
+ Re log ψ(U u(i,j) )
2
2
> exp − U |Σ|∞ + 3T U β−2 .
Therefore, for γ and U satisfying the conditions of the theorem,
2
β
2
eRU +3T U log(ep) 1/2
−2+β
P |σ̂i,j − σi,j | > 6γ
+
3T
U
6 12(ep)−γ .
2
U
n
A union bound concludes the proof.
the familiar factor
p The first term in τ (U ) is an upper bound for the stochastic error. We recover
(log p)/n which is due to a sub-Gaussian bound on the maximum of the p2 entries (σ̂i,j ). The
term exp(RU 2 + 3T U β ) is an upper bound for ϕ(u)−1 appearing in the linearization (3). Note
that for β < 2 this bound can be written aspexp(RU 2 (1 + o(1))) for U → ∞. This suggests the
choice of spectral radius in the form U∗ = c log(n/ log(ep)) for some sufficiently small constant
c > 0. The second term in (6) bounds the deterministic error and determines the resulting rate
U∗−2+β = O((log(n/ log(ep)))−1+β/2 ), cf. Theorem 5.
3. Thresholding
Based on the spectral covariance estimator, we can now propose estimators of high-dimensional
sparse covariance matrices. We consider the following sparsity classes of matrices:
n
o
G0 (S, R) := Σ > 0 : Σ = Σ> , |Σ|0 6 S, |Σ|∞ 6 R
and
(7)
n
o
Gq (S, R) := Σ > 0 : Σ = Σ> , |Σ|qq 6 S, |Σ|∞ 6 R
for q ∈ (0, 2),
Covariance estimation in high-dimensional deconvolution
7
where S > 0 denotes the sparsity parameter and R > 0 bounds the largest entry of Σ. We also
consider larger classes Gq∗ (S, R) that differ from Gq (S, R) only in that the condition Σ > 0 is
dropped. Note that S > p for the classes Gq (S, R), since otherwise the condition Σ > 0 does
not hold. This restriction on S does not apply to the classes Gq∗ (S, R), for which the unknown
effective dimension of Σ can be smaller than p. However, for the classes Gq∗ (S, R), the overall
model remains, in general, p-dimensional since the distribution of the noise can be supported on
the whole space Rp .
The sparsity classes considered by Bickel and Levina [3] and in many subsequent papers are
given by
p
n
o
X
Uq (s, R) := Σ > 0 : Σ = Σ> , max
|σi,j |q 6 s, max σi,i 6 R
i
i
j=1
for s, R > 0, q ∈ (0, 1) and with the usual modification for q = 0. We have Uq (s, R) ⊆ Gq (sp, R),
so that our results can be used to obtain upper bounds on the risk for the classes Uq (s, R).
Based on the spectral covariance estimator, we define the spectral hard thresholding estimator
for Σ as
H
Σ̂H
τ : = (σ̂i,j )i,j=1,...,p
with
H
σ̂i,j
:= σ̂i,j 1{|σ̂i,j |>τ } ,
(8)
for some threshold value τ > 0. The following theorem gives an upper bound on the risk of this
estimator in the Frobenius norm.
Theorem 2. Let R, T, S > 0, β ∈ [0, 2), and q ∈ [0, 2). Let τ (U ) be defined in (6) with parameters
p
√
2
β
γ > 2 and U > 1 satisfying 8γ (log(ep))/n 6 e−RU −3T U . Then
sup
Σ∈Gq∗ (S,R),ψ∈Hβ (T )
1/2 1−q/2
PΣ,ψ (kΣ̂H
τ
) 6 c∗ p2−γ
τ − Σk > 3S
2
2
provided that τ > τ (U ) for q = 0, and τ > 2τ (U ) for q ∈ (0, 2). Here, c∗ = 12e−γ .
Proof. First, consider the case q = 0 and τ > τ (U ). In view of Theorem 1, the event A =
2
maxi,j=1,...,p |σ̂i,j − σi,j | < τ is of probability at least 1 − c∗ p2−γ for all τ > τ (U ). On A we
have the inclusion j : |σ̂i,j | > τ ⊆ j : σi,j 6= 0 , so that |Σ̂H |0 6 |Σ|0 . Therefore, on the
event A,
2
H
H
2
H
2
H
2
kΣ̂H
τ − Σk 6 |Σ̂τ − Σ|0 |Σ̂τ − Σ|∞ 6 2|Σ|0 |Σ̂τ − Σ|∞ 6 2S|Σ̂τ − Σ|∞ .
H
Note that, again on A, we have |Σ̂H
τ − Σ|∞ 6 |Σ̂τ − Σ̂|∞ + |Σ̂ − Σ|∞ 6 2τ . Combining this with
the last display implies the assertion of the theorem for q = 0.
Consider now the case q ∈ (0, 2) and τ > 2τ (U ). We use the following elementary fact: If
|y − ϑ| 6 r for some y, ϑ ∈ R and r > 0, then |y 1{|y|>2r} − ϑ| 6 3 min{|ϑ|, r} (cf.[51]). Taking
y = σ̂i,j , ϑ = σi,j , and r = τ /2, and using Theorem 1 we obtain that, on the event of probability
2
at least 1 − c∗ p2−γ ,
H
|σ̂i,j
− σi,j | 6 3 min{|σi,j |, τ /2},
i, j = 1, . . . , p.
2
Thus, for any q ∈ (0, 2), with probability at least 1 − c∗ p2−γ ,
X
X
2
H
2
kΣ̂H
(σ̂i,j
− σi,j )2 6 9
min{σi,j
, τ 2 /4} 6 9(τ /2)2−q |Σ|qq 6 9τ 2−q S.
τ − Σk =
i,j
i,j
Since all bounds hold uniform in Σ ∈ Gq∗ (S, R) and ψ ∈ Hβ (T ), the theorem is proven.
In the direct observation case where εj = 0 we have ψ(u) = 1 for all u ∈ Rp , so that the
deterministic error term in (6) disappears. In this case, U can be fixed and the threshold can be
8
D. Belomestny, M. Trabs and A.B. Tsybakov
p
chosen as a multiple of (log p)/n, analogously to [3]. Together with the embedding Uq (s, R) ⊆
Gq (sp, R), we recover Theorem 2 from Bickel and Levina [3]. In Section 4 we will discuss in detail
the optimal choice of the spectral radius and the threshold in the presence of noise.
The spectral soft thresholding estimator is defined as
S
S
Σ̂Sτ := (σ̂i,j
)i,j=1,...,p with σ̂i,j
:= sign(σ̂i,j ) |σ̂i,j | − τ +
with some threshold τ > 0. It is well known, cf., e.g. [51], that
Σ̂Sτ = arg min |A − Σ̂|22 + 2τ |A|1 .
(9)
A∈Rp×p
Adapting the proof of Theorem 2 in Rigollet and Tsybakov [44], we obtain the following oracle
inequality, which is sharp for q = 0 and looses a factor 2 otherwise.
Theorem 3. Assume that |Σ|∞ 6 R, and ψ ∈ Hβ (T ) for some β, R, T > 0.pLet τ > τ (U )
√
where τ (U ) is defined in (6) with parameters γ > 2 and U > 1 such that 8γ (log(ep))/n 6
2
β
e−RU −3T U . Then,
n
o
√ 2 2
2
2)
τ
|A|
kΣ̂Sτ − Σk2 6 min
kA
−
Σk
+
(1
+
(10)
0
p×p
A∈R
2
2
with probability at least 1 − c∗ p2−γ where c∗ = 12e−γ . For any q ∈ (0, 2) we have, with probability
2
at least 1 − c∗ p2−γ ,
n
o
(11)
kΣ̂Sτ − Σk2 6 min 2kA − Σk2 + c(q)τ 2−q |A|qq
A∈Rp×p
where c(q) > 0 is a constant depending only on q.
Proof. Starting from the characterization (9), we use Theorem 2 by Koltchinskii et al. [33]. To this
end, we write σ̂i,j = σi,j + ξi,j , i, j ∈ {1, . . . , p}, where ξi,j are random variables with exponential
concentration around zero due to Theorem 1. Observing σ̂i,j is thus a sequence space model in
>
dimension p2 and a special case of the trace regression model Yj = tr(Zi,j
A0 ) + ξi,j considered
in [33]. Namely, A0 is the diagonal matrix with diagonal entries σi,j and Zi,j are diagonalisations
of the canonical basis in Rp×p . In particular, Assumption 1 in [33] is satisfied for µ = p, i.e.,
kBk2L2 (Π) = p−2 |B|22 where we use the notation of [33]. Note also that the rank of a diagonal
matrix B is equal to the number of its non-zero elements. Consequently, Theorem 2 in [33] yields
with λ = 2τ
p2 that
o
n
√ 2 2
2
|Σ̂Sτ − Σ|22 6 min
|A
−
Σ|
+
(1
+
2)
τ
|A|
0
2
p×p
A∈R
on the event that A = {maxi,j |σ̂i,j − σi,j | < τ }. To estimate the probability of A, we apply
Theorem 1. Inequality (11) follows from (10) using the same argument as in Corollary 2 of [44].
This theorem shows that the soft thresholding estimator allows for estimating matrices that
are a not exactly sparse but can be well approximated by a sparse matrix. Choosing A = Σ in the
oracle inequalities (10) and (11) we obtain the following corollary analogous to Theorem 2.
Corollary 4. Let R, T, S > 0, β ∈ (0, 2), and q ∈ [0, 2). Let τ > τ (U ) where τ (U ) is defined in
p
√
2
β
(6) with parameters γ > 2 and U > 1 such that 8γ (log(ep))/n 6 e−RU −3T U . Then
sup
Σ∈Gq∗ (S,R),ψ∈Hβ (T )
where C = 1 +
√
PΣ,ψ (kΣ̂Sτ − Σk > CS 1/2 τ 1−q/2 ) 6 c∗ p2−γ
2 for q = 0, and C =
p
c(q) for q ∈ (0, 2).
2
Covariance estimation in high-dimensional deconvolution
9
4. Minimax optimality
In this section, we study minimax optimal rates for the estimation of Σ on the class Gq (S, R) ×
Hβ (T ). We first state an upper bound on the rate of convergence of the hard thresholding estimator
in this high-dimensional semiparametric problem. It is an immediate consequence of Theorem 2.
Due to Corollary 4, the result directly carries over to the soft thresholding estimator.
Theorem 5.
Let R, T, S > 0, β ∈ (0, 2), and q ∈ [0, 2). For γ >
s
1
n
U∗ =
log
.
4R
64γ 2 log(ep)
√
2, set
(12)
1/(2−β)
Let n be large enough such that U∗ > ( 3T
∨(c̄/T )1/β ∨1 for some numerical constant c̄ > 0.
R )
Then for any τ > τ (U∗ ) where τ (·) is defined in (6) we have
2−γ 2
sup
PΣ,ψ kΣ̂H
with
τ − Σk > C̄1 r̄n,p 6 C̄0 p
(Σ,ψ)∈Gq (S,R)×Hβ (T )
r̄n,p := S 1/2 R1−β/2 T log
n −1+β/2 1−q/2
log(ep)
(13)
for some numerical constants C̄0 , C̄1 > 0.
Proof. It follows from the assumption on U∗ that 3T U∗β 6 RU∗2 . This and the definition of U∗
p
2
β
imply that 8γ (log(ep))/n 6 e−RU∗ −3T U∗ . Therefore, we can apply Theorem 2, which yields the
result since
2
τ (U∗ ) 6 6γ
2c̄
e2RU∗ log(ep) 1/2
+ 3T U∗−2+β 6
+ 3 T U∗−2+β .
2
U∗
n
3
It is interesting to compare Theorem 5 with the result of Butucea and Matias [5] corresponding
to p = 1, S = 1, and establishing a logarithmic rate for estimation of the variance in deconvolution
model under exponential decay of the Fourier transform of εj . Butucea and Matias [5] have shown
that, if log |ψ(u)| = O(|u|β ), their estimator achieves asymptotically a mean squared error of the
order (log n)−1+β/2 . This coincides with the case p = 1 and q = 0 of the non-asymptotic bound
in (13). A similar rate for p = 1 has been obtained by Matias [40] under the assumptions on the
decay of the Laplace transform.
We now turn to the lower bound matching (13) for q = 0. Intuitively, the slow rate comes from
the fact that the error distribution can mimic the Gaussian distribution up to some frequency in
the Fourier domain. A rigorous application of this observation to the construction of lower bounds
goes back to Jacod and Reiß [30], though in quite a different setting. For the multidimensional
case that we consider here the issue becomes particularly challenging.
Theorem 6. Let β ∈ (0, 2) and assume that C1 p 6 S 6 C2 p, T (log n)−1+β/2 6 C3 Rβ/2 ,
T (log n)c∗ > 1 ∨ Rβ/2 for some constants C1 , C2 , C3 > 0, and c∗ > 0. Then, there are constants
c1 , c2 > 0 such that
inf
sup
PΣ,ψ kΣ̃ − Σk > c1 rn,p > c2 with
Σ̃ (Σ,ψ)∈G0 (S,R)×Hβ (T )
rn,p := S 1/2 R1−β/2 T log n
where the infimum is taken over all estimators Σ̃.
−1+β/2
10
D. Belomestny, M. Trabs and A.B. Tsybakov
The proof of this theorem is postponed to Section 8. We use the method of reduction to testing
of many hypotheses relying on a control of the χ2 -divergence between the corresponding distributions, cf. Theorem 2.6 in [50]. The present high-dimensional setting introduces some additional
difficulties. When the dimension p of the sample space is growing, an increasing number of derivatives of the characteristic functions has to be taken into account for the χ2 -bound. Achieving
bounds of the correct order in p causes difficulty when p is arbitrarily large. We have circumvented
this problem by introducing a block structure to define the hypotheses. The construction of the
family of covariance matrices of Xj used in the lower bounds relies on ideas from Rigollet and
Tsybakov [44], while the error distributions are chosen as perturbed β-stable distributions. To
bound the χ2 -divergence, we need a lower bound on the probability density of Yj . It is shown by
Butucea and Tsybakov [7] that the tails of the density of a one-dimensional stable distribution are
polynomially decreasing. We generalize this result to the multivariate case (cf. Lemma 15 below)
using properties of infinitely divisible distributions.
We now give some comments on the lower bound of Theorem 6. Assuming S of order p means
that we consider quite a sparse regime. We always have S 6 p2 . Recall also that S > p as the
diagonal of the covariance matrix is included in the definition of S for the class G0 (S, R). An
alternative strategy pursued in the literature is to estimate a correlation matrix, i.e., to assume
that all diagonal entries are known and equal to one. However, this seems not very natural in the
present noisy observation scheme. On the other hand, Theorem 6 shows that even in the sparse
regime S = O(p) the estimation error tends to ∞ as n → ∞ for dimensions p growing polynomially
in n. The logarithmic in n rate reflects the fact that the present semiparametric problem is severely
ill-posed.
Comparing the lower bound rn,p with the upper bound r̄n,p from Theorem 5, we see that they
coincide if the dimension satisfies p = O(exp(cnγ )) for some γ ∈ [0, 1) and some c > 0. Thus, we
have established the minimax optimal rate under this condition. Note also that we only loose a
factor of order log log p for very large p, for instance, if p = en/ log n .
5. Discussion and extensions
5.1. The adaptivity issue
Since the threshold τ (U∗ ) in Theorem 5 depends on unknown parameters R, T , and β, a natural
question is whether it is possible to construct an adaptive procedure independent of these parameters that achieves the same rate. One possibility to explore consists in selecting τ in a data-driven
way. Another option would be to construct estimators corresponding to values of R, T , and β on
a grid, and then to aggregate them.
For direct observations an adaptive choice of the threshold, more precisely a cross-validation
criterion, has been proposed by Bickel and Levina [3] and was further investigated by Cai and Liu
[8]. For noisy observations that we consider here, the adaptation problem turns out to be more
delicate since not only an optimal constant has to be selected but also the order of magnitude of
τ (U ) depends on the unknown parameter β.
Often an upper bound R on the maximal entry of Σ is known, so that one does not need
considering adaptation
to R. Ignoring the issue of unknown R, the choice of the spectral radius U∗
p
of the order R−1 log(n/ log(ep)) is universal, which reflects the fact that the estimation problem
is severely ill-posed with dominating bias. Indeed, U∗ in Theorem 5 corresponds to undersmoothing
such that the deterministic estimation error dominates the stochastic error without deteriorating
the convergence rates. To construct an adaptive counterpart of τ , we need either an estimator
of the error of an optimal procedure for estimating Σ under the | · |∞ -loss or an estimator of the
“regularity” β. Therefore, extrapolating the argument of Low [38] to our setting, it seems plausible
that an adaptive choice of τ cannot, in general, lead to the optimal rate. This does not exclude that
optimal adaptive estimators can be constructed by other type of procedure, such as aggregation
of estimators on the grid as mentioned above.
Covariance estimation in high-dimensional deconvolution
11
5.2. Low-rank covariance matrix
Alternatively to the above setting where the covariance matrix Σ is sparse, we can consider a lowrank matrix Σ. This is of particular interest in the context of factor models where, as discussed
by Fan et al. [25, 26], an additional observation error should be taken into account. While [25,
26] estimate the covariance matrix of the noisy observations assuming that the errors have a
sparse covariance structure, a spectral approach analogous to the one developed above allows for
estimating directly the low-rank covariance matrix of X without sparsity restrictions on the error
distribution.
Such an approach, which is at first sight quite natural, would be to use the spectral covariance
estimator Σ̂ from (5) together with a nuclear norm penalization. The following oracle inequality
is an easy consequence of Theorem 1 in Koltchinskii et al. [33].
Proposition 7. Assume that M ⊆ Rp×p is convex and let τ > 0. On the event {2kΣ̂−Σk∞ 6 τ },
2
the estimator Σ̂R
τ := arg minS∈M {kS − Σ̂k + τ kSk1 } satisfies
√
o
n
1 + 2 2 2
R
2
2
τ rank(S) .
kΣ̂τ − Σk 6 inf kS − Σk +
S∈M
2
To use this proposition, we need to find a bound on the spectral norm kΣ̂ − Σk∞ that hold with
high probability. The techniques from Cai et al. [11] designed for the case of direct observations
allow us to obtain an upper bound on this quantity of order p up to a logarithmic in n/ log(p)
factor. Thus, the convergence rate of this estimator is rather slow.
Let us show now that another estimator can be constructed based the approach from Belomestny
and Trabs [2], which allows for a better dependence on p. To this end, we write
−
hu, Σui
uu>
= hΘ(u), Σi with design matrix Θ(u) := − 2 ,
2
|u|
|u|
u ∈ Rp \ {0}.
For a weight function w : Rp → R+ supported on the annulus {u ∈ Rp : 14 6 |u| 6 21 } and a
spectral radius U > 1, we set wU (u) := U −p w(u/U ), u ∈ Rp . Motivated by (2), we define the
weighted Lasso-type estimator
n ˆ Re log ϕ (u)1
2
o
n
{|ϕn (u)|>ι}
Σ̃λ := arg min
−
hΘ(u),
M
i
w
(u)du
+
λkM
k
(14)
U
1
|u|2
M ∈M
Rp
for a convex set M ⊆ {M ∈ Rp×p : M > 0} and with nuclear norm penalisation for some λ > 0.
We have inserted a truncation function 1{|ϕn (u)|>ι} for some threshold ι > 0 which increases the
stability of the estimator by cutting
off frequencies with too small point estimates ϕn (u). Under
√
the universal choice ι = 1/(2 n) this indicator function will be one with high probability. The
estimator Σ̃λ is associated to the weighted scalar product which replaces the classical empirical
scalar product:
ˆ
hA, BiU :=
hΘ(u), AihΘ(u), BiwU (u)du and kAk2U := hA, AiU ,
Rd
for matrices A, B ∈ Rp×p . As in [2, Lemma 3.2] we have for any for any positive semi-definite
matrix A ∈ Rp×p an isometry with respect to the Frobenius norm
ˆ
|v1 |4
w(v)dv, κ w := kwkL1 .
κ w kAk2 6 kAk2U 6 κ w kAk2 with κ w :=
4
Rp |v|
Adapting slightly the proof of Theorem 1 in [33], we obtain the following oracle inequality.
12
Theorem 8.
D. Belomestny, M. Trabs and A.B. Tsybakov
Let M be convex. Define
ˆ
Re log ϕn (u)1{|ϕn (u)|>ι}
Rn :=
− hΘ(u), Σi Θ(u)wU (u)du.
2
|u|
Rp
The estimator Σ̃λ from (14) satisfies on the event {kRn k∞ 6 λ}
kΣ̃λ − Σk2U 6 inf kM − Σk2U + C∗2 λ2 rank(M )
M ∈M
for the constant C∗ = (1 +
√
2)/(2κ w ) depending only w.
We omit the proof of this theorem as it is analogous to Theorem 3.4 in [2]. In combination with
the isometry property we obtain an oracle inequality with respect to the Frobenius norm:
kΣ̃λ − Σk2 6 inf C1∗ kM − Σk2 + C2∗ λ2 rank(M )
M ∈M
√
with C1∗ = κ w /κ w and C2∗ = (1 + 2)2 /(4κ 3w ). The best leading constant in this oracle inequality
can be obtained by minimizing C1∗ with respect to w. We do not detail it here.
To apply Theorem 8, we need a sharp probabilistic bound for kRn k∞ . At first sight, this might
look similar to bounding kΣ̂ − Σk∞ in Proposition 7. However, the dependence on the dimension
is much better because the design matrix satisfies kΘ(u)k∞ = 1.
Consider the error distributions in the subclass of Hβ (T ) defined as follows:
n
o
Hβ0 (T ) := ψ characteristic function : log |ψ(u)| 6 T 1 + |u|β2 , u ∈ Rp ⊆ Hβ (T ).
Theorem 9. Let T > 0, β ∈ [0, 2) and ψ ∈ Hβ0 (T ) and choose ι = 2√1 n . Then there are constants
Ci = Ci (w) > 0, i = 1, 2, depending only on w, such that for any γ > 1 and any U > 1 satisfying
√
2
β
2
ekΣk∞ U /8+2T U 6 n we have P(kRn k∞ > λ) 6 3e−γ if
2
λ > C1 γ 2
β
ekΣk∞ U /4+4T U
√
+ C2 T U −2+β .
U2 n
(15)
The proof√is given in the appendix. The right-hand side of (15) is similar to the threshold (6),
but without log p. Hence, this upper bounds depends on the dimension p only via spectral norm
kΣk∞ .pIn the well-specified case, Σ ∈ M and optimizing over the spectral radius yields U of the
−1+β/2
order (log n)/kΣk∞ and the corresponding λ of the order (kΣk−1
. The error bound
∞ log n)
takes the form
p
1−β/2
kΣ̃λ − Σk 6 C rank(Σ) kΣk∞
(log n)−1+β/2
with high probability. Here, C > 0 is a constant depending only on w and T . Note that this bound
for the estimation error improves a corresponding result in [2]. In the direct observation case, we
p
−1/2
can choose U = kΣk∞ and obtain kΣ̃λ − Σk 6 CkΣk∞ rank(Σ)/n with high probability.
5.3. Elliptical distributions
Most of the literature on high-dimensional covariance estimation relies on a sub-Gaussian assumption on the distribution of Xj . To relax the moment assumption and allow for heavy-tailed
distributions, the rich class of elliptical distributions has been studied, see the review paper by Fan
et al. [24]. We refer to Fang et al. [27] for an introduction to the theory of elliptical distributions.
We will now outline how our approach can be generalized to the case where Xj follow a centered
elliptical distribution, that is the characteristic function of Xj is of the form
E[eihu,Xj i ] = Φ(u> Σu),
u ∈ Rp ,
Covariance estimation in high-dimensional deconvolution
13
for some scalar function Φ : R → R and some positive definite matrix Σ, which is proportional
to the covariance matrix. The function Φ is called the characteristic generator. It is easy to see
that E[Xj Xj> ] = −2Φ0 (0)Σ provided that Φ is differentiable. We impose the mild assumption
that Φ(·) = exp(−η(·)) for some function η : R+ → R+ . Then, the characteristic function of the
observations Yj has the form
ϕ(u) = exp − η(u> Σu) + log ψ(u) , u ∈ Rp .
We recover the Gaussian case with η(x) = x2 . Other important examples are multivariate α-stable
distributions where η(x) = xα/2 for α ∈ (0, 2] or normal mixtures. To adapt the estimation strategy
from Section 2, we assume that | Re log ψ(u)| decays slower than η(u> Σu). If η is differentiable
and strictly monotone with inverse function η −1 , a first order Taylor approximation and the fact
that (η −1 )0 = 1/(η 0 ◦ η −1 ) yield
log |ψ(u)|
.
η −1 − log |ϕ(u)| = η −1 η u> Σu − log |ψ(u)| ≈ u> Σu − 0 >
η (u Σu)
If the last term is of smaller order than u> Σu = hu, Σui for |u| → ∞, we can use these heuristics
to estimate Σ. The argument is made rigorous by the following lemma proved in the appendix.
Lemma 10. Let E[eihu,Xj i ] = exp(−η(u> Σu)) for a positive-definite matrix Σ and a strictly
monotone function η : R+ → R+ which is twice continuously differentiable outside a neighbourhood
of the origin. Assume further that
log |ψ(u)|
6 T (1 + |u|)β
η 0 (hu, Σui)
and
|xη 00 (x)| 6 T |η 0 (x)|,
for all u ∈ Rp , x ∈ R+ ,
for some β < 2 and T > 0. For all u ∈ Rp with |u| > (2β+1 T 2 /λmin )1/(2−β) ∨ 1 we then have
log |ψ(u)|
4T 2
η −1 − log |ϕ(u)| − hu, Σui − 0
6
|u|2β−2 ,
η (hu, Σui)
λmin
where λmin > 0 is the smallest eigenvalue of Σ.
A major consequence of this lemma for our purposes is that |u|−2 η −1 − log |ϕ(u)| =
−2+β
hu,Σui
|u|2
Φ
O(|u|
) as |u| → ∞. Thus, we can act as in Section 2. This leads to the estimator Σ̂
Φ
(σ̂i,j
)i,j=1,...p for Σ where
+
=
1 −1
η
− Re log ϕn (U u(i) ) ,
U2
Φ
Φ
σ̂i,i
+ σ̂j,j
1
Φ
σ̂i,j
:= 2 η −1 − Re log ϕn (U u(i,j) ) −
for i 6= j.
U
2
Applying an argument as in Lemma 10 together with the linearization for log ϕn , we can bound the
Φ
stochastic error of the estimators σ̂i,j
. We obtain the following proposition analogous to Theorem 1.
The proof is again postponed to the appendix.
√
Proposition 11. Let the assumptions ofpLemma 10 be satisfied. Let γ > 2 and suppose that
U > (22+β T 2 /λmin )1/(2−β) ∨ 1 satisfies 8γ (log(ep))/n < ∆Σ,U for
Φ
σ̂i,i
:=
∆Σ,U := min η 0 (U 2 hu(i,j) , Σu(i,j) i)|ϕ(U u(i,j) )|.
i,j
Set
12γ
τ (U ) = 2
U ∆Σ,U
r
log(ep)
+ 4(T + 1)U −2+β .
n
2
Then, for c∗ = 12e−γ ,
PΣ,ψ
2
Φ
max |σ̂i,j
− σi,j | < τ (U ) > 1 − c∗ p2−γ .
i,j=1,...,p
14
D. Belomestny, M. Trabs and A.B. Tsybakov
Under more specific assumptions on η it is possible to derive a uniform bound for ∆Σ,U . Since
|ϕ(u)| > exp(−c Re η(u> Σu)) for some constant c > 0, the stochastic error may not explode as fast
as for normal distributions resulting in possibly faster convergence rates depending on η. Relying
on Σ̂Φ , hard and soft thresholding estimators can be constructed with similar behaviour as for the
Gaussian case.
For the estimator Σ̂Φ , the function η is assumed to be known. It would be interesting to extend
the approach of this section to the case where η belongs to a parametric family introducing an
additional nuisance parameter.
6. Numerical example
In this section we numerically analyse the performance of the soft thresholding estimator for the
convolution model Y = X +ε, where X follows a p-dimensional normal distribution with zero mean
and covariance matrix Σ and ε is independent of X and has an elliptical distribution. Specifically,
we study the model
d √
ε = W AZ,
where Z ∼ N (0, Ip ) has a standard p-dimensional normal distribution, A is a p × p matrix and
W is a nonnegative random variable with a Laplace transform L. As can be easily seen, the
characteristic function of ε is given by
>
u AA> u
ψ(u) = E eihu,εi = L
.
2
Thus ε has indeed an elliptical distribution. We assume that W follows a Gamma distribution
with the density pW (x) = Γ(ϑ)−1 xϑ−1 e−x , x ≥ 0, for some ϑ > 0. Then we have
−ϑ
u> AA> u
ψ(u) = 1 +
.
2
Our aim is to compare several estimators of the covariance matrix Σ based on n independent
copies Y1 , . . . , Yn of Y . In the direct observations case where ε = 0 we may apply the sample
covariance matrix
n
1X
Σcov := Σ∗Y =
Yj Yj> .
(16)
n j=1
Adapting to sparsity in a high-dimensional framework, a soft thresholding estimator based on Σcov
is given by the solution of the optimisation problem, cf. Rothman et al. [46],
Σsτ := arg min |S − Σcov |22 + 2τ |S|1 ,
(17)
S∈Rp×p
with threshold parameter τ > 0. In some situations positive definiteness of the covariance matrix
estimate is desirable when the covariance estimator is, for example, applied to supervised learning
or if one needs to generate samples from the underlying normal distribution. In order to achieve
positive definiteness, Rothman [45] proposed to use the following modification of (17):
Σpds
:= arg min
|S − Σcov |22 + 2τ |S|1 − λ log |S| ,
(18)
τ
S∈Rp×p , S 0
where |S| denotes the determinant of the matrix S and λ is a fixed small number. The logarithmic
barrier term in (18) ensures the existence of a positive definite solution, since log |S| =
Pp
j=1 log(σj (S)), where σj (S) is the jth largest eigenvalue of S > 0. In order to solve (18), an
algorithm similar to the graphical lasso algorithm can be applied, see Friedman et al. [28].
15
0.1710
0.1695
0.1700
0.1705
Q(τ)
0.1715
0.1720
Covariance estimation in high-dimensional deconvolution
0.1
0.2
0.3
0.4
0.5
τ
Figure 1. The objective function Q100 (τ ) for the choice of the tuning parameter τ
Turning back to our deconvolution problem, we have already seen that the estimators (16), (17)
and (18) fail to deliver a consistent estimator for Σ unless ε is zero. Hence, we finally introduce
the positivity preserving version of the spectral soft thresholding estimator from (9):
Σsps
arg min
|S − Σ̂|22 + 2τ |S|1 − λ log |S| .
(19)
τ :=
S∈Rp×p , S 0
The tuning parameter τ can be chosen using a method introduced in [3]. The data is randomly
partitioned N times into a training set of size n1 and a validation set of size n2 with n2 =
bn/ log(n)c and n1 = n − n2 . The tuning parameter is then selected as τb = arg minτ QN (τ ), where
QN (τ ) =
N
X
1)
kΣsps,(m,n
− Σ̂(m,n2 ) k2 ,
τ
m=1
sps,(m,n )
1
where Στ
is the estimator, with penalty parameter τ, computed with the training set of
the mth split and Σ̂(m,n2 ) is the estimator (5) computed with the validation set of the mth split.
First, we consider a tridiagonal model where the population covariance matrix Σ has entries
σij = 0.4 · 1(|i − j| = 1) + 1(i = j), i, j ∈ {1, 2, . . . , p}. Using this covariance model with p = 20,
we generate n = 50 realizations of independent normal random vectors with mean zero and the
covariance matrix Σ. Adding an independent noise ε with the above elliptical distribution with
A = Id , depending on the parameter ϑ, we compute three estimates Σcov , Σpds
and Σsps
τ
τ . This
procedure was repeated 500 times. The parameters of the algorithms are τ = 0.25, λ = 10−4 ,
where the parameter τ is selected as a minimum of the function Q100 (τ ) shown in Figure 1.
The results are presented in Figure 2 for the case of direct observations and for three different
noise specifications corresponding to the values ϑ ∈ {0.5, 1, 2}. The used values of the tuning
parameter U are 1, 3, 3, respectively. While in the case of direct observations, the estimator Σsps
τ
has no advantages over Σcov and Σpds
τ , it significantly outperforms these two estimators in the
case of non-zero noise. We do not only observe a strong bias for Σcov and Σpds
in the presence of
τ
noise, but also a much better concentration of the spectral estimator Σsps
compared
to the other
τ
two procedures. The higher is the variance of the noise, the stronger are these bias and variance
effects.
16
D. Belomestny, M. Trabs and A.B. Tsybakov
ϑ = 0.5
●
0.40
0.18
Direct
0.35
0.16
●
●
●
●
●
●
●
●
0.12
0.25
0.14
0.30
●
●
cov
pds
0.20
●
0.15
0.10
●
●
●
●
●
●
●
●
sps
cov
sps
ϑ=2
0.30
0.5
0.35
0.6
0.40
0.7
0.45
0.8
0.50
ϑ=1
pds
●
0.15
0.2
0.20
0.3
0.25
0.4
●
●
cov
pds
sps
cov
pds
sps
Figure 2. Tridiagonal Σ : box plots of the estimation errors kΣoτ − Σk for o ∈ {cov, pds, sps} in the case of the
d √
convolution model Y = X + ε with ε = W Z, where Z ∼ N20 (0, I20 ) and W ∼ Gamma(ϑ).
Now, let us consider the case of normal noise. Note that this situation corresponds to β = 2 and
is not covered (at least formally) by our theoretical study. Specifically we generate samples from
the model Y = X + ε, where X follows a p-dimensional normal distribution with zero mean and
covariance matrix Σ and ε is independent of X and has also normal distribution with zero mean
and covariance matrix ρ2 I. We again consider tridiagonal model where the population covariance
matrix Σ has entries σij = 0.4 · 1(|i − j| = 1) + 1(i = j), i, j ∈ {1, 2, . . . , p}. In Figure 3 the
corresponding estimation errors for three methods are presented in the case of p = 20, τ = 0.4,
n = 50 and ρ ∈ {0.1, 0.5}. As one can see, even in the case of misspecified models the spectral
estimator continues to perform reasonably well.
Finally, we study the situation where the matrix Σ is block diagonal with the elliptical error
distribution from above. In particular, we generate positive definite matrix with randomly-signed,
non-zero elements. A shift is added to the diagonal of the matrix so that its condition number
equals p. Using this covariance model, we generated n = 100 realizations of independent 20dimensional normal random vectors with mean zero and covariance Σ. We then proceed as before
considering the case of direct observations and ϑ ∈ {0.5, 1, 2}. The tuning parameter U was taken
to be 3 for all three cases. The errors show a similar behaviour as in the first case, see Figure 4.
Covariance estimation in high-dimensional deconvolution
17
0.20
0.18
0.18
0.20
0.22
ρ = 0.5
0.22
ρ = 0.1
●
0.14
●
●
0.12
0.12
0.14
●
0.16
●
0.16
●
cov
pds
sps
cov
pds
sps
Figure 3. Tridiagonal Σ : box plots of the estimation errors kΣoτ − Σk for o ∈ {cov, pds, sps} in the case of the
d
convolution model Y = X + ε with ε = Z, where Z ∼ N20 (0, ρI20 ).
7. Proofs
7.1. Concentration of the spectral estimator
For the proof of Theorem 1, we need the following lemmas. Set S(u) = Re(log ϕn (u) − log ϕ(u)).
Lemma 12.
For any x ∈ (0, 1], and any u ∈ Rp such that ϕ(u) 6= 0,
x
P |S(u)| > x 6 3P |ϕn (u) − ϕ(u)| > |ϕ(u)| .
2
Proof. We have
S(u) = log
ϕ (u) − ϕ(u)
ϕn (u)
ϕn (u) − ϕ(u)
n
6 log
+1 6
.
ϕ(u)
ϕ(u)
ϕ(u)
n
Thus, P S(u) > x 6 P |ϕn (u) − ϕ(u)| > x|ϕ(u)| for all x > 0. Next, on the event |ϕn (u) −
o
ϕ(u)| 6 12 |ϕ(u)| we have
−S(u) = log
ϕ (u) − ϕ(u)
ϕ (u) − ϕ(u)
ϕ(u)
ϕn (u) − ϕ(u)
n
n
6 log
+1 6 log 2
+1 6 2
.
ϕn (u)
ϕn (u)
ϕ(u)
ϕ(u)
Therefore, for any x > 0,
1
P − S(u) > x 6 P 2|ϕn (u) − ϕ(u)| > x|ϕ(u)| + P |ϕn (u) − ϕ(u)| > |ϕ(u)| .
2
Since x ∈ (0, 1], we obtain P − S(u) > x 6 2P |ϕn (u) − ϕ(u)| > (x/2)|ϕ(u)| and hence the
lemma.
Lemma 13.
√
For any κ ∈ (0, n/8] we have
2
3κ
P |ϕn (u) − ϕ(u)| > √
6 4e−κ .
n
18
D. Belomestny, M. Trabs and A.B. Tsybakov
ϑ = 0.5
0.60
0.65
●
0.40
0.45
Direct
●
●
●
●
0.45
0.35
0.50
0.55
●
cov
pds
sps
cov
ϑ=1
pds
sps
ϑ=2
●
1.0
0.8
1.2
0.9
●
●
●
0.8
0.7
●
●
0.6
●
●
0.6
●
●
●
cov
pds
sps
cov
pds
sps
Figure 4. Block diagonal Σ : box plots of the estimation errors kΣoτ − Σk for o ∈ {cov, pds, sps} in the case of the
d √
convolution model Y = X + ε with ε = W Z, where Z ∼ N20 (0, I20 ) and W ∼ Gamma(ϑ).
Proof. We decompose ϕn − ϕ into real and imaginary part. Both can be estimated analogously,
such that we consider only the real part. We write
n
1X
Re ϕn (u) − ϕ(u) =
ξk (u) with
n
ξk (u) := Re eihu,Yk i − Re ϕ(u).
k=1
The independent and centred random variables ξk (u), k = 1, . . . , n, satisfy
|ξk (u)| 62 and Var(ξk (u)) 6 1 − |ϕ(u)|2 6 1.
√
Using the fact that κ ∈ (0, n/8] and then applying Bernstein’s inequality we find
√
2
3κ
2κ 2κ2
6 2e−κ .
P Re ϕn (u) − ϕ(u) > √ 6 P Re ϕn (u) − ϕ(u) > √ +
3n
2 n
n
Corollary 14.
p
For any γ > 0 and u ∈ Rp such that γ (log(ep))/n 6 |ϕ(u)|/8 we have
p
2
6γ log(ep)
P |S(u)| > √
6 12(ep)−γ .
n|ϕ(u)|
Covariance estimation in high-dimensional deconvolution
19
√
p
6γ log(ep)
Proof. We use Lemma 12 with x = √n|ϕ(u)| and then Lemma 13 with κ = γ log(ep). To apply
q
q
Lemma 12 we need 6γ lognep 6 |ϕ(u)|, while Lemma 13 requires 8γ lognep 6 1. Since |ϕ(u)| 6 1
both conditions are satisfied.
8. Proof of the lower bound: Theorem 6
Since C1 p 6 S 6 C2 p it is enough to assume that 2p 6 S (otherwise we consider a (C1 p/2)dimensional subspace). Furthermore, we will assume without loss of generality that S = p + 2k
for some integer k > 1 corresponding to p non-zero diagonal entries and 2k non-zero off-diagonal
entries of the covariance matrix. Note that under our assumptions, S, k and p are of the same
order up to constants:
S
S−p
S
C2 p
6k=
6 6
.
(20)
4
2
2
2
Let PΣ,ψ denote the distribution of Yj corresponding to the covariance matrix Σ ∈ Gq (S, R) and
to the error distribution with characteristic function ψ ∈ Hβ (T ). Set
1
ϕΣ,ψ (u) := EΣ,ψ [eihu,Yj i ] = exp − hu, Σui + log ψ(u) .
2
Applying Theorem 2.6 in [50], it is sufficient to construct a finite number of pairs (Σi , ψi ) with
Σ0 = RIp , ψ0 ∈ Hβ (T ) and (Σi , ψi ) ∈ Gq (p + 2k, R) × Hβ (T ) for i = 1, . . . , M , such that the
following two conditions hold:
−1+β/2
(i) kΣi − Σj k > CS 1/2 T R−1 log n
for all 0 6 i < j 6 M and some constant C > 0,
⊗n
⊗n
2
(ii) χ (PΣj ,ψj , PΣ0 ,ψ0 ) 6 M/3 for all j = 1, . . . , M .
Step 1: Constructing the pairs (Σi , ψi ). Without loss of generality, consider p that can be
decomposed as p = Lb where b and L are integers. For a block size b ∈ N and L = p/b ∈ N let
B ⊆ Rp×p denote the set of symmetric block diagonal matrices B = diag(A1 , . . . , AL ) satisfying:
• B = (bij ) has exactly k non-zero over-diagonal entries, all equal to 1;
• bii = 0 for i = 1, . . . , n;
• Al ∈ Rb×b for l = 1, . . . , L.
There are N := Lb(b − 1)/2 = p(b − 1)/2 positions over the diagonal of B where the entry 1
can possibly appear. Since k 6 C2 p/2, we have k < N for b > C2 + 1. In what follows, we
select b > C2 + 1, which is a fixed integer independent of k and p. Lemma A.3 in Rigollet and
Tsybakov [43] yields that there is a subset {B1 , . . . BM } ⊆ B such that for any i 6= j we have
kBi − Bj k2 > (k + 1)/4 and for some constants C10 , c01 > 0,
eN
c0 bp
log M > C10 k log 1 +
> C10 k log 1 + 1
.
(21)
4k
k
We consider now the following family of matrices
Σ0 = RIp ,
Σj = RIp +
where
δn,p = R1/2 6 log
ρT 2−β
δ
Bj ,
b n,p
j = 1, . . . , M,
−1/2
n
,
ρ0 log(1 + c01 bp/k)
(22)
and ρ, ρ0 > 0 are small enough constants to be chosen later. By construction and using (20) we
have
−(1−β/2)
T 2−β 1/2
T
n
kΣi − Σj k > ρ δn,p
k
> k 1/2 6R−1 log
0
c bp
2b
2b
ρ0 log 1 + 1k
−(1−β/2)
> c02 T S 1/2 R−1 log n
20
D. Belomestny, M. Trabs and A.B. Tsybakov
2−β
is
where c02 > 0 is a constant. Moreover, since by assumption of the theorem, R−1 T b−1 δn,p
uniformly bounded, the matrices Σi are diagonally dominant and thus positive semi-definite for
sufficiently small ρ. We conclude that the Σi thus defined are covariance matrices satisfying the
lower bound in (i) above.
We now turn to the construction of characteristic functions ψj . To have an as small as possible
L2 -distance between the characteristic functions, we choose ψj such that log ψj (u) − log ψ0 (u)
mimics hu, (Σj − Σ0 )ui/2 for small frequencies, keeping the block structure. In what follows, we
denote by F the Fourier transform operator. On each block of the matrix Bj = diag(Aj,1 , . . . , Aj,L ),
for j = 1, . . . , M, l = 1, . . . , L, we define
2−β
ρT δn,p
hu, Aj,l uiFK(δn,p u) + log ψ0,l (u),
ˆ 2b
T
dx,
log ψ0,l (u) :=
eihu,xi − ihu, xi1{β>1} − 1
ξb |x|β+b
b
R
u ∈ Rb ,
log ψj,l (u) :=
u ∈ Rb ,
where ξb > 0 is a constant depending only on b, and K ∈ L1 (Rb ) ∩ C 2 (Rb ) is a function satisfying
FK ∈ C ∞ (Rb ), and
FK(u) = 1 for |u| 6 1,
FK(u) = 0 for |u| > 2,
and 0 6 FK(u) 6 1
∀ u.
Writing ul := (ub(l−1)+1 , . . . , ubl ) for 1 6 l 6 L and u ∈ Rp , we then set
ψj (u) :=
L
Y
ψj,l (ul ),
j = 0, . . . , M.
l=1
Note that ψ0,l is the characteristic function of a b-dimensional symmetric stable distribution, cf.
Sato [48, Thm. 14.3]. To check that ψ0 ∈ Hβ (T ) is satisfied, we use Theorem 14.10 in [48], which
yields
log |ψ0 (u)| 6
L
X
log |ψ0,l (u)| 6
L
X
l=1
l=1
Cβ
T 2π b/2 l β
T 2π b/2
|u | 6 Cβ
|u|β ,
ξb Γ(b/2)
ξb Γ(b/2) β
where Cβ > 0 is a constant depending only on β and where
(b − 1)-dimensional sphere. Thus, choosing
ξb = c
2π b/2
Γ(b/2)
is the surface area of the
2π b/2
Γ(b/2)
(23)
for some sufficiently large c > 0 guarantees that ψ0 ∈ Hβ (T ). Note that ξb is bounded uniformly
in b.
We also have ψj ∈ Hβ (T ) for sufficiently small ρ since maximal singular value kAj,l k∞ 6 b and
L
L
2−β
X
ρT δn,p
ρkKkL1 T 2−β X
hul , Aj,l ul iFK(δn,p ul ) 6
δn,p
kAj,l k∞ |ul |2 1{|ul |62/δn,p }
2b
2b
l=1
l=1
6
ρkKkL1 T 2−β
δn,p
2b
L
X
l=1
b
2 2−β l β
|u |
δn,p
(24)
6 2ρkKkL1 T |u|ββ .
It remains to verify that ψj,l , l = 1, . . . , L are indeed characteristic functions. Denoting Aj,l =
(aj,l
k,m )k,m=1,...,b and
1 X j,l
νj,l :=
ak,m (∂k ∂m K),
2b
k,m
Covariance estimation in high-dimensional deconvolution
21
where ∂k stands for the derivative with respect to kth coordinate, we rewrite the characteristic
exponent as
b
2−β X
ρT δn,p
aj,l
k,m uk ul FK(δn,p u) + log ψ0 (u)
2b
k,l=1
h
i
−β−b
−1
= −F ρT δn,p
νj,l (δn,p
·) (u) + log ψ0 (u)
ˆ
T
−β−b
−
ρT
δ
ν
(x/δ
)
dx
=
eihu,xi − ihu, xi1{β>1} − 1
j,l
n,p
n,p
ξb |x|β+b
Rb
´
where
in the last line we have used the relations Rb νj,l (x)dx = Fνj,l (0) = 0 and, if β > 1,
´
ihu, xiνj,l (x)dx = hu, ∇(Fνj,l )(0)i = 0 for any u ∈ Rb . Consequently, ψj,l is the characteristic
Rb
−β−b
function of an infinitely divisible distribution with Lévy density T ξb−1 |x|−β−b −ρT δn,p
νj,l (x/δn,p )
provided that the latter is non-negative. To check this, it is enough to verify the equivalent condition ρξb νj,l (x) 6 |x|−β−b for all x ∈ Rb \ {0} and some sufficiently small ρ. We have
log ψj,l (u) =
k|x|β+b νj,l (x)k∞ 6 kνj,l k∞ + k|x|2d(β+b)/2e νj,l (x)k∞
6 kFνj,l kL1 + k∆d(β+b)/2e Fνj,l kL1
(25)
where ∆ denotes the Laplace operator, dxe is the minimal integer greater than x, and k·kLq stands
1
hu, Aj,l uiFK(u), and thus
for the Lq (Rb )-norm. By construction, Fνj,l (u) = 2b
kFνj,l kL1 6
kAj,l k∞
|u|2 FK(u)
2b
L1
6
1
|u|2 FK(u)
2
L1
where we have used the inequality kAj,l k∞ 6 b. Since the support of FK is compact the last
expression is bounded. The second term in (25) admits an analogous bound.
Step 2: Bounding the χ2 -divergence. Due to the block structure, for any pair (Σi , ψi ) we
have
L
Y
PΣi ,ψi =
Pi,l
l=1
for all i = 1, . . . , M , l = 1, . . . , L, where Pi,l is the convolution of the normal distribution N (0, RIb +
ρT 2−β
b
b δn,p Ai,l ) on R with a distribution given by the characteristic function ψi,l . We also denote by
P0 the convolution of N (0, RIb ) with the stable distribution given by ψ0,l . We have
n
⊗n
2
χ2 (P⊗n
−1
Σi ,ψi , PΣ0 ,ψ0 ) = 1 + χ (PΣi ,ψi , PΣ0 ,ψ0 )
=
L
Y
1 + χ2 (Pi,l , P0 )
n
− 1.
(26)
l=1
Thus, to check condition (ii) stated at the beginning of this subsection, we need to bound from
above the value
2
ˆ
fi,l (x) − f0 (x)
2
dx
(27)
χ (Pi,l , P0 ) =
f0 (x)
f0 (x)>0
where fi,l and f0 are the densities of Pi,l and P0 , respectively. To this end, we first establish a lower
bound for f0 , which is the density of the convolution of a normal distribution on Rb with zero mean
and covariance matrix RIb and a stable distribution on Rb . If there is no Gaussian component, we
write R = 0 referring to a convolution of the stable distribution with a Dirac measure in zero.
Lemma 15. In the special case of a standard stable density f0 (R = 0, T = 1) and β ∈ (0, 2)
we have f0 (x) > Cb (1 + |x|β+b )−1 for a constant Cb > 0 depending only on b. If R > 0 and T > 0
are such that T (log n)−c 6 CRβ/2 for some C, c > 0, we have the lower bound
f0 (x) > Cb0 R−b/2 (log n)−cb/β
1
(1 + T −1−b/β |x|b+β )
22
D. Belomestny, M. Trabs and A.B. Tsybakov
for another constant Cb0 > 0.
Proof. Step 1: We first consider the case R = 0, T = 1 and start with β ∈ (0, 1). We have
f0 = hc ∗ hf where
h
1 ˆ
i
1
−
1
dy (x),
hc (x) :=F −1 exp
eihu,yi − 1
ξb |y|61
|y|β+b
h
1 ˆ
i
1
hf (x) :=F −1 exp
eihu,yi − 1
dy
(x), x ∈ Rb ,
ξb Rb
|y|β+b ∨ 1
are the densities of an infinitely divisible distribution with Lévy density νc (x) := ξ1b ( |x|1β+b −
1
, x ∈ Rb ,
1)1{|x|61} and an infinitely divisible distribution with Lévy density νf (x) := ξb (|x|β+b
∨1)
respectively. Since νf is integrable, hf is the density of a compound Poisson distribution which
can be written as convolution exponential, cf. [48, Remark 27.3],
hf = e
−νf (Rb )
∞
X
νf∗j
j=0
(28)
j!
where νf∗j denotes the j-fold convolution of νf , and νf∗0 := δ0 is the Dirac measure in zero.
Therefore,
∗j
∞
b X hc ∗ νf
b
f0 = e−νf (R )
> e−νf (R ) hc ∗ νf ,
(29)
j!
j=0
where, with some abuse of notation, νf (Rb ) stands for the total mass of νf . Due to the compactness
of the support of the Lévy measure corresponding to hc , the density
´ hc admits a finite exponential
moment [48, Theorem 26.1], that is there exists α > 0 such that Rb eα|y| hc (y)dy < ∞.
For any x 6= 0, we have
ˆ
νf (x − y) − νf (x) hc (y)dy
|hc ∗ νf (x) − νf (x)| =
b
R
ˆ
ˆ
6
νf (x − y) − νf (x) hc (y)dy +
νf (x − y) − νf (x) hc (y)dy
|y|6|x|/2
|y|>|x|/2
Rewriting νf (x) = µ(|x|) with µ(r) := ξb−1 (r−β−b ∧ 1), we see that the expression in the last
display does not exceed
ˆ
ˆ
0
sup |µ (|v|)|
|v|6|x|
6
|y|hc (y)dy + 2kνf k∞ e
Rb
sup |µ0 (|v|)| + 2kνf k∞ e−α|x|/2
|v|6|x|
−α|x|/2
ˆ
eα|y| hc (y)dy
|y|>|x|/2
(|y| ∨ eα|y| )hc (y)dy.
(30)
Rb
By the polynomial decay of µ we have |µ0 (|x|)| = o(νf (x)) as |x| → ∞ implying that |hc ∗ νf (x) −
νf (x)| = νf (x)o(1) as |x| → ∞. Combining (29) and (30) yields
f0 (x) > e−νf (R
b
)
b
hc ∗ νf (x) = e−νf (R ) νf (x) 1 + o(1) >
C
,
|x|β+b
∀|x| > r,
(31)
where C > 0 and r > 0 are constants depending only on b.
From the representation (28) we see that f0 is strictly positive. By the decay of its characteristic
function, f0 is also continuous. Together with (31), we conclude that
f (x) > Cb (1 + |x|β+b )−1
for β ∈ (0, 1), R = 0, T = 1.
Covariance estimation in high-dimensional deconvolution
23
In the case β ∈ [1, 2), R = 0, T = 1 the proof
´ is analogous with
´ the only difference that the
convolution exponential hf is shifted by a := ( Rb x1 νf (x)dx, . . . , Rb xb νf (x)dx) ∈ Rb , i.e.,
b
hf = e−νf (R ) δa ∗
∞
X
ν ∗j
f
j=0
j!
where δa is the Dirac distribution at a. Thus, we replace everywhere above gb ∗ hc by gb ∗ hc ∗ δa .
Clearly, the argument remains valid with such a modification.
Step 2. We now denote by f the density f0 from Step 1 corresponding to R = 0, T = 1.
β
Thus, f is a density with characteristic function e−C|u| for some C > 0. With this notation, for
β
R = 0, T > 0 we have f0 (x) = F −1 [e−CT |u| ](x) = T −b/β f (T −1/β x). We now turn to the case
R > 0, T > 0. Denoting the density of the normal distribution N (0, Ib ) by g and using the lower
bound from Step 1, we obtain
f0 (x) = T −b/β f (T −1/β ·) ∗ R−b/2 g(R−1/2 ·) (x)
ˆ
2
T 2/β
Cb
e− 2R |y| dy
> (2πR)−b/2
−1/β
b+β
x − y|
Rb 1 + |T
ˆ
2/β
1
1
− T2R |y|2
−b/2
e
> Cb (2πR)
dy
(1 + 2b+β−1 T −1−b/β |x|b+β ) Rb 1 + 2β+b−1 |y|b+β
ˆ
T 2/β 2
rb−1
1
2π b/2 Cb
(2πR)−b/2
e− 2R r dr,
> b+β
b+β
−1−b/β
b+β
4
Γ(b/2)
(1 + T
|x|
) R+ 1 + r
where in the third line we have used the fact that b + β > 1 and the convexity of |x|b+β ). Using
the assumption (log n)−c 6 CRβ/2 T −1 , we deduce that with some constants C̄b , Cb0 > 0 depending
only on b,
ˆ
c 2/β 2
rb−1
1
−b/2
f0 (x) > C̄b R
e−(C(log n) ) r /2 dr
b+β
−1−b/β
b+β
(1 + T
|x|
) R+ 1 + r
ˆ
2
rb−1
1
> C̄b R−b/2 (C(log n)c )−b/β
e−r /2 dr
−1−b/β
b+β
(1 + T
|x|
) R+ 1 + (C(log n)c )−1−b/β rb+β
ˆ
2
rb−1 e−r /2
1
> Cb0 R−b/2 (log n)−cb/β
dr.
(1 + T −1−b/β |x|b+β ) R+ 1 + rb+β
Since the last integral is finite and positive we obtain the result of the lemma for R > 0, T > 0.
Due to Lemma 15 and the assumption T (log n)−1+β/2 6 C3 Rβ/2 , the χ2 -divergence (27) satisfies
ˆ
2
χ2 (Pj,l , P0 ) 6CRb/2 (log n)(1−β/2)b/β
fj,l (x) − f0 (x) dx
(32)
Rb
ˆ
2
1
+ 1+b/β
|x|β+b fj,l (x) − f0 (x) dx .
T
Rb
We now bound separately the first and the second integral in (32). Using Plancherel’s identity,
we rewrite the first integral as
ˆ
2
1
2
fj,l (x) − f0 (x) dx =
ϕj,l − ϕ0 L2 ,
(33)
b
(2π)
b
R
where ϕj,l and ϕ0 are the characteristic functions corresponding to fj,l and f0 , respectively. We
now consider the difference of the characteristic exponents
1
ηj (u) := log ϕj,l (u) − log ϕ0 (u) = − hu, Āj,l ui 1 − FK(δn,p u) ,
2
24
D. Belomestny, M. Trabs and A.B. Tsybakov
2−β
where Āj,l = (ρT δn,p
/b)Aj,l . A first order Taylor expansion yields
ˆ
1
etηj (u) dt
ϕj,l (u) − ϕ0 (u) = ηj (u)ϕ0 (u)
0
= ηj (u)e−R|u|
2
ˆ
/2
0
1
t
1−t
t
ψ0,l
(u)ψj,l
(u) exp − hu, Āj,l ui dt.
2
−1
Due to the property 1 − FK(δn,p u) = 0 for |u| 6 δn,p
and the elementary inequality x2 e|x| 6
exp(3|x|), ∀x ∈ R, we obtain
ˆ
kϕj,l −
ϕ0 k2L2
1
R
2
t
ηj (u) exp − |u|2 − hu, Āj,l ui
dt
2
2
L2
0
ˆ 1ˆ
1
6
|hu, Āj,l ui|2 exp − R|u|2 − thu, Āj,l ui du dt
4 0 |u|>1/δn,p
ˆ
1
6
exp − R|u|2 + 3|hu, Āj,l ui| du
4 |u|>1/δn,p
ˆ
1
2−β
6
exp − (R − 3ρT δn,p
)|u|2 du,
4 |u|>1/δn,p
6
where we have used the bound kAj,l k∞ 6 b. Finally, we choose ρ > 0 sufficiently small to satisfy
2−β
3ρT δn,p
6 R/4. Then,
kϕj,l −
ϕ0 k2L2
2
1
6 e−R/(4δn,p )
4
ˆ
e−R|u|
−1
|u|>δn,p
2
/2
du 6
2
1 2π b/2 −R/(4δn,p
)
e
.
4 R
(34)
To take into account the first factor in (32), we note that the definition of δn,p in (22) imply
2
Rb/2 (log n)(1−β/2)b/β R−b/2 e−R/(12δn,p ) 6 (log n)(1−β/2)b/β
ρ0 log(1 + c0 bp/k) 1/2
1
n
,
(35)
where the last expression is uniformly bounded by a constant. Combining this remark with (33)
and (34), we finally get that there is a constant C > 0 such that
ˆ
2
R
Rb/2 (log n)(1−β/2)b/β
fj,l (x) − f0 (x) dx 6 C exp − 2 .
(36)
6δn,p
Rb
To bound the second integral in (32) we use the following proposition proved in the supplementary material in Appendix C.
Proposition 16. There is a constant C > 0 depending only on the kernel K and on b such that,
for all β ∈ (0, 2) and j = 1, . . . , M , l = 1, . . . , L,
ˆ
2
R
(37)
ξb |x|β+b fj,l (x) − f0 (x) dx 6 C(1 ∨ Rβ/2 ) exp − 2 .
5δn,p
Rb
This proposition and the assumption T (log n)c∗ > 1 ∨ Rβ/2 yield, via an argument analogous
to (35), that
(1−β/2)b/β
(log n)
Rb/2
T 1+b/β
ˆ
2
|x|β+b fj,l (x) − f0 (x) dx
Rb
6 C 0 (log n)(1−β/2)b/β
Rb/2 ∨ R(b+β)/2
R
R
exp
−
6
C
exp
−
2
2
5δn,p
6δn,p
T (b+β)/β
(38)
Covariance estimation in high-dimensional deconvolution
25
where C, C 0 > 0 are constants. Combining (32), (36) and (38) and using the definition of δn,p in
(22), we find
ρ0 log(1 + c01 bp/k)
Cρ0 log M
ρ00 log M
R
6C
χ2 (Pj,l , P0 ) 6 C exp − 2
6
:=
,
(39)
6δn,p
n
C10 kn
kn
where the last inequality follows from (21). Taking into account (26) and (39) we get
pρ00 log M
00
⊗n
2
− 1 6 M 2ρ /b − 1, (40)
χ2 (P⊗n
Σj ,ψj , PΣ0 ,ψ0 ) 6 exp Ln max χ (Pj,l , P0 ) − 1 6 exp
l
bk
where we have used that, by construction, L = p/b and p 6 2k. Finally, choose ρ0 (and thus
00
ρ00 ) small enough to guarantee that M 2ρ /b − 1 6 M/3. Hence, condition (ii) is verified, which
concludes the proof of the theorem.
Appendix A: Proof of Theorem 9
For the later reference we set ξU := inf |u|6U/2 |ϕ(u)| and introduce the events
Ω(u) := |ϕn (u) − ϕ(u)| 6 |ϕ(u)|/2 ,
u ∈ Rp .
Due to the decomposition (2), we obtain the bound
kRn k∞
where Sn(1)
Sn(2)
Dn
S (1) + Sn(2) + Dn ,
(41)
ˆn
kΘ(u)k∞
wU (u)du,
:=
Re log ϕn (u)1{|ϕn (u)|>ι} − log ϕ(u) 1Ω(u)
|u|2
p
R
ˆ
kΘ(u)k∞
:=
log |ϕn (u)|1{|ϕn (u)|>ι} − log |ϕ(u)| 1Ω(u)c
wU (u)du,
|u|2
Rp
ˆ
kΘ(u)k∞
log |ψ(u)|
wU (u)du.
:=
|u|2
Rp
6
(i)
Here, Sn are stochastic error terms and Dn is a deterministic error term. Using the decay of
ψ ∈ Hβ0 (T ), the form of the support of wU and the fact that kΘ(u)k∞ = 1, we obtain
ˆ
w(v)
Dn 6 16U −2 sup log |ψ(u)|
dv 6 C(w)T U −(2−β)
(42)
2
|u|6U/2
Rp |v|
for a constant C(w) > 0 depending only on w.
√
(1)
To bound Sn in (41), we first note that we have on Ω(u) under the assumption ξU > 1/ n
1
|ϕn (u)| > |ϕ(u)| − |ϕn (n) − ϕ(u)| > |ϕ(u)|/2 > √ = ι,
2 n
for all u in the support of wU . Thus, the indicator function 1{|ϕn (u)|>ι} in Sn can be omitted.
Linearizing the logarithm yields
ϕ (u) − ϕ(u)
ϕ (u) − ϕ(u)
n
n
log ϕn (u) − log ϕ(u) = log
+1 =
+ rn (u)
ϕ(u)
ϕ(u)
(1)
with a residual rn satisfying on Ω(u)
|rn (u)| 6 c̄
ϕn (u) − ϕ(u)
ϕ(u)
2
(43)
where c̄ > 0 is a constant. Hence, we have
Sn(1)
6 Ln + Tn
(44)
26
D. Belomestny, M. Trabs and A.B. Tsybakov
where
ˆ
Ln :=
Rp
|ϕn (u) − ϕ(u)|
wU (u)du,
|u|2 |ϕ(u)|
ˆ
Tn :=
Rp
|rn (u)|
1Ω(u) wU (u)du.
|u|2
Here, Ln is the linearized stochastic error and Tn is a remainder. By the Cauchy-Schwarz inequality
ˆ
1/2
16
16κw
Ln 6 2
|ϕn (u) − ϕ(u)|wU (u)du 6 2
Z
(45)
U ξU Rp
U ξU
with κ̄w = kwkL1 and
Z = Z(Y1 , . . . , Yn ) =
ˆ
1/2
|ϕn (u) − ϕ(u)|2 wU (u)du
.
Rp
Similarly, we deduce from (43)
Tn 6
16c̄
U2
ˆ
Rp
|ϕn (u) − ϕ(u)|2
16c̄
wU (u)du 6 2 2 Z 2 .
2
|ϕ(u)|
U ξU
(46)
Note that Z satisfies the bounded difference condition
∀Yi , Yi0 ∈ Rp :
1/2
|Z(Y1 , . . . , Yi−1 , Yi0 , Yi+1 , . . . , Yn ) − Z(Y1 , . . . , Yn )| 6 2κ̄w
/n
2
By the bounded difference inequality [29, Theorem 3.3.14] we get P(Z > E(Z) + t) 6 exp(− 4nt
κ̄w ),
1/2
for all t > 0. Since E(Z) 6 (κ̄w /n)
this implies
1/2
2
κ̄w
P Z > √ (2γ + 1) 6 e−γ ,
n
∀γ > 0.
Using (45),(46), and the assumption that γ > 1 we find that there exists a numerical constant
c∗1 > 0 such that
2
c∗ κ̄w γ
γ
P Sn(1) > 21 √ 1 + √
6 2e−γ .
U ξU n
ξU n
Since ξU 6 1, this implies
2
2c∗ κ̄w γ 2
P Sn(1) > 21 2 √
6 2e−γ .
(47)
U ξU n
Using lower bounds ι and ξU for |ϕn (u)| and |ϕ(u)|, respectively, and applying the elementary
(2)
bound 1{a>1} < a for any a > 0, the term Sn in (41) is bounded as follows:
ˆ
wU (u)
Sn(2) 6
log ι−1 + log ξU−1 1Ω(u)c
du
|u|2
p
ˆR
|ϕn (u) − ϕ(u)| wU (u)
62
log ι−1 + log ξU−1
du
|ϕ(u)|
|u|2
Rp
√
ˆ
32(2 log ξU−1 + log 2)
32 κ̄w
6
|ϕ
(u)
−
ϕ(u)|w
(u)du
6
Z.
n
U
U 2 ξU
U 2 ξU2
Rp
Hence, for some numerical constant c∗2 > 0 we have
2
c∗ κ̄w
P Sn(2) > 22 2 √ γ 6 e−γ ,
U ξU n
∀γ > 0.
(48)
Combining (41), (42), (47) and (48) and using the fact that γ > 1 we obtain
2
(2c∗1 + c∗2 )κ̄w γ 2
−2+β
√
P kRn k∞ >
+
C(w)T
U
6 3e−γ .
2
2
U ξU n
Finally, we use the bound ξU > exp(−kΣk∞ U 2 /8−2T U β ) that is shown similarly to the analogous
bound in the proof of Theorem 1.
Covariance estimation in high-dimensional deconvolution
27
Appendix B: Proofs for Section 5.3
Proof of Lemma 10. By Taylor’s formula we have for some ξ ∈ [0, 1] that
log |ψ|(u)
R(u) :=η −1 − log |ϕ|(u) − hu, Σui − 0
η (hu, Σui)
(log |ψ|(u))2 −1 00
=
(η ) η hu, Σui − ξ log |ψ|(u)
2
Since (η −1 )00 (x) = −η 00 (η −1 (x))/η 0 (η −1 (x))3 and thus |(η −1 )00 (x)| 6 T |η −1 (x)|−1 |η 0 (η −1 (x))|−2 we
have
|R(u)| 6
T | log |ψ|(u)|2
− ξ log |ψ|(u)))|2
2|g(u, ξ)||η 0 (η −1 (η(hu, Σui
(49)
with g(u, ξ) := η −1 (η(hu, Σui − ξ log |ψ|(u)). Since η −1 is non-negative and monotone increasing
and log |ψ(u)| 6 0, we have
|g(u, ξ)| = g(u, ξ) > η −1 η(hu, Σui) = hu, Σui.
Another Taylor expansion for the second term in the denominator in (49) yields for some ξ 0 ∈ [0, 1]
η 0 (η −1 (η hu, Σui − ξ log |ψ|(u))
= η 0 (hu, Σui) − ξ(log |ψ(u)|) (η 0 ◦ η −1 )0 (η(hu, Σui − ξ 0 log |ψ|(u))
η 00 ◦ η −1
(η(hu, Σui − ξ 0 log |ψ|(u))
= η 0 (hu, Σui) + ξ(log |ψ(u)|)
0
−1
η ◦η
> η 0 (hu, Σui) − T | log |ψ|(u)|/g(u, ξ 0 )
(50)
0
2
β
> η (hu, Σui) 1 − T (1 + |u|) /hu, Σui .
If |u| > (2β+1 T 2 /λmin )1/(2−β) , then we conclude
|R(u)| 6
2| log |ψ|(u)|2
4T 2
6
|u|2β−2 .
0
2
hu, Σuiη (hu, Σui)
λmin
Proof of Proposition 11. Due to Lemma 10 and the mean value theorem, the estimation error
can be bounded for any U > (2β+1 T 2 /λmin )1/(2−β) by
4T 2 −2+β −2+β
Φ
|σ̂i,i
− σi,i | 6 U −2 η −1 − log |ϕn (U u(i) )| − η −1 − log |ϕ(U u(i) )| + 2T +
U
U
λmin
|S(U u(i) )|
= 2 0 −1
+ 2T + 2 U −2+β
U η (η (− log |ϕ(U u(i) )| + ξS(U u(i) )))
for some ξ ∈ [0, 1] and S(u) = log |ϕn (u)| − log |ϕ(u)| from Lemma 12. As in (50) (for any u with
g(u, ξ) > hu, Σui > 1), we deduce
η 0 η −1 (− log |ϕ(u)| + ξS(u)) > η 0 (hu, Σui) − T (| log |ψ|(u)| + S(u)|.
On the event {| log |ψ|(u)| + S(u)| 6 η 0 (U 2 σii )/2}, we thus obtain
Φ
|σ̂i,i
− σi,i | 6
2|S(U u(i) )|
+ 2T + 2 U −2+β .
2
0
2
U η (U σii )
From this line, the argument is analogous to the proof of Theorem 1.
28
D. Belomestny, M. Trabs and A.B. Tsybakov
Appendix C: Supplementary material. Proof of
Proposition 16
Our aim is to prove the bound (37). We fix l, and we will omit for brevity the index l throughout
the proof. This will cause a slight change of notation as compared to the proof of Theorem 6 since,
in what follows, ψj := ψj,l , while ψj was a product of ψj,l ’s in the proof of Theorem 6. In addition,
we will use the notation Sj = RIb + Āj , S0 = RIb , j = 1, . . . , M . Throughout C > 0 will denote a
generic constant whose value may change from line to line.
By construction, the characteristic functions ϕj (u) and ϕ0 (u) coincide on {|u| 6 1/δn,p }. For
2−β
sufficiently small ρ we have hu, Āj ui 6 ρT δn,p
|u|2 6 R|u|2 /2 implying hu, Sj ui = R|u|2 +
2
hu, Āj ui > R|u| /2. Therefore, we can write
ϕj (u) − ϕ0 (u)
= ϕSj (u)ψj (u)1{|S 1/2 u|>(R/2)1/2 δ−1 } − ϕS0 (u)ψ0 (u)1{|S 1/2 u|>(R/2)1/2 δ−1 }
n,p
j
n,p
0
1{|u|>1/δn,p }
with ϕSj denoting the characteristic function of N (0, Sj ). We have
ˆ
2
ξb
|x|β+b fj (x) − f0 (x) dx 6 2ξb (Tj + T0 )
Rb
where (also for j = 0)
ˆ
2
Tj =
|x|β+b F −1 ϕSj 1{|S 1/2 ·|>(R/2)1/2 δ−1 } ∗ F −1 ψj 1{|·|>1/δn,p } (x)dx
n,p
j
Rb
F −1 ϕSj 1{|S 1/2 ·|>(R/2)1/2 δ−1 } ∗ F −1 ψj 1{|·|>1/δn,p }
= |x|(β+b)/2
n,p
j
6 2β+b−1 |x|(β+b)/2 F −1 ϕSj 1{|S 1/2 ·|>(R/2)1/2 δ−1 }
j
+ F −1 ϕSj 1{|S 1/2 ·|>(R/2)1/2 δ−1 }
n,p
j
2
L1
2
L2
2
−1
F
ψ
j 1{|·|>1/δn,p }
n,p
L1
L2
2
|x|(β+b)/2 F −1 ψj 1{|·|>1/δn,p }
,
2
L2
using Young’s inequality in the last estimate. Let us simplify this upper bound. We have the
identity
−1/2
1
−1
F −1 ϕIb 1{|u|>(R/2)1/2 δn,p
x)
F −1 ϕSj (u)1{|S 1/2 u|>(R/2)1/2 δ−1 } (x) = p
} (Sj
n,p
j
det Sj
such that
|x|(β+b)/2 F −1 ϕSj 1{|S 1/2 ·|>(R/2)1/2 δ−1 }
n,p
j
1
1/2
|Sj x|(β+b)/2 F
=p
det Sj
(β+b)/2
6
and
kSj k∞
p
det Sj
−1
2
L2
−1
ϕIb 1{|·|>(R/2)1/2 δn,p
}
−1
|x|(β+b)/2 F −1 ϕIb 1{|·|>(R/2)1/2 δn,p
}
2
L2
2
L2
2
−1 ]k 1 .
kF −1 [ϕSj 1{|S 1/2 ·|>(R/2)1/2 δ−1 } ]k2L1 = kF −1 [ϕIb 1{|·|>(R/2)1/2 δn,p
} L
j
n,p
Since ρδn,p is small, the construction of Sj as pertubation of RIb implies that kSj k∞ is of the
(β+b)/2
order R and det Sj is of the order Rb . Hence, (det Sj )−1/2 kSj k∞
6 CRβ/2 . We deduce
2
2
−1
−1
F −1 ψj 1{|·|>δn,p
Tj 6 C2β+b Rβ/2 |x|(β+b)/2 F −1 ϕIb 1{|·|>(R/2)1/2 δn,p
}
}
2
L
L1
2
2
(β+b)/2 −1
−1
+ F −1 ϕIb 1{|·|>(R/2)1/2 δn,p
|x|
F
ψ
1
.
j {|·|>1/δn,p }
}
1
2
L
L
Covariance estimation in high-dimensional deconvolution
29
By the Cauchy-Schwarz inequality we further estimate
2
−1 ]k 1
kF −1 [ϕIb 1{|·|>(R/2)1/2 δn,p
} L
2
(1
L2
2
|x|(b+β)/2 )−1 L2
6 (1 + |x|(b+β)/2 )−1
−1 ]
+ |x|(b+β)/2 )F −1 [ϕIb 1{|·|>(R/2)1/2 δn,p
}
2
L2
6 2 (1 +
2
(b+β)/2 −1
2
−1 ]k 2 + k|x|
−1 ]k 2 .
× kF −1 [ϕIb 1{|·|>(R/2)1/2 δn,p
F
[ϕ
1
1/2
I
b
L
L
}
{|·|>(R/2)
δn,p }
2
−1
An analogous estimate can be applied to kF −1 ψj 1{|·|>δn,p
} kL1 . By a polar coordinate transformation we have
ˆ ∞
ˆ ∞
−2
2π b/2
rb−1
2π b/2
2
(1 + |x|(b+β)/2 )−1 L2 =
dr
6
1{r<1} + r(1+β)/2 dr
(b+β)/2
2
Γ(b/2) 0 (1 + r
Γ(b/2) 0
)
b/2
2π
Recalling the definition of ξb = c Γ(b/2)
, we conclude with Plancherel’s identity and the Laplace
operator ∆
ˆ
2
ξb
|x|β+b fj (x) − f0 (x) dx
Rb
2
2
β+b
β/2
(β+b)/2 −1
−1 ]
−1
6C2
ξb F −1 [ϕIb 1{|·|>(R/2)1/2 δn,p
+
(1
∨
R
)
|x|
F
[ϕ
1
]
1/2
I
b {|·|>(R/2)
} L2
δn,p } L2
X
2
2
−1
(β+b)/2 −1
×
ξb F
ψk 1{|·|>1/δn,p } L2 + |x|
F
ψk 1{|·|>1/δn,p } L2
k∈{0,j}
ˆ
ˆ
2
Cξb 2β+b
2
β/2
d(β+b)/4e
ϕ
6
(u)
du
+
(1
∨
R
)
∆
ϕ
(u)
du
I
I
b
b
−1
−1
(2π)b
|u|>(R/2)1/2 δn,p
|u|>(R/2)1/2 δn,p
ˆ
ˆ
X
ξb
2
2
|ψ
(u)|
du
+
∆d(β+b)/4e ψk (u) du .
(51)
×
k
b
(2π)
Rb
Rb
k∈{0,j}
To bound these integrals, we apply the following proposition concerning tail integrals for stable
distributions.
Proposition 17.
on α such that
ˆ
Let α, β ∈ (0, 2] and t > 0. Then there is a constant Cα > 0 depending only
2
α
∆d(b+β)/4e exp(−|x|α /α) dx 6 Cα Γ(b/2)(13π + 25πα)b/2 b8 e−t
/α
.
|x|>t
Proof. Set q = d(b + β)/4e. Verified
p by induction, we have for any function g : R → R and for
the radius r : Rb → R, x 7→ |x| = x21 + . . . + x2b the following identity for the Laplace operator
applied to the radial function g ◦ r
2q−k
q
X
1
1 ∂
k 2q
∆ g(r) =
∆ r
·
g(r).
2k k!
r ∂r
q
k=0
Since ∆(x21 + . . . + x2b )k = γb,k (x21 + . . . + x2b )k−1 with γb,k := 2k(b + 2(k − 1)), we get
k−1
Y
∆k r2q =
γb,q−j r2(q−k)
j=0
30
D. Belomestny, M. Trabs and A.B. Tsybakov
and as a result
γb,q−j r2(q−k) 1 ∂ 2q−k
∆q g(r) =
·
g(r)
2k k!
r ∂r
k=0
2q−k
q k−1
X
q Y
1 ∂
2(q−k)
=
g(r).
(b + 2(q − j − 1)) r
·
k
r ∂r
j=0
q
X
Q
k−1
j=0
k=0
Since b 6 4q, we can estimate for all k 6 q
k−1
Y
k
Y
(b + 2(q − j − 1)) = 2k
j=0
(b/2 + q − j) 6 2k
j=1
k
Y
(3q − j)
j=1
= 2k
k
Y
1+
j=1
k
q Y
(2q − j)
2q − j
j=1
k
6
and thus
4 (2q)!
(2q − k)!
2q−k
q k
X
q 4 (2q)! 2(q−k)
1 ∂
∆ g(r) 6
r
·
g(r)
k (2q − k)!
r ∂r
q
(52)
k=0
Let g(r) = exp(−rα /α) for α ∈ (0, 2). For any m ∈ N we define a polynomial Pm via
Pm (r)g(r) :=
1 ∂ m
g(r).
r ∂r
It is easy to see that
Pk (r) =
k
X
(−1)l λk,l rlα−2k ,
l=1
with coefficients
λk,1 = −
k−1
Y
(α − 2j),
λk,k = 1
j=1
λk,l = (lα − 2(k − 1))λk−1,l + λk−1,l−1 ,
for l = 2, . . . , k − 1.
From this recursion formula, we obtain the upper bound
|λk,l | 6 (1 + A−1 )k Al
k−l
Y
(2j + l(2 − α) − 2)
(53)
j=1
for all k > 1, l = 1, . . . , k and any A > 0. Indeed, (53) is satisfied for λk,1 , λk,k for all k and we
check inductively, given (53) holds true for λk−1,l ,
|λk,l | 6 |2(k − l) + l(2 − α) − 2| |λk−1,l | + |λk−1,l−1 |
6 (1 + A−1 )k−1
k−l
Y
(2j + l(2 − α) − 2) Al + Al−1
j=1
6 (1 + A−1 )k Al
k−l
Y
j=1
(2j + l(2 − α) − 2).
Covariance estimation in high-dimensional deconvolution
31
From (53), we obtain the upper bound
m
m−l
1 ∂ m
X
Y
α
Al
(2j + l(2 − α) − 2)rlα−2m
g(r) 6 e−r /α (1 + A−1 )m
r ∂r
j=1
6 e−r
α
/α
(2 + 2A
l=1
m
X
−1 m
)
l=1
m
A l Y
(j − lα/2)rlα−2m
2
j=l+1
m
X
α
m! A l lα−2m
6 e−r /α (2 + 2A−1 )m
r
.
l! 2
l=1
Therefore, we get from (52)
q
X
q
2q−k
X (2q − k)! A l
(2q)!
k 2(q−k) −r α /α
−1 2q−k
|∆ g(r)| ≤
4 r
e
(2 + 2A )
rlα−2(2q−k)
k (2q − k)!
l!
2
l=1
k=0
q 2q−k
lα−2q
X
X
α
q k
A l r
= e−r /α (2q)!
4 (2 + 2A−1 )2q−k
.
k
2
l!
k=0 l=1 |
{z
}
q
=:γ(k,l)
Using b 6 4q, we find
ˆ
ˆ ∞
2π b/2
2
|∆q g(|x|)|2 dr =
rb−1 |∆q g(r)| dr
Γ(b/2) t
|x|>t
ˆ
q 2q−k q 2q−k
2π b/2 (2q)!2 X X X X γ(k, l)γ(k 0 , l0 ) ∞ p−1−4q+(l+l0 )α −2rα /α
6
r
e
dr
Γ(b/2)
l! l0 !
t
0
0
0
k=0 l=1 k =0 l =1
ˆ
q 2q−k q 2q−k
2π (2q)! X X X X γ(k, l)γ(k 0 , l0 ) ∞ (l+l0 )α−1 −2rα /α
r
e
dr.
6
Γ(b/2)
l! l0 !
t
0
0
b/2
0
2
k=0 l=1 k =0 l =1
To bound the tail integrals, we apply the following: Substituting s = rα /α, we obtain for any
m>0
ˆ ∞
ˆ ∞
α
α
α
rmα−1 e−2r /α dr 6 e−t /α
rmα−1 e−r /α dr
t
0
ˆ ∞
α
−tα /α m−1
=e
α
sm−1 e−s ds = e−t /α αm−1 Γ(m).
(54)
0
Γ(l+l0 )
l! l0 !
Together with
6
l+l0
l
ˆ
q
2
|∆ g(|x|)| dr 6 e
0
6 2l+l , we deduce from (54)
0
q 2q−k q 2q−k
0
0 Γ(l + l )
(2q)!2 X X X X
γ(k, l)γ(k 0 , l0 )αl+l
.
Γ(b/2)
l! l0 !
0
0
−tα /α 2π
|x|>t
b/2
k=0 l=1 k =0 l =1
6e
0
q 2q−k q 2q−k
0
(2q)! X X X X
γ(k, l)γ(k 0 , l0 )(2α)l+l
Γ(b/2)
0
0
−tα /α 2π
b/2
2
/α 2π
b/2
q 2q−k
2X
X
k=0 l=1 k =0 l =1
α
= e−t
(2q)!
Γ(b/2)
γ(k, l)(2α)l
2
.
k=0 l=1
The sum in the last bound is given by
q 2q−k
X
X
k=0 l=1
l
γ(k, l)(2α) =
q
X
q
k=0
k
4k (2 + 2A−1 )2q−k
2q−k
X
l=1
Aα
l
32
D. Belomestny, M. Trabs and A.B. Tsybakov
where for all A > α−1
2q−k
X
Aα
l
=
l=1
(Aα)2q−k − 1
(Aα)2q−k
6
.
1 − 1/(Aα)
1 − 1/(Aα)
Hence,
ˆ
|∆q g(|x|)|2 dr 6
|x|>t
α
q
2q−k 2
2e−t /α π b/2 (2q)!2 X q k
4 (2 + 2A−1 )2q−k Aα
α − 1/A Γ(b/2)
k
k=0
−tα /α
=
b/2
2
2q
π (2q)!
2e
(2Aα + 2α)2q 4 + 2(A + 1)α
α − 1/A Γ(b/2)
Finally, we apply (2q)! = (b/2 + 3)! 6 (b/2 + 3)4 Γ(b/2), due to q 6 b/4 + 3/2, to conclude
ˆ
b/2 8 −tα /α
b e
|∆q g(|x|)|2 dr 6 Cα,A Γ(b/2)π b/2 (2Aα + 2α)b/2 4 + 2Aα + 2α
|x|>t
for some constant Cα,A depending only on α and A. Since α 6 2 there there is some A > α−1
such that (Aα + α)(4 + Aα + α) 6 13 + 25α we obtain the asserted upper bound
ξb
(2π)b
=
nents in (51) and using
ˆ
2
ξb
|x|β+b fj (x) − f0 (x) dx
´
2
2
e−|u| du 6 ξb 2b/2−1 Γ(b/2)e−t /2
|u|>t
c
from (23), we obtain
2b π b/2 Γ(b/2)
Applying Proposition 17 and
to the Gaussian compo-
Rb
−2
Cξb 2b
(1 ∨ Rβ/2 )Γ(b/2)(63π)b/2 b8 e−Rδn,p /4
b
(2π)
ˆ
X Cξb ˆ
2
2
d(β+b)/4e
|ψ
(u)|
du
+
∆
ψ
(u)
du
×
k
k
(2π)b Rb
Rb
k∈{0,j}
ˆ
ˆ
X
2
−2
C
2
d(β+b)/4e
du
+
∆
ψ
(u)
du .
6 C8b (1 ∨ Rβ/2 )e−Rδn,p /4
ψ
(u)
k
k
2b π b/2 Γ(b/2) Rb
Rb
6
k∈{0,j}
We now use that ψj (u) coincides with ψ0 (u) for |u| > 2/δn,p and thus
ˆ
ˆ
2
2
d(β+b)/4e
∆
ψj (u) du +
∆d(β+b)/4e ψ0 (u) du
b
b
R
ˆR
ˆ
2
2
2
d(β+b)/4e
62
∆
ψ0 (u) du +
∆d(β+b)/4e ψj (u) + ∆d(β+b)/4e ψ0 (u) du.
|u|>2/δn,p
|u|62/δn,p
Since ψ0 is the characteristic function of a stable distribution, the first integral can be estimated by
Proposition 17. The multiplicative perturbation of ψ0 in the second integral is a smooth function
which can be uniformly bounded by Db for some suitable constant D > 0. An analogous argument
applies to the L2 -norms of ψk . Since the ball {u ∈ Rb : |u| 6 2/δn,p } has Lebesgue measure
2π b/2
(2/δn,p )b bΓ(b/2)
, we get
ˆ
2
ξb
|x|β+b fj (x) − f0 (x) dx
Rb
b/2
c
b b/2
b
b 2π
Γ(b/2)8
π
+
D
(2/δ
)
n,p
bΓ(b/2)
2b π b/2 Γ(b/2)
b −b
2CD
δ
−2
n,p
6 C8b (1 ∨ Rβ/2 )e−Rδn,p /4 1 + C4b +
.
T bΓ(b/2)2
−2
6 C8b (1 ∨ Rβ/2 )e−Rδn,p /4 1 +
The last term is uniformly bounded in b thus that we finally have
ˆ
2
−2
ξb
|x|β+b fj (x) − f0 (x) dx 6 C32b (1 ∨ Rβ/2 )e−Rδn,p /5 .
Rb
Covariance estimation in high-dimensional deconvolution
33
Acknowledgements
D. Belomestny acknowledges the financial support from the Russian Academic Excellence Project
“5-100” and from Deutsche Forschungsgemeinschaft (DFG) through the SFB 823 “Statistical modelling of nonlinear dynamic processes”. M. Trabs gratefully acknowledges the financial support by
the DFG research fellowship TR 1349/1-1. The work of A.B. Tsybakov was supported by GENES
and by the French National Research Agency (ANR) under the grants IPANEMA (ANR-13-BSH10004-02) and Labex Ecodec (ANR-11-LABEX-0047). This work has been started while M.T. was
affiliated to the Université Paris-Dauphine.
References
[1] Belomestny, D. and Reiß, M. (2006). Spectral calibration of exponential Lévy models. Finance
Stoch., 10(4):449–474.
[2] Belomestny, D. and Trabs, M. (2017). Low-rank diffusion matrix estimation for highdimensional time-changed Lévy processes. Ann. Inst. Henri Poincaré Probab. Stat. To appear.
ArXiv preprint arXiv:1510.04638.
[3] Bickel, P. J. and Levina, E. (2008a). Covariance regularization by thresholding. Ann. Statist.,
36(6):2577–2604.
[4] Bickel, P. J. and Levina, E. (2008b). Regularized estimation of large covariance matrices. Ann.
Statist., 36(1):199–227.
[5] Butucea, C. and Matias, C. (2005). Minimax estimation of the noise level and of the deconvolution density in a semiparametric convolution model. Bernoulli, 11:309–340.
[6] Butucea, C., Matias, C., and Pouet, C. (2008). Adaptivity in convolution models with partially
known noise distribution. Electron. J. Stat., 2:897–915.
[7] Butucea, C. and Tsybakov, A. B. (2008). Sharp optimality in density deconvolution with
dominating bias. i. Theory Probab. Appl., 52(1):24–39.
[8] Cai, T. and Liu, W. (2011). Adaptive thresholding for sparse covariance matrix estimation. J.
Amer. Statist. Assoc., 106(494):672–684.
[9] Cai, T. T., Ren, Z., and Zhou, H. H. (2016). Estimating structured high-dimensional covariance
and precision matrices: optimal rates and adaptive estimation. Electron. J. Stat., 10(1):1–59.
[10] Cai, T. T. and Zhang, A. (2016). Minimax rate-optimal estimation of high-dimensional
covariance matrices with incomplete data. arXiv preprint arXiv:1605.04358.
[11] Cai, T. T., Zhang, C.-H., and Zhou, H. H. (2010). Optimal rates of convergence for covariance
matrix estimation. Ann. Statist., 38(4):2118–2144.
[12] Cai, T. T. and Zhou, H. H. (2012). Minimax estimation of large covariance matrices under
`1 -norm. Statist. Sinica, pages 1319–1349.
[13] Carroll, R. J. and Hall, P. (1988). Optimal rates of convergence for deconvolving a density.
J. Amer. Statist. Assoc., 83(404):1184–1186.
[14] Comte, F. and Lacour, C. (2011). Data-driven density estimation in the presence of additive
noise with unknown distribution. J. R. Stat. Soc. Ser. B Stat. Methodol., 73(4):601–627.
[15] Cressie, N. and Wikle, C. K. (2011). Statistics for Spatio-Temporal Data. John Wiley & Sons.
[16] Dattner, I., Reiß, M., and Trabs, M. (2016). Adaptive quantile estimation in deconvolution
with unknown error distribution. Bernoulli, 22(1):143–192.
[17] Delaigle, A. and Hall, P. (2016). Methodology for non-parametric deconvolution when the
error distribution is unknown. J. R. Stat. Soc. Ser. B. Stat. Methodol., 78(1):231–252.
[18] Delaigle, A., Hall, P., and Meister, A. (2008). On deconvolution with repeated measurements.
Ann. Statist., 36(2):665–685.
[19] Delaigle, A. and Meister, A. (2011). Nonparametric function estimation under Fourieroscillating noise. Statist. Sinica, 21(3):1065–1092.
[20] Eckle, K., Bissantz, N., and Dette, H. (2016). Multiscale inference for multivariate deconvolution. arXiv preprint arXiv:1611.05201.
34
D. Belomestny, M. Trabs and A.B. Tsybakov
[21] El Karoui, N. (2008). Operator norm consistent estimation of large-dimensional sparse covariance matrices. Ann. Statist., 36(6):2717–2756.
[22] Fan, J. (1991). On the Optimal Rates of Convergence for Nonparametric Deconvolution
Problems. Ann. Statist., 19(3):1257–1272.
[23] Fan, J., Li, Y., and Yu, K. (2012). Vast volatility matrix estimation using high-frequency
data for portfolio selection. J. Amer. Statist. Assoc., 107(497):412–428.
[24] Fan, J., Liao, Y., and Liu, H. (2016). An overview of the estimation of large covariance and
precision matrices. Econom. J., 19(1):C1–C32.
[25] Fan, J., Liao, Y., and Mincheva, M. (2011). High-dimensional covariance matrix estimation
in approximate factor models. Ann. Statist., 39(6):3320–3356.
[26] Fan, J., Liao, Y., and Mincheva, M. (2013). Large covariance estimation by thresholding
principal orthogonal complements. J. R. Stat. Soc. Ser. B. Stat. Methodol., 75(4):603–680.
With 33 discussions by 57 authors and a reply by Fan, Liao and Mincheva.
[27] Fang, K., Kotz, S., and Ng, K. W. (1990). Symmetric multivariate and related distributions,
volume 36. Chapman & Hall/CRC.
[28] Friedman, J., Hastie, T., and Tibshirani, R. (2008). Sparse inverse covariance estimation with
the graphical lasso. Biostatistics, 9(3):432–441.
[29] Gine, E. and Nickl, R. (2016). Mathematical Foundations of Infinite-Dimensional Statistical
Models. Cambridge Univ. Press, Cambridge.
[30] Jacod, J. and Reiß, M. (2014). A remark on the rates of convergence for integrated volatility
estimation in the presence of jumps. Ann. Statist., 42(3):1131–1144.
[31] Johannes, J. (2009). Deconvolution with unknown error distribution. Ann. Statist.,
37(5A):2301–2323.
[32] Kappus, J. and Mabon, G. (2014). Adaptive density estimation in deconvolution problems
with unknown error distribution. Electron. J. Stat., 8(2):2879–2904.
[33] Koltchinskii, V., Lounici, K., and Tsybakov, A. B. (2011). Nuclear-norm penalization and
optimal rates for noisy low-rank matrix completion. Ann. Statist., 39(5):2302–2329.
[34] Lam, C. and Fan, J. (2009). Sparsistency and rates of convergence in large covariance matrix
estimation. Ann. Statist., 37(6B):4254–4278.
[35] Lepski, O. and Willer, T. (2017a). Estimation in the convolution structure density model.
Part I: oracle inequalities. arXiv preprint arXiv:1704.04418.
[36] Lepski, O. and Willer, T. (2017b). Estimation in the convolution structure density model.
Part II: adaptation over the scale of anisotropic classes. arXiv preprint arXiv:1704.04420.
[37] Lounici, K. (2014). High-dimensional covariance matrix estimation with missing observations.
Bernoulli, 20(3):1029–1058.
[38] Low, M. G. (1997). On nonparametric confidence intervals. Ann. Statist., 25(6):2547–2554.
[39] Masry, E. (1993). Strong consistency and rates for deconvolution of multivariate densities of
stationary processes. Stochastic Process. Appl., 47(1):53–74.
[40] Matias, C. (2002). Semiparametric deconvolution with unknown noise variance. ESAIM
Probab. Stat, 6:271–292.
[41] Meister, A. (2008). Deconvolution from Fourier-oscillating error densities under decay and
smoothness restrictions. Inverse Problems, 24(1):015003, 14.
[42] Neumann, M. H. (1997). On the effect of estimating the error density in nonparametric
deconvolution. J. Nonparametr. Statist., 7(4):307–330.
[43] Rigollet, P. and Tsybakov, A. (2011). Exponential screening and optimal rates of sparse
estimation. Ann. Statist., 39(2):731–771.
[44] Rigollet, P. and Tsybakov, A. B. (2012). Comment:" minimax estimation of large covariance
matrices under `1 -norm”. Statist. Sinica, 22(4):1358–1367.
[45] Rothman, A. J. (2012). Positive definite estimators of large covariance matrices. Biometrika,
99(3):733–740.
[46] Rothman, A. J., Levina, E., and Zhu, J. (2009). Generalized thresholding of large covariance
matrices. J. Amer. Statist. Assoc., 104(485):177–186.
[47] Sanandaji, B. M., Tascikaraoglu, A., Poolla, K., and Varaiya, P. (2015). Low-dimensional
models in spatio-temporal wind speed forecasting. In American Control Conference (ACC),
Covariance estimation in high-dimensional deconvolution
35
2015, pages 4485–4490. IEEE.
[48] Sato, K.-i. (2013). Lévy processes and infinitely divisible distributions, volume 68 of Cambridge
Studies in Advanced Mathematics. Cambridge University Press, Cambridge. Translated from
the 1990 Japanese original, Revised edition of the 1999 English translation.
[49] Tao, M., Wang, Y., and Zhou, H. H. (2013). Optimal sparse volatility matrix estimation for
high-dimensional Itô processes with measurement errors. Ann. Statist., 41(4):1816–1864.
[50] Tsybakov, A. B. (2009). Introduction to nonparametric estimation. Springer Series in Statistics. Springer, New York. Revised and extended from the 2004 French original, Translated by
Vladimir Zaiats.
[51] Tsybakov, A. B. (2013). Aggregation and high-dimensional statistics. Saint Flour Lecture
notes.
| 10math.ST
|
An Online Development Environment for
Answer Set Programming
Elias Marcopoulos1 and Christian Reotutar2 and Yuanlin Zhang3
1
Department of Computer Science, Tufts University, USA
Department of Computer Science, Johns Hopkins University, USA
3
Texas Tech University, Lubbock, TX, USA
[email protected], [email protected], [email protected]
arXiv:1707.01865v1 [cs.OH] 20 Jun 2017
2
Abstract. Recent progress in logic programming (e.g., the development
of the Answer Set Programming paradigm) has made it possible to teach
it to general undergraduate and even high school students. Given the
limited exposure of these students to computer science, the complexity
of downloading, installing and using tools for writing logic programs
could be a major barrier for logic programming to reach a much wider
audience. We developed an online answer set programming environment
with a self contained file system and a simple interface, allowing users to
write logic programs and perform several tasks over the programs.
1
Introduction
Answer Set Programming (ASP) [8] is becoming a dominating language in the
knowledge representation community [15,12] because it has offered elegant and
effective solutions not only to classical Artificial Intelligence problems but also
to many challenging application problems. Thanks to its simplicity and clarity
in both informal and formal semantics, Answer Set Programming provides a
“natural” modeling of many problems. At the same time, the fully declarative
nature of ASP also cleared a major barrier to teach logic programming, as the
procedural features of classical logic programming systems such as PROLOG are
taken as the source of misconceptions in students’ learning of Logic Programming
[16].
ASP has been taught to undergraduate students, in the course of Artificial
Intelligence, at Texas Tech for more than a decade. We believe ASP has become
mature enough to be a language for us to introduce programming and problem
solving to high school students. We have offered many sessions to students at
New Deal High School and a three week long ASP course to high school students
involved in the TexPREP program (http://www.math.ttu.edu/texprep/). In our
teaching practice, we found that ASP is well accepted by the students and the
students were able to focus on problem solving, instead of the language itself. The
students were able to write programs to answer questions about the relationships
(e.g., parent, ancestor) amongst family members and to find solutions for Sudoku
problems.
2
Marcopoulos, Reotutar and Zhang
However, we have some major issues while using existing tools: installation
of the tools to computers at a lab or at home is complex, and the existing
tools are sensitive to the local settings of a computer. As a result, the flow of
teaching the class was often interrupted by the problems associated with the use
of the tools. Strong technical support needed for the management and use of the
tools is prohibitive for teaching ASP to general undergraduate students or K-12
students.
During our teaching practice, we also found the need for a more vivid presentation of the results of a logic program (more than just querying the program or
getting the answer sets of the program). We also noted observations in literature
that multimedia and visualization play a positive role in promoting students’
learning [9,3].
To overcome the issues related to software tool management and use, we have
designed and built an online development environment for Answer Set Programming. The environment provides an editor for users to edit their programs, an
online file system for them to store and retrieve their program and a few simple
buttons allows querying the program inside the editor or getting answer sets of
the program. The environment uses SPARC [2] as the ASP language. SPARC
is designed to further facilitate the teaching of logic programming by introducing sorts (or types) which simplify the difficult programming concept of domain
variables in classical ASP systems such as Clingo [7] and help programmers to
identify errors early thanks to sort information. Initial experiment of teaching
SPARC to high school students is promising [18]. To promote students’ interests
and learning, our environment also introduces predicates for students to present
their solutions to problems in a more visually straightforward and exciting manner (instead of the answer sets which are simply a set of literals). The URL for
the online environment is http://goo.gl/ukSZET.
The rest of the paper is organized as follows. Section 2 recalls SPARC. The
design and implementation of the online environment are presented in Section 3.
The design and rendering of the drawing and animation predicates are presented
in Section 4. The paper is concluded by Section 5.
2
Answer Set Programming Language – SPARC
SPARC is an Answer Set Programming language which allows for the explicit
representation of sorts. A SPARC program consists of three sections: sorts, predicates and rules. We will use the map coloring problem as an example to illustrate
SPARC: can the USA map be colored using red, green and blue such that no
two neighboring states have the same color?
The first step is to identify the objects and their sorts in the problem. For
example, the three colors are important and they form the sort of color for this
problem. In SPARC syntax, we use #color = {red, green, blue} to represent the
objects and their sort. The sorts section of the SPARC program is
sorts % the keyword to start the sorts section
An Online Development Environent
3
#color = {red,green,blue}.
#state = {texas, colorado, newMexico, ......}.
The next step is to identify relations in the problem and declare in the predicates section the sorts of the parameters of the predicates corresponding to the
relations. The predicates section of the program is
predicates % the keyword to start the predicates section
% neighbor(X, Y) denotes that state X is a neighbor of state Y.
neighbor(#state, #state).
% ofColor(X, C) denotes that state X has color C
ofColor(#state, #color).
The last step is to identify the knowledge needed in the problem and translate
it into rules. The rules section of a SPARC program consists of rules in the typical
ASP syntax. The rules section of a SPARC program will include the following.
rules % the keyword to start the rules section
% Texas is a neighor of Colorado
neighbor(texas, colorado).
% The neighbor relation is symetric
neighbor(S1, S2) :- neighbor(S2, S1).
% Any state has one of the three colors: red, green and blue
ofColor(S, red) | ofColor(S, green) | ofColor(S, blue).
% No two neighbors have the same color
:- ofColor(S1, C), ofColor(S2, C), neighbor(S1, S2), S1 != S2.
3
3.1
Online Development Environment Design and
Implementation
Environment Design
The principle of the design is that the environment, with the simplest possible
interface, should provide full support, from writing programming to getting the
answer sets of the program, for teaching Answer Set Programming.
The design of the interface is shown in Figure 1. It consists of 3 components:
1) the editor to edit a program, 2) the file navigation system and 3) the operations
over the program.
4
Marcopoulos, Reotutar and Zhang
Fig. 1. User Interface of the System (the red numbers indicate the areas/components
in the interface)
One can edit a SPARC program directly inside the editor which has syntax
highlighting features (area 1). The file inside the editor can be saved by clicking
the “Save” button (2.4). The files and folders are displayed in the area 2.1. The
user can traverse them using the mouse like traversing a file system on a typical
operating system. Files can be deleted and their names can be changed. To
create a folder or a file, one clicks the “New” button (2.3). The panel showing
files/folders can be toggled by clicking the “Directory” button (2.2) (so that users
can have more space for the editing or result area (4)). To ask queries to the
program inside the editor, one can type a query (a conjunction of literals) in the
text box (3.1) and then press the “Submit” button (3.1). The answer to the query
will be shown in area 4. For a ground query (i.e., a query without variables), the
answer is yes if every literal in the query is in every answer set of the program,
is no if the complement (p and ¬p, where p is an atom, are complements) of
some literal is in every answer set of the program, and unknown otherwise. An
answer to a query with variables is a set of ground terms for the variables in the
query such that the answer to the query resulting from replacing the variables
by the corresponding ground terms is yes. Formal definitions of queries and
answers to queries can be found in Section 2.2 of [8]. To see the answer sets of
a program, click the “Get Answer Sets” button (3.2). When “Execute” button
(3.3) is clicked, the atoms with drawing and animation in the answer set of the
program will be rendered in the display area (4). (For now, when there is more
than one answer set, the environment displays an error.)
A user can only access the full interface discussed above after login. The user
will log out by clicking the “Logout” button (5). Without login, the interface is
much simpler, with all the file navigation related functionalities invisible. Such
an interface is convenient for testing or doing a quick demo of a SPARC program.
3.2
Implementation
.
The architecture of the online environment follows that of a typical web
application. Is consists of a front end component and a back end component.
The front end provides the user interface and sends users’ request to the back
An Online Development Environent
5
end, and the back end fulfills the request and returns results, if needed, back
to the front end. After getting the results from the back end, the front end will
update the interface correspondingly (e.g., display query answers to the result
area). Details about the components and their interactions are given below.
Front End. The front end is implemented by HTML and JavaScript. The editor
in our front end uses ACE which is an embeddable (to any web page) code
editor written in JavaScript (https://ace.c9.io/). The panel for file/folder
navigation is based on JavaScript code by Yuez.me.
Back End and Interactions between the Front End and the Back End.
The back end is mainly implemented using PHP and is hosted on the server side.
It has three components: 1) file system management, 2) inference engine and 3)
drawing/animation rendering.
The file system management uses a database to manage the files and
folders of all users of the environment. The ER diagram of the system is shown
below:
Fig. 2. The ER diagram for file/folder management. Most names have a straightforward meaning. The Folderurl and Fileurl above refer to the full path of the folder/file
in the file system.
The SPARC files are saved in the server file system, not in a database table.
The sharing is managed by the sharing information in the relevant database
tables. In our implementation, we use mySQL database system.
6
Marcopoulos, Reotutar and Zhang
The file management system gets request such as creating a new file/folder,
deleting a file, saving a file, getting the files and folders, etc, from the front end.
It then updates the tables and local file system correspondingly and returns the
needed results to the front end. After the front end gets the results, it will update
the graphical user interface (e.g., display the program returned from the back
end inside the editor) if needed.
The inference engine gets the request of answering a query or obtaining
all answer sets of a program. It calls the SPARC solver [2] to find all answer
sets. Then in terms of these answer sets, it returns requested information to the
front end. After the front end gets the response from the back end, it will show
the result in the display area of the web page.
Details of the design and implementation of drawing/animation rendering can be found in Section 4.2.
4
4.1
Drawing and Animation Design and Implementation
Drawing and Animation Design
To allow programmers to create drawings and animations using SPARC, we
simply design two predicates, called display predicates: one for drawing and one
for animation. The atoms using these predicates are called display atoms. To use
these atoms in a SPARC program, a programmer needs to include sorts (e.g.,
sort of colors, fonts and numbers) and the corresponding predicate declaration
which are predefined. In the following, we only focus on the atoms and their use
for drawing and animation.
Drawing. A drawing predicate is of the form: draw(c) where c is called a drawing command. Intuitively the atom containing this predicate draws texts and
graphics as instructed by the command c. By drawing a picture, we mean a
shape is drawn with a style. We define a shape as either text or a geometric line
or curve. Also, a style specifies the physical visual properties of the shape it is
applied to. For example, visual properties include color, thickness, and font. For
modularity, we introduce style names, which are labels that can be associated
with different styles so that the style may be reused without being redefined. A
drawing is completed by associating this shape and style to a certain position in
the canvas, which is simply the display board. Note, the origin of the coordinate
system is at the top left corner of the canvas.
Here is a an example of drawing a red line from point (0, 0) to (2, 2). First,
we introduce a style name redline and associate it to the red color by the
style command line color(redline, red). With this defined style we then
draw the red line by the shape command draw line(redline, 0, 0, 2, 2).
Style commands and shape commands form all drawing commands. The SPARC
program rules to draw the given line are
draw(line color(redline, red)).
draw(draw line(redline, 0, 0, 2, 2)).
An Online Development Environent
7
The style commands of our system include the following:
linewidth(sn, t) specifies that lines drawn with style name sn should be
drawn with a line thickness t. textfont(sn, fs, ff) specifies that text drawn
with style name sn should be drawn with a font size fs and a font family ff.
linecap(sn, c) specifies that lines drawn with style name sn should be drawn
with a capping c, such as an arrowhead. textalign(sn, al) specifies that text
drawn with style name sn should be drawn with an alignment on the page al.
line color(sn, c) specifies that lines drawn with style name sn should be
drawn with a color c. textcolor(sn, c) specifies that text drawn with style
name sn should be drawn with a color c.
The shape commands include the following:
draw line(sn, xs, ys, xe, ye) draws a line from starting point (xs, ys)
to ending point (xe, ye) with style name sn; draw quad curve(sn, xs, ys,
bx, by, xe, ye) draws a quadratic Bezier curve, with style name sn, from the
current point (xs, ys) to the end point (xe, ye) using the control point (bx,
by); draw bezier curve(sn, xs, ys, b1x, b1y, b2x, b2y, xe, ye) draws
a cubic Bezier curve, using style name sn, from the current point (xs, ys) to
the end point (xe, ye) using the control points (b1x, b1y) and (b2x, b2y);
draw arc curve(sn, xs, ys, r, sa, se) draws an arc using style name sn
and the arc is centered at (x, y) with radius r starting at angle sa and ending
at angle se going in the clockwise direction; draw text(sn, x, xs, ys) prints
value of x as text to screen from point (xs, ys) using style name sn.
Animation. A frame, a basic concept in animation, is defined as a drawing.
When a sequence of frames, whose content is normally relevant, is shown on the
screen in rapid succession (usually 24, 25, 30, or 60 frames per second), a fluid
animation is seemingly created. To design an animation, a designer will specify
the drawing for each frame. Given that the order of frames matters, we give
a frame a value equal to its index in a sequence of frames. We introduce the
animate predicate animate(c, i) which indicates a desire to draw a picture at
the ith frame using drawing command c and i starts from 0. The frames will be
shown on the screen at a rate of 60 frames per second, and the ith frame will be
showed at time (i ∗ 1/60) (in a unit of second) from the start of the animation
for a duration of 1/60 of a second.
As an example, we would like to elaborate on an animation where a red box
(with side length of 10 pixels) moves from the point (1, 70) to (200, 70). We will
create 200 frames with the box (whose bottom left corner is) at point (i + 1, 70)
in ith frame.
Let the variable I be of a sort called frame, defined from 0 to some large
number. In every frame I, we specify the drawing styling redline:
animate(line color(redline, red), I).
To make a box at the I th frame, we need to draw the box’s four sides using
the style associated with style name redline. The following describes the four
sides of a box at any frame: bottom - (I +1, 70) to (I +1+10, 70), left - (I +1, 70)
to (I + 1, 60), top - (I + 1, 60) to (I + 1 + 10, 60) and right - (I + 1 + 10, 60) to
(I + 1 + 10, 70). Hence we have the rules
8
Marcopoulos, Reotutar and Zhang
animate(draw
animate(draw
animate(draw
animate(draw
line(redline,I+1,70,I+11,70),I).
line(redline,I+1,70,I+1,60),I).
line(redline,I+1,60,I+11,60),I).
line(redline,I+11,60,I+11,70),I).
Note that the drawing predicate produces the intended drawing throughout
all the frames creating a static drawing. On the other hand, the animate predicate
produces a drawing only for a specific frame.
4.2
Algorithm and Implementation
We first define our input and output: The input to the main algorithm is a
SPARC program P . The output is an HTML5 program segment containing a
canvas element which will be rendered by an Internet browser. A key part of our
algorithm is to render the display atoms (specified in the answer set of P ) using
canvas methods.
HTML5 canvas element is used to draw graphics via scripting using JavaScript.
In the following, we will use an example to demonstrate how a drawing command
is implemented by JavaScript code using canvas methods. Consider again
draw(line color(redline, red)).
draw(draw line(redline, 0, 0, 2, 2)).
When we render the shape command draw line, we need to know the meaning of the redline style. From the style command line color, we know it
means red. We first create an object ctx for a given canvas (simply identified
by a name) where we would like to render the display atoms. The object offers methods to render the graphics in the canvas. We then use the following
JavaScript code to implement the shape command to draw a line from (0,0) to
(2,2):
ctx.beginPath();
ctx.moveTo(0,0);
ctx.lineTo(2,2);
ctx.stroke();
To make the line in red color, we have to insert the following JavaScript
statement before the ctx.stroke() in the code above:
ctx.strokeStyle="red";
The meaning of the canvas methods in the code above is straightforward. We
don’t explain them further. Now we are in a position to present the algorithm.
Algorithm:
– Input: a SPARC program P with display predicates.
– Output: a HTML program segment which allows the rendering of the display
atoms in the answer set of P in an Internet Browser.
– Steps:
1. Call SPARC solver to obtain an answer set S of P .
2. Let script be an array of empty strings. script[i] will hold the JavaScript
statements to render the graphics for ith frame.
3. For each display atom a in S,
An Online Development Environent
9
• If any error is found in the display atoms, present an error to the
user detailing the incorrect usage of the atoms.
• If a contains a shape command, let its style name be sn, find all style
commands defining sn. For each style command, translate it into the
corresponding JavaScript code Ps on modifying the styling of the
canvas pen. Then translate the shape command into JavaScript code
Pr that renders that command. Let Pd be the proper combination
of Ps and Pr to render a.
∗ if a is an drawing atom, append Pd to script[i] for every frame i
of the animation.
∗ if a is an animation atom, let i be the frame referred to in a.
Append Pd to script[i].
4. Formulate the output program P as follows:
• add, to P , the canvas element <canvas id="myCanvas" width="500"
height="500"> </canvas>.
• add, to P , the script element <script> </script> whose content
includes
∗ the JavaScript code to associate the drawings in this script element with the canvas element above.
∗ an array drawings initialized by the content of script array.
∗ Javascript code executing the statements in drawings[i] when
the time to show frame i starts.
End of algorithm.
Implementation. The “Execute” button in the webpage (front end) of the online SPARC environment is for programmers to render the display atoms in the
answer set of their programs. The Java program implementing our algorithm
above is at the server side. When the “Execute” button is clicked, the programmer’s SPARC program will be sent to the server side and the algorithm will be
invoked with the program. The output (i.e., the canvas and script elements) of
the algorithm will be sent back to the front end and the JavaScript in the front
end will catch the output and insert it into the result display area of the front
web page (See Figure 1). The Internet browser will then automatically render
the updated web page and the drawing or animation will be rendered as a result.
Example SPARC programs with drawing and animation can be found at
https://goo.gl/nLD4LD.
5
Discussion and Related Work
As ASP has been applied to more and more problems, the importance of ASP
software development tools has been realized by the community. Some integrated
development environment (IDE) tools, e.g., APE [19], ASPIDE[6], iGROM[10]
and SeaLion [17] have previously been developed. They provide a graphical user
interface for users to carry out a sequence of tasks from editing an ASP program
to debugging that program, easing the use of ASP significantly. However, the
10
Marcopoulos, Reotutar and Zhang
target audience of these tools is experienced software developers. Compared with
the existing environments, our environment is online, self contained (i.e., fully
independent of the users’ local computers) and provides a very simple interface,
focusing on teaching only. The interface is operable by any person who is able
to use a typical web site and traverse a local file system.
As for drawing and animation, our work is based on the work of Cliffe et
al. [4]. They are the first to introduce, to ASP, a design of display predicates
and to render drawings and animations using the program ASPviz. Our drawing commands are similar to theirs. The syntax of their animation atoms is not
clear from their paper. It seems (from examples on github at goo.gl/kgUzJK
accessed on 4/30/2017) that multiple answer sets may be needed to produce an
animation. In our work we use a design where the programmers are allowed to
draw at any frame (specifying a range of the frames) and the real time difference between two neighboring frames is 1/60 second. Another clear difference is
that our implementation is online while theirs is a standalone software. A more
recent system, Kara, a standalone software by Kloimullner et al. [11], deals with
drawing only. Another system ARVis [1] offers method to visualize the relations
between answer sets of a given program. We also note an online environment for
IDP (which is a knowledge representation paradigm close to ASP) by Dasseville
and Janssens [5]. It also utilizes a very simple interface for the IDP system and
allows drawing and animation using IDP through IDPD3 (a library to visualize
models of logic theories) by Lapauw et al. [14]. In addition to drawing and animation, IDPD3 allows users’ interaction with the IDP program (although in a
limited manner in its current implementation), which is absent from most other
systems including ours. Our environment is also different from the online IDP
environment in that ours targets ASP and offers an online file system. Both
DLV and Clingo offer online environments (http://asptut.gibbi.com/ and
http://potassco.sourceforge.net/clingo.html respectively) which provide
an editor and a window to show the output of the execution of dlv and clingo
command, but provide no other functionalities. We also noted the SWISH
(http://lpsdemo.interprolog.com) which offers an online environment for
Prolog and a more recent computer language Logic-based Production Systems
[13]. A unique functionality of our online environment is to query a program.
It allows to teach (particular to general students) basics of Logic Programming
without first touching the full concept of answer sets.
When we outreached to a local high school before, we needed an experienced
student to communicate with the school lab several times before the final installation of the software on their computers could be completed. A carefully
drafted document is prepared for students to install the software on their computers. There are still unexpected issues during lab or when students use/install
the software at home. These difficulties made it almost impossible to outreach to
the high school with success. With the availability of our online environment, we
only need to focus on the teaching content of ASP without worrying about the
technical support. We hope our environment, and other online environments, for
knowledge representation systems will expand the teaching of knowledge repre-
An Online Development Environent
11
sentation to a much wider audience in the future. The drawing and animation
are new features of the online environment and was not tested in high school
teaching. We have used the drawing and animation in a senior year course –
special topics in AI – in spring 2017. Students demonstrated interests in drawing and animation and they were able to produce interesting animation. We also
noted that it can be very slow for ASP solvers to produce the answer set of an
animation program when the ground program is big.
In the future, it will be interesting to have a more rigorous evaluation of the
online environment.
6
Acknowledgments
The authors were partially supported by National Science Foundation (grant#
CNS-1359359). We thank Evgenii Balai, Mbathio Diagne, Michael Degraw, Peter Lee, Maede Rayatidamavandi, Crisel Suarez, Edward Wertz and Shao-Lon
Yeh for their contribution to the implementation of the environment. We thank
Michael Gelfond and Yinan Zhang for their input and help.
References
1. Ambroz, T., Charwat, G., Jusits, A., Wallner, J.P., Woltran, S.: Arvis: visualizing
relations between answer sets. In: International Conference on Logic Programming
and Nonmonotonic Reasoning. pp. 73–78. Springer (2013)
2. Balai, E., Gelfond, M., Zhang, Y.: Towards answer set programming with sorts. In:
Logic Programming and Nonmonotonic Reasoning, 12th International Conference,
LPNMR 2013, Corunna, Spain, September 15-19, 2013. Proceedings. pp. 135–147
(2013), http://dx.doi.org/10.1007/978-3-642-40564-8 14
3. Clark, D., Nelson, B., Sengupta, P., DAngelo, C.: Rethinking science learning
through digital games and simulations: Genres, examples, and evidence. In: Learning science: Computer games, simulations, and education workshop sponsored by
the National Academy of Sciences, Washington, DC (2009)
4. Cliffe, O., De Vos, M., Brain, M., Padget, J.: Aspviz: Declarative visualisation and
animation using answer set programming. In: International Conference on Logic
Programming. pp. 724–728. Springer (2008)
5. Dasseville, I., Janssens, G.: A web-based ide for idp. arXiv preprint
arXiv:1511.00920 (2015)
6. Febbraro, O., Reale, K., Ricca, F.: ASPIDE: integrated development environment for answer set programming. In: Logic Programming and Nonmonotonic
Reasoning - 11th International Conference, LPNMR 2011, Vancouver, Canada,
May 16-19, 2011. Proceedings. pp. 317–330 (2011), http://dx.doi.org/10.1007/
978-3-642-20895-9 37
7. Gebser, M., Kaufmann, B., Kaminski, R., Ostrowski, M., Schaub, T., Schneider, M.: Potassco: The potsdam answer set solving collection. Ai Communications
24(2), 107–124 (2011)
8. Gelfond, M., Kahl, Y.: Knowledge Representation, Reasoning, and the Design of
Intelligent Agents. Cambridge University Press (2014)
12
Marcopoulos, Reotutar and Zhang
9. Guzdial, M.: Use of collaborative multimedia in computer science classes. ACM
SIGCSE Bulletin 33(3), 17–20 (2001)
10. iGROM: http://igrom.sourceforge.net/
11. Kloimüllner, C., Oetsch, J., Pührer, J., Tompits, H.: Kara: A system for visualising and visual editing of interpretations for answer-set programs. In: Applications
of Declarative Programming and Knowledge Management, pp. 325–344. Springer
(2013)
12. Kowalski, R.: Logic programming. Computational Logic, Volume 9 (Handbook of
the History of Logic) (2014)
13. Kowalski, R., Sadri, F.: Programming in logic without logic programming. Theory
and Practice of Logic Programming 16(03), 269–295 (2016)
14. Lapauw, R., Dasseville, I., Denecker, M.: Visualising interactive inferences with
idpd3. arXiv preprint arXiv:1511.00928 (2015)
15. McIlraith, S.: What’s hot in knowledge representation and reasoning. Talk in
the AAAI-12 SUBAREA SPOTLIGHTS TRACK on Knowledge Representation
(2011)
16. Mendelsohn, P., Green, T., Brna, P.: Programming languages in education: The
search for an easy start. Psychology of programming pp. 175–200 (1990)
17. Oetsch, J., Pührer, J., Tompits, H.: The sealion has landed: An ide for answer-set
programmingpreliminary report. In: Applications of Declarative Programming and
Knowledge Management, pp. 305–324. Springer (2013)
18. Reyes, M., Perez, C., Upchurch, R., Yuen, T., Zhang, Y.: Using declarative programming in an introductory computer science course for high school students. In:
Thirtieth AAAI Conference on Artificial Intelligence (2016)
19. Sureshkumar, A., De Vos, M., Brain, M., Fitch, J.: APE: an ansprolog* environment. Proc. SEA 7, 101–115 (2007)
| 2cs.AI
|
Evolving the Incremental λ Calculus into a Model of Forward AD∗
Robert Kelly†
Barak A. Pearlmutter‡
Jeffrey Mark Siskind§
April 2016
arXiv:1611.03429v1 [cs.PL] 10 Nov 2016
Introduction
Formal transformations somehow resembling the usual derivative are surprisingly common in computer science, with
two notable examples being derivatives of regular expressions [1] and derivatives of types [2, 3]. A newcomer to this list
is the incremental λ-calculus, or ILC, a “theory of changes” that deploys a formal apparatus allowing the automatic
generation of efficient update functions which perform incremental computation [4]. An example of this would be
using the ILC derivative-like operator D to alter a function f : B → B, which performs some major reorganization
on a database (of type B), into the update function D f : B → ∆B → ∆B. Here ∆B is the type of changes to B.
So D f , given an initial database, maps a change to that input database to a change to the output database. This in
principle, and as shown in their work also in practice, allows enormous savings when the change to the input is small
compared to the size of the input itself. Resemblance to the standard derivative can be exhibited by a simple example
D (λ x . f (g x))
(λ x x′ . D f (g x) (D g x x′ ))
(1a)
D f (g x) ◦ D g x
(1b)
or
D (f ◦ g) x
which seems suspiciously similar to the familiar Calculus 101 chain rule.
The ILC is not only defined, but given a formal machine-understandable definition—accompanied by mechanically
verifiable proofs of various properties, including in particular correctness of various sorts. Here, we show how the
ILC can be mutated into propagating tangents, thus serving as a model of Forward Accumulation Mode Automatic
Differentiation.1
This mutation is done in several steps. These steps can also be applied to the proofs, resulting in machine-checked
proofs of the correctness of this model of forward AD.
The Mutagenic Steps
There are two differences between the incremental λ calculus and forward AD. First, changes rather than tangents
are propagated. These changes are elements of change sets, and constitute finite (i.e., not infinitesimal) modifications.
(For example, a change to a list might consist of swapping the first two elements, and a change to a number might
consist of increasing its value by 5.) In numerics, these would be “differences” rather than “differentials”, and ∆
rather than ∂. Second, the changes are passed as additional arguments instead of being bundled together with primal
values. Passing changes in additional arguments makes great sense in the domain of incremental computation, where
the whole point of the construction is to partially evaluate a function D f : α → ∆α → ∆β with respect to f ’s original
input, yielding a mapping of changes to changes: ∆α → ∆β. But in the context of forward AD, we wish to propagate
tangent values in parallel with primal values, which necessitates both bundling the “new” values with the original
ones, and including the original output in the output of the transformed function.
We proceed to eliminate these two differences. This is done in two stages. First, considering only power series
change sets to the base type R. And second, uncurrying the outputs of the derivative operator and causing it to
propagate change sets and primal values bundled together all the way through to its output. Truncating the power
series changes them into Dual Numbers [7], yielding the familiar Forward AD. A commutative diagram of these steps
∗ Extended
abstract presented at the AD 2016 Conference, Sep 2016, Oxford UK.
Author, Dept of Computer Science, National University of Ireland Maynooth, funded by the Irish Research Council,
[email protected]
‡ Dept of Computer Science, National University of Ireland Maynooth, [email protected]
§ School of Electrical and Computer Engineering, Purdue University, [email protected]
1 The approach detailed here stands in contrast to the Simply Typed λ-Calculus of Forward Automatic Differentiation [5]. Aside from
some issues with confluence, that work folded together levels of the hierarchy by not distinguishing numeric basis functions which operate
on R from those which are lifted to operate on Dual numbers, while here these are distinguished. Moreover, here we have a framework
for machine-readable machine-verified proofs of various correctness and efficiency properties. This approach differs from the Differential
Lambda-Calculus [6] in analogous ways: complexity, machine-checked proofs, and explicit segregation of levels of differentiation.
† Corresponding
∆R = R[ε]
Incremental
λ-Calculus
Power
Series
Dual
Numbers
∆R = R[ε]
Higher-Order
Forward AD
uncurry
uncurry
uncurry
Bundled
ILC
∆R = R[ε]/ε2
∆R = R[ε]/ε2
Forward
AD
Figure 1: Mutating the Incremental λ-Calculus (ILC) into Forward-Mode Automatic Differentiation (Forward AD).
is shown in Figure 1. The original ILC is in the top left, with relevant changes indicated with transitions to new
states or nodes. Each of these edges leads to a different combination of forward AD in the ILC. The power series and
uncurry steps can be taken in either order, so the diagram should commute.
Let us describe these two steps in a bit more detail.
Step One: Power Series
To see how power series change sets are introduced, we note that the ILC allows change sets to be defined for any base
type τ . These change sets need only obey a particular set of axioms, which in our context amounts to associativity
of addition of real numbers. We constrain ourselves to consider only change sets to reals: the base type R. We then
represent these change sets not as differences, but instead as power series (in some variable ε) with a zero constant
term. This means that the change set of x : R is a term of the form hzpsε i, where
hzpsε i ::= 0 | ε ∗ hpsε i
hpsε i ::= R | R + hzpsε i
(2a)
(2b)
∆R ≡ hzpsε i
(2c)
For a specific value of ε (possibly subject to conditions of convergence) this would take on a particular numeric value.
We further define an operator coeff which takes a nonnegative integer index and a power series in ε wrapped in a λ
expression, i.e., (λε . hpsε i), and yields the requested coefficient of the given power series.
coeff 0 (λε . r)
coeff 0 (λε . r + ε ∗ e)
r
r
coeff 0 (λε . ε ∗ e)
0
coeff i (λε . r + ε ∗ e)
coeff i (λε . ε ∗ e)
(where ε 6∈ FV(r))
(where ε ∈
6 FV(r))
(3a)
(3b)
(3c)
coeff (i − 1) (λε . e)
coeff (i − 1) (λε . e)
(where i > 0 and ε 6∈ FV(r))
(where i > 0)
(3d)
(3e)
For instance,
coeff 2 (λε . 0.1 + ε ∗ (0.2 + ε ∗ (0.3 + ε ∗ (0.4 + ε ∗ (0.5 + · · · )))))
0.3
Useful properties of such a change set are straightforward to establish: closure under the derivatives of the numeric
basis functions, and dependence during such operators of coefficients only on coefficients of the same or lower order.
The first property is necessary for consistency, while the second allows these power series to be truncated at ε2 ,
thus yielding the tangents of standard forward AD. With this machinery, we could define the familiar derivative
diff : (R → R) → (R → R), for instance diff sin = cos, as
diff f x ≡ coeff 1 (λε . (D f x (ε ∗ 1)))
(4)
By defining coeff to distribute over algebraic datatypes
coeff i (λε . Constructor e1 · · · en )
Constructor (coeff i (λε . e1 )) · · · (coeff i (λε . en ))
(5a)
and post-compose over functions
coeff i (λε . (λx . e))
(λx . coeff i (λε . e))
(where x 6= ε)
(5b)
this machinery can find directional derivatives of functions with non-scalar output, including Church-encoded output.
In this formulation, the tagging necessary to distinguish distinct nested invocations of derivative-taking operators
[8, 9] is handled by the standard λ-calculus mechanisms for avoiding variable capture during β-substitution, e.g.,
α-renaming.
Step Two: Uncurrying and Bundling
The second step is uncurrying arguments, and bundling the output. We need to change the type of the derivative
operator from
D : (t1 → t2 → · · · → tn → u) → (t1 → ∆t1 → t2 → ∆t2 → · · · → tn → ∆tn → ∆u)
(6)
D̂ : (t1 → t2 → · · · → tn → u) → (F t1 → F t2 → · · · → F tn → F u)
(7)
to
where F t is isomorphic to t × ∆t, a primal value bundled with its change set. If we define F (t1 → t2 ) = F t1 → F t2
then this yields a simpler type signature,
D̂ : t → F t
(8)
The mechanics of this change are straightforward, requiring that the ILC reductions be modified to take the new
shape. Note that, thus uncurried and carrying primal and change set values in tandem, the chain rules of Equation 1
are simplified: D̂ (f ◦ g)
D̂ f ◦ D̂ g.
Acknowledgments
This work was supported, in part, by Science Foundation Ireland grant 09/IN.1/I2637 and by NSF grant 1522954-IIS.
Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and
do not necessarily reflect the views of the sponsors.
References
[1] Janusz A. Brzozowski. Derivatives of regular expressions. Journal of the ACM (JACM), 11(4):481, 1964.
[2] Conor McBride.
The derivative of a regular type is its type of one-hole contexts, 2001.
URL
http://strictlypositive.org/diff.pdf. Available online.
[3] Michael Abbott, Neil Ghani, Thorsten Altenkirch, and Conor Mcbride. ∂ for data: Differentiating data structures. Fundamenta Informaticae, 65(1-2):1–28, August 2004. URL http://strictlypositive.org/dfordata.pdf.
[4] Yufei Cai, Paolo G. Giarrusso, Tillmann Rendel, and Klaus Ostermann. A theory of changes for higher-order languages: Incrementalizing λ-calculi by static differentiation. In Proceedings of the 35th ACM SIGPLAN Conference
on Programming Language Design and Implementation, pages 145–55, 2014. doi: 10.1145/2594291.2594304. URL
https://inc-lc.github.io/resources/pldi14-ilc-author-final.pdf. See arXiv:1312.0658.
[5] Oleksandr Manzyuk. A simply typed λ-calculus of forward automatic differentiation. In Mathematical Foundations of Programming Semantics Twenty-eighth Annual Conference, pages 259–73, Bath, UK, June 6–9 2012. URL
http://dauns.math.tulane.edu/~ mfps/mfps28proc.pdf.
[6] Thomas Ehrhard and Laurent Regnier. The differential lambda-calculus. Theoretical Computer Science, 309(1-3):1–41,
December 2003.
[7] William Kingdon Clifford. Preliminary sketch of bi-quaternions. Proceedings of the London Mathematical Society, 4:381–95,
1873.
[8] Jeffrey Mark Siskind and Barak A. Pearlmutter. Nesting forward-mode AD in a functional framework. Higher-Order and
Symbolic Computation, 21(4):361–76, 2008. doi: 10.1007/s10990-008-9037-1.
[9] Oleksandr Manzyuk, Barak A. Pearlmutter, Alexey Andreyevich Radul, David R. Rush, and Jeffrey Mark Siskind. Confusion of tagged perturbations in forward automatic differentiation of higher-order functions. Higher-Order and Symbolic
Computation, 2015. To appear. See also arXiv:1211.4892.
| 2cs.AI
|
Computational impact of hydrophobicity in
protein stability
Geetika S. Pandey¹
Dr. R.C Jain²
Research Scholar,
CSE dept., RGPV,
Bhopal (M.P), India
[email protected]
Director, SATI(D),
Vidisha(M.P), India
[email protected]
Abstract- Among the various features of amino
acids, the hydrophobic property has most visible
impact on stability of a sequence folding. This is
mentioned in many protein folding related work,
in this paper we more elaborately discuss the
computational impact of the well defined
‘hydrophobic aspect in determining stability’,
approach with the help of a developed ‘free
energy computing algorithm’ covering various
aspects - preprocessing of an amino acid sequence,
generating the folding and calculating free energy.
Later discussing its use in protein structure
related research work.
proteins are categorized as extracellular and
intracellular. So basically through the various studies
[2] could conclude that the core of protein contains
hydrophobic amino acids forming certain bonds and
thus structures. the stability of the structures is
determined by the free energy change , as mentioned
by Zhang et. al [3] i.e.
Keywords- amino acids,
energy, protein stability.
hydrophobicity,
ΔG(folding)=G(folded)-G(unfolded) [3]
Later in this paper various aspects of folding and
stability are discussed in detail.
free
II.
BACKGROUND
A. Features
I.
Shaolei Teng et.al.[4] mentioned twenty amino acid
features which they used to code each amino acid
residue in a data instance. They obtained these
features
from
Protscale
(http://expasy.org/tools/protscale.html)
[5]
and
AAindex (http://www.genome.jp/aaindex/) [6]. They
further mentioned these features into four categories -
INTRODUCTION
Since the earliest of proteomics researches, it has
been clear that the positioning and properties of
amino acids are key to structural analysis [1].
According to Betts et.al. in the protein environment a
feature of key importance is cellular location.
Different parts of cells have very different chemical
environments with the consequence that many amino
acids behave differently. The biggest difference as
mentioned by Betts et.al. is between soluble proteins
and membrane proteins. The soluble proteins tend to
be surrounded by water molecules i.e have polar or
hydrophilic residues on their surface whereas
membrane proteins are surrounded by lipids i.e they
tend to have hydrophobic residues on the surface that
interact with the membrane. Further the soluble
Biochemical features – includes M, molecular
weight, this is related to volume of space that a
residue occupies in protein structure. K, side chain
pka value, which is related to the ionization state of a
residue and thus plays a key role in pH dependent
protein stability. H, hydrophobicity index, which is
important for amino acid side chain packing and
protein folding. The hydrophobic interactions make
non-polar side chains to pack together inside proteins
1
and disruption of these interactions may cause protein
destabilization. P, polarity, which is the dipole-dipole
intermolecular interactions between the positively
and negatively charged residues. Co, overall amino
acid composition, which is related to the evolution
and stability of small proteins.
Structural features- this includes A, alpha-helix. B,
beta-sheet. C, coil. Aa, average area buried on
transfer from standard state to folded protein. Bu,
bulkiness, the ratio of the side chain volume to the
length of the amino acid.
Empirical Features- this includes, S1, protein stability
scale based on atom atom potential of mean force
based on Distance Scaled Finite Ideal-gas Reference
(DFIRE). S2, relative protein stability scale derived
from mutation experiments. S3, side-chain
contribution to protein stability based on data from
protein denaturation experiments.
Other biological features- F, average flexibility
index. Mc, mobility of an amino acid on
chromatography paper. No, number of codons for an
amino acid. R, refractivity, protein density and
folding characteristics. Rf, recognition factor,
average of stabilization energy for an amino acid.
Rm, relative mutability of an amino acid. Relative
mutability indicates the probability that a given
amino acid can be changed to others during
evolution. Tt, transmembrane tendency scale. F,
average flexibility index of an amino acid derived
from structures of globular proteins.
C. Factors affecting protein stability
Protein stability is the net balance of forces
which determine whether a protein will be in
its native folded conformation or a
denatured state. Negative enthalpy change
and positive entropy change give negative
i.e. stabilizing, contributions to the free
energy of protein folding, i.e. the lower the
∆G, the more stable the protein structure is
[7]. Any situation that minimizes the area of
contact between H₂O and non-polar, i.e
hydrocarbon, regions of the protein results
in an increase in entropy [9].
B. Protein folding
Protein folding has been considered as one of the
most important process in biology. under the various
physical and chemical conditions the protein
sequences fold forming bonds , when these
conditions are favourable the folding leads to proper
biological functionality. But some conditions could
lead to denaturation of the structures thus giving
unfolded structures. protein denaturants could be [7]
–
Heavy metals(e.g. lead, cadmium etc),
highly toxic, efficiently induce the ‘stress
response’.
Proteotoxic agents(e.g. alcoholc, crosslinking agents etc.)
Oxygen radicals, ionizing radiation- can
cause permanent protein damage.
Chaotropes (urea, guandine hydrochloride
etc.), highly potent at denaturing proteins,
often used in protein folding studies.
Protein folding considers the question of
how the process of protein folding occurs,
i.e how the unfolded protein adopts the
native state. Very often this problem has
been described as the second half of the
genetic code. Studies till date conclude the
following steps as the solution for this
problem [8] –
3D structure prediction from primary
sequence.
Avoiding misfolding related to human
diseases.
Designing proteins with novel functions.
∆G = ∆H - T∆S
High temperatures, can cause protein
unfolding, aggregation.
Low temperatures, some proteins are
sensitive to cold denaturation.
2
Following are the factors affecting protein
stability [8]:
pH : proteins are most stable in the vicinity
of their isoelectric point, pI. In general, with
some exceptions, electrostatic interactions
are believed to contribute to a small amount
of the stability of the native state.
The data in this case is a protein sequence loaded
from protein data bank with pdb id 5CYT, heme
protein, using Matlab 7.
Ligand binding: binding ligands like
inhibitors to enzymes, increases the stability
of the protein.
Disulphide bonds: it has been observed that
many extracellular proteins contained
disulphide bonds, whereas intracellular
proteins usually did not exhibit disulphide
bonds. Disulphide bonds are believed to
increase the stability of the native state by
decreasing the conformational entropy of the
unfolded state due to the conformational
constraints imposed by cross linking (i.e
decreasing the entropy of the unfolded
state).
Dissimilar properties of residues: not all
residues make equal contributions to protein
stability. Infact, studies say that the interior
ones, inaccessible to the solvent in the native
state make a much greater contribution than
those on the surface.
III.
Pro=
'XGDVAKGKKTFVQKCAQCHTVENGGKHKVG
PNLWGLFGRKTGQAEGYSYTDANKSKGIVWN
NDTLMEYLENPKKYIPGTKMIFAGIKKKGERQ
DLVAYLKSATS'
C. Methods
In brief the steps are as follows:
1) Preprocessing of the input primary protein
sequence using the hydrophobicity scale
developed by Kyte & Doolittle [9], i.e.
developing a vector with hydrophobic amino
acids represented by 1 and hydrophilic by 0.
2) Calculating the free energy of this initial
sequence
3) Now generating various foldings through
iteration, using complex number ‘i’.
4) Calculating the free energy for all these
foldings.
5) Now further these free energy values could
be used to check the stable structures.
EXPERIMENTAL PROCEDURE
A. Approach
As per the amino acid features mentioned previously,
the hydrophobic property is most responsible for the
folding, as well as stability related issues. Hence in
the algorithm mentioned later this property is taken
as the key in preprocessing of the input sequence, i.e.
the binary representation where ‘1’ denotes the
hydrophobic amino acids and others as ‘0’, as per the
hydrophobicity scales proposed by Kyle et. al [9].
Then using the complex plane the folding
configurations are formed and their combinations
denote various turns [10]. The cumulative sum of the
configuration is calculated which gives the direction
of each fold. Later the free energy of each folding is
calculated using Euclidean distance between the
hydrophobic amino acids i.e. all 1s and as per the
study the folding having lower free energy value
would be stable hence the stable structures could be
obtained.
D. Algorithm
Input – an amino acid sequence, Pro.
Output- an array of free energy of each structure
predicted, E.
1) Preprocessing of the input protein sequence
a) N ← length(Pro)
b) bin ← Pro
c) for idx ← 1:N
d) if Pro(idx)= hydrophobic
e) then bin(idx)← 1
f) else bin(idx) ← 0
g) end
h) end
2) folding formation
a) conf ← ones(length(bin)-1,1)
b) e ← Free_energy(conf)
c) for k ← 2:length(conf)
d) f(1:k) ← i
e) f(k+1:end)←1
B. Data
3
f) conf ← conf*f
g) F(:,count) ← conf
h) count = count+1
i) end
3) free energy of all the structures in F(m,n)
a) for j ← 1:n
b) q ← F(:,j)
c) p ← Cumulative_sum(q)
d) E(j) ← Free_Energy(p)
e) End
4) Algorithm for Cumulative Sum
Cumulative_sum (a)
a) for x ← 1 : length(a)
b) sum ← sum + a(x)
c) end
5) Algorithm for Free energy
Free_Energy(a)
a) a ← a * (bin with only
hydrophobic elements)
b) for x ← 1 : length(a)
c) d ← abs( a(x) –a(x+1))
d) sum ← sum + d
e) end
f) energy ← sum
IV.
hydrophobicity could be coupled with any other
amino acid feature.
REFERENCES
[1]
Matthew J. Betts and Robert B. Russell , Amino Acid
Properties and Consequences of substitutions, Chap. 14,
‘Bioinformatics for Geneticists’, 2003.
[2] Cuff JA. Barton G.J. “Evaluation and Improvement of
Multiple Sequence Methods for Protein Secondary
Structure Prediction, PROTEINS: Structure, Function,
and Genetics, 1999; 34: 508-19, Available from:
httop://binf.gmu.edu/vaisman/csi731/pr99-cuff.pdf.
[3] Zhe Zhang, Lin Wang, Daquan Gao, Jie Zhang, Maxim
Zhenirovskyy and Emil Alexov, “Predicting folding
free energy changes upon single point mutations”.
Bioinformatics Advance Access published, Jan. 2012.
[4] Shaolei Teng, Anand K. Srivastava, and Liangjiang
Wang, “Biological Features for Sequence-Based
Prediction of Protein Stability Changes upon Amino
Acid Substitutions”. International Joint Conference on
Bioinformatics, Systems Biology and Intelligent
Computing, 2009.
[5] H.C. Gasteiger E., Gattiker A., Duvaud S., Wilkins
M.R., Appel R.D. and Bairoch A. , The Proteomics
Protocols Handbook, Humana Press, 2005.
[6] S. Kawashima and M. Kanehisa, "AAindex: amino acid
index database," Nucleic Acids Res, vol. 28, Jan 1.
2000, pp. 374.
[7] Lecture
2,
Proteins:
structure,
translation,
http://www.sfu.ca/~leroux/class_L02.ppt.
[8] Protein stability, Protein Folding, misfolding –
chemistry,
http://www.chemistry.gsu.edu/faculty/.../Protein
/lecture6_foldingprotein_stability.ppt
[9] 76-456/731 Biophysical Methods- Protein structure
component, Lecture 2: Protein interactions leading to
folding http://www.chembio.ugo.
[10] Jack Kyte and Russell F. Doolittle,’ A simple method
for displaying the hydropathic character of a protein ‘,
J. Mol. Biol. 157, 105-132, 1982.
[11] www.mathworks.in/matlabcentral/contest/contests/11/r
ules.
Results
The length of the sequence in this case was 104,
hence as the algorithm total number of folding
created is 103, each column of matrix F (fig. 1)
shows a folding. And each row of array E (fig. 2)
shows the free energy for each folding. Here the free
energy of the unfolded structure is ‘e= 45194’.
V.
Discussion and futurework
The result from this approach provides the practical
aspect of the impact of hydrophobicity on stability,
the various outcomes could be used for further
research or with some modifications could lead the
ultimate solution. With the help of this method the
folding could be generated at any structure level,
these folding could be used for further research work
like in machine learning or neural networks. The free
energy calculated could be further used for clustering
or classification purposes, thus could enhance the
study of the stability factors. In the future work
4
AUTHORS PROFILE
R. C. Jain, M.Sc., M. Tech., Ph. D., is a Director of
S.A.T.I. (Engg. College) Vidisha (M. P.) India. He has 37 years of
teaching experience. He is actively involved in Research with area
of interest as Soft Computing, Fuzzy Systems, DIP, Mobile
Computing, Data Mining and Adhoc Networks. He has published
more than 125 research papers, produced 7 Ph. Ds. and 10 Ph. Ds
are under progress.
Geetika S. Pandey obtained her B.E degree in
Computer Science and Engineering from University Institute of
Technology, B.U, Bhopal in 2006. She obtained Mtech degree in
Computer Science from Banasthali Vidyapith, Rajasthan in 2008.
She worked as Assistant Professor in Computer Science and
Engineering Department in Samrat Ashok Technological Institute,
Vidisha (M.P). She is currently pursuing Ph.D. under the
supervision of Dr. R.C Jain, Director, SATI, Vidisha. Her research
is centered on efficient prognostication and augmentation of
protein structure using soft computing techniques.
5
Appendix
Fig. 1, F(103x103) , various folding of sequence pro.
Fig. 2, E(103x1), free energy of each folding.
6
| 5cs.CE
|
Object-oriented solutions †
V.E.Wolfengagen‡
arXiv:cs/0106021v1 [cs.LO] 11 Jun 2001
Vorotnikovsky per., 7, bld. 4
Institute for Contemporary Education “JurInfoR-MSU”
Moscow, 103006, Russia
[email protected]
Abstract
The resulting model seems to be a kind of object evaluator.
The object evaluator feature is to incorporate the schematic elements which are subdivided into individuals and individual concepts. Both of the entities are based on the notion of the variable
domain. This is a schematic construction and is equipped with both
the cloning and transactional means to capture dynamics.
All the parts of object evaluator share the same functor model
with the parameterized types and assignments. The logical part has
been supplied with both the atomic and non-atomic formulae with
the variables ranging over the variable domains. The categorical
part assists the evaluation to enable the extraction of a substitutional
part. The applicative part is capable of separating the computation
paths for function and its argument.
In this paper are briefly outlined the motivations, mathematical
ideas in use, pre-formalization and assumptions, object-as-functor
construction, ‘soft’ types and concept constructions, case study for
concepts based on variable domains, extracting a computational
background, and examples of evaluations.
In this paper are briefly outlined the motivations, mathematical
ideas in use, pre-formalization and assumptions, object-as-functor
construction, ‘soft’ types and concept constructions, case study for
concepts based on variable domains, extracting a computational
background, and examples of evaluations.
1
Introduction
An early incite in a theory of computations was to incorporate objects for a variety of purposes. They were assumed to represent the
existent - actual, possible or virtual objects in a problem domain.
The nature of existence was also under the concentrated study. The
recent years have generated a lot of object assumptions and discussions. Nevertheless, the initial notion of an object became overloaded by the mismeaning and not significant features. Every new
research in the area added the excessive troubles to understand the
clear sense and meaning of the object paradigm.
An attempt to rearrange the useful ideas will be done here. The
main attention is paid to establishing the parallelism between a theory of computations and the object-oriented notions.
1.1
1.2
Motivation for object evaluator
Object can be represented by embedding in a host computational
environment. An embedded object is accessed by the laws of the
host system. A pre-embedded object is observed as the decomposition into substitutional part and access function part which are
generated during the object evaluation. They assist to easy extract
of the result.
Subsumption is an usual theory-of-computation technique. Counterparts of the entire method – logic, functor category, and applicative computations, – are attached to generate an intermediate
computational framework. This intermediate representation is indirectly based on the categorical combinatory logic. The needed
optimizations may be obtained equationally.
†
This research is supported by the Russian Foundation for Basic Research (project
93-012-943)
‡
also: Kashirskoe Avenue, 31, Cybernetics Department, Moscow Engineering
Physical Institute, Moscow, 115409, Russia
Proceedings of the 2nd International Workshop on
Advances in Databases and Information Systems
(ADBIS’95). Vol. 1: Regular papers.
Moscow, June 27–30, 1995.
c 1995–2001 V.E. Wolfenagegen.
235
Evolution of the ideas
A technical intuition for an object is approximately as follows: object is proclaimed to be an entity (by default) with the strictly attached attributes – ‘internal state’ and ‘behavior’ [EGS93]. Some
of the objects are called the ‘dynamic objects’, that communicate
with each other (note that communication is presupposed of great
importance). Next step is to classify objects by their type, to collect objects into classes, to superimpose the various inheritances, to
compose for generating complex objects. Note that computational
intuition tends to establish object as a mathematical process.
1.2.1
Logic to incorporate objects
An approach to apply logic to a phenomena of object seems to be
clear and natural (e.g., [HC89], [Jac92] [Gab93]). Nevertheless,
adoption of more or less traditional approach of logic is distant of
the essence of the initial task to be solved. When the researcher
was to pick this kind of science it would combine some significant
elements.
1. The conditions of reasoning that transcend not only logic, but
both the mathematics and the theory of computation(s).
2. The traditions of observation and insight that led into the foundations of these sciences.
The semantics of traditional logics is often used but this argument is not sound. The most important for the notation is to be
usable by the computational tool that applies it to the environment
to produce the results.
To share the concern for the rigorous theory it is not necessary to adopt all the amount of any particular formalism. The more
1.2.5
prominent approach seems to be based more on the constraints that
can be superimposed by the problem. If the existing formalism
turns off to match these conditions then that means a perspective to
find out for meaningful thing.
If to confine the search for the theory of objects to areas were
formalism has already succeed in the answer may to be missed. A
necessary theory is likely to be found where the logic meets the incompleteness, troubles of intensions etc. As a rule, the traditional
logical machinery seems to be well applied to pre-formalized reality and is not suitable equipped with the means for more dynamic
occasions. When we go back to the generic principal ideas we have
more possibilities to expand the predefined tools to deal with the
problem as it is arose and used by the computational devices.
1.2.2
A ‘phenomena’ of object was discussed many times with a lot of
attitudes. Some selected and superimposed questions seem to be as
follows:
how new individuals come into existence and go away as situation changed;
how concepts get their semantics in realistic conditions e.g.,
with a tremendous set of possible worlds;
traditional (logical) machineries are usable to prove the existence of an individual (under some properties) but give no
equipment to name that, possibly generated, individual and
refer to it by name in further consideration;
Manifesting a category theory
An early trouble was the suitability of a theory for the working
researcher. The same is for a category theory.
The theoretician position (see [Law75], [Gog89], [EGS91]) seems
to have embraced category theory as the pre-eminent universal science, to adopt its more or less traditional approach with a possible
missing the significant initial features.
The term ‘arrow thinking’ as it is used in a category theory
refers to the standardized notion – within this theory, – that prescribes a mapping of the terms and expressions of the initial system into a world of abstract functions. But it is only one element
of categorical philosophy. In most significant applications of category theory such a thinking does not map symbolic expressions
into real objects with the substantial properties, and such models
become only imaginable.
For some systems of logic model is described by the theory,
e.g. in the form of cartesian closed category (c.c.c.). The need is to
manipulate the elements. A domain T is said to have an element if
there is a map h : I → T (here: I is a domain of assignments). If
f is a constant function f : T → S, then f ◦ h : I → S. Thus,
maps in c.c.c. can behave as functions on elements.
1.2.3
first-order logic provides a tool to support the necessary truths
and their consequences. It provides no machinery to deal
with the relationships with the creation or destruction of individuals;
what is the machinery to characterize the way individuals
change over time;
what is an ability to reify the notion of the state (in different
contexts);
how to talk about both the direct and side effects of the actions;
...
All of this place the state-of-things in the proper perspective.
All of this clearly indicate that the long term hoped-for unified
logic, categorical framework, or computational system is not yet
reached. The variety of logics, theories, models, and approaches
tends to more growth.
2
Applicative computational systems
Restricting the topic: pre-formalization
Some efforts to encircle the task will be needed. Both direct and indirect solutions are substantially based on putting the ideas in a certain order. Subsumption is a common technique shared by distinct
‘dimensions’ – logical, categorical, and computational ([Wol93]).
A lot of theories (not necessary logic or category theory) have the
ultimate goal to develop a notion or construction which suits for
the interpretation of abstract objects. For instance, λ−calculus and
combinatory logic contain in their foundation a concept of object
to suit the computational needs ([Sco80]).
Moreover, an isomorphism between intuitionist logics and typed
λ−calculi was established. An original Curry’s theory of functions
generated formula-as-type notion under which a proof of a statement ∃xB is a pair < a, b > consisting of an object a and a proof
B[a]. In practice, a type is regarded as an abstract object whereas a
formula is the name of a type.
All of this is in harmony with a principal feature of the applicative computations, namely: (1) the symbols of function and its
argument are treated separately as the distinct objects; (2) the first
object is applied to the second under an application metaoperator.
The advantages of this approach are not yet entirely observed.
1.2.4
Introducing abstract objects
2.1
The starting assumptions
Most of the approaches start with the notion of a problem domain.
The problem domain is viewed as a part of physical or imaginable
(perceptual) reality, or external world. This is a natural starting
point. As a result the observer is to operate with a representation.
The represented domain is inhabited by the (atomic) entities, or
individuals. A safety reason is to set up individual as a primary
concept that is not assumed to be definable. In fact, the observer
operates with the constructs that represent the individuals.
Important: The possibility does exist to gather the individuals
into a single domain D, and this D is given from the beginning.
Intermediate theoretical framework
All of the theories above seem to have an universality. The method
is to add the restrictions to enrich the pure theory by the needed
sensitivity.
For instance, the connection between λ−calculus and c.c.c.
([CCM85]) has generated the variants of categorical combinatory
logic. A basic concept for the approach was given by the set of
abstract objects, namely, categorical combinators. This kind of objects is within both a category and computational system. They
share the clear advantages of the distinct subsystems.
The advanced studies in a theory of computations prescribe D
as a domain of potential (or: schematic) individuals. To the contrast the recent object-oriented studies almost ignore this fact. This
ignorance does omit namely the feature of potentiality, or possibility of individual. The individual is possible with respect to some
theory (of individuals).
Advance: The individual may be relativized and gives a family of object-oriented strategies.
236
Filling in the gap: The gap between the observer (and his
language) on the one hand and the individuals on the other
hand does exist in object-oriented modelling.
E.g., ‘this theory of objects is similar to usual’. The individuals
(theories) enter the domain and leave it cancelling their own existence. The ‘flow of events’ in the example may be based on a time
flow. Any two theories are to be compared in spite of their existence in different ‘moments’. The theories are not necessary fixed,
thus all amount of the possible individuals is involved.
An abridgement is given by the evaluation map:
k·k· :
Further advance: The individuals are separated, at least, into
possible and virtual.
descriptions
λ − expressions
× assignments → individuals.
(Here: an assignment is temporary viewed as an index ranging the
families.) The abridged concepts are an attribute a and property
Φ(·) (via the description):
Only the virtual individuals are completely ideal objects. So the
regularity of the observer’s language is increased. In mathematical
practice to be a possible individual means to be described, but the
virtual individual (objects) does need the axioms.
a =k Ix.Φ(x) ki for i ∈ I
Effect: The virtual objects increase the structure regularity of
the (initial) domain D.
(Attr)
An attribute thus defined indicates the set of individuals with a
property Φ(·). In usual terms the functional representation of attribute is established (attribute is a mapping from a set of things and
a set of ‘observation points’ into a set of values). Note that a ‘thing’
is represented by the ‘description’.
As a result, clear distinction between actual, possible and virtual individuals induces the inclusion:
A ⊆ D ⊆ V,
Principle adopted: The attribute is defined by (Attr). The
addition of the uniqueness
where A is a set of actual individuals, D is a set of possible individuals, and V is a set of virtual individuals.
¯ i = true} (Singleton)
{a} = {d ∈ D | kΦ(d)k
Advance: The central computational proposal is to generate
actual individuals as the different families of D,
as necessary and sufficient condition
kIx.Φ(x)ki = a ⇔
¯ i = true} (Unique)
⇔ {a} = {d ∈ D | kΦ(d)k
Ai ⊆ D for i ∈ I.
enforces the observer to conclude: fixing the family i ∈ I and
¯ i relatively to every d ∈ D, he verifies the
evaluating kΦ(d)k
uniqueness of d.
Trouble: The object-oriented approaches propose to operate a
fuzzy notion of a thing and property ignoring the distinctions
between generic and derived concepts. The language of the
observer is likely mixed with the domain D. Thus, the meaning of an individual is violated.
2.2
Here the individual is called as a and is adopted as an evaluation
of the description relatively to i.
Other generic notions
2.3
Starting with things and properties the observer builds the composite things and establishes for his objects the attributes (is there any
object without attribute ?). Thus, an observer actually needs a (logical) language, even overcoming his own initial desire. The obvious
approach is getting started with a choice of logics.
Functional scheme
A general solution for attributes attracts the set of attribute functions (Attr) that is called as a functional scheme.
Advance: Equation (Attr) is to be revised as follows:
Ix.Φ(x) = h̄
kh̄k = h
Trouble: The logics is not homogeneous. Its branches, especially for a theory of computations contain the suitable advances. They do not suit the amorphous idea of a thing and
property.
h(i) = a
Instead of overcoming this barrier theory of computations enables the regular and working logics of the descriptions. The descriptions directly illustrate the difficulties and tend to general operators.
Operating with things and properties gives a specific property
- law. The law is essentially the constraint superimposing to the
properties of a thing.
Recall that in application the observer assigns attributes to things
(they are not the intrinsic to things in contrast to properties).
in a language of observer
is an individual concept
in a domain
is an individual
in a domain
Further advance: Previously given scheme has a universe of
discourse as ‘concept-individual’. An undevoted observer if
needed may prefer the ‘individual-state’ universe.
Thus, if h is an individual, then a is its state under the forcing
condition i.
Advantage: The generalized individuals (or: concepts) are
schematic:
h : I → C,
Important: Both the logical formula Φ(x) and λ−expression
λx.Φ(x) give the property, but the direct assignment of the
property Φ(·) to the individual x is given by the description:
where h is a mapping from the ‘observation points’ into the
(subset of) attribute C. The latter undoubtedly is the set of
individuals.
Ix.Φ(x),
with a sense ‘the (unique) x that Φ(x)’ (compare with λx.Φ(x),
‘those x that Φ(x)’).
There is a clear reason to call h as a concept. Thus a concept
really represent the functional scheme.
237
2.5
Effect: The (individual) functional schemes are to be gathered
into a greater stock:
Dynamics via evolvent
The more dynamics may be added to an object. The task under
solution is a behavior of a thing (= state evolution ‘in a time’).
Note that the state will change due to both the external and internal
events.
{h | h : I → C} = HC (I) (VDom).
Certainly, HC (I) is and idealized object.
Important: The evolvent of stages is needed:
Important: The object HC (I) is a representation, and what is
specific the feature of a variable domain is captured.
f : B → I,
The possibilities and the advantages of a notion of variable domain are applied mostly to the dynamics.
2.4
where stages are evolved from I to B (note the reversed order,
so B is later than I).
Computationally are given: Hg : HC → HE for g : C → E
(C, E are the attributes) and f : B → I for stages I, B. The
combined transformation is generated both by f and g:
Dynamics of objects
The state in an object-oriented approach is viewed as the value of
the functions in the functional scheme at a given point among the
‘observation points’. This agrees with the computational framework.
h ∈ HC (I)
{h(i)} ⊆ C
Important: Computationally a set of individuals is generated
by:
HC ({i}) ⊆ C for i ∈ I.
f
g
2.6
Computationally, an object has: (1) attributes C, E, . . .; (2) transformations g : C → E, . . .; and (3) composable transformations
(possibly, they are closed under composition). In particular, objects
with exclusively interface attributes are viewed as the static objects. This can be modelled by g = 1C : C → C, f = 1I : I → I
etc.
Composition: As usually, the composite object is assumed to
be combined from the other objects.
To cover the possible effects the natural transformations Hg :
HC → HE are added. The element-wise consideration gives:
7→
7→
Object characteristics
Encapsulation: An object contains: (1) state, (2) capability of
transitions (state changes; actions; services), and (3) interface.
Generalization: The notion of a variable domain gives the
natural observation of the dynamics in an object-oriented approach. Even more, it gives a suitable metatheoretic framework.
: h ∈ HC (I)
: {h(i)} ⊆ C
= 1I : I → I,
= 1C : C → C.
The commonly used in object studies are encapsulation, composition, classification, and communication/transaction.
= {h(i)} ⊆ C
= {h(i)} ⊆ E
...
Transformations g : s1 7→ s2 are the counterparts of the events
(they are triples):
< s1 , s2 ; g > .
Hg (I)
Hg ({i})
g ◦ h ◦ f ∈ HE (B),
((g ◦ h) ◦ f )(b) ⊆ E.
for b ∈ B.
In particular, a stable state is generated by:
This set is a state of a variable domain HC (I), where C gives
the local universe of possible individuals. The pointer i marks the
family of individuals that is ‘observed’ from i. The states s1 , s2 , . . .
of a functional scheme have a representation by the stages of the
variable domain:
HC ({i})
HE ({i})
...
7→
7→
This means the following: (1) logics (of the properties) is attached, (2) composition (possibly, in a category) is added etc. All
of this is in a full harmony with the theory of computations.
g ◦ h ∈ HE (I),
(g ◦ h)(i) ⊆ E.
Important: The set of transformations gives the laws of things
in object-oriented reasoning.
Classification: Traditionally, the objects with the same set of
properties (attributes, actions) are gathered into a class.
The immediate result gives a clear understanding of interaction
of things (via state variable common to interacting things). Thus,
the set of natural transformations is a representation of the laws of
. . . . And here is a short diagram of what of:
The computational generalization attracts the concept of a variable domain HC (I) = {h | h : I → C} that is defined over the
schematic objects.
Communication/interaction: Ordinarily communication mainly
implies the changes of the object attributes (change is the same
as a request). A request may cause a state transition (change of
the non-interface attributes; change the state of the receiver/sender
via interface attributes).
{h(i)} ⊆ C
x1 ∈ {h(i)}; x2 ∈ {h(i)}
. . . Φ(x1 )&Ψ(x2 )&x1 = x2 (= z) . . . ,
where z is a common variable (joint state variable).
238
3
3.1.3
Construction of object
Functorial properties
op
Let functor HT in S C be treated as a variable domain: (1) for
every I ∈ C an associated domain HT (I) is the set; (2) the maps
f : B → I in C give transitions from stage I to stage B.
Every transition clones elements in HT (I) into elements in
HT (B) along the map f .
The verification of functorial properties of HT is straightforward. The properties of the restriction come down to the following:
A point of importance to determine an object is the notion of type.
The known results either illustrate the analogy between typed and
type-free models, or establish their real connection. In particular,
untyped models contain object-as-types via embedding. The computation, e.g., in type-free λ-calculus has a goal to derive an object
with the pre-defined properties (dynamic typing). To the contrary,
the same computation in a typed λ-calculus has to obtain the derived type by the rules (static typing).
To conform types with dynamics they are to be fitted the dynamical considerations. The initial set of ‘hard’ types is usually
predefined. To the contrary the ‘soft’ types are derived from the
generic to give rise to a more flexible ground.
Untyped models naturally combine type and its implementation (embedded objects). Sometimes the researcher may prefer to
separate them. As a working hypothesis the thesis ‘to represent
means to classify properly ’ meets the opposition from the alternative approach. This second way tends to the ‘slight’ variations of
the initially formed objects.
3.1
h⌉1I
(h⌉f )⌉g
4
Fragment of a theory of types
Many possible theories of types are known, and the need is of getting down to some details of object-as-functor for types.
The domains A of C are associated to the type symbols, and they
are basic types. The derived types are generated by constructions:
1 (empty product), T × S (cartesian product), T → S (functional
space), [T ] (power type).
In the functor category an arbitrary type T is indicated as HT ,
and an evaluation mapping k · k needs an additional parameter, so
that k·k·. And this is an important stage to treat the functor category
as an interpretation for a higher order theory.
Embedding objects into functor category
4.1
Dynamics: further understanding via logic
The construction of a logical framework reflects the adopted object
solutions.
Object-as-functor
4.1.1
Let a mapping F : C → S be the association to arbitrary domain I
of C a set F (I) of S and to every map f : B → I of C a function
F (f ) : F (I) → F (B) so that:
Logical language
A language contains a supply of variables for every type. Atomic
formulae are the equations:
x = y, where x, y are of the same type;
F (1I ) = 1F (I) , andF (f ◦ g) = F (g) ◦ F (f ),
y = gx, where g is a constant g : T → S of C, x and y have
the types T and S respectively;
provided f : B → I and g : C → B in C.
So defined functor F determines the family of objects parameterized by I.
3.1.2
(HT (1I ))(h)
h ◦ 1I
h,
h⌉(f ◦ g),
where h⌉f = (HT (f ))(h) = h ◦ f is an abbreviation.
Give a construction to accumulate the intuitive reasons above. Let
to consider more than one category. At first, given category C is a
set and is assumed as c.c.c. Let S be the category of all sets and
arbitrary functions, a c.c.c. Construction of the functor category (it
op
is a c.c.c.) S C give all the (contravariant) functors from C into S.
The known result is that the functor category is a model for higher
order logic.
3.1.1
=
=
=
=
z = [x, y], where x, y of types T, S respectively, z of type
T × S;
z = x(y), where x has type T → S, y type T , z type S;
Object-as-domain
y ∈ x, where y is of type T , x of type [T ].
To construe an object that models the meaning of the variable domain an example of functor category is used.
For every T of C let
Formulae Φ are generated as usually by the connectives and
quantifiers.
4.1.2
HT (I) = {h|h : I → T }
Interpretation
Assume the following: A is a domain of C, Φ is a formula, k · k is
an evaluation of the non-bound variables of Φ.
An evaluation of the variable makes k·k relative to the domains
of C (e.g., to A) and needs the explanation.
Visible objects are percepted by the observer via his machinery
in spite of the doctrine of the predefined objects.
The events evolve from A to B. The inhabitants of the world
A evolve, so they inhabit the world B. The world B contains the
clones of A-inhabitants, and also some other inhabitants, if any.
and if f : B → I in C, let HT (f ) be the map taking h ∈ HT (I)
into h ◦ f ∈ HT (B). It is easy to verify HT is a contravariant
functor.
Transactions. Let g : T → S in C. There is a natural transformation Hg : HT → HS . Every h ∈ HT (I) can be mapped to
g ◦ h ∈ HS (I). So defined mapping g determines a rectified idea
of transaction.
Clones. The composite map for f : B → I takes h ∈ HT (I)
into g ◦ h ◦ f ∈ HS (B). Thus, the individuals from HT (I) are
f -cloned into HS (B).
op
It is easy to verify H : C → S C is a covariant functor, and C
may be assumed to be c.c.c.
A
f
←−−−−
H(f )
B
⊆
kx̄kA ∈HT (A) −−−−→ (HT )f −−−−→ HT (B)∋ kȳkB
239
f
A
kx̄kA = kȳkB = kx̄kf B
HC (f )
HC (A) −−−−→ HC (B)
non-cloned, g-transacted:
Hg : HC (A) ∋ h 7→ g ◦ h ∈ HD (A)
A
C
A
C
gy
D
C
gy
D
The ‘transaction-clone’ notion having been applied to the functor
op
category H : C → S C has a benefit of explicate arrow-thinking.
In the following family of diagrams the mapping f : B → A
clones the individual from A into B. Besides that, the mapping
g : C → D represents the transition (an explanatory system is of
free choice):
HC (f )
HD (f )
B
C
HC (f )
Hg (A)y
⊆
HD (f )
H (B)
y g
HD (A) −−−−→ HD (B)
: HC (A) ∋ h 7→ h ◦ f ∈ HC (B)
: HD (A) ∋ h 7→ h ◦ f ∈ HD (B)
HC (A)
⊆
HT (A) −−−−→ (HT )f −−−−→ HT (B)
HT (f )
B
HC (A) −−−−→ HC (B)
A
HT (f )
f
←−−−−
f -cloned, non-transacted:
general diagram:
Hg (A)y
HD (1A )
H (A)
y g
HD (A) −−−−−→ HD (A)
A
Case study for variable domains
T
HC (1A )
Hg (A)y
HD (f ) ◦ Hg : HC (A) ∋ h 7→ g ◦ h ◦ f ∈ HD (B)
(here: C1A = C(A); Cf ⊆ C(B); C = HT )
gy
A
f -cloned, g-transacted:
C(A) −−−−→ Cf −−−−→ C(B)
f
1
A
←−−
−−
HC (A) −−−−−→ HC (A)
⊆
←−−−−
HD (A)
HD (1A ) ◦ Hg : HC (A) ∋ h 7→ g ◦ h ◦ 1A ∈ HD (A)
f
A
y
1A -cloned, g-transacted:
←−−−− B
C(f )
T
Hg (A)
D
The notion of a ‘concept’ depends on a set of conditions and was
studied under the various assumptions. The following matches an
intuition for a ‘variable domain’.
A notational remark. In the below k · k(t/y) means the fixed
evaluation where t matches y of the same type. The evaluation
k · k(t/y) ⌉f = k · kf matches kykf with every relevant variable y.
Any case the restriction ⌉ is superimposed to the functor HT with
T is the type of y.
Let concepts C(A), C(B), and Cf be the different restrictions
of the HT :
C(A) = {t ∈ HT (A) | kΦ(y)k1A (t/y) A}
(Conc(A))
C(B) = {t ∈ HT (B) | kΦ(y)k1B (t/y) B}
(Conc(B))
Cf = {t ∈ HT (B) | kΦ(y)kf (t/y) B}
(Concf )
Their relationships correspond to the diagram:
4.2
HC (A)
gy
Construction of concept
A
B
(1)
The evaluation of the atomic formulae is getting down to the
case study (are given for atomic case).
Variables.
kx = ykA ⇐⇒ kxkA = kykA
(Var)
Constant function.
ky = gxkA ⇐⇒ kykA = g ◦ kxkA
(CFun)
Ordered pair.
kz = [x, y]kA ⇐⇒ kzkA = [kxkA, kykA]
(DPair)
Application (variable function).
(ε)
kz = x(y)kA ⇐⇒ kzkA = kxk1A A(kykA)
Powerset.
(PSet(A))
ky ∈ xkA ⇐⇒ kykA ∈ kxk1A A
(PSet(B))
ky ∈ xkB ⇐⇒ kykB ∈ kxk1B B
ky ∈ xkf ⇐⇒ kykB ∈ kxkf B
( PSetf )
4.1.3
←−−−−
D
H (B)
y g
f
←−−−−
B
HC (f )
−−−−→ HC (B)
HD (f )
HD (AD (A) −−−−→ HD (B)
The functorial properties of HT come down to the case study
given above.
HT (A) −−−−→ (HT )f −−−−→ HT (B)
singular:
4.3
HC (A) = {h | h : A → [C]}
HC (A)
Evaluation mapping
The functor category in use may enrich the intuition concerning an
evaluation mapping. In particular, the diagram given below reflects
f -cloned:
f -cloned evaluation mapping:
HC (f ) : HC (A) ∋ h 7→ h ◦ f ∈ HC (B)
240
Applicator
A
f
←−−−−
εBC : (B → C) × B → C
B
kΦkf
which applies function f to its argument x: εBC : [f, x] 7→ f (x).
Currying
⊆
kΦkA −−−−→ kΦkf
−−−−→ kΦkB
ΛABC : (A × B → C) → (A → (B → C))
{y}
HT (f )
−−−−→ {y ◦ f }
∈y
t
∈y
HT (f )
which shifts variables.
More exactly, if h : A × B → C, then ΛABC h : A → (B →
C).
For k : A → (B → C) and h : A × B → C mapping Λ gives
a correspondence. Equationally, it means
∈y
⊆
HT (A) −−−−→ (HT )f −−−−→ HT (B)
ε◦ < (Λh) ◦ F st, Snd > =
Λ(ε◦ < k ◦ F st, Snd >) =
Similarly, g-transacted, f -cloned evaluation mappings shown
in Fig. 1.
(N.B. Possibly, Ψ may be equal to Φ; y = u, and t = v.)
The interpretation of previous diagram depends on the evailable
engineering machinery.
An advance in the representation may be achieved with the concepts C1 , C2 corresponding to Φ, Ψ respectively.
The previous diagram is comprehenced to:
A
{y}
C
gy
D
∈y
f
←−−−−
C1 (f )
C1 (A) −−−−→
Hg (A)y
C2 (f )
C2 (A) −−−−→
x
∈
{u}
C2 (f )
F st : A × B → A, Snd : A × B → B.
Note, that the equation (ε) may be rewritten:
kzk = ε◦ < kxk1A , kyk >
Next step will be done to determine the meaning of an expression.
−−−−→ {y ◦ f }
C1 (f )
for the first projection F st and second projection Snd:
B
∈y
C1f
C2f
5.1
t
∈y
⊆
−−−−→ C1 (B)
H (B)
y g
⊆
−−−−→ C2 (B)
x
∈
x
∈
−−−−→ {u ◦ f }
f
←−−−−
C1 (f )
C1 (A) −−−−→
C
gy
D
⊆y
v
5.1.1
HC (f )
B
Env = (E × Dy ) × Dx
C1f
⊆y
HD (f )
5.1.2
C1 (B)
⊆y
⊆
kyk : Env → Dy ,
H (B)
y g
⊆
x
⊆
C2 (f )
C2 (A) −−−−→
x
C2f
⊆
Case study
Atomic parts. An object λx.yx contains atoms y, x, and nonatomic part yx:
kxk : Env → Dx .
Non-atomic parts. A non-atomic part yx is evaluated as follows:
HD (A) −−−−→ (HD )f −−−−→ HD (B)
⊆
Building an access
The values of the variables are available via access functions from
an environment. The representation of an environment is given by
the domains Dy , Dx , . . . which are the ranges of possible values
of y, x, . . .. The domains Dy , Dx give the explicit part of an environment Env, and its implicit rest (not be detailed for the current
consideration) is denoted by E:
HC (A) −−−−→ (HC )f −−−−→ HC (B)
Hg (A)y
Meaning of expression
The goal is to determine the meaning of an expression F (x), or F x
where F is the description of a function and x is a formal parameter.
Thus, x is bound, or substitutional variable.
A treatment may be simplified with the λ-notations. The expression above is to be denoted as λx.yx where the description of
a function F is associated to a variable y.
The meaning of a function depends on the meanings of its subparts y, x, yx. Those components, in turn, depend on the value of
y.
The only ‘transaction-clone’ dependencies are visible, so an explicit object is extracted.
Note in addition, that the concept-image of g-transacted, f cloned evaluation mapping:
A
h, and
k
the pair < kyk, kxk > is composed, and
x
< kyk, kxk >: Env → Dy × Dx ;
C2 (B)
the metaoperator ε is applied to the pair:
ε◦ < kyk, kxk > .
is in a harmony with the “logical” diagram in Fig. 1.
D
5
To exemplify let Dy = (Dx → Dy′ ); thus, ε : (Dy′ ) x ×
Dx → Dy′ is determined by ε[u, v] = u(v) = uv, and Dy′ is the
range for kyxk, i.e.
Extracting a computational background
In applications a theory of functions is based on some additional
objects.
kyxk = ε◦ < kyk, kxk >: Env → Dy′
241
A
f
←−−−−
kΦkf
kΦkA −−−−→
{y}
C
gy
D
∈y
HC (f )
−−−−→
HC (f )
HC (A) −−−−→
Hg (A)y
HD (f )
HD (A) −−−−→
x
∈
B
⊆
kΦkf
−−−−→ kΦkB
{y ◦ f }
t
∈y
∈y
−−−−→ HC (B)
(HD )f
−−−−→ HD (B)
⊆
(HC )f
H (B)
y g
⊆
x
∈
HD (f )
∈
{g ◦ y} −−−−→ {(g ◦ y) ◦ f }
k
k
{u}
{u ◦ f }
kΨkf
kΨkA −−−−→
x
v
k
v
⊆
kΨkf
−−−−→ kΨkB
Figure 1: g-transacted, f -cloned evaluation mapping
g
5.1.3
✲
✯
✟
✟✟
✟
✟
ε
✟✟
✟
❄ ✟✟
✟
D
Env × Dx
Substitution
The expression λx.yx contains y (free variable) and does not contain x (bound, or substitutional variable; x may be renamed, if
needed). To take into account this reason the modified environment Env × Dx is temporary generated to support the substitution
Substx :
Substx : Env × Dx → Env,
ĝ × idDx
(Dy′ )
where for i ∈ Env, h′ ∈ Dx the result is
x
Dy′
× Dx
Figure 2: Commutative diagram for kyxk ◦ Substx = g
Substx [i, h′ ] = i(h′ /x) .
5.2
It means that substitution Substx for every ordered pair [i, h′ ] gives
a correspondent environment i(h′ /x) which differs from i exclusively in a point x (x is substituted by h′ ).
An access function for Substx is generated by the equation:
Let kyxk ◦ Substx = g, and
g([i, h′ ]) ∈ Dy′
for g : Env × Dx → Dy′ .
For i ∈ Env and every h′ ∈ Dx the function g is determined
by gi (h′ ) = g([i, h′ ]). Now the function ĝ is defined by the equation ĝ(i) = gi for h′ ∈ Dx . For arbitrary pair [i, h′ ] ∈ Env × Dx
the equation
Substx =< F st ◦ F st, Snd >
5.1.4
Correspondence of the meanings
Composition
An observation is as follows: the function kyxk and Substx are
composed:
ε[ĝ(i), h′ ] = gi (h′ ) = g([i, h′ ])
is valid.
Note, that an operation ˆ· generates the additional metaoperator
Λ of currying:
(Λ(g)(i))(h′ ) = g([i, h′ ])
Hence, a curried version of g = kyxk ◦ Substx is exactly
kλx.yxk, and finally the needed equation is obtained:
kyxk ◦ Substx : Env × Dx → Dy′
The meaning of λx.yx depends on Env for y (y has a free occurrence in λx.yx, and x is bound). Thus, kλx.yxk is a function
that associate to y the function associating yx to x. A type consideration gives:
D
kλx.yxk : Env → (Dy′ ) x
kλx.yxk = Λ(kyxk ◦ Substx )
Let to summarize the above reasons in Fig. 2. In this figure the
following notations are used:
To the contrast kyxk is a function from (E × Dy ) and Dx :
kyxk : (E × Dy ) × Dx → Dy′
g : Env × Dx → Dy′ , i ∈ Env, h′ ∈ Dx
gi : Dx → Dy′ , gi (h′ ) = g([i, h′ ]), ĝ(i) = gi
[i, h′ ] ∈ Env × Dx
ε([ĝ(i), h′ ]) = gi (h′ ) = g([i, h′ ])
Some difficulties exist to establish the correspondence between
meanings kλx.yxk, kyxk, Substx .
242
kλx.yxk = Λ((ε◦ < Snd ◦ F st, Snd >)◦ < F st ◦ F st, Snd >)
ΛkΦk · ×idT
It is easy to verify an optimized version of the access function:
kλx.yxk = Λ(ε◦ < Snd ◦ F st ◦ F st, Snd >)
kΦk[·, ·]
✲ []
✯
✟
✟
✟✟
✟✟
ε
✟
✟
✟✟
❄✟✟
I×T
At last, an access function for kλx.yxk is generated in accordance with the equation:
[T ] × T
from the properties of pairs < ·, · > and composition.
5.3
Figure 3: Evaluation in c.c.c.
Examples
Some examples of computation are briefly given below.
Constant c.
1
2
3
4
kcki =
= k0!k[i, c′ ]
= Snd[i, c′ ]
= c′ (= c)
= ε[kf k[i, h′ ], h′ ]
i ∈ Env, c′ ∈ {c} for singleton {c}
k0!k – a.f. to {c} in Env
= ε[(S ◦ F )(i), h′ ]
= f ′ h′
3 = Λ(kf xk ◦ Substx )ih′ Replace by a.f.
4 = Λ((ε◦ < S ◦ F, S >)◦ < F ◦ F, S >)ih′
(for Substx =< F ◦ F, S >, Substx : Env × Dx → Env)
5 = (ε◦ < S ◦ F, S > ◦ < F ◦ F, S >)[i, h′ ]
(for [i, h′ ] ∈ Env × Dx )
6 = (ε◦ < S ◦ F, S >)(< F ◦ F, S > [i, h′ ])
(Substitution)
7 = (ε◦ < S ◦ F, S >)[F (i), h′ ]
(for F (i) ∈ E × Df , h′ ∈ Dx )
8 = ε[< S ◦ F, S > [F (i), h′a.f.
]]
9 = ε[(S ◦ F )(i), h′ ]
ε;
((S ◦ F )(i) extracts value of Df )
10 = ε[f ′ , h′ ]
11 = f ′ h′
Variable x. The evaluation of a variable gives one of the possible atomic cases. The abbreviations F for F st and S for Snd are
used.
1
kxki =
Generation of a.f. :
= k0!ki(h′ /x)
2
= k0!k[i, h′ ]
= S[i, h′ ]
= h′
= (kxk ◦ Substx )[i, h′ ]
3
4
5
6
= (S◦ < F ◦ F, S >)[i, h′ ]
= S(< F ◦ F, S > [i, h′ ])
= S[F (i), h′ ]
= h′
i ∈ Env
h′ ∈ Dx ;
k0!k – a.f. to Dx in Env
Env = E × Dx
Substx : Env × Dx → Env
Substx =< F ◦ F, S >
Replace by a.f.
Substitution
a.f.
h′ ∈ Dx
5.4
3
4
5
6
7
8
kΦk[i, hi]
ΛkΦki(hi)
k(λx.x)hki=
i ∈ Env
= kλx.xkih′
h′ ∈ Dx
Generation of direct access:
= Λk0!kih′
k0!k – a.f. to Dx in Env,
x – bound variable,
Env = E × Dx
= S[i, h′ ]
[i, h′ ] ∈ Env × Dx
′
=h
= Λ(kxk ◦ Substx )ih′
Using a.f.
= Λ(S◦ < F ◦ F, S >)ih′ Substx : Env × Dx → Env
Substx =< F ◦ F, S >
= (S◦ < F ◦ F, S >)[i, h′ ] [i, h′ ] ∈ Env × Dx
= S(< F ◦ F, S > [i, h′ ])
Substitution
= S[F (i), h′ ]
a.f.
h′
kΦk
= ΛkΦk(F st[i, hi])(Snd[i, hi])
= ΛkΦk(F st[i, hi])(Snd[i, hi])
= ε[ΛkΦk(F st[i, hi]), (Snd[i, hi])]
= (ε◦ < ΛkΦk ◦ F st, id ◦ Snd >)[i, hi]
= kΦk[i, hi]
= ε◦ < ΛkΦk ◦ F st, id ◦ Snd >
An abbreviation
kΦk = kΦ(x)k ◦ Substx
is used if there is no ambiguity. Hereafter T is a type of substitutional variable x, and environment Env is renamed by I.
Evaluation in c.c.c. The diagram in Fig. 3 illustrates an idea.
• ΛkΦk : I → [T ];
• kΦk : I × T → [ ];
Compound evaluation.
1
2
Advanced examples
The additional examples of generalized nature involve more complicated objects.
Evaluation of formula. This kind of object has the following
equations:
Identity transformation. The evaluation of an identity transformation gives a clear separation of access functions (a.f.) and
substitution.
1
2
kf k – a.f. to Df in Env × Dx ,
i.e. kf k = S ◦ F ◦ F
(S ◦ F )(i) ∈ Df
• For i ∈ I and hi ∈ T an evaluation ε[ΛkΦki, idT (hi)] generates the truth values from [ ].
k(λx.f x)hki =
i ∈ Env
= k(λx.f x)kih′
h′ ∈ Dx
Generation of access:
= Λkf 0!kih′
k0!k – a.f. to Dx in Env
Env = (E × Df ) × Dx
= kf 0!k[i, h′ ]
[i, h′ ] ∈ Env × Dx
′
= (ε◦ < kf k, S >)[i, h ] S(i) ∈ Dx
Individuals in c.c.c. A correspondence of the distinct forms of
individuals shows their similarities.
R ❀ hR . Given the relation R ⊆ I × T a function hR : I →
[T ] is determined by the equality hR (i) = {h′ | h′ ∈ T ∧ iRh′ }.
In fact, this defines the correspondence R ❀ hR .
243
✄
✂
R
✲
I×T
Λ(Snd) · ×idT
g
hR × idT
❄✄
✂
❄
✲
∈T
Snd
✲
✯
✟✟
✟
✟
ε
✟✟
✟
✟
❄ ✟✟
✟
I×T
T
TT × T
[T ] × T
Figure 7: Free variable
Figure 4: Variants of individuals
R
✄
✂
✲
I×T
ΛkΦk · ×idT
C · ×idT
❄✄
✂
✲
∈T
kΦk[·, ·]
ε
✚
✚
❄✚
✲
❃
✚
✚
✚
✚
✚
[ ]
right-part computations are to be started at the same ‘moment’. An
additional mappings canT of canonical embedding of the constants
are also used.
The more exact correspondences are as follow:
R
ΛkΦki
ΛkΦk
C({i})
C(I)
[T ] × T
Figure 5: Computational properties
Λ(f ◦ Snd) · ×idT
Snd
✲
T
S
6
Important: The notion of a variable domain gives a sound
ground of the communication analysis (see, e.g.: [WW94],
[Jac92]). As may be shown they generate the specific diagrams
to consider the variety of transition effects.
Figure 6: Built-in function f
Open discussion: The questions arise:
1. Is the language of categories adequate to database dynamics even though the object-oriented approach successively applied?
h ❀ Rh . Given the sets I, T the bijection between functions
from I into [T ] and the relations from I to T is defined as follows.
The function h : I → [T ] determines the relation Rh ⊆ I × T by
the biconditional iRh h′′ ⇐⇒ h′′ ∈ h(i) for i ∈ I and h′′ ∈ T .
∈T . The domain ∈T = {< U, h′ >| U ⊆ T, h′ ∈ T ∧ h′ ∈
U } is the relation containing all the necessary information concerning element-subset inclusions. The following biconditionals
are valid:
′
Conclusions
A common object technique shared by distinct ‘dimensions’ – logical, categorical, and computational is outlined.
ST × T
′
{[i, h′ ] | kΦk[i, h′ ] = 1}
[T ]
I → [T ]
{h(i) | ΛkΦki(hi) = 1}
{h | kΦ(x)k[h/x] : I → [T ]}
(here: x, h : I → [T ], so h(i) ⊆ T ; x is a free variable.)
f
✲
✯
✟✟
✟
✟
ε
✟✟
✟✟
❄ ✟✟
✟
I×T
=
∈
:
=
=
References
[Bro94]
M. L. Brodie. Interoperable information systems: Motivations, challenges, approaches, and status. In Second International Conference on Cooperative Information Systems, CoopIS-94, Tutorial Notes. May 17–20,
1994, Toronto, Ontario, Canada, May 1994.
[Bro95]
M.L. Brodie. Interoperable Information Systems: Motivations, Challenges, Approaches, and Status. Russian
Basic Research Foundation, Moscow, Russia, April, 67, 1995.
′
[i, h ] ∈ R ⇐⇒ h ∈ hR (i) ⇐⇒ [hR (i), h ] ∈ ∈T
Hence, R is a domain and ∈T is a range for mapping hR × 1T
where hR × 1T : [i, h′ ] 7→ [hR (i), h′ ].
The diagram in Fig. 4 reflects the ideas given above.
Here: g is an R-restricted version of hR × idT . Note that all of
this is quite elementary.
Computational properties of the individuals. The combined diagram in Fig. 5 establishes not so evident correspondences. What
is important that the functor C · ×idT includes as a left counterpart
the mapping C· : I → [T ]. This mapping is relative to relation R
and this relation is induced by the evaluation of the restriction Φ.
In particular, a built-in function for the given (and evaluated)
argument in a category results in the diagram in Fig. 6.
A free variable is evaluated according to the diagram in Fig. 7.
A simplified example of computation (note that both the operands
are to be embedded into the computational environment) like
+[[2/x1 ]x1 , [3/x2 ]x2 ]
is in Fig. 8. The entry points for the computations of the distinct
operands are, in general, independent. Thus, both the left-part and
244
[CCM85] G. Cousineau, P.-L. Curien, and M. Mauny. The categorical abstract machine. In Functional programming
languages computer architecture, volume 201 of Lecture Notes in Computer Science, pages 50–64. Heidelberg, Springer-Verlag, 1985. This is a detaled paper on
programming in a category-style.
[EGS91] H.-D. Ehrich, M. Gogola, and A. Sernadas. A categorial theory of objects as observed processes. In J.W.
deBakker et. al., editor, Proceedings of the REX/FOOL
School/Workshop, volume 489 of Lecture Notes in Computer Science, pages 203–228. Berlin, Heidelberg, New
York, Springer Verlag, 1991.
entry point 1
[i, 3]
I×T
I×T
entry point 2
❍
❍❍Snd
❍❍
❥
❍
✟
Snd ✟
✟
✟
✙✟
✟
2
✟✟
✟
✟
✟
✙
✟
✛
[i, 2]
T
canT
2
T
3
T
Λ(canT ◦ Snd) · ×idT
ε
Λ(canT ◦ Snd) · ×idT
❄ ❄
TT × T
TT × T
❍❍ canT
❍❍
❍❍
❥
ε
✲
3
T
✿
❳
②
✘✘
❳❳
✘✘
❳❳❳
✘
✘
❳❳ F st
Snd ✘✘✘
❳❳❳
✘✘
✘
❳❳❳
✘
[2, 3]
5
✘
❳❳
✘✘
+
Snd
✲ T ×T
✲ T
I × (T × T )
✶
✏✏
✏✏
✏
✏
✏✏
✏
✏
ε ✏
Λ(+ ◦ Snd) · ×idT ×T
✏✏
✏
✏✏
❄ ✏✏✏
✏
T T ×T × (T × T )
Figure 8: An example of computation +[2, 3]
[EGS93] H.-D. Ehrich, M. Gogola, and A. Sernadas. Objects and
their specification. In M. Bidoit and C. Choppy, editors,
Recent Trends in Data Type Specification. 8th Workshop on Specification of Abstract Data Types joint with
the 3rd COMPASS Workshop, Dourdan, France, August
26–30, 1991, Selected Papers, volume 655 of Lecture
Notes in Computer Science, pages 40–65. Berlin, Heidelberg, New York, Springer Verlag, 1993.
[Gab93]
and Information Systems, ADBIS’95, page this volume.
Russian Academy of Sciences, Institute for Problems of
Informatics, 1995.
P. Gabriel. The object-based specification language Π:
Concepts, syntax, and semantics. In M. Bidoit and
C. Choppy, editors, Recent Trends in Data Type Specification. 8th Workshop on Specification of Abstract Data
Types joint with the 3rd COMPASS Workshop, Dourdan, France, August 26–30, 1991, Selected Papers, volume 655 of Lecture Notes in Computer Science, pages
254–270. Berlin, Heidelberg, New York, Springer Verlag, 1993.
[Gog89]
J. Goguen. A categorical manifesto. Technical report
PRG-72, Programming Research Group, Oxrofd University, March 1989.
[HC89]
F. Heyes and D. Coleman. Objects and inheritance: an
algebraic view. Technical Memo, HP labs, Information
Management Lab, Bristol, 1989.
[Jac92]
I. Jacobson. Object-Oriented Software Engineering: A
Use-Case Driven Approach. Addison-Wesley, Reading,
Massachusetts, 1992.
D.S. Scott. Relating theories of the λ-calculus. In
J. Hinhley and J. Seldin, editors, To H.B. Curry: Essays on combinatory logic, lambda calculus and formalism, pages 403–450. New York and London, Academic
Press, 1980.
[Wol93]
V.E. Wolfengagen. Computational aspects of data objects. In Proceedings of the workshop on Advances in
DataBase and Information Systems, ADBIS’93, May 1114, Moscow, 1993, Moscow, May 1993.
[WW94] Y. Wand and C. C. Woo. Object oriented analysis of organizational activities: A CoopIS tutorial. In Second
International Conference on Cooperative Information
Systems, CoopIS-94, Tutorial Notes. May 17–20, 1994,
Toronto, Ontario, Canada, May 1994.
[Law75] F.W. Lawvere. Continuously variable sets: algebraic geometry = geometric logic. In H.E. Rose and J.C. Shepherdson, editors, Logic Colloquium ’73, pages 135–156.
North Holland, Amsterdam, 1975.
[NR95]
[Sco80]
D.A. Nelson and B.N. Rossiter. Prototyping a categorical database in P/FDM. In L.A. Kalinichenko, editor,
Proceedings of the workshop on Advances in DataBases
245
| 6cs.PL
|
Automatic Test Data Generation and Model
Checking with CHR
Ralf Gerlich
BSSE, Auf dem Ruhbuehl 181, D-88090 Immenstaad, Germany,
[email protected]
arXiv:1406.2122v1 [cs.SE] 9 Jun 2014
Abstract. We present an example for application of Constraint Handling
Rules to automated test data generation and model checking in verification
of mission critical software for satellite control.
1
Overview
Verification and validation of software takes up a large proportion of project effort
and cost, especially in the area of mission and safety critical software. This is one of
the driving forces for automation of these aspects of software engineering, besides
the reduction of the potential for errors due to manual labor.
Automation of software testing requires among others the automatic generation
of test data. Some options for this are random test data generation [3,8,9,15] or
constraint-based test-data generation (CBDTG) [4,6,10,12,14,16,19].
We have designed and implemented a toolchain for CBTDG using Constraint
Handling Rules (CHR), which we have already presented in the past [10].(in [13]
and [4] the two examples of the paper are noted using typica
In a recent case-study on the effectiveness of source-code-based random testing [11], we have also seen that the basic elements of this toolchain can be used for
support in manual verification of defects in mission critical satellite control software
and that a use for model checking seems plausible.
2
Introduction
Constraint-based test data generation is concerned with the generation of program
inputs for use in software test using constraint programming techniques. The goal
is to find program inputs which fulfill specific criteria, typically from structural
coverage goals such as executing a specific portion of the program under test.
In a first step, in our approach – similar to that of others – the control flow
graph of the function under test is used to construct a path that fulfills the request.
For each path an associated path constraint can be constructed, which is actually a
set of constraints.
An example control-flow graph for an implementation of Euclid’s algorithm for
determination of the greatest common divisor of two positive integers is given in
Fig. 1. Execution starts at node 1 and continues along the edges until node 6 is
reached. Nodes and edges are annotated by statements and conditions, respectively.
An edge may be traversed only if the condition attached to it is fulfilled.
The path constraint is a constraint over the input variables of the function, and
its solutions are the inputs that fulfill the given criteria. Thus, in a second step, a
solution to the path constraint is sought.
There are paths in the control flow graph which are associated with a path
constraint without solution. These paths are called infeasible and – unfortunately –
may represent a large portion of all paths, so that a simple trial-and-error-approach
to path construction is insufficient.
Proceedings of 11th Workshop on Constraint Handling Rules (CHR 2014), paper CHR/2014/1.
1
a=b
6
2
a<b
a>b
3
4
a:=a-b
b:=b-a
5
Fig. 1. Control Flow Graph
Instead, infeasible paths should be detected early in the path construction phase.
This requires both a useful strategy of path construction and a constraint solver
which is efficient at detecting inconsistencies.
The criteria applied to the desired input can also be described in the form of
constraints, allowing integration of the structural goals for the test input with, e.g.,
arithmetic constraints on the state of program variables at specific stations during
execution. This combination also makes the use of the same approach for static
verification using symbolic enumeration of the state space possible, as will become
apparent from our example.
This paper is structured as follows: In Section 3 we detail the constraint solver
used and the diverse requirements it must fulfill. This is followed by a short discussion on the reasons for applying CHR for implementation in Section 4 and the
presentation of open problems in Section 5. Finally, we present an example of use
in Section 6 and briefly conclude in Section 7.
3
Constraint Solver Approach
There are three constraint-based functional aspects of our approach, all of which we
implemented using CHR [1,7], based on the CHR-compiler included in SWI-Prolog:
– path construction,
– satisfiability checking, and
– constraint solving.
Due to the issue of infeasible paths, at least path construction and satisfiability
checking need to be combined in order to facilitate detection of infeasible paths
early during path construction. However, the coupling can be loose in that the path
construction phase may be implemented as a separate CHR solver in which the
constraints handled by the satisfiability checker are modelled as builtin constraints.
The constraints to be handled represent all relevant expressions and conditions
from the underlying language, also covering the different arithmetic types, typically
consisting of bounded integer and floating point types. Information has to be propagated bidirectionally across the borders of these arithmetic theories, for example,
when integer values are converted to floating point values or vice versa.
Control-flow is completely represented in the structure of the control-flow graph.
As such, control-flow is current only considered in the path construction module,
but not in any of the other constraint solving modules.
The same is true for boolean operations such as conjunction, disjunction and
negation. Most if not all practically relevant programming languages define the
evaluation of boolean constructs such that whenever the value of the first operand
2
of a boolean operator completely determines the result, the second operand is not
evaluated. For example, in the expression A && B, the second operand B is not
evaluated if the first, A evaluates to false, as in this case the result of the whole
expression is already determined to be false.
In code generation, this is achieved by so-called short-circuit code, which is also
applied in our test-data generator.
Because interaction between arithmetic domains may be necessary in presence
of type conversions, all arithmetic constraints are handled by a combined constraint
handler.
Currently, floating-point constraints are handled by interval filtering [2,17].
3.1
Path Construction
Our approach to path construction has been previously described in detail [10] and
shall be only briefly revisited here.
The path construction approach is centered around a constraint of the form
path(In,A,B,Out), meaning: There is a path from node A to B transforming memory state In to memory state Out. The constraint is limited in that A and B must
be ground.
Construction may proceed in one of several ways:
– forward construction: Construct the path edge-by-edge in the normal direction
of execution, starting at A.
– backward construction: Construct the path edge-by-edge in reverse direction,
starting at B.
– splitting: Given a node C that is (topologically) reachable from A and from which
B can be reached, split up the path from A to B into two paths, one from A to C
and one from C to B.
All of the three strategies can be described in terms of CHR simplification rules.
However, non-determinism is present in that there may be more than one candidate
for the successor, predecessor or intermediate node, respectively.
In absence of any further information indicating paths which may be more likely
to help detecting coding defects, bias must be avoided. To achieve this, alternatives
are selected randomly, but not necessarily according to a uniform distribution to
achieve appropriate distribution, e.g., of loop iteration counts and total length of
the constructed path.
Experience so far indicates that backward construction is most effective in terms
of expected time to a feasible path.
3.2
Linear Integer Constraints
Linear integer arithmetic is expressed in equations and inequations
Pn of the form e = 0
and e ≥ 0, where e is a Presburger expression of the form c0 + i=1 ci vi with integer
constants ci and logical variables vi .
For example, the assignment a = 2b + c is represented as a − 2b − c = 0 or
−a + 2b + c = 0. Similarly, 2a + b > 3c is represented as −1 + 2a + b − 3c ≥ 0.
Over the reals, constraint systems of this form could be solved by a combination of Gaussian elimination for the equations and Fourier-Motzkin-elimination for
the inequations. However, Gaussian elimination requires the existence of the multiplicative inverse and Fourier-Motzkin-elimination requires the compactness of the
underlying number set, both of which are not given for the integers.
3
We therefore use an algorithm known as the Omega Test [18]. In this algorithm
first the equations are simplified using suitable parameter substitutions. For example, the equation a − 2b = 0 is trivially processed by substituting a = 2b in all other
constraints.
Equations which cannot be transformed trivially this way are modified by introducing suitable parameter substitions. For example, the equation 3a − 2b = 0 is
equivalent to a − 2α = 0 ∧ b = 3α, with α ∈ Z. This can be further simplified by
substiting a = 2α, leading to the properly parameterised solution a = 2α ∧ b = 3α.
Inequations are simplified using variable elimination very similar to FourierMotzkin-elimination. There,
a variable w can be eliminated
from a pair of inequaP
P
tions of the form c0 + i=1 ci vi < aw P
and bw < d0 + j=1
Pdj vj – with a, b > 0 –
by introducing a new inequation bc0 + i=1 bci vi < ad0 + j=1 adj vj .
Processing all such pairings in this way, all occurrences of w can be eliminated,
and the new set of inequations is equivalent to the original set in terms of satisfiability.
Repeating this process for all variables except for one, the original problem is
reduced to two inequations of the form xl < x < xu . Selecting a value for x from this
range and substituting it back into all inequations eliminates the inequations over x
and leaves two inequations yl < y < yu with y being the variable eliminated secondto-last. A solution for the original set of inequations can be found by repeating this
process of selection and substition until all free variables are bound.
Unfortunately, this only works for compact number sets, i.e. sets where for any
α < γ there is a β with α < β < γ. This is not the case for the integers.
As a consequence, neither the equivalence regarding satisfiability nor the process
for labelling applies when using Fourier-Motzkin-elimination in its usual form. It is
therefore possible that the range for the last variable x is found to be non-empty,
but none of the values from this range are part of a solution.
The Omega-Test [18] therefore uses a modification of Fourier-Motzkin, underapproximating the set of solutions on each elimination step. Now any solution of
the new set of inequations can be extended to a solution of the original set of
inequations.
However, satisfiability of the original set of inequations does not generally imply
satisfiability of the inequations after elimination: As the approximation step may
exclude some solutions, these have to be considered separately and in addition to
the usual elimination approach.
Depending on the value of the coefficients a and b, this set of solutions removed
by approximation may grow arbitrarily large. However, by carefully selecting the
order in which the variables are eliminated, the number of solutions to be considered
in addition can be reduced significantly.
The algorithm processing equality constraints can be implemented as an onlinesolver, processing each constraint when it is added to the goal store. This is of
advantage for path construction, where each new step in the program introduces
new constraints and requires a new satisfiability check.
Because the quality of elimination results for inequations depends on the order
in which variables are eliminated, this process is at best difficult to implement for
online processing. New constraints – both equations and inequations – may impact
the order in which variables need to be eliminated to reduce the impact of the
approximation. Ad-hoc reordering may thus be necessary.
3.3
Non-Linear Integer Constraints
Non-linear integer constraints are handled by a combination of interval filtering and
dynamic linear relaxations [5].
4
Linear relaxations are approximations of non-linear constraints. For example, the
intervals x ∈ [xl ; xu ] and y ∈ [yl ; yu ] imply that the inequation (x − xl ) (y − yl ) ≥ 0
is fulfilled. Expanding the expression on the left hand side and using the relationship
z = xy, this can be transformed to z−xl y−xyl +xl yl ≥ 0, which is a linear inequation
in x, y and z.
In a similar manner, the constraints −z + xyl + xu y − xu yl ≥ 0, −z + xyu +
xl y − xl yu ≥ 0 and z − xyu − xu y + xu yu ≥ 0 can be derived.
These constraints overapproximate the set of solutions for x ∈ [xl ; xu ] ∧ y ∈
[yl ; yu ] ∧ z = xy, but they contain more information about the relationship between
x, y and z than the simple interval constraint z ∈ [zl ; zu ], where zl and zu are
derived from xl , xu , yl and yu .
Dynamic linear relaxations are updated whenever the source information – in
this case the intervals of x and y – change.
4
Application of CHR
As has been discussed in previous sections, the constraint solver required for a
constraint-based test data generator has to handle a variety of constraint types
from different constraint theories normally considered in isolation. Also, different
handling strategies are required at different stages of the test data generation process: During path construction, the focus is on satisfiability checking. As soon as
a path which is expected to be feasible with sufficient probability is identified, the
focus shifts to selection of an actual solution.
Due to the loose coupling possible by way of a rule-based specification concept,
CHR allows almost straight-forward integration of the different constraint handlers
and of different approaches for solving constraints from the same theory.
One example is the integration of dedicated solution strategies for linear integer
constraints besides the more general domain filtering approach for non-linear integer
constraints.
This advantage, however, can only be realised for mostly local propagation and
simplification strategies, such as the approach to solving linear equations over the
integers or domain filtering.
Other strategies such as the elimination procedure of the Omega-Test for linear
integer inequations are highly sequential in nature. Naturally, for implementation
of these strategies, imperative languages – which CHR is not – are more suited.
5
Open Problems
So far the approach has not yet been used for test data generation on industrialgrade software due to several open issues, mainly lack of scalability of the constraint
solvers and missing support for arithmetic constraints over floating point numbers.
Industrial software may be large and contain many interactions between different
functions, leading to a large number of constraints being generated for a single testdata generation run. The constraints may be highly coupled, because many of them
refer to a small set of variables, namely those variables representing the input values
to the function. Therefore, the issue of scalability is inherent in the problem itself.
Floating point arithmetic plays an increasing role also in embedded software
systems in general and in space control software in particular, as more and more
embedded hardware platforms have builtin floating point units, thereby gradually
replacing the old fixed-point arithmetic implementations.
The theory of constraints over the floats is strictly separate from the theory of
constraints over the reals. This becomes obvious when comparing the magnitude of
the underlying sets of numbers: While the reals are non-countable infinite, the floats
5
are countable and even finite. This means that a constraint system over the reals
may have a solution while the same system is inconsistent when expressed over the
floats, and vice versa.
Also, due to the significand-exponent-representation, all operations are nonlinear. In addition, only elementary arithmetic, remainder and square-root have
standardised results. Others, such as the trigonometric functions, are not standardised, the Table Maker’s Dilemma being among the reasons [13].
Some exact domain-filtering approaches to the solution of floating point constraints exist [2,17], but their filtering efficiency is insufficient in many situations.
It is quite conceivable that from a theoretical point of view, these problems do
not have efficient solutions or any solution at all. After all, the problem of analytical
test-data generation itself cannot be solved in general, as it can be reduced to the
halting problem.
However, in practice not the whole set of theoretically applicable
operations and
p
their combinations is used. For example, inequations such as x2 + y 2 < 10−7 may
be expected to occur much more often in a practical context than x = x + y ∧ y 6= 0
– which, due to rounding, has a floating point solution, but requires much more
computation time to solve than the latter with current filtering approaches.
As another example, the issue of missing scalability is inherent in the problem,
but only when purely focusing on the theoretical description of the problem. It is
likely that in practice the path taken is not influenced by all arithmetic operations
performed. Thus, lazy evaluation – i.e. introduction of constraints only if they are
part of conditions or one or more of their variables are used directly or indirectly
in a condition – may allow for a considerable reduction of complexity in practice,
although clearly it would not directly solve the scalability issue once and for all.
Unfortunately, it is quite difficult to define the actual domain of constraint systems to be expected in practice.
6
Example of Use
The approach has not yet been applied for test data generation in practice on
industry-grade source-code, mostly due to the open problems stated in Section 5.
However, its basic elements were used in the context of a study on the effectiveness of sourcecode-based random test data generation [11]. In the course of this
study, random test inputs were injected into functions found in the sourcecode of the
control software of an earth observation satellite using the tool DCRTT (Dynamic
C Random Test Tool) developed and maintained in-house at BSSE. Notably, that
control software had previously gone through the rigorous testing and validation
stages typical for mission-critical systems, i.e. systems the failure of which could
lead to a loss of the satellite of the complete loss of its functionality.
The code was instrumented by the random test tool to monitor for non-specific
indications of failures, such as memory access violations, time outs and similar. Such
indications were seen as hints at faults, each of which had to be verified manually to
determine whether there was an actual defect or whether the indication was actually
a false positive.
In one case, a memory access failure hinted at a critical defect in code related
to the installation of in-flight software updates. A much simplified and anonymised
excerpt of the code is given in Listing 1. The goal of the function is to store data
of a given length in a contiguous block at the next free position in the buffer. If the
data block does not fit anymore at the end of the buffer, it shall be stored at the
beginning.
This code seems short and simple enough to analyse. However, the defect is
non-obvious and the presence of the remainder-operation in combination with a
6
#define MAX_BUFFER_SIZE ...
char buffer[MAX_BUFFER_SIZE];
void store_into_buffer(char* data, unsigned int length) {
const unsigned int last_entry_start = ..., last_entry_length = ...;
unsigned int next_entry_start = last_entry_start+last_entry_length;
unsigned int space_available;
if ((MAX_BUFFER_SIZE - (length-1u)) < next_entry_start)
next_entry_start = 0;
space_available = (last_entry_start - next_entry_start) % MAX_BUFFER_SIZE;
if (space_available >= length)
memcpy(&buffer[next_entry_start],data,length);
...
}
Listing 1: Relevant portions of the faulty function
set of choices introduces complexity. Manual analysis had led to a suspicion for
overflowing the buffer in the call to memcpy, but was inconclusive both regarding
the validity of the suspicion and the possible extent of the overflow. Due to the
complexity, it was not clear whether the results of the analysis were to be trusted.
The matter was settled by providing the path construction part of the test
data generator with the goal to reach the call to memcpy, adding the constraint
next_entry_start+length>MAX_BUFFER_SIZE. Within a second, a descriptive solution was provided, and that solution indicated the potential for a buffer overflow by one byte, namely when next_entry-start==MAX_BUFFER_SIZE-length+1,
last_entry_start>0 and last_entry_start<length-1.
In this case, there is not sufficient space at the end of the buffer, but the condition (MAX_BUFFER_SIZE-(length-1u))<next_entry_start is false. Therefore the
algorithm fails to reset next_entry_start to the start of the buffer, and the call
to memcpy leads to a buffer overflow by exactly one byte.
Further manual analysis led to the conclusion that the one-byte buffer overflow could lead to corruption of volatile data and non-volatile program memory.
Although there was fallback software present for this case – the so-called safe-mode
software – the satellite could have unexpectedly become unresponsive for at least
some significant time frame.
We know that in the same project, static verification tools using abstract interpretation have been used to verify the code. In principle, these tools should have
detected the defect by themselves. It is still not clear whether the tools failed to
flag that defect or whether the message got lost in a large number of messages of
possible false positives and was therefore not considered.
As a consequence of our report, the defect was fixed. Also, the instrumentation
of the random testing tool was extended to check for such cases, which has led to
detection of further defects of the same kind.
7
Conclusion
The experience so far has shown that CHR is well-suited for developing complex
constraint solvers based on local solution strategies, as detailed in Section 4. The
correspondence between declarative and operational semantics is helpful in verification of the constraint solver itself.
However, that correspondence often has to be broken in workarounds whenever
global strategies need to be implemented, as is the case for the Omega Test.
7
Further research is required for solving the open problems regarding scalability
and handling of floating point constraints.
8
Acknowledgements
BSSE is currently performing research on open aspects of industrial-grade CBTDG,
supported by a grant by the German federal government under the grant number
50RA1339.
References
1. Abdennadher, S., Schütz, H.: CHR∨ : A flexible query language. In: Proceedings of
Third International Conference on Flexible Query Answering Systems. Lecture Notes
in Artificial Intelligence, vol. 1495. Springer Verlag (May 1998)
2. Botella, B., Gotlieb, A., Michel, C.: Symbolic execution of floating-point computations.
Softw. Test. Verif. Reliab. 16(2), 97–121 (2006)
3. Chen, T.Y., Huang, D.H., Kuo, F.C.: Adaptive random testing by balancing. In: RT
’07: Proceedings of the 2nd international workshop on Random testing. pp. 2–9. ACM
(2007)
4. Denise, A., Gaudel, M.C., Gouraud, S.D.: A generic method for statistical testing.
In: Proceedings of the 15th IEEE International Symposium on Software Reliability
Engineering (ISSRE). pp. 25–34. IEEE (2004)
5. Denmat, T., Gotlieb, A., Ducassé, M.: Improving constraint-based testing with dynamic linear relaxations. In: Proceedings of the The 18th IEEE International Symposium on Software Reliability. pp. 181–190. ISSRE ’07, IEEE Computer Society,
Washington, DC, USA (2007)
6. Ferguson, R., Korel, B.: The chaining approach for software test data generation. ACM
Trans. Softw. Eng. Methodol. 5(1), 63–86 (1996)
7. Fruehwirth, T.: Constraint Handling Rules. Cambridge University Press, 1st edn.
(2009)
8. Gerlich, R., Fercher, G.: A random-testing environment for Ada programs. In: Eurospace Symposium ”Ada in Aerospace” (1993)
9. Gerlich, R.: Size-optimising automatic random testcase generation for non-formal conditions. Diploma thesis, Universität Ulm (2005)
10. Gerlich, R.: Generic and extensible automatic test data generation for safety critical
software in CHR. In: Van Weert, P., De Koninck, L. (eds.) CHR’10: Proceedings of
the 7th Workshop on Constraint Handling Rules. K.U. Leuven (July 2010)
11. Gerlich, R., Gerlich, R., Kvinnesland, K., Johansen, B.S., Prochazka, M.: A case study
on automated source-code-based testing methods. In: Proceedings of DASIA 2013
DAta Systems in Aerospace (2013)
12. Godefroid, P., Klarlund, N., Sen, K.: DART: directed automated random testing.
In: PLDI ’05: Proceedings of the 2005 ACM SIGPLAN conference on Programming
language design and implementation. pp. 213–223. ACM (2005)
13. Goldberg, D.: What every computer scientist should know about floating-point arithmetic. ACM Comput. Surv. 23(1), 5–48 (1991)
14. Gotlieb, A., Botella, B., Rueher, M.: Automatic test data generation using constraint
solving techniques. SIGSOFT Softw. Eng. Notes 23(2), 53–62 (1998)
15. Hamlet, R.: Random testing. In: Marciniak, J. (ed.) Encyclopedia of Software Engineering, pp. 970–978. Wiley (1994)
16. Korel, B.: Automated software test data generation. IEEE Trans. Softw. Eng. 16(8),
870–879 (1990)
17. Michel, C.: Exact projection functions for floating point number constraints. In: Proceedings of seventh AIMA Symposium (2002)
18. Pugh, W.: The Omega Test: a fast and practical integer programming algorithm for
dependence analysis. Communication of the ACM 8, 102–114 (August 1992)
19. Sy, N.T., Deville, Y.: Automatic test data generation for programs with integer and
float variables. In: Proceedings of the 16th IEEE International Conference on Automated Software Engineering (ASE). pp. 13–21. IEEE (2001)
8
| 6cs.PL
|
Analytical solution of transient scalar wave and diffusion problems of
arbitrary dimensionality and geometry by RBF wavelet series
W. Chen*
Department of Informatics, University of Oslo, P.O.Box 1080, Blindern, 0316 Oslo, Norway
Email: [email protected]
Summary
This study applies the RBF wavelet series to the evaluation of analytical solutions of linear
time-dependent wave and diffusion problems of any dimensionality and geometry. To the best
of the author’s knowledge, such analytical solutions have never been achieved before. The
RBF wavelets can be understood an alternative for multidimensional problems to the standard
Fourier series via fundamental and general solutions of partial differential equation. The
present RBF wavelets are infinitely differential, compactly supported, orthogonal over
different scales and very simple. The rigorous mathematical proof of completeness and
convergence is still missing in this study. The present work may open a new window to
numerical solution and theoretical analysis of many other high-dimensional time-dependent
PDE problems under arbitrary geometry.
Key words: RBF wavelet series, Helmholtz equation, Fourier series, Gibbs phenomenon,
time-dependent problem, high-dimensional problems, geometric complexity.
*
This research is supported by Norwegian Research Council.
Contents
1. Fourier series, a historic retrospect…………………………………………..1
2. Radial basis function and wavelets…………………………………………..3
3. Analytical solution of transient wave problem with RBF wavelet series……7
3.1. Helmholtz eigenvalues and eigenfunctions with RBF…………………….…8
3.2. Analytical solution with RBF wavelet series………………………….……10
4. Applications to inhomogeneous problems…………………………………..13
5. Extension to other time-dependent equations……………………………….14
6. Generalized RBF wavelet series and transforms……………………………14
7. Promises and open problems………………………………………………..16
References……………………………………………………………………...18
1. Fourier series, a historic retrospect
Many of the important concepts of analysis and computation have their origins in the
study of physical problems leading to the partial differential equation (PDE) systems [1].
The currently ubiquitous Fourier series and transform came from Fourier’s original
exploration of the solution of a bar heat transmission problem in the early 1800s. What
Fourier proposed due to this quest is that an arbitrary 1D function f(x) over a bounded
interval, even if not differentiable, can be represented by an infinite sum of sinusoids
∞
f ( x ) = ∑ c k sin
k =1
kπ
x,
a
(1)
where sinusoids are the eigenfunctions of this PDE problem. Despite a lack of rigorous
proof, Fourier was quite confident of the basic truth of his assertion for obvious physical
and geometric grounds. Nevertheless, the implications of this discovery go well beyond
Fourier’s wildest imaginations. The unanswered mathematical points forcefully gave
birth to many more new problems and consequently motivated the development of many
important mathematic concepts and techniques such as Riemann integration, Sturm-
1
Liouville eigenvalue problem, set theory, Laplace transform, Lebesgue integration,
Green’s function and distribution theory, functional analysis, and most recently, wavelets
theory [1] as well as enormous applications in numerous ramifications of science and
engineering.
Despite the widespread applicability [1-3], the Fourier analysis approach suffers some
drawbacks. Most noticeably, for more than one-dimension problems, the direct use of the
Fourier series becomes very mathematically complicated and is only feasible for such
regular geometry as rectangular, circle, sphere, cylindrical domains, etc., where we can
separate the space variables (in Cartesian, polar, or some other coordinate systems [4]).
Otherwise the tensor product approach, very costly in high dimensions, must be applied.
However, when the scattered data are involved, the tensor product Fourier analysis also
immediately fails.
Because of the great success of the polynomials, splines, and tensor product methods,
mathematician and engineers alike grow accustomed to expressing a function in terms of
coordinate variables. To majority of scientific and engineering community, the radial
basis function (RBF), which uses the one-dimensional distance variable irrespective of
dimensionality, is a quite brand-new and exotic concept. In high dimensional scattered
data cases, the RBF approach, however, is the method of the choice [5]. It is also found
that the RBF is very efficient in handling lower-dimensional problems [5-7]. In the
following sections, we will try to clarify an underlying connection between the RBF and
PDE problem, which unveils that the RBF approach, just like Fourier series, is no more
than a natural technique originating from the solution of PDE problems.
Another notorious drawback of the Fourier approach is to lack efficient localization in
both time and frequency or scale, where the promise of wavelets technique comes into
play. In recent decade fast development and widespread applications of the wavelets have
been one of the most significant achievements in mathematics and many physical areas
[8]. Unlike the most traditional expansion systems such as Fourier one [9], the basis
functions of the wavelet analysis, however, are not solutions of differential equation. This
2
comes without a surprise since the wavelets have mainly been developed in signal
processing area, where the PDE system is rarely involved. By means of the fundamental
solution and general solution of partial differential equation, the next section concerns a
convergence of the RBF and wavelets while attainting the compactness and infinite
smoothness.
2. Radial basis function and wavelets
Since pioneer works of Frankle [10], Michaelli [11] and Kansa [7], the research into the
RBF theory and applications becomes very active. In parallel, Daubechies’ breakthrough
orthogonal compact wavelets [12] lead revolutionary advances in multiscale analysis. The
RBF is well known for its striking effectiveness in multivariate scattered data
approximation [5]. However, in general the RBFs available now lack critical multiscale
analysis capability. To handle high-dimensional multiscale scattered data and PDE
systems, the RBF wavelets are mostly wanted to combine the strengths of both. In last
decade much effort has been devoted to various non-orthogonal prewavelets RBF theory
and applications by using some constructive approximation strategies [13-17]. Very
recently Chen [18] could develop the orthogonal RBF wavelet series and transform by
using the fundamental solution and general solution of some typical PDE’s. The work
given in [18], however, is more conjecture and speculation than a complete theory with
some evident errors. Next is summarized the discrete orthogonal RBF wavelet series
versus the Fourier series. Some corrections to ref. 18 are also given.
v
Consider a real-value Riemann-integral multidimensional function f( x ) on the ndimension spherical domain Ω of radius R,
∞ ∞
v
v v
f ( x ) = a 0 2 + ∑∑ α jk ϕ n (η j x − x k
j =1 k =1
3
)
(2)
is called its RBF-based wavelet expansion series, where ϕn represents the wavelet basis
v v
function, x − x k means Euclidean distance; and αjk are the expansion coefficients. The
reason why α0 is multiplied by ½ is technical as in the Fourier series and will be
explained below. To discrete harmonic analysis, we choose the nonsingular general
solutions of n-dimension Helmholtz equation
ϕ n (rk ) = 1 ,
ϕ1 (η j rk ) =
1 ηj
ϕ n (η j rk ) =
4 2πrk
( n 2 )−1
η = 0,
(3a)
1
sin (η j rk ) ,
2η j
η≠0,
J (n 2 )−1 (η j rk ) ,
n≥2,
(3b)
η≠0
(3c)
v v
as wavelet basis function, where rk = x − x k , J(n/2)-1() is the (n/2-1)th order Bessel
function of the first kind; and ηj are the zeros of ϕn(Rr). The eigenfunctions (3) form an
orthonormal set of basis functions over different scales. In 3D case, the present RBF is
the renowned Sinc function.
Despite mathematical beauty and simplicity of the RBF expansion (2), a theoretical proof
of convergence and completeness under certain conditions is missing in [18]. Levesley et
al [19] provides an important clue to integrate the RBF wavelet approach with
convolution operator theory. Considering the fundamental solution in the distribution
theory, the general solution of n-dimensional Helmholtz equation may
(may not)
establish
∫
Ω∞
ϕ n ( x − x k )dΩ k ≠ 0 p ∞ ,
v
v
and then for a radially symmetric domain
4
(4)
∫
∞
0
r n −1ϕ n (r )dr p ∞ .
(5)
v
Note that the above condition is very stringent. Thus, for f ( x ) ∈ R n we have
lim c j ∫Ω ϕ n ( j x − x k ) f (x )dΩ k = f (x ) ,
v
v
v
(6)
j →∞
where cj are the coefficients depending on the RBF ϕn. After a numerical discretization of
the integral in Eq. (6), we get RBF approximate expression (2). Its convergence is
therefore guaranteed uniform and compact [19]. The condition (4) is satisfied by RBF
wavelet transform using singular fundamental solution given in later section 6. Like the
Fourier expansion, the function expressible in RBF wavelet series (2) should subject to
some conditions, especially if ϕn does not satisfy the condition (4). For now, we do not
have an explicit answer for this. A more loose condition on convergence of the Helmholtz
RBF wavelet series may exist.
On the other hand, very recently the present author became aware of a relevant important
paper by Mourou and Trimeche [20], which originally introduces the so-called
generalized wavelet transform using the solution of the Bessel operator. Although
Mourou and Trimeche [20] use the term of the radial function, the generalized wavelet
transform given there is in fact based on 1D coordinate variable rather than Euclidean
distance variable and therefore is in fact not a RBF wavelet, which can only handle onedimensional cases if without tensor product approach. Noteworthily, the admissibility
condition of their generalized wavelet transform is also applicable to the present RBF
wavelet series due to the underlying connection between Helmholtz equation and Bessel
operator. Namely, we have
∞
0 p C g = ∫ h(ϕ n )(λ )
0
5
2
dλ
λ
p ∞,
(7)
where h(ϕ n ) is the Fourier transform of general solution or fundamental solution ϕn of
Helmholtz equation.
Another important issue is how to calculate the expansion coefficients. It is known that
the Bessel function possesses the orthogonality:
∫
R
0
0 η≠µ
rJ v (ηr )J v (µr )dr = 2
2
R J v +1 (Rη ) 2 η = µ
(8)
where Jv(Rη)=Jv(Rµ)=0. However, the present RBF wavelet series is not necessary
orthogonal over translation of the same scale. Ref. [18] made wrong assumption that the
present RBF wavelets are orthogonal over both scale and translation. In this regard the
Gram-Schmidt orthogonality method may enforce the present RBF wavelets orthogonal
over translation. Thus we can directly determine the expansion coefficients.
In the 1D case, the coefficient formulas are quite similar to those of the corresponding
Fourier series:
α0 =
α jk =
2 jπ
R2
2
f (ζ )dΩ ζ ,
R ∫Ω R
jπ
∫ f (ζ )sin R r ζ dζ ,
R
k
−R
(9a)
(9b)
j = 1,2,K , k = 1,2, K
For multidimensional problems, we have
α0 =
2
Vn
∫
ΩR
f (ζ )dΩ ζ ,
6
(10a)
α jk =
8
S n J n 2 (η j )
2
ηj
2π
1− n 2
∫
Ωζ
rk1ζ−n 2 f (ζ )J (n 2 )−1 (η j rkζ )dΩ ζ ,
(10b)
j = 1,2, K , k = 1,2, K , n≥2,
v
v
where rkζ = x k − xζ , R is the radius of the spherical domain centering node k, and Sn
represents the surface parameter; Vn is the volume of n-dimensional sphere. The reason
for the coefficient ½ in (2) is that this will facilitate to streamline formulas (9a) and (9b)
or (10a) and (10b) if we choose proper constant in Eq. (3a).
3. Analytical solution of transient wave problem with RBF wavelets
The following n-dimension homogeneous wave problem with homogeneous boundary
conditions serves as an illustrative example of the present strategy:
∇ 2u =
1 ∂ 2u
,
c 2 ∂t 2
v
u ( x , t ) = 0,
v
∂u ( x , t ) = 0,
∂n
v
v
u ( x ,0 ) = φ ( x ),
v
u t (x ,0 ) = ψ ( x )
x∈Ω,
v
x ⊂ Su ,
t ≥ 0,
v
x ⊂ ST ,
v
x ⊂ S,
v
x ⊂ S,
(11)
(12)
(13)
v
where x means multi-dimensional independent variable, and n is the unit outward
normal. S = S u U S Γ , The separation of the time variable yields
v
v
u ( x , t ) = T (t )v( x )
7
(14)
Substitution into the wave equation (11) allows the separation of the time part
∇ 2 v + λ2 v = 0 ,
(15)
d 2T
+ λ2T = 0 .
dt 2
(16)
Here λ is the separation constant and actually eigenvalue of the system. The problem has
only nonnegative eigenvalues λj [22, pp. 248] and the solutions of the wave equation are
[
(
)
(
)]
∞
1
1
v
v
u ( x , t ) = A0 + B0 t + ∑ A j cos ct λ j + B j sin ct λ j v j (x ) ,
2
2
j =1
(17)
v
where vj( x ) is the corresponding eigenfunctions. For the Robin (radiation) boundary
condition ( ∂u ∂n + au = 0 ), the nonnegative eigenvalues holds provided that a≥0 [22].
3.1. Helmholtz eigenvalues and eigenfunctions with RBF
The challenging issues are to evaluate the spatial eigenvalues and eigenfunctions from the
Helmholtz equation (15). The standard approaches based on the coordinate variables do
not work in each case of irregular geometry, scattered data and high dimensions. So we
have to resort the radial basis function methodology. In the case of the n-dimensional
Helmholtz problem with eigenvalues λj, the complete RBF solution contains a complex
argument [23]:
1 λ
u (r ) =
4 2πr
n 2 −1
*
[Y (λr ) − iJ (λr )],
n 2 −1
n 2 −1
n≥2,
(18)
where rk = x − x k , J(n/2)-1 and Y(n/2)-1 are respectively the (n/2-1)th order Bessel function
of the first and second kinds. The Bessel function of the first kind is C∞ smooth while the
8
Bessel function of the second kind encounters a singularity at the origin. Thus, the
nonsingular general solution (regular distribution) is the imaginary part of the above
complex RBF solution (18) and is seen as the eigenfunctions. Namely, we have RBF
eigenfunctions as follows
ϕ n (rk ) = 1 ,
ϕ1 (λ j rk ) =
1
sin (λ j rk ),
2λ j
1 λj
ϕ n (λ j rk ) =
4 2πrk
( n 2 )−1
η = 0,
(19a)
n=1, λ ≠ 0 ,
(19b)
J (n 2 )−1 (λ j rk ),
n≥2,
λ≠0,
(19c)
Since the eigenfunctions must be finite at the origin, scraping the singular part of the RBF
solution (18) does not raise the completeness issue in general.
Note that the general solution is independent of geometry. In other words, the general
solution satisfies the Helmholtz equation (15) irrespective of the boundary shape of
interests, which is validated by the computer software “Maple” in terms of 2D and 3D
Cartesian coordinates.
The next issue is to calculate the eigenvalues. We present the two boundary knot method
(BKM) [24,25] schemes for this task. The first scheme is to employ the BKM
discretization of the Helmholtz equation (15). The approximate representation of the
symmetric BKM is
LD
LD + L N
s =1
s = LD +1
v( x ) = ∑ β sϕ n (rs ) −
∑
βs
∂ϕ n (rs )
,
∂n
(20)
where k is the index of source points on boundary, βk are the desired coefficients; n is the
9
unit outward normal as in boundary condition (12), and LD and LN are respectively the
numbers of knots on the Dirichlet and Neumann boundary surfaces. The minus sign
associated with the second term is due to the fact that the Neumann condition of the first
order derivative is not self-adjoint.
In terms of representation (20), we have the homogeneous collocation analogue Hv=0 of
boundary condition equations (12), where H is the symmetric BKM interpolation matrix.
And then, just as in the traditional Fourier solution of 1D wave problem, the determinant
of interpolation matrix has to be zero to attain the nontrivial solution, namely,
det (H ) = 0 .
(21)
The infinitely many roots of the above algebraic equation are the eigenvalues of the
Helmholtz equation (15). Note that since we use the symmetric BKM [25], all the
solutions of eigenvalues will be real valued.
The solution of transcendental equation (21) is often a daunting job and hence less
attractive in practical use. So we develop the second strategy of the BKM for this task.
The Helmholtz equation (15) can be rewritten as
(
)
∇ 2v + δ 2v = − λ2 + δ 2 v ,
(22)
where δ is a small artificial real parameter (around 0.1) and insensitive to the boundary
shape and dimensionality [25]. In terms of the symmetric BKM expression (20), the
above equation (22) can be discretized into the standard algebraic eigenvalue problem
(
)
Κv = − λ 2 + δ 2 v ,
(23)
where the interpolation matrix K is symmetric irrespective of the boundary conditions.
Many algorithm packages are readily available to calculate this standard eigenvalue
10
problem. Unlike the previous first strategy, the second scheme, however, requires using
some inner nodes to guarantee the stability and accuracy of the BKM solution [25].
3.2. Analytical solution with RBF wavelet series
In terms of solution (17), the resulting analytical solution of wave equations (1,2,3) can
be expressed as
[
(
)
(
)]
∞ ∞
1
1
v
v v
u ( x , t ) = A0 + B0 t + ∑∑ A jk cos ct λ j + B jk sin ct λ j ϕ n (λ j x − x k ) .
2
2
j =1 k =1
(24)
The RBF series solution (24) is valid for any dimensionality and geometry since the RBF
approach is independent of dimensionality and geometry. Note that this solution is an
inseparable wavelet series, where eigenvalue λj is understood scale parameter (dilation)
v
and source node x k are seen as location parameter (shift) in wavelet terminology.
In terms of initial conditions (13), we have
v
∞ ∞
1
v
v v
v
u ( x ,0 ) = A0 + ∑∑ A jk ϕ n (λ j x − x k ) = φ ( x ) ,
2
j =1 k =1
(25a)
∞
∞
1
v
v v
v
u t (x ,0 ) = B0 + ∑ B jk c λ j ∑ ϕ n (λ j x − x k ) = ψ ( x ) .
2
j =1
k =1
(25b)
v
φ ( x ) and ψ ( x ) are expressible in the RBF wavelet series (2) of Helmholtz general
solution provided that in this case they have the first order differential continuity and
equal zero at boundary. Like the Fourier series approach, the coefficients Aj and Bj can
most efficiently calculated via orthogonality. Applying the divergence theorem and
Green’s second identity, it is proved [22, pp. 246] that for Helmholtz equation (15), the
real eigenfunctions corresponding to distinct real eigenvalues are necessarily orthogonal.
Concerning the eigenfunctions (19), the orthogonality over scale is
11
∫
Ω
for s ≠ t ,
X sp X tq dΩ = 0 ,
(26)
where s and t denote scales, p and q are the translation locations. It is noted that the RBF
wavelet series is not necessarily orthogonal over the translation. By the Gram-Schmidt
orthogonality method, we can enforce orthogonization of general solution eigenfunctions
of the same scale. Otherwise, we have to solve the simultaneous equations for
coefficients Aj and Bj.
In [22, pp. 254-256] the method of separation of variables is used to solve vibrations of a
regular drumhead, where the eigenfunctions using coordinate variable is a general Fourier
series combining Bessel function and trigonometric functions. Due to the orthogonality,
all expansion coefficients, the number of which is equal to that of the present RBF
wavelet coefficients, are calculated directly. By analogy with this strategy, the following
formulas are given without rigorous proof under translation-orthogonal assumption and
just for reference. A relevant Gram-Schmidt procedure should be developed in the future.
A0 =
B0 =
A jk =
B jk =
2
S n J n 2 (λ j )
2
2
Vn
∫ φ (ζ )dΩζ ,
ΩR
2
∫
Vn c λ j
λj
2π
8
c λ j S n J n 2 (λ j )
2
∫
Ωζ
λj
2π
ψ (ζ )dΩ ζ ,
(28)
rk1ζ− n 2φ (ζ )J (n 2 )−1 (λ j rkζ )dΩ ζ ,
(29)
ΩR
1− n 2
1− n 2
∫
Ωζ
rk1ζ− n 2ψ (ζ )J (n 2 )−1 (λ j rkζ )dΩ ζ ,
j = 1,2, K , k = 1,2, K
12
(27)
(30)
It is very interesting to note that the 1D nonsingular general solution (19b) is the same
sinusoids as in the 1D Fourier series solution (1). The distinctness of the RBF wavelet
and Fourier series solutions is that when applied to practical problems with a truncated
finite series, the former can locally adjust the scale parameter to avoid the Gibbs
phenomena. Therefore, the RBF wavelet series solution are much more robust than the
Fourier one.
It is worth pointing out that the present RBF wavelets using the general solution of
Helmholtz equation have periodic (harmonic) property and are much natural than those
periodic RBF developed by [26,27], where the sine and cosine functions are employed as
the RBF basis without considering dimensional affect and PDE eigensolution.
4. Applications to inhomogeneous problems
In practical engineering, the governing equation [11] and boundary conditions [12] are
often not homogeneous. For instance,
∇ 2u =
1 ∂ 2u
v
+ f (x ) ,
2
2
c ∂t
v
v
u ( x ) = D( x ),
v
∂u ( x ) = R( xv ),
∂n
v
x ⊂ Su ,
v
x ⊂ ST ,
v
v
u ( x ,0 ) = φ ( x ),
v
u t (x ,0 ) = ψ ( x )
v
x ⊂ S,
v
x ⊂ S,
x∈Ω
(31)
t ≥ 0,
(32)
(33)
v
where f ( x ) is outside forcing function. Just like the Fourier solution of inhomogeneous
PDE’s, we have two strategies to apply the method of the separation of variables [22, pp.
13
140-143] with RBF wavelet series. The method of shifting the data makes the
inhomogeneous boundary conditions homogeneous by subtracting any known function
that satisfies them, while the expansion method expand everything in the eigenfunctions
of the corresponding homogeneous problems and then we can get a set of ordinary
differential equations which is very easy to solve.
5. Applications to other problems
Besides the wave equation considered previously, the following equations also frequently
appear in electrical, magnetic, thermal, gravitational, vibration, hydrodynamics and
acoustics problems [4]:
1. The diffusion equation
∇ 2u =
1 ∂u
,
h 2 ∂t
(34)
∇ 2u =
∂u
1 ∂ 2u
+R ,
2
2
∂t
c ∂t
(35)
∇ 2u =
∂u
1 ∂ 2u
+R
+ Su ,
2
2
∂t
c ∂t
(36)
2. The damped wave equation
3. Transmission line equation
The solution of all these equations can be reduced to solutions of the scalar Helmholtz
equation (15) and the corresponding very simple ordinary differential equation in time.
Thus, the extension of the present RBF wavelet series solution to these equations is very
straightforward and omitted here for brevity.
14
6. Generalized RBF wavelet series and transforms
In the foregoing sections, we only consider the RBF wavelet series using the general
solution of Helmholtz equation. Those RBF wavelets describe periodic behaviors of
physical systems as the trigonometric Fourier series. The higher order fundamental
solutions of Laplace operator can construct the RBF wavelet series corresponding to the
common polynomial interpolation and approximation, where the order of fundamental
solution corresponds to the scale or polynomial order. Moreover, they are orthogonal
over scales since the lower-order fundamental solutions satisfy the higher-order operator.
The drawbacks of the Laplacian RBF wavelets are limited smoothness.
On the other hand, we can create continuous RBF wavelet transform using the
fundamental solution (irregular distribution) of PDE’s. For example,
f (ζ )g n (λrξζ )dΩ ζ ,
F (λ , ξ ) = ∫
Ω∞
(37)
and
+∞
v
f ( x ) = C g−1 ∫ ∫ F (λ , ξ )g n (λrxξ )λ2 n −1 dΩ ξ dλ ,
0
Ω∞
(38)
where Cg is decided by formula (7),
1
g n (λrk ) =
2π
− iλ
2πrk
λ
2πrk
( n 2 )−1
K (n 2 )−1 (− iλrk )
(39)
K (n 2 )−1 (λrk )
(40)
for Helmholtz harmonic wavelets and
1
g n (λrk ) =
2π
( n 2 )−1
for modified Helmholtz wavelets, K is the modified Bessel function of the first kind. The
15
corresponding dual wavelet basis function is its conjugate function g n (λrk ) , which
satisfies the condition (4). The modified Helmholtz wavelets are the counterpart of the
Laplace transform. Note that here we correct some errors of Eqs. (15) and (16) in ref. 18
without using rn-1 weight. For more details see refs. [18,21]. The harmonic RBF wavelet
transform enjoys a nice feature of Fourier transform:
∇ 2 F (λ , ξ ) = −λ2 F (λ , ξ ) .
(41)
It is also highly likely to develop the RBF wavelets transforms based on the general and
fundamental solutions of the other typical partial differential equation such as convectiondiffusion equation [18].
7. Promises and open problems
Promises:
1. Since the present series solution is wavelets, the Gibbs phenomenon long
bothering the Fourier series is eliminated. By adapting the arbitrarily scaling
parameter (dilations) rather than the dyadic multiresolution analysis, we get
locally supported RBF basis both in scale and location. Compactly-supported
wavelets using spline, especially orthogonal such as the popular Daubechies
wavelets [12], have a limited degree of smoothness in compromise to the
compactness. In contrast, the RBF wavelets are not only orthogonal over scale but
also infinitely differentiable.
2. Due to the use of the RBF, we avoid using the tensor-product approach for high
dimensional problems with irregular geometry. In addition, the present method is
meshfree and feasible to handle scattered data problems.
3. The strategy presented here is expected feasible to other type of problems such as
data processing and edge detection etc., where an efficient description of
multidimensional multiscale scattered data is crucial. In addition, the present
16
method may be employed to evaluate the fractional differential equation
corresponding to fractal geometry by using fractional dimensionality.
Problems:
The present form of the RBF wavelets, resembling the immature status of Fourier’s early
work, which used concepts and theories as yet undeveloped or underdeveloped [1], may
provide some of the refresh impulse on new advances in applied and basic analysis. Some
worrisome points of this study are stated below (the distribution theory may be useful to
research them).
1. In terms of Fourier series, there are three kinds of convergence: pointwise,
uniform and L2 convergences. This study will not go further into the convergence
issue of the RBF wavelet series solution of transient PDE’s. Ref. [22] defines that
an orthogonal system is called complete if and only if it is not a proper subset of
another orthogonal system. The present general solution eigenfunctions satisfy
this condition. In addition, the key issue is if it satisfies Parseval’s identity.
Section 2 provides some proof of the convergence and completeness of the RBF
wavelet series via convolution operator theory and admissibility condition of
wavelets. The proof of completeness and convergence of RBF series solution of
PDE’s, however, is still lacking now. Chen and Tanaka [17] and Hon and Chen
[28] also physically discussed the completeness issue of the general solution
expansion series within the framework of the boundary knot method.
2. It is unknown now if the RBF wavelet transform given in section 6 [18] can be
employed to get the analytical solution of time-dependent problems in the integral
form. If OK, this integral solution should be equal to the limit of the present RBF
wavelet series solution.
3. The computing formulas (9), (10) and (27-30) of expansion coefficients based on
orthogonality may be error prone.
17
References
1. E.A. Gonzalez-Velasco, Fourier Analysis and Boundary Value Problems, Academic
Press, 1995.
2. O. Erosy, Fourier-Related Transforms, Fast Algorithms and Applications, PrenticeHall, 1997.
3. D.G. Duffy, Transform Methods for Solving Partial Differential Equations, CRC
Press, Florida, 1994.
4. P. Moon and D.E. Spencer, Field Theory Handbook, Springer, 1961.
5. M.D. Buhmann, Radial basis function, Acta Numerica, 1-38, 2000.
6. R.L. Hardy, Theory and applications of the multiquadric-biharmonic method: 20
years of discovery, Comput. Math. Appl. 19, 163-208, 1990.
7. E.J. Kansa, Multiquadrics: A scattered data approximation scheme with applications
to computational fluid-dynamics, Comput. Math. Appl. 19,147-161, 1990.
8. C.K. Chui, Wavelets: A Mathematical Tool for Signal Analysis. SIAM, Philadelphia,
1997.
9. C.S. Burrus, R.A. Gopinath and H. Guo, Introduction to Wavelets and Wavelet
Transforms, Prentice Hall, 1998.
10. R. Franke, Scattered data interpolation: tests of some methods, Math. Comp. 38, 181199, 1982.
11. C.A. Micchelli, Interpolation of scattered data: distance matrices and conditionally
positive definite functions, Constr. Approx. 1, 11-22, 1986.
12. I. Daubechies, Orthonormal bases of compactly supported wavelets, Comm. Pure
Appl. Math., XLI(7), 909-996, 1988
13. M.D. Buhmann, Multiquadric prewavelets on nonequally spaced knots in one
dimension, Math. Comp., 64, 1611-1625, 1995.
14. C.K. Chui, J. Stockler and J.D. Ward, Analytical wavelets generated by radial
functions, Adv. Comp. Math., 5, 95-123, 1996.
15. C.A. Michaelli, C. Rabut, and F.L. Utreras, Using the refinement equation for the
construction of pre-wavelets III: elliptic splines, Numer. Algorithms, 1, 331-352,
1991.
18
16. C. Li and N. Zheng, A theory of constructing locally supported radial wavelet frame
and its application, Science in China (Series E), 42(6), 584-594, 1999.
17. W. Chen and M. Tanaka, Relationship between boundary integral equation and radial
basis function. In the 52th Symposium of Japan Society for Computational Methods in
Engineering (JASCOME) on BEM, Tokyo, Sept. 2000.
18. W. Chen, Orthonormal RBF wavelet and ridgelet-like series and transforms for highdimensional problems, Int. J. Nonlinear Sci. Numer. Simulation, 2, 163-168, 2001.
19. J. Levesley, Y. Xu, W. Light and W. Cheney, Convolution operators for radial basis
approximation, SIAM J. Math. Anal., 27(1), 286-304, 1996.
20. M.A. Mourou and K. Trimeche, Inversion of the Weyl integral transform and the
radon transform on Rn using generalized wavelets, C.R. Math. Rep. Acad. Sci.
Canada, XVIII(2-3), 80-84, 1996.
21. W. Chen, Errata and supplements to: Orthonormal RBF Wavelet and Ridgelet-like
Series
and
Transforms
for
High-Dimensional
Problems,
CoRR
preprint:
http://xxx.lanl.gov/abs/cs.NA/0105014, April, 2001.
22. W.A. Strauss, Partial Differential Equations, an Introduction, John Wiely & Sons,
1992.
23. P.K. Kythe, Fundamental Solutions for Differential Operators and Applications,
Birkhauser, Boston, 1996.
24. W. Chen and M. Tanaka, New Insights into Boundary-only and Domain-type RBF
Methods, Int. J. Nonlinear Sci. & Numer. Simulation, 1(3), 145-151, 2000.
25. W. Chen, Boundary knot method for Laplace and biharmonic problems, Proc. of the
14th Nordic Seminar on Comp. Mech., Lund, Sweden, (L. Beldie, O. Dahlblom, A.
Olsson, N.S. Ottosen and G. Sandberg eds.), pp. 117-120, 2001.
26. Y. Xu and E.W. Cheney, Interpolation by periodic radial functions, Comput. Meth.
Appl., 24(2), 201-215, 1992.
27. W.A. Light and E.W. Cheney, Interpolation by periodic radial basis functions, J.
Math. Anal. Appl., 168, 111-130, 1992.
28. Y.C. Hon and W. Chen, Boundary knot method for 2D and 3D Helmholtz and
convection-diffusion problems under complicated geometry, Int. J. Numer. Methd.
Engng., (submitted), 2001.
19
| 5cs.CE
|
Uniqueness Trees: A Possible Polynomial Approach to the Graph
Isomorphism Problem
Jonathan Gorard
arXiv:1606.06399v1 [cs.DM] 21 Jun 2016
Department of Mathematics, King’s College London, Strand, London, WC2R 2LS
Abstract
This paper presents the novel ‘uniqueness tree’ algorithm, as one possible method for determining whether two finite, undirected graphs are isomorphic. We prove that the algorithm
has polynomial time complexity in the worst case, and that it will always detect the presence
of an isomorphism whenever one exists. We also propose that the algorithm will equivalently
discern the lack of an isomorphism whenever one does not exist, and some initial justifications are given for this proposition, although it cannot yet be rigorously proven. Finally,
we present experimental evidence for both the effectiveness and efficiency of the uniqueness
tree method, using data gathered from a practical implementation of the algorithm. Some
consequences and directions for further research are discussed.
Keywords: graph isomorphism, computational complexity, canonical labelling
1. Introduction
1.1. Background
The graph isomorphism problem is the decision problem of determining whether two
finite graphs, G = (V, E) and H = (U, F ) are isomorphic, denoted G ∼
= H. The graphs are
isomorphic if and only if there exists a bijection f between the two sets of vertices V and U
such that, for every pair of vertices (u, v) in V , the edge f (u)f (v) exists in F if and only if
the corresponding edge uv exists in E.[1] Formally:
G∼
= H ⇐⇒ ∃f : V → U | ∀(u, v) ∈ V, f (u)f (v) ∈ F ⇐⇒ uv ∈ E
(1)
The graph isomorphism problem is of particular interest in the field of computational
complexity theory, since it is one of only a few problems whose complexity class is not
solidly classified: it is not known to be solvable in polynomial time, yet neither has it been
shown to be NP-complete. Thus, it is often placed in the theoretical complexity class of ‘N P intermediate’[2] (which, by Ladner’s theorem, exists if and only if P 6= N P [3]). In addition,
efficient algorithms for detecting graph isomorphism are of great practical importance across
a variety of fields, including network analysis, organic chemistry, condensed matter physics,
chemical engineering, electronic engineering, computational biology, and others.[5][4][6]
Preprint submitted to Journal of Discrete Algorithms
February 26, 2018
There exist many algorithms which run in polynomial (or even sub-polynomial) time in all
practical cases, but which degenerate to exponential time in the worst-case[7], making them
unsatisfactory from a complexity-theoretic point of view. We provide a brief outline of one
such algorithm (McKay’s NAUTY), for the purpose of demonstrating how its approach, and
the approach adopted by many similar state-of-the-art algorithms, differs from the uniqueness
tree method proposed in this paper.
We then provide both a formal and an informal statement of the uniqueness tree algorithm
itself, along with an illustrative example of its application to a pair of non-isomorphic graphs.
Next, we give a formal proof that the algorithm runs in septic polynomial (O(n7 )) time in the
worst case, and that the algorithm will always correctly detect the presence of an isomorphism
between two graphs, whenever one exists. We propose that the converse statement is also
true (i.e. that the algorithm will correctly discern the lack of an isomorphism, whenever one
does not exist), and give a brief sketch of a possible proof method, though this statement
remains a conjecture. Finally, we supply experimental evidence of both the algorithm’s
effectiveness in determining isomorphism/non-isomorphism between random graphs, and its
polynomial efficiency.
For the purposes of this paper, we will consider only simple graphs (i.e. unweighted,
undirected graphs containing no loops or multiple edges). However, it is possible to generalise
these methods to directed graphs, as well as to graphs in which multiple edges and loops are
permitted, as will be shown in a future work.
1.2. NAUTY
Most practical graph isomorphism algorithms work by reducing graphs to a so-called
‘canonical form’ - an object whose structure is independent of the particular ordering of
the vertices, but dependent upon all other properties of the graph.[8] Thus, if the canonical
forms for two graphs are equivalent, then the graphs must be isomorphic; conversely, if
the canonical forms differ, then the graphs must be non-isomorphic. NAUTY, as with most
similar algorithms, operates by producing a so-called ‘search tree’ as its canonical graph form,
which is a rooted tree in which each vertex corresponds to a distinct ‘partition’ (colouring)
of the graph’s vertices. Loosely, the process of partition refinement is as follows:[9]
Figure 1: A partition π.
Figure 2: A refined partition π1 .
2
1. The root of the tree is an initial (uniform) colouring of the graph.
2. If two vertices share the same colour in a particular partition, but have neighbourhoods with different colourings, then ‘refine’ the partition by assigning each vertex a
new colour. This refined partition is a child vertex of the original partition.
3. If a particular partition cannot be refined further, then that vertex of the tree becomes
a leaf, with no children.
However, the uniqueness tree algorithm does not produce a single search tree to represent
a graph. Rather, it produces a set of ‘uniqueness trees’: one for each vertex of the graph.
In turn, each vertex of a uniqueness tree represents a single vertex of the graph, as opposed
to the richer structure of an entire vertex partition. This paper conjectures that the set of
uniqueness trees is a canonical graph form, and provides some initial justification for that
assertion.
2. The Uniqueness Tree Algorithm
2.1. Brief Outline
For two finite, simple graphs G = (V, E) and H = (U, F ), the uniqueness tree algorithm
associates a rooted tree T (v) with every vertex v ∈ V , and T (u) with every vertex u ∈ U .
This paper proposes that, if every tree associated with a vertex of G is uniquely isomorphic
to a tree associated with a vertex of H, then G and H are isomorphic:
G∼
= H ⇐⇒ ∀v ∈ V, ∃!u ∈ U | T (v) ∼
= T (u)
(2)
Checking this criterion is an efficient process, since an isomorphism between two rooted
trees may be computed in linear (O(n)) time, where n is the number of vertices in each
tree.[10]
The process for generating the uniqueness tree for a vertex v is as follows:
1. The root of the tree is the vertex v itself.
2. Every vertex in the current level of the tree which is not ‘unique’ (i.e. every vertex
which appears in the current level more than once) becomes a leaf, producing no children.
3. Every unique vertex in the current level of the tree produces 1 child for every adjacent
vertex in the graph.
4. This process continues until either the tree self-terminates (i.e. there are no more
unique vertices on the current level), or the height of the tree reaches n (the size of the
graph).
3
2.2. An Example Case
As an illustrative example, we shall test for isomorphism between the following pair of
graphs:
Figure 3: A graph, G = (V, E).
Figure 4: Another graph, H = (U, F ).
Clearly, G and H are not isomorphic (G is planar and H is not). However, they are
equivalent in every other respect, since H is simply an embedding of G from a plane onto
a Möbius strip. In the interests of brevity, we shall not apply the entire algorithm, since to
do so would require generating 16 uniqueness trees, and then making up to 36 comparisons
between them. Rather, we shall simply show that vertex A in G cannot be equivalent to
vertex A in H, and the non-isomorphism of G and H follows trivially.
For graph G, the first level of A’s uniqueness tree contains A’s immediate adjacencies:
B, C and G (all of which are unique).
Figure 5: The first level of A’s uniqueness tree for graph G.
Similarly, B’s children on the second level of the tree are A, D and H (B’s immediate
adjacencies), C’s children are A, D and E, and G’s children are A, E and H. Since A, D, E
and H all appear multiple times on the second level, all of the vertices becomes leaves and
the tree terminates.
4
Figure 6: The complete uniqueness tree for vertex A in graph G.
On the other hand, the first level of A’s uniqueness tree for graph H contains the adjacent
vertices B, C and H, which are, again, unique.
Figure 7: The first level of A’s uniqueness tree for graph H.
B’s children on the second level of the tree are then A, D and G, C’s children are A, D
and E, and H’s children are A, F and G. Vertices A, D and G become leaves, since they
are not unique at this level.
Figure 8: The second level of A’s uniqueness tree for graph H.
Since vertices E and F are both unique on the second level, their adjacencies (C, F and
G for vertex E, and D, E and H for vertex F ) are carried to the third level of the tree.
Since C, D, E, F , G and H are all unique on the third level, their adjacencies will, in turn,
be carried up to the fourth level, and so on.
5
Figure 9: The third level of A’s uniqueness tree for graph H.
Already, we can see that the two uniqueness trees cannot possibly be isomorphic (since
the first tree has only two levels, whilst the second has at least four), and so vertex A in
graph G cannot be equivalent to vertex A in graph H.
2.3. Formal Statement
The uniqueness tree algorithm may be divided into two distinct stages: the tree generation
stage, and the tree comparison stage.
Algorithm 1 Uniqueness tree generation
1: for each graph G = (V, E) do
2:
for each vertex v ∈ V do
3:
create a new tree T (v), with v as its root
4:
while ∃ u, a vertex which appears only once in the current level of T (v), and
height(T (v)) < n do
5:
for each w ∈ neighbourhood(u) do
6:
add w to the next level of the uniqueness tree.
7:
end for
8:
end while
9:
end for
10: end for
Then, for two graphs G = (V, E) and H = (U, F ):
6
Algorithm 2 Uniqueness tree comparison
1: for each unmapped vertex v ∈ V do
2:
for each unmapped vertex u ∈ U do
3:
if height(T (v)) 6= height(T (u)) then
4:
v and u are not equivalent
5:
end if
6:
for each level of T (v) do
7:
if width(currentlevel(T (v))) 6= width(currentlevel(T (u))) then
8:
v and u are not equivalent
9:
end if
10:
for i ← 1, (n − 1) do
11:
if vertices with i children in currentlevel(T (v)) 6= vertices with i children
in currentlevel(T (u)) then
12:
v and u are not equivalent
13:
end if
14:
end for
15:
end for
16:
if v and u are equivalent then
17:
map v onto u
18:
end if
19:
end for
20: end for
21: if all vertices v ∈ V and u ∈ U have been mapped then
22:
G∼
=H
23: else
24:
G H
25: end if
3. Rigorous Results
3.1. Proof of Polynomial Time Complexity
Theorem 3.1. For two finite, simple graphs G = (V, E) and H = (U, F ), each of size n, the
uniqueness tree algorithm runs in septic polynomial time (O(n7 )) in the worst case.
Since the uniqueness tree algorithm can be divided into two sequential stages (tree generation and tree comparison), we shall analyse the time complexity of each stage separately,
and then add the two complexities together.
Lemma 3.1.1. The tree generation algorithm runs in sextic polynomial time (O(n6 )) in the
worst case.
Proof 3.1.1.
7
1. The total number of uniqueness trees which must be generated is 2n = O(n).
2. The maximum width of a single uniqueness tree is O(n2 ) (since the maximum number
of unique vertices which can appear in a single level is n, and each vertex can have a
maximum of (n − 1) adjacencies, giving a maximum of (n2 − n) children on the next
level, all of which would be leaves).
3. ∴ The maximum number of vertices in each uniqueness tree is O(n3 ) (since the maximum width is O(n2 ), and the maximum height is O(n)).
4. The maximum number of operations required to generate each vertex in the tree is
O(n2 ) (i.e. a maximum of (n − 1) comparisons with other vertices in the graph, plus a
maximum of (n2 − n) comparisons with other vertices in the current level of the tree,
in order to test uniqueness).
5. ∴ The worst case time complexity of the tree generation algorithm is
O(n) ∗ O(n3 ) ∗ O(n2 ) = O(n6 ).
Lemma 3.1.2. The tree comparison algorithm runs in septic polynomial time O(n7 ) in the
worst case.
Proof 3.1.2.
1. The first tree of graph G must be compared with a maximum of n trees from graph H,
the second with (n − 1) trees, and so on.
n
X
∴ The maximum number of tree comparisons which must be made is
i = O(n2 ).
i=1
2. Determining an isomorphism between the two rooted trees requires comparing the heights
of both trees, comparing the total number of vertices in each level of each tree, and comparing the total number of children possessed by each vertex in each level of each tree.
3. Comparing the heights of two rooted trees requires O(1) operation.
4. Comparing the total number of vertices in each level of two rooted trees requires O(n)
operations in the worst case (since each tree may hold a maximum of n levels).
5. Each level may contain a maximum of (n2 − n) vertices, each of which must be compared with a maximum of (n2 − n) vertices in the corresponding level of the other tree,
for each of a possible n levels of the tree.
∴ Comparing the total number of children possessed by each vertex in each level of each
tree requires
O(n2 ) ∗ O(n2 ) ∗ O(n) = O(n5 ) operations in the worst-case.
8
6. ∴ The worst case time complexity for detecting an isomorphism between the two rooted
trees is
O(n5 ) + O(n) + O(1) = O(n5 ).
7. ∴ The worst case time complexity of the tree comparison algorithm is
O(n2 ) ∗ O(n5 ) = O(n7 ).
Since O(n7 ) + O(n6 ) = O(n7 ), the desired theorem follows directly from these two lemmas.
3.2. Proof of Effectiveness (Positive Case)
Theorem 3.2. If two graphs G = (V, E) and H = (U, F ) are isomorphic, then the uniqueness tree algorithm will correctly determine that G ∼
= H.
Proof 3.2.1.
The key point is that the uniqueness tree algorithm is based entirely around vertex adjacencies (which are graph-invariant), and does not depend at all upon vertex ordering. Suppose
that uv ∈ E and f (u)f (v) ∈ F .
Figure 10: A section of a uniqueness tree for
graph G.
Figure 11: A section of the equivalent uniqueness
tree for graph H.
1. If a vertex u is not unique to a particular level of a particular uniqueness tree for graph
G, then the vertex f (u) will be also non-unique to the equivalent level of the equivalent
vertex tree for graph H.
2. Conversely, if vertex u is unique to that level, then vertex v in the next level of G’s
tree will have the same uniqueness property as vertex f (v) in H’s tree, and so on.
3. Thus, it follows that if f (u)f (v) ∈ F ⇐⇒ uv ∈ E, then graphs G and H will produce
an equivalent set of uniqueness trees.
4. Since an isomorphism is defined as a bijection f between the sets V and U which
satisfies the above property, the desired theorem follows naturally.
9
3.3. Proposition of Effectiveness (Negative Case)
Proposition 3.2.1. If two graphs G = (V, E) and H = (U, F ) are not isomorphic, then the
uniqueness tree algorithm will correctly determine that G H.
Justification 3.2.1.
One possible approach to proving this proposition may be an analogy to the positive case,
as in the following informal sketch:
1. If G H, then (without loss of generality) there must exist at least one f (u)f (v) ∈ F ,
such that uv ∈
/ E (by negation of the definition of isomorphism).
2. If vertex u (in a particular level of a particular uniqueness tree for graph G) has a different uniqueness property to vertex f (u) (in the equivalent level of the equivalent uniqueness tree for graph H), then the uniqueness trees produced will be non-isomorphic, since
one vertex will become a leaf, whilst the other will not.
3. Conversely, if vertices u and f (u) are both unique, then there will be an instance of
the vertex f (v) in the next level of H’s tree, without a corresponding instance of the
vertex v in G’s. Thus, the uniqueness trees produced will be non-isomorphic, since the
vertices in the next level will have different uniqueness properties as a result.
4. Additionally, if vertices u and f (u) are both non-unique, then the non-uniqueness of
vertex f (u) may be the result of having f (v) as its parent vertex in H’s tree, whilst v
cannot be a parent vertex of u in G’s tree. If this is the case, then the uniqueness trees
produced will, again, be non-isomorphic.
Clearly, further work will be required to either formalise a proof of this proposition, or
to demonstrate why it is incorrect.
4. Practical Implementation
For the purposes of experimentally verifying both the effectiveness and efficiency of the
uniqueness tree method, we present a practical implementation of the algorithm in Java; the
complete source code for this algorithm, along with all of the tests described in this section,
may be found in the Appendix. Independent verification of these experimental findings is
strongly encouraged.
4.1. Experimental Verification of Effectiveness (Positive Case)
The algorithm was tested on 10,000 pairs of random graphs, ranging in size from 1 to
100 (100 pairs of each). The graph pairs were known to be isomorphic, since the second
graph was generated by randomly permuting the vertices of the first. The presence of an
isomorphism was correctly detected in all 10,000 cases.
10
4.2. Experimental Verification of Effectiveness (Negative Case)
When testing proposed algorithms for graph isomorphism, generating pairs of graphs
which are known to be non-isomorphic is a fundamentally difficult problem, since doing so
pre-supposes an effective criterion for detecting isomorphisms in the first place. It is trivial
to generate pairs of graphs with differing graph invariant properties, but then one is simply
testing the capability of the algorithm to discern those particular graph invariants, rather
than its ability to detect non-isomorphism in general.
For the purpose of the present experimental test, we will attempt to generate a pair of
similar but non-isomorphic graphs, by creating a second graph that is a vertex-permutation of
the first (as above), but with one edge randomly replaced. Clearly, there is a small probability
of inadvertently creating an isomorphic pair, particularly when the graphs involved are small.
Indeed, when tested on the first 900 pairs of graphs with n < 10, some apparent false-positives
were produced. However, when these possible exceptions were tested (either by hand, or by
cross-checking with NAUTY), it was determined that all of the problematic graphs had been,
by chance, isomorphic. For the remaining 9,100 pairs of graphs with n ≥ 10, there were no
false positives.
4.3. Experimental Verification of Efficiency
The total computation time required for the algorithm to test for isomorphism between
10,000 pairs of random isomorphic test graphs (100 pairs of each size n, with n ranging from
1 to 100) was recorded, and then plotted as a function of n. The equivalent process was
repeated for 10,000 pairs of random non-isomorphic test graphs, also:
Figure 12: The computation time required to test
for isomorphism between 100 pairs of known isomorphic graphs, as a function of the number of
vertices.
Figure 13: The computation time required to test
for isomorphism between 100 pairs of known nonisomorphic graphs, as a function of the number
of vertices.
11
For large n, taking the natural logarithm of both axes produces a straight line, demonstrating the asymptotically polynomial efficiency of the algorithm (since if y = xk , then
ln(y) = kln(x), with the gradient of the line thus representing the degree of the polynomial):
Figure 14: The natural logarithm of the computation time required to test for isomorphism between 100 pairs of isomorphic graphs, as a function of the natural logarithm of the number of
vertices. The plot fits a straight line of gradient
4.174996, with R2 = 0.968827.
Figure 15: The natural logarithm of the computation time required to test for isomorphism
between 100 pairs of non-isomorphic graphs, as
a function of the natural logarithm of the number of vertices. The plot fits a straight line of
gradient 4.187919, with R2 = 0.966003.
The seemingly anomalous points observed for low values of n are due to intrinsic computational overheads, which naturally smooth out for larger graphs. Thus, removing the
results for the first 2,000 graph pairs from each plot (i.e. the values 1 ≤ n ≤ 20) produces a
much better linear fit in both cases:
12
Figure 16: The same as the plot above, but
with the results for the smallest 2,000 isomorphic
graph pairs removed. The plot now fits a straight
line of gradient 5.124289, with R2 = 0.995931.
Figure 17: The same as the plot above, but with
the results for the smallest 2,000 non-isomorphic
graph pairs removed. The plot now fits a straight
line of gradient 5.202082, with R2 = 0.996664.
These plots suggest that the uniqueness tree algorithm runs in approximately quintic
polynomial time (O(n5 )) on average for random graphs, which is consistent with our proof
that the time complexity should never exceed O(n7 ).
5. Concluding Remarks
This paper has presented the polynomial time ‘uniqueness tree’ algorithm, and used a
combination of rigorous results and heuristic verification to propose that it may represent
an effective approach to tackling the graph isomorphism problem.
The importance placed upon the graph isomorphism problem within computational complexity theory is primarily due to its present status as a prime candidate for an ‘N P intermediate’ problem[2] (a problem that is neither in the complexity class P , nor N P complete), and many other proposed N P -intermediate problems are known to be reducible
to GI.[11] Thus, if the graph isomorphism problem could be shown to be solvable in polynomial time, it would constitute evidence against the existence of the N P -intermediate set
(and therefore, by Ladner’s theorem, evidence that P = N P [3]). As such, any indication of
the possibility of a polynomial time solution for GI would have significant implications for
the entire field.
Further research into the effectiveness of the uniqueness tree algorithm will centre around
attempting to either prove or disprove the proposition that it is able to correctly discern the
lack of an isomorphism in all relevant cases. Extensions of the methods described in this
paper to non-simple and directed graphs will appear in a future work, along with possible
generalisations of the uniqueness tree approach to related problems in graph theory (including
graph automorphism, subgraph homeomorphism, etc.).
13
6. Acknowledgements
The author is grateful to Richard Bridges, for many helpful suggestions and clarifications
during the early stages of this research.
7. References
References
[1] A. Prolubnikov, Reduction of the graph isomorphism problem to equality checking of
n-variables polynomials and the algorithms that use the reduction, http://arxiv.org/
pdf/1512.03139v4.pdf (2016) 1
[2] A. Dawar, Report on The Graph Isomorphism Problem, Dagstuhl Seminar 15511 on the
Graph Isomorphism Problem (2015) 1-2
[3] M. Bodirsky, M. Grohe, Non-dichotomies in Constraint Satisfaction Complexity, Automata, Languages and Programming LNCS 5126 (2008) 184
[4] W. Fan, Graph pattern matching revised for social network analysis, Proceedings of the
15th International Conference on Database Theory (2012) 8
[5] T. Akutsu, H. Nagamochi, Comparison and Enumeration of Chemical Graphs, Computational and Structural Biotechnology Journal 5 (6) (2013) 2
[6] J. Whitham, A Graph Matching Search Algorithm for an Electronic Circuit Repository,
MEng Thesis, University of York (2004) 5
[7] P. Foggia, C. Sansone, M. Vento, A performance comparison of five algorithms for
graph isomorphism, Proceedings of the 3rd IAPT TC-15 Workshop on Graph-based
Representations in Pattern Recognition (2001) 2
[8] B. D. McKay, A. Piperno, Practical graph isomorphism, II, https://arxiv.org/pdf/
1301.1493.pdf (2013) 2
[9] B. D. McKay, Practical graph isomorphism, Congressus Numerantium 30 (1981) 54
[10] S. R. Buss, Alogtime algorithms for tree isomorphism, comparison, and canonization,
Computational Logic and Proof Theory 1289 (2005) 19
[11] G. L. Miller, Graph isomorphism, general remarks, Journal of Computer and System
Sciences 18 (2) (1979) 129
14
Appendix: Source Code
import java . util . Random ;
import java . io .*;
public class Main {
private static Random rnd ;
private static FileWriter fileWriter ;
private static BufferedWriter bufferedWriter ;
private static int [] vertexCount ;
private static int [][] vertexDegree ;
private static int [][][] vertexAdjacency ;
private static int [] vertexMap ;
private
private
private
private
static
static
static
static
int [][] u n i q u e n e s s T r e e H e i g h t ;
int [][][] u ni q u en e s sT r e eW i d th ;
int [][][][] u n i q u e n e s s T r e e V e r t i c e s ;
int [][][][] u n i q u e n e s s T r e e V e r t e x C h i l d C o u n t ;
private static int [][][][] u n i q u e n e s s T r e e V e r t e x C h i l d C o u n t O c c u r r e n c e ;
private static boolean [][] vertexMapped ;
public static void initialise ( int n ) {
rnd = new Random () ;
vertexCount = new int [2];
vertexDegree = new int [2][ n ];
vertexAdjacency = new int [2][ n ][ n ];
vertexMap = new int [100];
u n i q u e n e s s T r e e H e i g h t = new int [2][ n ];
u ni q u en e s sT r e eW i d t h = new int [2][ n ][ n ];
u n i q u e n e s s T r e e V e r t i c e s = new int [2][ n ][ n ][ n * n ];
u n i q u e n e s s T r e e V e r t e x C h i l d C o u n t = new int [2][ n ][ n ][ n * n ];
u n i q u e n e s s T r e e V e r t e x C h i l d C o u n t O c c u r r e n c e = new int [2][ n ][ n ][ n ];
vertexMapped = new boolean [2][ n ];
}
public static void g en e r at e R an d o mG r a ph ( int size ) {
vertexCount [0] = size ;
for ( int i = 0; i < size ; i ++) {
for ( int j = 0; j < size ; j ++) {
if ( i != j && rnd . nextDouble () <= 0.5) {
boolean verticesAdjacent = false ;
15
for ( int k = 0; k < vertexDegree [0][ i ]; k ++) {
if ( vertexAdjacency [0][ i ][ k ] == j ) {
verticesAdjacent = true ;
}
}
if (! verticesAdjacent ) {
vertexAdjacency [0][ i ][ vertexDegree [0][ i ]] = j ;
vertexAdjacency [0][ j ][ vertexDegree [0][ j ]] = i ;
vertexDegree [0][ i ] += 1;
vertexDegree [0][ j ] += 1;
}
}
}
}
}
public static void g e n e r a t e I s o m o r p h i c G r a p h () {
vertexCount [1] = vertexCount [0];
for ( int i = 0; i < vertexCount [0]; i ++) {
vertexMap [ i ] = i ;
}
for ( int i = 0; i < vertexCount [0]; i ++) {
int swap = rnd . nextInt ( vertexCount [0]) ;
int temp = vertexMap [ i ];
vertexMap [ i ] = vertexMap [ swap ];
vertexMap [ swap ] = temp ;
}
for ( int i = 0; i < vertexCount [0]; i ++) {
vertexDegree [1][ vertexMap [ i ]] = vertexDegree [0][ i ];
for ( int j = 0; j < vertexDegree [0][ i ]; j ++) {
vertexAdjacency [1][ vertexMap [ i ]][ j ] = vertexMap [ vertexAdjacency
[0][ i ][ j ]];
}
}
}
public static void g e n e r a t e N o n I s o m o r p h i c G r a p h () {
g e n e r a t e I s o m o r p h i c G r a p h () ;
for ( int i = 0; i < vertexCount [1]; i ++) {
if ( vertexDegree [1][ i ] > 0) {
vertexDegree [1][ i ] -= 1;
vertexDegree [1][ vertexAdjacency [1][ i ][ vertexDegree [1][ i ]]] -= 1;
for ( int j = 0; j < vertexCount [1]; j ++) {
if ( i != j && j != vertexAdjacency [1][ i ][ vertexDegree [1][ i ]]) {
boolean verticesAdjacent = false ;
16
for ( int k = 0; k < vertexDegree [1][ j ]; k ++) {
if ( vertexAdjacency [1][ j ][ k ] == i ) {
verticesAdjacent = true ;
}
}
if (! verticesAdjacent ) {
vertexAdjacency [1][ i ][ vertexDegree [1][ i ]] = j ;
vertexAdjacency [1][ j ][ vertexDegree [1][ j ]] = i ;
vertexDegree [1][ i ] += 1;
vertexDegree [1][ j ] += 1;
j = vertexCount [1];
i = vertexCount [1];
}
}
}
}
}
}
public static void g e n e r a t e U n i q u e n e s s T r e e s () {
for ( int i = 0; i < 2; i ++) {
for ( int j = 0 ; j < vertexCount [ i ]; j ++) {
u n i q u e n e s s T r e e V e r t i c e s [ i ][ j ][0][0] = j ;
u n iq u e ne s s Tr e e Wi d t h [ i ][ j ][0] = 1;
int u niqueV ertexC ount = 1;
while ( uniqu eVerte xCount > 0 && u n i q u e n e s s T r e e H e i g h t [ i ][ j ] < (
vertexCount [ i ] - 1) ) {
u n i q u e n e s s T r e e H e i g h t [ i ][ j ] += 1;
uniq ueVert exCoun t = 0;
for ( int k = 0; k < u n iq u e ne s s Tr e e Wi d t h [ i ][ j ][
u n i q u e n e s s T r e e H e i g h t [ i ][ j ] - 1]; k ++) {
boolean vertexUnique = true ;
for ( int l = 0; l < u n iq u e ne s s Tr e e Wi d t h [ i ][ j ][
u n i q u e n e s s T r e e H e i g h t [ i ][ j ] - 1]; l ++) {
if ( k != l && u n i q u e n e s s T r e e V e r t i c e s [ i ][ j ][
u n i q u e n e s s T r e e H e i g h t [ i ][ j ] - 1][ k ] ==
u n i q u e n e s s T r e e V e r t i c e s [ i ][ j ][ u n i q u e n e s s T r e e H e i g h t [ i ][ j ] 1][ l ]) {
vertexUnique = false ;
}
}
if ( vertexUnique ) {
uni queVer texCou nt += 1;
17
for ( int l = 0; l < vertexDegree [ i ][ u n i q u e n e s s T r e e V e r t i c e s [ i
][ j ][ u n i q u e n e s s T r e e H e i g h t [ i ][ j ] - 1][ k ]]; l ++) {
u n i q u e n e s s T r e e V e r t i c e s [ i ][ j ][ u n i q u e n e s s T r e e H e i g h t [ i ][ j ]][
u n iq u e ne s s Tr e e Wi d t h [ i ][ j ][ u n i q u e n e s s T r e e H e i g h t [ i ][ j ]]]
=
vertexAdjacency [ i ][ u n i q u e n e s s T r e e V e r t i c e s [ i ][ j ][
u n i q u e n e s s T r e e H e i g h t [ i ][ j ] - 1][ k ]][ l ];
u n i q u e n e s s T r e e V e r t e x C h i l d C o u n t [ i ][ j ][ u n i q u e n e s s T r e e H e i g h t [
i ][ j ] - 1][ k ] += 1;
u n iq u e ne s s Tr e e Wi d t h [ i ][ j ][ u n i q u e n e s s T r e e H e i g h t [ i ][ j ]] +=
1;
}
}
}
}
}
}
}
public static boolean c om pu te Is om or ph is m () {
for ( int i = 0; i < 2; i ++) {
for ( int j = 0; j < vertexCount [ i ]; j ++) {
for ( int k = 0; k < u n i q u e n e s s T r e e H e i g h t [ i ][ j ]; k ++) {
for ( int l = 0; l < u n iq u e ne s s Tr e e Wi d t h [ i ][ j ][ k ]; l ++) {
u n i q u e n e s s T r e e V e r t e x C h i l d C o u n t O c c u r r e n c e [ i ][ j ][ k ][
u n i q u e n e s s T r e e V e r t e x C h i l d C o u n t [ i ][ j ][ k ][ l ]] += 1;
}
}
}
}
for ( int i = 0; i < vertexCount [0]; i ++) {
for ( int j = 0; j < vertexCount [1]; j ++) {
if (! vertexMapped [0][ i ] && ! vertexMapped [1][ j ]) {
boolean ve rt ice sE qu iv al en t = true ;
if ( u n i q u e n e s s T r e e H e i g h t [0][ i ] != u n i q u e n e s s T r e e H e i g h t [1][ j ]) {
v ert ic es Eq ui va le nt = false ;
}
for ( int k = 0; k < u n i q u e n e s s T r e e H e i g h t [0][ i ]; k ++) {
if ( un i q u en e s sT r e eW i d th [0][ i ][ k ] != u ni q u en e s sT r e eW i d th [1][ j ][
k ]) {
v ert ic es Eq ui va le nt = false ;
}
for ( int l = 0; l < vertexCount [0]; l ++) {
if ( u n i q u e n e s s T r e e V e r t e x C h i l d C o u n t O c c u r r e n c e [0][ i ][ k ][ l ] !=
u n i q u e n e s s T r e e V e r t e x C h i l d C o u n t O c c u r r e n c e [1][ j ][ k ][ l ]) {
v ert ic es Eq ui va le nt = false ;
18
}
}
}
if ( v er ti ce sE qu iv al en t ) {
vertexMapped [0][ i ] = true ;
vertexMapped [1][ j ] = true ;
}
}
}
}
boolean graphsIsomorphic = true ;
for ( int i = 0; i < vertexCount [0]; i ++) {
if (! vertexMapped [0][ i ]) {
graphsIsomorphic = false ;
}
}
return graphsIsomorphic ;
}
public static void main ( String [] args ) {
int falseResults = 0;
for ( int i = 1; i < 100; i ++) {
System . out . println ( i ) ;
long startTime = System . cu rrentT imeMil lis () ;
for ( int j = 1; j < 100; j ++) {
initialise ( i ) ;
g e ne r a te R a nd o m Gr a p h ( i ) ;
g e n e r a t e I s o m o r p h i c G r a p h () ;
// g e n e r a t e N o n I s o m o r p h i c G r a p h () ;
g e n e r a t e U n i q u e n e s s T r e e s () ;
if (! c om pu te Is om or ph is m () ) {
// if ( c om pu te Is om or ph is m () && i > 9) {
falseResults += 1;
System . out . println (" False result ! " + falseResults ) ;
}
}
try {
fileWriter = new FileWriter (" output . csv " , true ) ;
bufferedWriter = new BufferedWriter ( fileWriter ) ;
bufferedWriter . write ( Integer . toString ( i ) + " , " + Integer . toString
(( int ) ( System . cu rrentT imeMil lis () - startTime ) ) ) ;
bufferedWriter . newLine () ;
bufferedWriter . close () ;
19
} catch ( IOException ex ) {
System . out . println (" Error writing to output file .") ;
}
}
}
}
20
| 8cs.DS
|
Maximal PSL2 subgroups of exceptional groups of Lie type
David A. Craven
arXiv:1610.07469v1 [math.GR] 24 Oct 2016
December 19, 2017
Abstract
In this article we study embeddings of PSL2 (q0 ) into exceptional groups G(q) for G = F4 , E6 , E7 ,
and q0 and q powers of the same prime p. With a few possible exceptions, we prove that there are no
maximal subgroups with socle such a simple group inside an almost simple group with socle G(q), except
for those that arise as fixed points of a maximal positive-dimensional subgroup of the corresponding
algebraic group.
In the few remaining cases we provide considerable information about a potential maximal subgroup.
Contents
1 Introduction
2
2 Notation and preliminaries
5
3 Maximal subgroups
9
4 Unipotent and semisimple elements
14
4.1
Actions of unipotent elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
4.2
Blueprints and element orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
4.3
Blueprints inside A1 s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
a
4.4
Traces for modules of PGL2 (p ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
4.5
The graph automorphism of F4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
4.6
sl2 -subalgebras of L(G) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
5 Modules for SL2 (pa )
5.1
25
a
25
a
Modules for SL2 (2 ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2
Modules for SL2 (3 ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
5.3
Modules for SL2 (p) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
5.4
a
Modules for SL2 (p ) for p > 5 and a > 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
6 Some PSL2 s inside E6 in characteristic 3
39
7 Proof of the theorems: strategy
41
1
8 F4
43
8.1
Characteristic 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
8.2
Characteristic 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
8.3
Characteristic at least 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
9 E6
51
9.1
Characteristic 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
9.2
Characteristic 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
52
9.3
Characteristic at least 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
52
10 E7 in characteristic 2
57
11 E7 in odd characteristic: PSL2 embedding
74
11.1 Characteristic 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
74
11.2 Characteristic at least 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
75
12 E7 in odd characteristic: SL2 embedding
82
12.1 Characteristic 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
82
12.2 Characteristic at least 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
86
A Actions of maximal positive-dimensional subgroups on minimal and adjoint modules
1
101
Introduction
Classifying the maximal subgroups of a finite group is one of the most fundamental problems in the field
of finite group theory. Michael Aschbacher and Len Scott [5] reduced the problem for all finite groups to
understanding H 1 (G, M ) for all simple modules M for all finite simple groups G, and classifying all maximal
subgroups of almost simple groups.
This paper is a contribution towards the latter, ambitious goal. For alternating and classical groups
there is in some sense no complete answer, since the dimensions of the classical groups (and degrees of
the alternating groups) tend to infinity, although there is substantial work in this direction. However, for
sporadic and exceptional groups there is a possibility of a complete answer being known.
For sporadic groups, a complete answer is known for all groups but the Monster, and here we concentrate
on exceptional groups of Lie type. There is a classification of maximal subgroups for exceptional groups
G = G(q) for G not of type F4 , E6 , 2E6 , E7 and E8 already, and so we focus on these cases. What is known
in the literature so far is summarized in Section 3, but broadly speaking, all maximal subgroups are known
in these groups apart possibly from various almost simple maximal subgroups, and these are either a small
list of simple groups that are not Lie type in defining characteristic, and if the potential maximal is Lie type
in defining characteristic then what is left are groups of small rank and small field size, together with a large
collection of possible subgroups PSL2 (pa ), the focus of this paper. We prove the following theorems, for any
almost simple group of the appropriate type.
Theorem 1.1 Let p be a prime, a > 1 be an integer, and let q be a power of p. Let G be an almost simple
group with socle F4 (q), and suppose that H is an almost simple group with F ∗ (H) = PSL2 (pa ). If H is
maximal in G then one of the following holds:
2
(i) pa = 9;
(ii) pa = 13, H = PSL2 (13) and is a Serre embedding;
(iii) q = pa , p > 13, F ∗ (H) = PSL2 (q), and H is the intersection of G with a maximal algebraic A1
subgroup of the algebraic group F4 .
The definition of a Serre embedding is given formally in Definition 4.7, but informally it is a copy of
PSL2 (h + 1) where h is the Coxeter number of G and this subgroup contains a regular unipotent element.
(This subgroup is named after Serre as he constructed copies of PSL2 (h + 1) (if h + 1 is a prime) over all
fields in [27].)
In recent work of Tim Burness and Donna Testerman, this case has been solved, and proved to be a
subcase of (iii) above, so pa = 9 is the only outstanding case. It seems difficult to remove the first possibility,
although it might be possible using more advanced geometric techniques than employed here; of course no
such maximal subgroup is known. In Section 8.2 we give more information about the case pa = 9, where we
give the action of a potential maximal subgroup on the minimal module; such a subgroup does exist, but is
contained inside a positive-dimensional subgroup, and representation-theoretic techniques do not seem able
to prove uniqueness.
Kay Magaard [24] proved Theorem 1.1 for p > 5 in his PhD thesis, with the extra condition that q = 13
in (ii).
For E6 we have a complete theorem, as the Serre embedding can be shown to lie in F4 .
Theorem 1.2 Let p be a prime, a > 1 be an integer, let q be a power of p, and let G be an almost simple
group with socle either E6 (q) or 2E6 (q). There does not exist an almost simple maximal subgroup H of G
with F ∗ (H) = PSL2 (pa ).
Almost all of this theorem was obtained by Aschbacher [4] using geometric techniques, where only the
case q = pa = 11 and H contains a semiregular unipotent element, from class E6 (a1 ), is left open; here we
remove it using the Lie algebra structure of the adjoint module L(G).
For E7 , here we again have some potential exceptions, but not the case pa = 9, which was completed in
[9]. This time the difficult cases are the Serre embedding and pa = 7, 8, 25.
Theorem 1.3 Let p be a prime, a > 1 be an integer and let q be a power of p. Let G be an almost simple
group with socle E7 (q), and suppose that H is an almost simple subgroup with F ∗ (H) = PSL2 (pa ). If H is
maximal in G then one of the following holds:
(i) pa = 7, pa = 8 or pa = 25;
(ii) pa = 19, H = PSL2 (pa ) and is a Serre embedding;
(iii) q = pa , p > 17, F ∗ (H) = PSL2 (q), and H is the intersection of G with a maximal algebraic A1
subgroup of the algebraic group E7 .
Again, Burness and Testerman have showed that (ii) is a subcase of (iii). In the case (i) where pa = 8,
we can give the composition factors of H on the minimal module, and can give the precise module structure
as well whenever 8 | q. For pa = 7, there are unresolved cases of potential copies of PSL2 (7) where the
preimage of the subgroup in the simply connected version of E7 is both 2 × PSL2 (7) and SL2 (7). In both
cases the module structures on the minimal module can be given precisely, but it seems difficult to progress
3
further using these techniques. In the case of pa = 25, this is a copy of SL2 (25) inside the simply connected
version of E7 with centres coinciding, and we have complete information about the module structures on
both the minimal and adjoint modules. If it exists then it is a maximal subgroup of E7 (q) for the smallest
q into which the group embeds.
We do not deal with maximal subgroups of E8 here, and only consider it for certain lemmas, which will
be useful in a later treatment of this case. A rough estimate is that, with current techniques, attempting E8
here would result in many unresolved cases and double the length of this work. For exceptional groups other
than E8 , the minimal module has dimension much smaller than the dimension of the group (as an algebraic
group) and we can use representation theory to analyse this module. We can still do things with the Lie
algebra for E8 , as we did in [9],
The strategy for the proofs of these theorems is given in Section 7, and relies heavily on computer
calculations in three ways:
(i) The first is to compute the traces of semisimple elements of large order on various modules for exceptional groups. Tables of these traces are available for elements of small order, but we need them for
very large orders, sometimes in the hundreds. For this we can use the program that Litterick produced
in his PhD thesis [22], or construct the normalizer of a torus explicitly in Magma and take the conjugacy classes, then compute their eigenvalues. (Litterick has produced a much faster algorithm for
computing traces of elements on fundamental modules, but we do not need this for our cases.)
(ii) The second is to do large linear algebra problems. To find all sets of composition factors that could
arise as the composition factors of the restriction of a G-module to a subgroup H involves checking
many possible combinations against these large lists of possible sets of composition factors. This is
done to reduce the possible module structures for the subgroup on the minimal and adjoint modules,
and was also used in [22].
(iii) The third is to construct explicit modules for finite groups, and show that certain module structures
cannot exist. This would be possible by hand, at least in some cases, but incredibly complicated and
prone to mistakes. In each case, a clear recipe is given for how to reproduce the module we construct
to ease verifiability.
With these three uses of a computer in mind, the rest of the argument is done by hand, in Sections 8 to 12.
The structure of this article is as follows: in the next section we give notation and some preliminary
results. In Section 3 we give information about maximal subgroups of finite and algebraic exceptional
groups, and in the following section we give lots of information about unipotent and semisimple elements of
exceptional groups, together with information about sl2 -subalgebras of exceptional Lie algebras. Section 5
gives information about modules for SL2 (pa ), and the section after gives some constructions of PSL2 s inside
E6 in characteristic 3.
We then launch into the proof proper, with Section 7 giving an outline of the strategy of the proof, Sections
8 and 9 proving the results for F4 and E6 , and then the three sections after doing E7 in characteristic 2, and
then E7 in odd characteristic, split into two sections according as the embedding into the simply connected
group is 2 × PSL2 (pa ) and SL2 (pa ).
The appendix gives some information about the composition factors of the reductive and parabolic
maximal subgroups of F4 , E6 and E7 on the minimal and adjoint modules, information that is well known
but given here for ease of reference.
4
2
Notation and preliminaries
In this section we give the notation that we need, both for groups and for modules, and give a few preliminary
results.
Throughout this paper, G = G(k) will denote an exceptional finite group of Lie type defined over k,
a field of characteristic p > 2. More specifically, let G be a simple, simply connected algebraic group of
¯
exceptional type, equipped with a Frobenius endomorphism σ and set G = Gσ . The precise types of G that
¯
we are interested in are those exceptional groups whose maximal subgroups are not yet known, i.e., F4 (q),
E6 (q), 2E6 (q), E7 (q) and E8 (q), although we do not do much in the case of E8 (q), and often will exclude it
from consideration.
Notice that we consider the simply connected version of G, so E7 (k) possesses a centre when p is odd.
We want the simply connected versions in order to work with the minimal module and the adjoint module
simultaneously. Where this is particularly important we will remind the reader, for example when considering
PSL2 (pa ) embedded in the simple group of type E7 , where in E7 (k) we can embed either SL2 (pa ) in E7 (k)
with the centres coinciding or 2 × PSL2 (pa ) into E7 (k) with the centres coinciding, representing the two
possible preimages of a copy of PSL2 (pa ) in the simple group. If G possesses a graph automorphism of order
2, denote this by τ ; we will remind the reader of this notation when we use it.
We let Ḡ be an almost simple group with socle G/Z(G). The maximal subgroups M of Ḡ split into three
categories: M ∩ G is a maximal subgroup of G, M ∩ G is not a maximal subgroup of G, and G 6 M . The
third collection are easily computed, and the first can be deduced from a list of maximal subgroups of G.
However, the second, called novelty maximal subgroups, cannot easily be seen from the maximal subgroups
of G. They arise in the following manner: let H be a subgroup that is not maximal in a simple group X,
but H is normalized by a group of outer automorphisms A of X while every proper subgroup of X properly
containing H is not normalized by it. In this case, H.A is a maximal subgroup of X.A. However, it is of
course very difficult to understand these if one is simply given a list of maximal subgroups of X, so we will
prove more than simply that a given subgroup is not maximal in the simple group, but that it is contained
in stabilizers of various subspaces of a given module, enough that we can see that it does not form a novelty
maximal subgroup.
The modules that we normally consider are the two smallest non-trivial ones. Write Vmin for one of the
minimal modules for G, namely L(λ4 ) for F4 , either L(λ1 ) or L(λ6 ) for E6 and 2E6 , L(λ7 ) for E7 and not
defined for E8 . We write L(G) for the Lie algebra or adjoint module, which is L(λ1 ), L(λ2 ), L(λ1 ) and
L(λ1 ) respectively. If L(G) has a trivial composition factor so is not irreducible, which occurs in E7 in
characteristic 2 and E6 in characteristic 3, let L(G)′ denote the other simple factor, and in other cases let
L(G)′ = L(G). These two modules have the following dimensions:
Group
dim(Vmin )
dim(L(G)′ )
F4
26 − δp,3
52
E6
27
78 − δp,3
E7
56
133 − δp,2
E8
248
248
τ
If G = F4 in characteristic 2, L(G) has factors Vmin and Vmin
, where τ denotes the graph automorphism
of G, so in this case we will not consider L(G) at all but these two modules. In all other cases, L(G)′ is
irreducible.
We now introduce some notation for modules. All modules will be finite dimensional and will normally
5
be defined over k, the field over which G is defined. If H is a group, let Irr(H) denote the set of irreducible
modules over the field, which is always k unless otherwise stated. As usual write ‘⊕’ and ‘⊗’ for the direct
sum and tensor product of two modules. Let Λi and S i denote the exterior and symmetric powers. We
write M ↓H for the restriction of M to H, and write soci (M ) for the ith socle layer and radi (M ) for the ith
radical layer of M . Write top(M ) for the top of M , i.e., M/ rad(M ), and cf(M ) for the composition factors
of M as a multiset. Let H 1 (H, M ) denote the 1-cohomology group of M , and in general Ext1 (M, M ′ ) denote
the group of extensions with submodule M ′ and quotient M . The projective cover of a module M will be
denoted by P (M ).
We will often have to talk about the structures of modules, as in their socle layers. If M is a module
with socle A and second socle B then we can write
B
A
for this structure; however, this is often too space-consuming when we have many socle layers, and so we
also write B/A for this module.
We also introduce the concepts of radical and residual. If I is a subset of Irr(H), then the I-radical of
M is the largest submodule of M whose composition factors lie in I, and let I ′ = Irr(H) \ I. The I-residual
of M is the smallest submodule whose quotient has composition factors in I.
One lemma that we occasionally use, that can be quite powerful, relates the minimal and adjoint modules
for exceptional groups. We place it here because there seems no more appropriate place.
Lemma 2.1 Let G be one of F4 , E6 and E7 .
(i) Let G = F4 . If p = 3 then L(G) is a submodule of Λ2 (Vmin ). If p > 5 then L(G) is a summand of
Λ2 (Vmin ).
∗
∗
(ii) Let G = E6 . If p = 2 then L(G) is a submodule of Vmin ⊗ Vmin
. If p = 3 then the socle of Vmin ⊗ Vmin
is 1-dimensional, and quotienting out by this, L(G)′ is a submodule. If p > 5 then L(G) is a summand
∗
of Vmin ⊗ Vmin
.
(iii) Let G = E7 . If p = 2 then the socle of Λ2 (Vmin ) is 1-dimensional, and quotienting out by this, L(G)′
is a submodule. If p = 3 then L(G) is a submodule of S 2 (Vmin ). If p > 5 then L(G) is a summand of
S 2 (Vmin ).
In many cases we want to prove that a module has a particular composition factor as a submodule or
quotient, often the trivial module. Thus we need a method of proving that a particular composition factor
is always a submodule or quotient in any module with those factors. This is the idea of pressure.
Suppose that H is a finite group such that Op (H) = H, and such that for all simple modules M over a
field k, H 1 (H, M ) = H 1 (H, M ∗ ). The pressure of a module V for H is the quantity
X
dim H 1 (H, M ) − δM,k .
M∈cf(V )
Results on pressure have occurred in the literature before, with the most general being in [9]. Another
generalization of this allows us to understand the situation of forcing a module from a collection M of
simple modules to be a submodule of a given module V . If M is a collection of simple modules for a group
6
H, with Ext1 (M, M ′ ) = 0 for all M, M ′ ∈ M, and such that Ext1 (A, M ) = Ext1 (M, A) for all simple
modules A and M with M ∈ M, then the M-pressure of a module V is the quantity
X
X
Ext1 (M, M ′ ) − δM,M ′ .
M ′ ∈cf(V ) M∈M
The lemma from [9] directly generalizes to M-pressure, with the exact same proof, and we give it now.
Lemma 2.2 Suppose that H is a finite group, and let M be a set of simple modules for H such that
Ext1 (M, M ′ ) = 0 for all M, M ′ ∈ M, and Ext1 (M, A) = Ext1 (A, M ) for all M ∈ M and all simple modules
A. Let V be a module for H of M-pressure n.
(i) If n < 0 then Hom(M, V ) 6= 0 for some M ∈ M, i.e., V has a simple submodule isomorphic to some
M ∈ M. If n = 0 then either Hom(M, V ) 6= 0 or Hom(V, M ) 6= 0, i.e., V has either a simple submodule
or quotient isomorphic to some member of M.
(ii) More generally, if a composition factor of V has M-pressure greater than n, then either Hom(M, V ) 6= 0
or Hom(V, M ) 6= 0 for some M ∈ M.
(iii) If Hom(M, V ) = Hom(V, M ) = 0 for all M ∈ M, then any subquotient W of V has M-pressure
between −n and n.
The concept of pressure can be used to prove that either Vmin or L(G) possesses a trivial submodule or
quotient when restricted to H. We therefore would like to know whether that is enough in some circumstances
to conclude that H is contained within a σ-stable, positive-dimensional subgroup of G.
Lemma 2.3 [[9, Lemma 1.4]] Let G = Gσ be one of F4 , E6 , 2E6 , E7 or E8 . Let H 6 Gσ , acting on a module
¯
V defined over k, where V is either Vmin or L(G)′ and k is the underlying field of G. If one of the following
holds, then H is contained in a σ-stable, positive-dimensional subgroup of G:
(i) H fixes a 1-space or hyperplane of Vmin or L(G);
(ii) G = F4 , E6 , 2E6 or E7 , and H fixes a 2-space or a space of codimension 2 in Vmin ;
(iii) G = E6 or 2E6 , and H fixes a 3-space or a 24-space of Vmin .
In the next section we consider the set of maximal positive-dimensional subgroups, and this lemma will
more or less translate across to the almost simple group Ḡ.
We end with giving the line stabilizers for the minimal modules for the finite groups E6 (k) and E7 (k).
These have appeared in the literature before, and we take these from [15, Lemmas 5.4 and 4.3].
Lemma 2.4 Let G = E6 (q). There are three orbits of lines of the action of G on Vmin , with line stabilizers
as follows:
(i) F4 (q) acting on Vmin as L(λ4 ) ⊕ L(0);
(ii) a D5 -parabolic subgroup; q 16 D5 (q).(q − 1), acting uniserially as L(λ1 )/L(λ4 )/L(0).
(iii) a subgroup q 16 .B4 (q).(q − 1) acting indecomposably as L(0), L(λ1 )/L(λ4 )/L(0).
Lemma 2.5 Let G = E7 (q). There are five orbits of lines of the action of G on Vmin , with line stabilizers
as follows:
7
(i) E6 (q).2 (the graph automorphism) acting semisimply with composition factors of dimensions 54, 1, 1;
(ii) 2E6 (q).2 (the graph automorphism) acting semisimply with composition factors of dimensions 54, 1, 1;
(iii) an E6 -parabolic subgroup q 27 .E6 (q).(q − 1) acting uniserially as L(0)/L(λ1 )/L(λ6 )/L(0);
(iv) a subgroup q 1+32 .B5 (q).(q − 1) acting uniserially as L(0)/L(λ1 )/L(λ5 )/L(λ1 )/L(0);
(v) a subgroup q 26 .F4 (q).(q − 1) acting indecomposably as L(0), L(0)/L(λ4 )/L(λ4 )/L(0), L(0).
8
3
Maximal subgroups
This section summarizes what is known about the maximal subgroups of the finite groups G and Ḡ, and
also the algebraic group G, about which complete information on positive-dimensional maximal subgroups
¯
is known.
The maximal subgroups of positive dimension in G are given in [20], and given G we denote by X this
¯
¯
collection; write X σ for the fixed-point sets Xσ for X ∈ X being σ-stable. Note that we also include in
¯
¯
X σ the fixed points of G under a field, graph, or field-graph automorphism of prime order (so, for example,
2
E6 (p2 ) and E6 (p) inside E6 (p2 )). If Ḡ is almost simple rather than merely simple, the set X σ shall be
taken to mean the normalizers in Ḡ of the elements of X σ for F ∗ (Ḡ).
While the maximal subgroups of G are known, the maximal subgroups of G and Ḡ are of course not. We
¯
start with a broad characterization of the maximal subgroups of Ḡ, given in [7] and [17, Theorem 2].
Theorem 3.1 Let M be a maximal subgroup of Ḡ not containing F ∗ (Ḡ). One of the following holds:
(i) M is a member of X σ ;
(ii) M is the normalizer of an elementary abelian r-group for some r 6= p;
(iii) M = (Alt(5) × Sym(6)) · 2 and G = E8 with p > 5;
(iv) M is almost simple.
The algebraic groups containing the subgroups in (i) are known and are the fixed points of those in [20];
the subgroups in (ii) are known and given in [8]; the subgroup (iii) was discovered by Borovik and is unique
up to conjugacy. The potential subgroups in (iv) have been steadily reduced over the last two decades. We
start with those almost simple groups that are not Lie type in defining characteristic. Here the list is fairly
short and given in [19], but note that a fair number of these have been eliminated in a variety of papers,
too numerous to list here, but we mention the papers [23] and [9] for all Lie type groups, and with F4 and
E6 having almost all possibilities for M removed by Magaard and Aschbacher in [24] and [4] respectively.
The author has also made progress on eliminating still more of this list and proving uniqueness of various
maximal subgroups, with details appearing elsewhere.
For M a group of Lie type in defining characteristic, define t(G) and v(G) to be the following integers:
t(G2 ) = 12,
t(F4 ) = 68,
t(E6 ) = 124,
t(E7 ) = 388,
v(G2 ) = 4,
v(F4 ) = 18,
v(E6 ) = 18,
v(E7 ) = 75,
t(E8 ) = 1312.
v(E8 ) = 1312.
The rank of M is at most half the rank of G by [16] and [21]. Furthermore, for those groups we have the
following possibilities by [18]:
(i) M (q) has semisimple rank at most half that of G, q 6 9, and M (q) is not one of PSL2 (q), 2B2 (q) and
2
B2 (q);
(ii) PSL3 (16) and PSU3 (16);
(iii) PSL2 (q), 2B2 (q) and 2 G2 (q) for q 6 gcd(2, q − 1) · t(G).
9
The paper [10] allows us to replace t(G) by v(G) in (iii). For (ii), note that PSL3 (16) has elements of order
9, and PSU3 (16) has elements of order 255, so neither case can occur for G 6= E8 by [10, Theorem 1.1]. The
author, Kay Magaard and Chris Parker have removed almost all of (i) for G 6= E8 in work in preparation,
as well as the Suzuki and Ree groups from (iii), leaving just PSL2 (pa ), which we consider in this paper.
We can therefore assume, from now on, that pa 6 gcd(2, p − 1) · v(G). We remind the reader at the start
of each section the value of v(G).
We wish to end this section with a result that states that if H is a copy of PSL2 (pa ) inside an exceptional
group of Lie type (other than E8 ), then NḠ (H) is either an almost simple maximal subgroup or is inside
a member of X σ . The rest of this paper will be spent proving that the latter case holds rather than the
former, but for this section we will need to have some exceptions. One source of possible exceptions is that
NḠ (H) is contained inside another maximal subgroup of Ḡ other than those in X σ , for example a copy of
PSp6 (p), which contains PSL2 (p3 ).
The statement of the next result, and the proofs of the next two results, use ideas, definitions and
techniques that will be introduced throughout this paper, but logically the results should be in this section.
As such, the author recommends that the reader does not read the proofs of these results until after they
have read the next few sections.
In order to reduce the list of exceptions that arise, we will remove some of the Lie type groups of medium
rank appearing in (i) above. We start with a table giving the largest possible order of semisimple elements
of various groups of Lie type. The group type appears on the left and the field size on the top. In each
entry, there is a number which is the largest order of a semisimple element in the simple group, and if this
is even we include the largest element of odd order in brackets, then for groups for which not all semisimple
elements are real (type A only) we place after that the largest order of a real semisimple element.
Group
2
3
4
5
7
8
9
PSL3
7, 3
13, 4
7, 5
31, 6
19, 8
73, 9
91, 10
PSL4
15, 5
20 (13), 10
85, 17
39, 13
200 (171), 50
585, 65
205, 41
PSU3
-
8 (7), 4
15, 5
8 (7), 6
48 (43), 8
21, 9
80 (73), 10
PSU4
9, 5
8 (7), 8
65, 17
63, 26
86 (75), 25
513, 65
365, 82
PSp4
5
5
17
13
25
65
41
PSp6
15
20 (13)
85
78 (63)
200 (171)
585
410 (365)
PΩ7
15
20 (13)
85
78 (63)
200 (171)
585
410 (365)
G2
7
13
21
31
57
73
91
(We omit PSU3 (2), which is not simple, and consider the derived subgroup when the group is not simple,
i.e., PSp4 (2)′ and G2 (2)′ .)
We now compare these numbers to v(G): for F4 if the first number is greater than 18 then the subgroup
is a blueprint for Vmin ; for E6 if the number in brackets is greater than 75 or the second number is greater
than 18 then the subgroup is a blueprint for Vmin ; for E7 the number in brackets needs to be greater than
75 for this. For example, subgroups isomorphic with G2 (9) and PSp6 (8) are always blueprints for Vmin when
inside F4 , E6 and E7 , whereas PSU4 (7) is always a blueprint for Vmin for F4 and E6 , but not necessarily for
E7 .
We prove an intermediate proposition that will help in our stated goal of producing the result we mentioned about NḠ (H) being either almost simple or in a member of X σ when H is PSL2 (pa ).
Proposition 3.2 Let G be one of F4 , E6 and E7 .
10
(i) For p = 5, 7, any copy of H = PSp4 (p) in G, or H = Sp4 (p) in G = E7 (k) with Z(H) = Z(G), is a
blueprint for Vmin .
(ii) For p = 5, 7, any copy of H = PSL4 (p) or PSU4 (p) in G, or H = 2 · PSL4 (p) or H = 2 · PSU4 (p) in
G = E7 (k) with Z(H) = Z(G), is a blueprint for Vmin .
(iii) Let p be an odd prime and a > 1. Any copy of H = PSp6 (pa ) in G, or H = Sp6 (pa ) in G = E7 (k)
with Z(H) = Z(G), is a blueprint for Vmin .
(iv) Let p be an odd prime and a > 1. Any copy of H = Ω7 (pa ) in G, or H = Spin7 (pa ) in G = E7 (k) with
Z(H) = Z(G), is a blueprint for Vmin .
Proof: We prove (i) and (ii) for p = 5 first. For the first part, we first compute the conspicuous set of
composition factors for Vmin ↓H in the case of G = E7 . The simple modules of dimension at most 56 are 1,
5, 10, 13, 30, 351 , 352 and 55. The only conspicuous set of composition factors is 102 , 56 , 16 , and since these
composition factors have no extensions with each other, Vmin ↓H is semisimple, thus a unipotent element u
from the conjugacy class of H with the largest centralizer acts on Vmin as 32 , 216 , 118 , a generic class. This
proves the result since generic classes are blueprints for Vmin . Of course, since the minimal modules for F4
and E7 are submodules of Vmin , the result holds for F4 and E6 as well.
If H = Sp4 (5) 6 E7 (k) with centres coinciding, then the involutions in H act on faithful modules with
trace 0, not allowed since the trace of an involution in E7 is ±8. This proves (i).
As SL4 (5) and SU4 (5) contain Sp4 (5), and the centres of SL4 (5) and SU4 (5) contain the centre of Sp4 (5),
we have that PSp4 (5) 6 PSL4 (5), PSU4 (5), and therefore (ii) holds as subgroups that contain blueprints are
themselves blueprints.
For p = 7 the exact same proof holds, except that the dimensions of the simple modules are now 1, 5,
10, 14, 25, 351 , 352 and 54.
We now prove (iii). For pa 6= 3, 5, the largest semisimple element of odd order has order greater than
75, so these are already blueprints for Vmin . If H = PSp6 (3) then there are only three simple modules of
dimension at most 56, with dimensions 1, 14 and 21. The traces of elements of orders 5 and 7 are enough to
prove that H does not embed in G = E7 , and hence not in its subgroups. If H = Sp6 (3) then the appropriate
simple modules have dimensions 6, 14 and 50, and traces of elements of order 5 are enough to prove that
the only conspicuous set of composition factors for Vmin ↓H is 14, 67 . There are no extensions between
composition factors, so this is semisimple, and the action of a unipotent element with largest centralizer in
H is 212 , 132 , a generic class. We conclude that H is a blueprint for Vmin and therefore so is any subgroup
containing H, as needed.
For p = 5 all of the same statements hold except we only need traces of elements of order 3 to prove that
PSp6 (5) does not embed, and for Sp6 (5) elements of orders 2 and 3 suffice.
Finally, we consider (iv). Since the semisimple elements have the same orders in Ω7 (pa ) as PSp6 (pa ), we
again need only consider pa = 3, 5. For pa = 3, the simple modules for H = Ω7 (3) of dimension at most 56
are 1, 7, 27 and 35. The traces of elements of orders 2 and 4 are enough to find the unique conspicuous set of
composition factors, 212 , 72 . and since there are no extensions between these modules Vmin ↓H is semisimple.
A unipotent element H with maximal centralizer size acts on this module with blocks 32 , 216 , 118 , which is
generic by [13, Table 7], so that H is a blueprint for Vmin . In the other case of H = Spin7 (3), a non-central
involution in H has trace 0 on all faithful modules, and so since an involution in E7 has trace ±8, we cannot
get this case.
11
The exact same proof works for p = 5 except we use traces of elements of orders 2 and 3 to eliminate all
but one set of composition factors.
From this we can see that if H is a potential maximal subgroup, and we prove that H is contained inside
a larger subgroup that is not G, then for almost all possibilities for H, either H must lie inside another
almost simple group or it lies inside a member of X σ . This is made formal with the following proposition.
Proposition 3.3 Let H = PSL2 (pa ), let G = G(k) be an exceptional group of Lie type in characteristic p
other than E8 , and let Ḡ be an almost simple group with socle G. If H 6 G then one of the following holds:
(i) NḠ (H) is contained in a member of X σ ;
(ii) NḠ (H) is an almost simple maximal subgroup of Ḡ with socle H;
(iii) one of the following holds:
(a) G = G2 , pa = 4, 7;
(b) G = F4 , pa = 4, 5, 7, 8, 9, 11;
(c) G = E6 , pa = 4, 5, 7, 8, 9, 11, 16, 25 (for pa = 25, H 6 2F4 (2)′ );
(d) G = E7 , pa = 4, 5, 7, 8, 9, 11, 16, 25, 64 (for pa = 25, H 6 Ru).
Proof: As we wrote above, the classification-so-far of maximal subgroups of G (not equal to E8 ) states that
if M is a maximal subgroup then one of the following holds:
(1) M is a member of X σ ;
(2) M is an exotic local subgroup;
(3) F ∗ (M ) is PSL2 (pa ) for pa 6 gcd(2, p − 1) · v(G);
(4) F ∗ (M ) is 2B2 (pa ) for pa 6 32 or 2 G2 (pa ) for pa 6 27;
(5) F ∗ (M ) is a simple group of Lie type in characteristic p, whose untwisted rank is at most half of that
of G, and whose field of definition is at most 9;
(6) F ∗ (M ) is a simple group not of Lie type in defining characteristic, and is one of the groups in [19,
Section 10].
If we assume, in the proposition, that neither (i) nor (ii) holds, then there must be a maximal subgroup,
M , containing H, and such that M arises in (2), (3), (4), (5) or (6). The Suzuki and small Ree groups cannot
contain H and so we can exclude (4), and (3) is dealt with as we get (i). For (5) we have the restrictions
imposed above, and from (6) we can exclude all alternating groups other than Alt(6) and Alt(7) by [9].
The exotic local subgroups for G 6= E8 have composition factors either cyclic groups, or SL3 (2) (G2 and
above), SL3 (3) (F4 and above) and SL3 (5) (E6 and above); the first two are minimal simple groups anyway,
and the third contains only SL2 (4), so we get pa = 4 for E6 , which is also contained in Alt(6), for example,
so we can exclude (2).
For G = G2 , the possible M that are not minimal simple are PSU3 (3) which contains PSL2 (7), J1
in characteristic 11, which contains PSL2 (11), and J2 in characteristic 2, which contains SL2 (4), which
completes the proof for this case. (We exclude pa = 11 > 2 · v(G2 ).)
12
For F4 , Alt(6) contains PSL2 (5) = SL2 (4), and from Lie type in defining characteristic we can place
PSL2 (pa ) inside another Lie type group in characteristic p, say PSL3 (pa ), for pa 6 9. All other possibilities
for M cannot include PSL2 (pa ) for pa 6= 4, 5, 7, 8, 9, 11.
For E6 and E7 , from (5) we get Sp4 (8), which contains SL2 (64) (but this fails v(E6 )), and Sp4 (4) which
contains SL2 (16). From (6), note that PSL2 (25) lies inside 2F4 (2)′ and the sporadic group Ru. The former
of these must act irreducibly on the minimal module for E6 , and with factors 272 , 12 on the minimal module
for E7 , thus lies inside a E6 -parabolic and is indeed a blueprint for Vmin in this case (see the table in [23]).
The latter only embeds in E7 , and as 28 ⊕ 28∗ .
13
4
Unipotent and semisimple elements
This section collects together a variety of facts about unipotent and semisimple elements in groups of Lie
type. We consider criteria for semisimple elements that are blueprints (see Definition 4.5 below), summarizing
results of [10] and providing one more example of the calculations performed there. We then move on to
considering modules for SL2 , and how the weight spaces of the module and the eigenvalues of elements of
SL2 interact, with the aim of finding blueprint elements and subgroups of SL2 .
4.1
Actions of unipotent elements
Let G be a simple algebraic group in characteristic p. The Bala–Carter–Pommerening labelling system for
the unipotent classes, as used in a slightly modified form (to deal with interpolation of extra classes in certain
bad characteristics) in our main reference [13] for unipotent classes of exceptional groups, gives us a way to
discuss unipotent classes that is independent of the characteristic p of G. We may therefore compare the
action of a unipotent class on a fixed simple module for different primes.
As is well known, any matrix of order a power of a prime p defined over a field of characteristic p can
be written in Jordan normal form, with the conjugacy class in the general linear group being determined
by the sizes of the Jordan blocks. Thus, if u is a unipotent element of an algebraic group G then for every
module for G of dimension n we can associate a partition of n, the sizes of the various Jordan blocks in the
action of u on the module. We use the notation for this, and unipotent classes, from [13], which determines
the Jordan block structure of the action of all unipotent classes of exceptional groups on the minimal and
adjoint modules.
The only cases we will need that are not covered in [13] are when the minimal or adjoint module is
not simple, e.g., F4 in characteristic 3. The next lemma gives the actions of the unipotent classes on the
25-dimensional simple module, on the 26-dimensional module 25/1, which is the ‘minimal module’ for other
characteristics, and on the 27-dimensional minimal module for E6 , which has structure 1/25/1.
Lemma 4.1 Let u be a unipotent element in F4 (3n ). The Jordan blocks of the action of u on the 25dimensional minimal module, together with the extension 25/1 and the minimal module for E6 is one of
those given in Table 4.1.
Proof: The actions of the unipotent elements on the 26-dimensional module are given in [13, Table 3], and
using a computer, a representative of each of the classes was constructed in F4 (3). The actions on the 25dimensional composition factors were then computed, and are as above. The classes on the 25/1 are exactly
those in [13, Table 3], and the corresponding classes for E6 are in [13, Table 5].
Using a computer and constructing classes manually is the method by which we prove the next two
lemmas, which we include for completeness.
Lemma 4.2 Let u be a unipotent element in E6 (3n ). The Jordan blocks of the action of u on the 77dimensional Lie algebra module L(G)′ are obtained from the action on L(G) by removing a Jordan block of
size 1, except in the cases listed in Table 4.2.
Lemma 4.3 Let u be a unipotent element in E7 (2n ). The Jordan blocks of the action of u on the 132dimensional Lie algebra module L(G)′ are obtained from the action on L(G) by removing a Jordan block of
size 1, for every unipotent class.
14
Class in F4
Action on 25
6
Action on 25/1
13
6
2 ,1
Action on 1/25/1
14
26 , 115
A1
2 ,1
Ã1
3, 28 , 16
3, 28 , 17
3, 28 , 18
A1 + Ã1
33 , 26 , 14
33 , 26 , 15
33 , 26 , 16
A2
36 , 17
36 , 18
36 , 19
A2 + Ã1
37 , 22
37 , 22 , 1
37 , 22 , 12
Ã2 and Ã2 + A1
38 , 1
38 , 2
39
B2
5, 44 , 14
C3 (a1 )
2
5 , 4 , 3, 2
F4 (a3 )
C3 , F4 (a2 )
5, 44 , 15
5 , 42 , 3, 22 , 12
53 , 33 , 1
53 , 33 , 12
53 , 33 , 13
73 , 14
73 , 15
73 , 16
9, 62 , 3, 1
9, 62 , 3, 2
9, 62 , 32
92 , 7
92 , 7, 1
92 , 7, 12
15, 9, 1
15, 9, 2
15, 9, 3
F4 (a1 )
F4
2
2
2
5, 44 , 16
5 , 4 , 3, 2 , 1
B3
2
2
2
Table 4.1: Actions of unipotent elements on Vmin and its extensions for F4 in characteristic 3
Class in E6
2A2
Action on L(G)′
23
3 ,1
8
Action on L(G)
323 , 2, 17
324 , 22 , 1
324 , 23
93 , 82 , 64 , 32 , 14
93 , 82 , 64 , 32 , 2, 13
E6 (a3 )
94 , 7, 64 , 33 , 1
94 , 7, 64 , 33 , 2
E6 (a1 )
98 , 5
98 , 6
19, 152, 93 , 1
19, 152 , 93 , 2
2A2 + A1
A5
E6
Table 4.2: Actions of unipotent elements on L(G)′ and L(G) for E6 in characteristic 3, where one does not
obtain the former from the latter by removing a trivial Jordan block
15
We can see from the tables in [13] that for every unipotent class there is a set of primes P such that, for
any prime p 6∈ P the partition describing the Jordan block structure is the same.
Definition 4.4 Let G be an algebraic group and let M be a highest weight module for G. Let u be a
unipotent element of G. If the Jordan block structure of u on M is the same as for cofinitely many primes,
then u is said to be generic on M .
Thus, informally, the non-generic classes are those where the prime is in the set P described above, where
the partition differs from the ‘usual’ one.
The reason that generic unipotent classes are interesting is that we can find ‘nice’ A1 subgroups containing
them, at least if the class has order p. To pin down the concept of ‘nice’, we introduce the following definition.
Definition 4.5 Let G be an algebraic group and let M be a module for G. A subgroup H of G is a blueprint
for M if there exists a positive-dimensional subgroup X of G such that X and H stabilize the same subspaces
of M . An element x is a blueprint for M if hxi is.
We may now state the lemma from [9] that defines ‘nice’.
Lemma 4.6 ([9, Lemma 1.2]) Suppose that G = F4 , E6 , E7 , and if G = F4 then p is odd. If H is a finite
∗
subgroup of Ḡ and H contains a non-trivial unipotent element whose action on Vmin , Vmin ⊕ Vmin
if G = E6 ,
∗
or L(G) is generic, then u and therefore H are blueprints for Vmin , Vmin ⊕ Vmin
, or L(G) respectively, and in
particular H is contained in a member of X σ .
Thus, if any subgroup H of an exceptional algebraic group G contains a unipotent element of order p
that is generic for either the minimal or adjoint module, then H is contained inside an element of X and
indeed X σ if H is inside G = Gσ .
¯
For large primes, we will often prove that H stabilizes a unique 3-dimensional submodule of L(G), which
must be a subalgebra of the Lie algebra. If this 3-dimensional submodule of L(G) ↓H is a summand then we
may apply Proposition 4.17, but if the 3-dimensional submodule is not a summand then we cannot easily
prove that it is an sl2 , as it need not be simple. There is one case in particular where this occurs, which we
refer to as a Serre embedding. These are embeddings of PSL2 (h + 1) into an algebraic group, where h is the
Coxeter number of the group.
Definition 4.7 Let G be an exceptional algebraic group with Coxeter number h, and let p = h + 1. A
subgroup H = PSL2 (p) is a Serre embedding if the following conditions hold:
(i) on L(G), H stabilizes a unique 3-dimensional subspace;
(ii) H contains a regular unipotent element.
In Section 4.6 we discuss subalgebras of L(G).
4.2
Blueprints and element orders
Here we give a brief account of [10]; the result from it that we will use is the following theorem.
Theorem 4.8 Let x be a semisimple element of the simply connected form of an exceptional group of Lie
type G. If one of the following holds then x is a blueprint for Vmin :
16
(i) G = F4 and o(x) > 18;
(ii) G = E6 , x is real, and o(x) > 18;
(iii) G = E7 , o(x) is odd and greater than 75.
Write v(G) for these bounds, so v(F4 ) = v(E6 ) = 18, v(E7 ) = 75. If Vmin extends to a module for Ḡ,
i.e., if Ḡ doesn’t induce a graph automorphism on G, then the same statements hold for Ḡ, and so any
subgroup of Ḡ containing an element of the appropriate order is a blueprint for Vmin . If Ḡ does induce
τ
a graph automorphism on Vmin then we need to consider either L(G) or Vmin ⊕ Vmin
, where τ is a graph
automorphism, and whether x is a blueprint on one of these modules. In [10] it is shown that for G = E6 and
τ
x real (so every semisimple element in SL2 (pa )), x is a blueprint for Vmin ⊕ Vmin
whenever it is a blueprint
for Vmin . For F4 in characteristic 2, [10] proves that there is no almost simple maximal subgroup with socle
SL2 (2a ) for a > 5 in Ḡ even if Ḡ induces a graph automorphism on G, so this case has also been dealt with.
We can push being a blueprint for the minimal module of F4 up into E6 and E7 .
Lemma 4.9 Let G be E6 or E7 , and let x be a semisimple element of G that lies in F4 . If x is a blueprint
for the minimal module for F4 , then x is a blueprint for the minimal modules of E6 and E7 , and also the
∗
module Vmin ⊕ Vmin
for E6 .
Proof: The restrictions of the modules in question to F4 are a sum of minimal modules and trivial modules.
Since 1 is always an eigenvalue of any semisimple element of F4 on its minimal module, any element of F4
with the same eigenspaces on the minimal module for F4 as x also has the same eigenvalues on the three
modules mentioned for E6 and E7 . This completes the proof.
If G = E7 then v(G) = 75 is fairly large, and in certain circumstances we can bring this down. Here is one
such circumstance.
Proposition 4.10 Let G be the simply connected form of E7 , and let x be a semisimple element of G. If
the 1-eigenspace of x on Vmin has dimension at least 6 then x lies inside a conjugate of either an F4 or A4
subgroup. If in addition o(x) > 30 then x is a blueprint for Vmin .
Proof: Since the 1-eigenspace is positive dimensional, x fixes a line on Vmin and so lies inside E6 or B5 . If
x lies inside E6 then it must fix a line on its minimal module again, and so lies inside F4 , as claimed, or D5 .
However, D5 centralizes only a 4-space on Vmin , and so x fixes a line on either the natural or 16-dimensional
spin module. If x fixes a line on the 10-dimensional module then x lies inside B4 6 F4 , and if x fixes a line
on the 16-dimensional module then x lies inside A4 , as needed.
Thus we may assume that x lies inside B5 , whence again it fixes a line on either the 11-dimensional
natural module or the 32-dimensional spin module: the first puts x inside D5 and we are done, and the
second puts x inside A4 again. Thus x lies inside either F4 or A4 .
If x ∈ F4 and o(x) > 30 then x is a blueprint for Vmin since x is a blueprint for the minimal module for
F4 by Lemma 4.9. On the other hand, if x ∈ A4 then one uses the proof of the results from [10], noting that
the composition factors of A4 on Vmin are, up to multiplicity, 0000, 1000, 0100, 0010 and 0001. A computer
program running the algorithm in [10] yields the answer 30.
Suppose we want to find the eigenvalues on Vmin of semisimple elements of order 63 inside E7 , which
we will need to do when considering SL2 (64). There are too many to construct them all and store them
all effectively, but we can take an element x of order 21 and consider all 37 = 2187 preimages x̂ of x in a
17
torus. Since we have the eigenvalues of all elements of order 21, given a potential multiset of eigenvalues
for an element x of order 63 in E7 , we take the eigenvalues of x3 , find all semisimple classes of elements of
order 21 with those eigenvalues, then consider all preimages of representatives of each of those classes. The
eigenvalues of x are valid for coming from E7 if and only if one of those elements of order 21 has a preimage
with those values.
This idea to get the eigenvalues of elements of large composite order will be called the preimage trick in
the rest of this paper.
Blueprints inside A1 s
4.3
We now prove that certain semisimple elements, and subgroups of the form SL2 (pa ) and PSL2 (pa ) of exceptional groups, are blueprints for a given module by examining the constituents of the restriction of the
module to an A1 subgroup containing the element or subgroup.
The first lemma deals with modules for the algebraic group SL2 , and when the eigenspaces of semisimple
elements match the weight spaces.
Lemma 4.11 Let M be a module for SL2 with composition factors highest weight modules L(λ1 ), . . . , L(λr ),
arranged so that λi 6 λi+1 . Let T be a maximal torus of SL2 , and let x ∈ T be a semisimple element of
order n. If λi < n/2 then the eigenvalues of x on M are the same as the weight spaces of T . In particular,
x is a blueprint for M .
Proof: Since all maximal tori are conjugate, we may assume that x is the matrix
!
ζ
0
,
0 ζ −1
where ζ is a primitive nth root of 1. The eigenvalues of x on L(1) are ζ, ζ −1 , and the eigenvalues of x on
L(λi ) are roots of unity ζ ±j for 0 6 j 6 λi . If λi < n/2 for all i then the eigenspaces of x are simply the
weight spaces of the L(λi ), and so x and T stabilize the same subspaces of M , thus x is a blueprint for M .
We will apply this lemma to A1 subgroups of algebraic groups. We often will end up with composition
factors that do not precisely satisfy the hypotheses of this lemma though: if one composition factor has
slightly larger highest weight, then although the eigenspaces do not correspond to weight spaces, with some
weight spaces being merged, these all take place within one composition factor of the module, and so the
finite subgroup A1 (q) of the A1 is still a blueprint for the module in question, even if the element of order n
is not.
Lemma 4.12 Let G be the simply connected form of an exceptional algebraic group of Lie type, and let X
be a positive-dimensional subgroup of G of type A1 . Let x be a semisimple element of X of order n. Let V
be a module for G.
(i) If the composition factors of X on V are (n/2 − 1)-restricted then x and a maximal torus T containing
x stabilize the same subspaces of V , so that x is a blueprint for V .
(ii) Suppose that the highest weights of X on V are λ1 , . . . , λr , with λi 6 λi+1 , and let H = A1 (q) be a
finite subgroup of X containing x. If λr−1 + λr < n then H and X stabilize the same subspaces of V ,
so that H is a blueprint for V .
18
Proof: The first part follows immediately from Lemma 4.11, so we concentrate on the second statement.
Letting T be a maximal torus of X containing x, if λ and mu are two weights of T on V that are equal
when taken modulo n (i.e., yield the same eigenvalue for the action of x), then λ and µ differ by a multiple
of n. By assumption on the λi , since λ − µ > n, both λ and µ must be weights for the composition factor
L(λr ), since if λ is a weight for one of the other L(λi ) then it lies between −λi and +λi , and cannot differ
by n from any other weight for any other λj .
Let W be any H-submodule of V . If W does not contain the factor L(λr ) then the eigenvectors of x on
W all come from weight vectors for T , by the previous paragraph, and so T stabilizes W . If W contains
L(λr ), then T also stabilizes W by taking duals, as T stabilizes the dual (V /W )∗ of V ∗ . Thus T stabilizes
every H-submodule of V , and so hT, Hi = X and H stabilize the same subspaces of V , as claimed.
4.4
Traces for modules of PGL2 (pa )
Here we produce a specialized result about extending simple modules for PSL2 (pa ) to PGL2 (pa ), and the
traces and eigenvalues of the elements on such an extension.
There are two extensions of any simple module for PSL2 (pa ) to PGL2 (pa ). We will give a way of telling
these apart if the dimension of the simple module is odd, which is all that we need in what follows.
If L(i) is a simple module for PSL2 (pa ) of odd dimension, then of the two extensions of L(i) to PGL2 (pa ),
for one all defining matrices have determinant 1 and for the other half have determinant −1: to see this,
note that all elements in PSL2 (pa ) act with determinant 1, and the non-trivial 1-dimensional representation
acts like −1 for elements outside of PSL2 (pa ), so a given extension and its product with this 1-dimensional
representation give us the two cases. Write L(i)+ for the module for PGL2 (pa ) for which all matrices have
determinant +1, and L(i)− for the other extension. This notation will be used in the proof of the next
lemma.
Lemma 4.13 Let p be an odd prime and a > 1 an integer. Let M be a simple module for H = PGL2 (pa )
with Brauer character φ, and let g be an element of order pa ± 1. Let t be an involution in PSL2 (pa ), and
let h be the involution in hgi.
(i) There are two conjugacy classes of involutions in G. If o(g) is twice an odd number then t and h are
representatives of these two classes, and otherwise t and h are conjugate.
(ii) If dim(M ) is even then φ(t) = φ(h) = 0.
(iii) If dim(M ) is odd, then the dimensions of the +1-eigenspace and (−1)-eigenspace differ by 1.
(iv) If dim(M ) is odd, then φ(t) = ±φ(h). If M has only one of ±1 as an eigenvalue, then φ(t) = φ(h) if
and only if either t and h are conjugate, or g has eigenvalue 1 on M .
Proof:
(i) That H has two classes of involutions is well known, and one is a class of complements, the
other is in PSL2 (pa ). Thus the second statement follows easily.
(ii) We use Steinberg’s tensor product theorem, lifting all modules to GL2 (pa ): M has even dimension if
and only if, as a tensor product, one of the factors has even dimension, and the Brauer character is 0
for a given element if and only if one of the factors has Brauer character 0 for the same element. Thus
we need to check this for the symmetric powers of the natural module S i (M ′ ) for 0 6 i 6 p − 1, where
it is trivial to see that the trace of an involution is 0 on even-dimensional modules and ±1 on modules
of odd dimension.
19
(iii) Since dim(M ) is odd and M is self-dual, of course one of ±1 is an eigenvalue for the action of g. It is
an easy exercise to compute the eigenvalues of g on the Steinberg module for PGL2 (pa ), and we see
that these are all distinct if g has order pa + 1, and if g has order pa − 1 then ±1 appears twice and
∓1 appears once.
For other modules, from the definition of the Steinberg module as a tensor product of twists of the
fundamental module L(p − 1), and the fact that the eigenvalues of g on L(i) all appear in L(p − 1),
we see that the eigenvalues of g on any simple module appear in the eigenvalues of g on the Steinberg.
Thus the result holds since the sum of the dimensions of the (+1)- and (−1)-eigenspaces must be odd.
(iv) Return to PGL2 (pa ), and suppose that t and h are not conjugate, so that g has twice odd order. As
we saw above, for a fundamental module L(i)± for 0 6 i 6 p − 1 an even integer, the Brauer character
values of L(i)+ on t and h have the same sign, and the Brauer character values of L(i)− on t and h
have opposite signs. Notice that +1 is an eigenvalue of M if and only if there are an even number
of minus-type modules in the tensor decomposition, and this happens if and only if φ(t) = φ(h), as
needed.
Using this, the following result is now clear.
Corollary 4.14 Let p be an odd prime and a > 1 an integer. Let G be the simply connected form of E7 ,
and let H be a copy of SL2 (pa ) in G with Z(G) = Z(H). Suppose that g is an element of G such that
o(g) = pa ± 1 is twice an odd number, and −1 is not an eigenvalue for the action of g on L(G). Then the
group H̄ = hH, gi does not satisfy H̄/Z(H) = PGL2 (pa ).
Proof: Suppose that H̄/Z(H) = PGL2 (pa ). Let t be an element of H that is an involution in H/Z(H),
so o(t) = 4. The trace of t on L(G) is −7 or 25, depending on the class of t in G. The involution h in hgi
has trace 5 on L(G), since it is an involution in G rather than G/Z(G). We now show that h and t must
in fact have the same trace, a contradiction. By Lemma 4.13, any even-dimensional composition factors
of L(G) ↓H yield trace 0 for both t and h, and they have the same trace on odd-dimensional composition
factors since −1 is not an eigenvalue of g on L(G). Thus the trace of t and h on L(G) is the same, but this
is a contradiction.
4.5
The graph automorphism of F4
In this short section we describe how semisimple elements of odd order in F4 react to the graph automorphism
in characteristic 2. Since the graph automorphism τ does not fix the minimal module Vmin , and L(G) has
τ
composition factors Vmin and Vmin
, we can see the effect of the graph automorphism on semisimple classes
by taking the eigenvalues of an element on x on L(G) and removing those from Vmin .
Since the graph automorphism squares to a field automorphism, however, it is slightly more complicated
to understand those classes that are left invariant under an outer graph automorphism, since we need to
check whether xτ = xi for some i, rather than whether the eigenvalues of x and xτ match. This is still not
difficult using a computer, however; we give two special cases, where a conjugacy class is stable under the
graph automorphism (up to powers) and where the classes have integral traces.
Lemma 4.15 Let k be a field of characteristic 2. Let x be a semisimple element in G = F4 (k) such that xτ
is conjugate to a power of x. If x has order at most 9, then a power of x has trace on Vmin given below.
20
o(x)
Possible traces on Vmin
3
−1
5
7
1
4(ζ7 +
ζ7−1 )
+
3(ζ72
+ ζ7−2 ) + 5, −(ζ7 + ζ7−1 )
2 − 3(ζ9 + ζ9−1 )
9
Lemma 4.16 Let k be a field of characteristic 2. Let x be a semisimple element in G = F4 (k) such that
the trace of both x and xτ is an integer, and x and xτ are not conjugate. If x has order at most 9, then the
traces of x and xτ are as below, where we give x up to graph automorphism.
o(x)
Trace of x on Vmin
Trace of xτ on Vmin
3
8
−1
5
None
None
7
9
4.6
−2
5
3
−1 (x has trace −1)
3
2 (x has trace −1)
sl2 -subalgebras of L(G)
In this section we consider subalgebras of L(G), specifically sl2 -subalgebras. The stabilizers of sl2 -subalgebras
will be shown to be positive dimensional, and so if a subgroup stabilizes such a subalgebra, it must be
contained inside an element of X σ .
To begin with, we give a proposition that gives us a criterion for a subgroup H to stabilize an sl2 subalgebra in the first place. This proposition is a restatement of results of Alexander Ryba from [25],
particularly Lemma 10 from that paper.
Proposition 4.17 Let V be a 3-dimensional subspace of L(G), and let H be a subgroup of G such that
HZ(G)/Z(G) = PSL2 (pa ) for some p > 5. If V is H-stable and a complement for V is also H-stable (i.e.,
V is a summand of L(G) ↓H ) and HomH (V, L(G)) is 1-dimensional (i.e., there are no other submodules of
L(G) ↓H isomorphic to V ) then V is a subalgebra of L(G) isomorphic to sl2 .
Proof: Suppose that L(G) ↓H has a unique submodule isomorphic to V , and that this is a summand, so
that the quotient L(G) ↓H /V has no quotient isomorphic to V ∗ ∼
= V . By [25, Lemma 6], we have that V
possesses a non-singular trace form, and then we apply Block’s theorem [6] to see that V is a simple Lie
algebra of type sl2 .
In order to use this proposition, we need to know something about sl2 -subalgebras of the Lie algebras of
exceptional groups. The following is a theorem of David Stewart and Adam Thomas [28], specialized to the
case of G = E6 , E7 , for which we need it.
Theorem 4.18 Let g = E6 and p > 7, or G = E7 , E8 and p > 11. The sl2 -subalgebras of L(G) are in
one-to-one correspondence with the nilpotent orbits of L(G), with a bijection being realized by sending an
sl2 -subalgebra to the nilpotent orbit of largest dimension intersecting it non-trivially.
Along with the proof of this theorem, Stewart has representatives for the nilpotent orbits intersecting
each of these sl2 s, in a GAP file. When the sl2 is in bijection with a nilpotent class not of order p, there are
two nilpotent classes that intersect the sl2 , and for p > 5 and the sl2 not restricted we give the other class
that intersects it for E6 , E7 and E8 in Tables 4.3, 4.4 and 4.5 respectively. All of these classes have order
21
Class
p=5
p=7
D4
3A1
A5
A3
D5 (a1 )
A2 + A1
E6 (a3 )
A3 + A1
D5
A3
A3 + A1
E6 (a1 )
2A2 + A1
A5
E6
A2 + 2A1
2A2 + A1
p = 11
A5
Table 4.3: Second nilpotent class intersecting an sl2 -subalgebra of L(E6 ) for p > 5 and not restricted
p, and hence e and f lie in different nilpotent orbits of L(G) when the sl2 is not restricted. This yields the
following corollary.
Corollary 4.19 Let g = E6 and p > 7, or G = E7 , E8 and p > 11. Let H be a copy of PSL2 (p) in G/Z(G).
If NḠ (H) stabilizes an sl2 -subalgebra h of L(G), then h is restricted and NḠ (H) is contained inside an
element of X σ .
Proof: If h is restricted then it is stabilized by a good A1 in the algebraic group (see [26]), and hH, A1 i is
positive dimensional and stabilizes h, so we are done. Thus h is not restricted, and therefore e and f lie in
different nilpotent classes of L(G).
Since there is a unique conjugacy class of subgroups PSL2 (p) inside PSL2 , we see that because the
standard PSL2 (p) inside PSL2 swaps e and f , H must swap the two nilpotent orbits of h, clearly contradicting
the fact that they lie inside different orbits of L(G). Hence h is restricted, as needed.
22
Class
p=5
D4
(3A1 )′
′′
(A5 )
A3
D4 + A1
4A1
D5 (a1 )
A2 + A1
′
p=7
(A5 )
A3
A5 + A1
(A3 + A1 )′
D5 (a1 ) + A1
A2 + 2A1
D6 (a2 )
D4 (a1 ) + A1
E6 (a3 )
(A3 + A1 )′
D5
A3
E7 (a5 )
A3 + A2
A6
A2 + 2A1
D5 + A1
(A3 + A1 )′
A3 + 2A1
D6 (a1 )
A3 + 2A1
D4 (a1 ) + A1
E7 (a4 )
D4 (a1 ) + A1
A3 + A2
D6
A3
A3 + 2A1
E6 (a1 )
2A2 + A1
(A5 )′
E6
A2 + 2A1
p = 11
p = 13
p = 17
(A3 + A1 )′
2A2 + A1
′
E7 (a3 )
(A3 + A1 )
D4 (a1 ) + A1
E7 (a2 )
2A2 + A1
A3 + 2A1
′
(A5 )′
D6 (a2 )
E7 (a1 )
A2 + 2A1
(A5 )
D6
D6 (a1 )
E7
A4 + A2
A6
A5 + A1
D5 + A1
D6
Table 4.4: Second nilpotent class intersecting an sl2 -subalgebra of L(E7 ) for p > 5 and not restricted
23
Class
p=5
D4
3A1
A5
A3
D4 + A2
A2 + 3A1
E6 (a3 )
A3 + A1
D5
A3
A5 + A1
A3 + A1
p=7
p = 11
p = 13
p = 17
p = 19
p = 23
p = 29
A3 + A1
D5 (a1 ) + A2
2A2 + A1
E6 (a3 ) + A1
A3 + 2A1
D5 + A1
A3 + A1
E8 (a7 )
A4 + A3
A6 + A1
A2 + 3A1
E6 (a1 )
2A2 + A1
A5
D5 + A2
A3 + A2
A3 + A2 + A1
E6
A2 + 2A1
2A2 + A1
D7 (a2 )
2A3
A4 + A1
A3 + 2A1
A5
A7
2A2 + A1
A5
E6 (a1 ) + A1
2A2 + 2A1
A5 + A1
E7 (a3 )
A3 + A1
D4 (a1 ) + A1
E8 (b6 )
D4 (a1 ) + A1
E7 (a5 )
D7 (a1 )
A3 + 2A1
A3 + A2
E6 + A1
A2 + 3A1
2A2 + 2A1
E8 (a6 )
2A3
A4 + 2A1
D7
A2 + 3A1
A5
D5 + A1
E8 (b5 )
2A2 + 2A1
D4 (a1 ) + A1
E7 (a5 )
A5 + A1
E7 (a1 )
A2 + 2A1
A5
D6
E8 (a5 )
2A2 + A1
E6 (a3 ) + A1
E7 (a4 )
D6 (a1 )
E8 (b4 )
A2 + 3A1
A5 + A1
E7 (a3 )
E7 (a4 )
E8 (a4 )
A4 + A3
A5
D6 (a2 )
E7 (a3 )
E8 (a3 )
A4 + A2 + A1
A6 + A1
E6 (a3 ) + A1
D6 (a1 )
E7 (a3 )
E8 (a2 )
2A3
A4 + A3
A5 + A1
D7
E7 (a1 )
E7 (a2 )
E8 (a1 )
2A2 + 2A1
A4 + A2 + A1
D5 (a1 ) + A2
A7
D7
E7
E7 (a1 )
E8
A3 + A2 + A1
A3 + A2 + A1
A4 + A3
A6 + A1
A7
E6 + A1
D7
Table 4.5: Second nilpotent class intersecting an sl2 -subalgebra of L(E8 ) for p > 5 and not restricted.
(Missing classes, D4 + A1 , D5 (a1 ), D5 (a1 ) + A1 , D6 (a2 ), E7 (a5 ), A6 , D6 (a1 ), E7 (a4 ), D6 , E7 (a2 ) and E7 ,
are exactly as in Table 4.4)
24
E7
Modules for SL2(pa )
5
The purpose of this section is to describe everything we need to know about the simple modules and extensions
between them for the groups SL2 (pa ) for p a prime and a > 1.
5.1
Modules for SL2 (2a )
We construct certain modules for H = SL2 (2a ) for some a 6 10, and prove that various configurations of
module do not exist. (The reason we choose a 6 10 is so that these results may be used in work for E8 ,
for which v(E8 ) = t(E8 ) = 1312.) The main motivation for this is to get better bounds on the number of
certain composition factors that are needed to prevent a particular simple module appearing in the socle of
a given module M . The pressure, and more general M-pressure of Section 2 proves that if M does not fix a
line, yet has a trivial composition factor, then it needs at least one more 2-dimensional composition factor
than trivial factor. We can do better than this in some circumstances.
We begin with some notation. Let u be an element of order 2 in H. By 21 we denote the natural module
for H, and define 2i by the equation
⊗2
2i−1
= 1/2i /1,
i.e., 2i is the twist under the field automorphism of 2i−1 . Given this, if I is a subset of {1, . . . , a}, of
cardinality b, we define,
2bI =
O
2i ,
i∈I
for example, 41,2 = 21 ⊗ 22 ; the modules 2bI for all I ⊆ {1, . . . , a} furnish us with a complete set of irreducible
modules for H, by Steinberg’s tensor product theorem.
We first recall a result of Alperin [1], that determines Ext1 (A, B) for A, B simple modules for H.
Lemma 5.1 Let A and B be simple H-modules, corresponding to the subsets I and J of {1, . . . , a}. The
dimension of Ext1 (A, B) is always 0, unless
(i) |I ∩ J| + 1 = |I ∪ J| < a, and
(ii) if i ∈ I ∪ J and i − 1 ∈
/ I ∩ J, then i − 1 ∈
/ I ∪ J,
and in this case the dimension is 1.
In particular, if Ext1 (A, B) 6= 0 then the dimension of A is either half or double that of B.
Given this, we know that if a module M has a trivial composition factor but does not fix a line, then it
has at least two 2-dimensional composition factors, i.e., has positive pressure.
If it has pressure 1, then we can say something about the module still. This is important for F4 and E6
because there are no involutions acting projectively there. We will generalize this result in the next lemma,
but provide a full proof in this simple case for the benefit of the reader.
Lemma 5.2 Let V be an H-module that has at least one trivial composition factor but no trivial submodules
or quotients. If V has pressure 1 then an involution in H acts projectively on V if dim(V ) is even and with
a single Jordan block of size 1 if dim(V ) is odd.
Proof: Note that, since V has pressure 1, it cannot have 2i ⊕ 2j or 1⊕2 as a subquotient without fixing a
line or hyperplane. We proceed by induction on dim(V ), starting with the even-dimensional case. We may
25
assume that soc(V ) = 2i for some i: firstly there are no composition factors of soc(V ) of dimension greater
than 2 because the quotient by one would still satisfy the hypotheses of the lemma, and 2i ⊕ 2j cannot be
in the socle by the note above.
The module V / soc(V ) has pressure 0, so H must fix a line or hyperplane, but cannot fix a hyperplane
by assumption, so V / soc(V ) has a trivial submodule, and it must be unique by the note at the start of the
lemma. Quotient out by any possible factors of dimension at least 4 in the socle of V / soc(V ) to get a module
W of pressure 0 and with soc(W ) = 1. (If there is a 2 in soc(W ) then we find a submodule of pressure 2,
not allowed.)
The socle of W/ soc(W ) must be 2j for some j, since 2j ⊕ 2l cannot be a subquotient and 1 only has
extensions with simple modules of dimension 2. Now W/ soc2 (W ) is again pressure 0, so has a trivial
submodule as it cannot have a trivial quotient, and we have constructed a submodule 1/2j /1 inside W .
Letting U be the quotient of W by this submodule, we have removed 2i , 2j , 12 from V , and possibly some
other modules, and so an involution acts projectively on U , but it also acts projectively on the kernel of the
map W → U , namely 1/2j /1, and on the kernel of the map V → W since that has no trivial factors at all,
so an involution acts projectively on all of V , as needed.
For odd-dimensional modules, we now simply find any submodule W with a single trivial composition
factor and such that it is a quotient of W . The module V /W must have even dimension and has no trivial
submodule as otherwise V would have 1 ⊕ 1 as a subquotient. Also, W has pressure 0 since otherwise W
with the 1 removed from the top has pressure 2, contradicting Lemma 2.2, and so V /W has pressure 1, thus
an involution acts projectively on V /W and with a single 1 on W , as needed.
We can generalize this result to modules of larger pressure.
Lemma 5.3 Let V be an H-module that has at least one trivial composition factor but no trivial submodule
or quotient. If V has pressure n then an involution in H acts on V with at most n Jordan blocks of size 1.
Proof: As with the previous lemma, choose V to be a minimal counterexample to the lemma, so that the
socle and top of V consist solely of 2-dimensional modules. Notice that, by choice of minimal counterexample,
there cannot exist a submodule W such that W has no trivial quotients and V /W has no trivial submodules,
since otherwise one of W and V /W would also be a counterexample.
Let W be a minimal submodule with a trivial composition factor but no trivial quotient, and note
that, since all simple modules with non-trivial 1-cohomology have dimension 1, W contains a single trivial
composition factor. If W = V then V itself must have a single trivial composition factor, and so V has a
single block of size 1, and the result holds since V must have pressure at least 1.
Thus W < V . If V /W has no trivial submodule then we have a contradiction by the statements above, so
soc(V /W ) has a trivial composition factor: let W2 denote the preimage of this trivial submodule of soc(V /W )
in V . We claim that W2 has a quotient 1/2/1. If this true then, since W2 has pressure at least 0 and u acts
projectively, and V /W2 has no trivial submodules or quotients, there must be at most n Jordan blocks of
size 1 in V , a contradiction.
Lemma 5.4 Let a = 3. If M is an even-dimensional module with 2n > 0 trivial composition factors and no
trivial submodule or quotient, then it has at least 3n composition factors of dimension 2.
Proof: Note that if M = M1 ⊕ M2 with the Mi even-dimensional, then by induction M satisfies the
conclusion of the lemma: thus M is either indecomposable or the sum of two odd-dimensional indecomposable
modules.
26
The projective cover of 21 is
21 /1, 41,3/21 , 22 , 23 /1, 1, 42,3/21 , 22 , 23 /1, 41,3 /21 .
Remove any 4-dimensional factors from the top and socle of M , so that M is a submodule of a sum of copies
of projectives P (2i ). If M has seven socle layers then P (2i ) is a summand of M , so M = P (2i ) and we are
done. Thus M has at most five socle layers. The number of 2s in the first and third socle layers must be at
least as many as the number of 1s in the second layer, and there are at least as many 2s in the third and
fifth socle layers as 1s in the fourth layer. We therefore must have that there are at least 3n 2-dimensional
factors in M , as claimed.
Our next result is the best result possible in this direction.
Lemma 5.5 Let H = SL2 (2a ) for 3 6 a 6 10. The largest submodule of P (21 ) whose composition factors
have dimension 1 or 2 has dimension 10, and structure
21 /1/22 , 2a /1/21,
where 2a is a quotient of this module.
Let M be a module for SL2 (2a ) for some a 6 10. Suppose that u acts on M with b Jordan blocks of size
1. If there are c > 0 trivial composition factors in M , but no trivial submodules or quotients, then M has
at least 2b + 3(c − b)/2 composition factors of dimension 2.
Proof: The first statement, of the structure of the largest submodule of P (21 ) with factors of dimension 1
and 2, is easily verified by computer in the range specified. Thus we concentrate on the second statement,
which we prove by induction on dim(M ). By removing all submodules and quotients of simple modules of
dimension at least 4, we may assume that M is a submodule of a sum of P (2i )s for various i.
The largest submodule N of M with factors of dimension 1 and 2, firstly has no trivial submodules or
quotients, and secondly is a submodule of sums of modules of the form in the first part of this lemma. Since u
acts on the subquotient 1/2/1 with two blocks of size 2, if u has b′ Jordan blocks of size 1 on N and c′ trivial
composition factors in total, then we see that modulo the third socle soc3 (N ) of N we must have exactly
(c′ − b′ )/2 trivial composition factors; hence we have (c′ − b′ )/2 2-dimensionals in the fifth socle layer, and
the socle and third socle layers have at least (c′ + b′ )/2 2-dimensional modules each, yielding 2b′ + 3(c′ − b′ )/2
2-dimensionals in total. The quotient M/N also has no trivial submodules or quotients, so we are done if
the number of Jordan blocks of size 1 in the action of u on M/N is at least b − b′ , but this is clear. This
completes the proof.
Obviously this improves the result that the pressure must be positive, i.e., if there are c trivials then
there must be at least (c + 1) 2-dimensionals; we easily see from the modules in the lemma above that this
bound of (3c + b)/2 is best possible.
Lemma 2.3 shows that, not only can we not fix a 1-space on either Vmin or L(G), but we cannot fix a
2-space on Vmin either for F4 , E6 and E7 . By Lemma 5.1 we see that 2s only have non-split extensions with
1 and 4s, so we would like a similar result to the previous one, counting the number of 4-dimensionals in a
module M that has 2-dimensional composition factors but no 2-dimensional submodules or quotients. We
start with the easier case, where there are no trivial composition factors in M at all. Notice that we can use
M-pressure here as well, but we can do a bit better using the structure of modules for SL2 (2a ).
(We do not need to consider a > 6 here as these lemmas are not of use for E8 .)
27
Lemma 5.6 Let H = SL2 (2a ) for 4 6 a 6 6. The largest submodule of P (4i,j ) whose composition factors
have dimension 2 and 4 is as follows: for j = i ± 1, we have a 10-dimensional module
4i−1,i+1 /2i+1 /4i,i+1 ;
for a = 4 we have a 28-dimensional module
41,3 , 41,3 /21 , 23 /41,4 , 42,3 /21 , 23 /41,3 ;
for j = i ± 2 and a > 4 we have a 32-dimensional module
4i,i+2 , 4i,i+2 /2i , 2i+2 /4i+1,i+2 , 4i,i+3 , 4i−1,i+2 /2i , 2i+2 /4i,i+2 ,
with 4i−1,i+2 as a quotient. In all other cases, we have the module
4i,j , 4i,j /2i , 2j /4i,j−1 , 4i,j+1 , 4i+1,j , 4i−1,j /2i , 2j /4i,j ,
with 4i,j−1 and 4i−1,j as quotients.
Consequently, if M is a module with no trivial composition factors, with c > 0 composition factors of
dimension 2, and no 2-dimensional submodule or quotient, then M has at least c composition factors of
dimension 4.
Proof: The proof follows that of Lemma 5.5: we cannot produce a module 4/2, 2/4/2, 2/4 since the 4s in
the middle of the modules above do not have extensions with both 2s by Lemma 5.1. Thus we have at least
4/2, 2/4, 4/2, 2/4, and so we need as many 4s as 2s.
Of course, unlike in the case of Lemma 5.5, the 4i,j s are not all the same up to field automorphism, and
so the largest constructible modules depend on which 4s and 2s we have.
The next lemma brings together the previous two results, in the sense that we want to know how many
1s and 2s we can stack on top of a given simple module of dimension 4. This lemma gives that answer, and
hence how many 4s one needs to hide all 1 and 2s inside the middle of the module.
Lemma 5.7 Let H = SL2 (2a ) for some 2 6 a 6 6. The largest submodule of P (41,2 ) whose composition
factors modulo the socle have dimensions 1 or 2 is
22 /1/23/1/22 /41,2 ,
and an involution acts projectively on this module.
For a = 4 and a > 5 we have
22 , 24 /1/21 , 23 /41,3 ,
and 22 /1/21, 23 /41,3
respectively. For a > 6 and i = 4, 5 we have 1/21 , 2i /41,i .
In particular, if M is a module for H with no trivial or 2-dimensional submodules or quotients, and it
has 2n trivial composition factors for some n > 0, then it has n′ > n + 1 4-dimensional factors, and between
2n + 1 and 4n′ 2-dimensional composition factors.
Proof: The facts about the largest submodule of P (4i,j ) can easily be checked with a computer. For the
conclusion, we proceed by induction. By removing submodules and quotients of dimension 8 and above, we
may assume that the socle and top of M consists entirely of 4-dimensional modules.
28
Let M1 = soc(M ) and M2 /M1 be the 4′ -radical of M/M1 , noting that M2 /M1 is the direct sum of its
{1, 2}-radical and {1, 2, 4}′-radical. By the first part, if M2 has 2m trivial modules then M1 has at least
m copies of 4-dimensional modules to suppose the 2m trivials, and from the structure above the number of
2-dimensionals is at most 4m. Applying induction to M/M2 , this proves there are at least n + 1 different
4-dimensional factors and at most 4n 2-dimensional factors; there are at least 2n + 1 2-dimensional factors
since M must have positive pressure.
5.2
Modules for SL2 (3a )
In this section we describe the simple modules for H = SL2 (3a ) for a 6 7, describe various extensions
between some of the simple modules, and prove the existence or non-existence of various indecomposable
modules.
Let L = SL2 (3) 6 H. The simple modules for L have dimension 1, 2 and 3, with only the 2 being
faithful. Therefore, the modules for H are tensor products of modules of dimension 2 and 3, with a module
of dimension 2m 3n being faithful if and only if m is odd.
Writing 2i for the image of 2 under i iterations of the Frobenius map, and similarly for 3i , the simple
modules for H can be labelled by 2m 3nr1 ,...,rm+n , where m, n > 0 are integers, {r1 , . . . , rm+n } ⊂ {1, . . . , a}
with the ri distinct, with
2m 3nr1 ,...,rm+n
m
m+n
O
O
=
2 ri ⊗
3 ri .
j=1
j=m+1
Hence for example 122,3,1 = 22 ⊗ 23 ⊗ 31 is a module for PSL2 (3a ) for any a > 3.
We need to understand the restrictions of these simple modules to L, in order to understand which ones
we can have in the restrictions of minimal modules for G = F4 , E6 , E7 .
Lemma 5.8 Let H = PSL2 (3a ), a > 1, and let M be a simple module of dimension at most 56. The
restriction of M to PSL2 (3) is as below.
Module
Restriction
Composition factors of restriction
1
1
1
3
3
3
3⊕1
3, 1
4=2⊗2
⊕ P (1)
32 , 13
12 = 2 ⊗ 2 ⊗ 3
3⊕3 ⊕ P (1)
33 , 13
16 = 2 ⊗ 2 ⊗ 2 ⊗ 2
3⊕4 ⊕ P (1) ⊕ 1
34 , 14
27 = 3 ⊗ 3 ⊗ 3
3⊕7 ⊕ P (1)2
37 , 16
48 = 2 ⊗ 2 ⊗ 2 ⊗ 2 ⊗ 3
3⊕12 ⊕ P (1)⊕4
312 , 112
9=3⊗3
3
⊕2
We now move on to extensions. With the labelling above, we have the following easy lemma, which can
be found for example in [3].
Lemma 5.9 For any a > 1, a simple module M has non-trivial 1-cohomology if and only if M = 4i,i+1 for
some 1 6 i 6 a, and
1 a > 3,
dim(Ext1 (1, 4i,i+1 )) =
2 a = 2.
29
We will need more detailed information about extensions between low-dimensional modules for H, and
we summarize that which we need now. We restrict to the case when a 6= 2, because in this case things are
slightly different, with that pesky 2-dimensional 1-cohomology group, and secondly because we describe the
full projectives for this group after the lemma anyway.
Lemma 5.10 Let H = PSL2 (3a ) for 3 6 a 6 6. The following extension groups have dimension 1, for all
1 6 i, j 6 a:
(4i,i+1 , 1),
(1, 4i,i+1 ),
(3i , 4i−1,i ),
(4i−1,i , 3i ),
(4i,j , 4i±1,j ),
(4i,j , 4i,j±1 ),
If A and B are simple modules for H of dimension at most 9 then Ext1 (A, B) = 0 unless (A, B) is on the
list above.
We now consider certain modules. For a = 2, the structures of the projective indecomposable modules
are as follows:
1
3i
4
4 4
4
1 1 31 32
1 1 1 31 32
1 33−i
4 4 4
4 4
4
1 1 31 32
1
3i
4
We see that if a module M has five socle layers then it has a projective summand. More generally, if M
has trivial composition factors, then we can use these to prove that M must have more 4s than pressure
arguments suggest.
Lemma 5.11 Let H = PSL2 (9) and let M be a module for H. If M has no trivial submodules or quotients,
and there are 2n − 1 or 2n trivial composition factors, then the number of 4-dimensional factors is at least
2n.
Furthermore, the only submodules of P (4) consisting of 4s and 1s are submodules of a self-dual module
4/1, 1/4. In particular, there is no uniserial module of the form 4/1/4.
Proof: Let M be a module for H. We may assume that M is indecomposable. If M is the 9-dimensional
projective simple then the claim is true. If M has any 3-dimensional submodules or quotients then we may
remove them without affecting the claim, and so we may assume that M is a submodule of copies of P (4).
If M is projective then the result holds, so M is not projective, in which case it has at most four socle
layers. Since the fourth socle layer consists solely of 1s and 3i s, M must actually have three socle layers. In
particular, the trivials are all in the second socle layer, so if there are 2n − 1 or 2n of them, there must be at
least n copies of the 4-dimensional module in the socle, and similarly in the top. This completes the proof
of the first claim.
The second is easy to see by a computer proof that 4/1, 1/4 is the largest such module. Since it is
self-dual, we cannot construct a 4/1/4 inside it, yielding the second statement.
Lemma 5.12 Let H = PSL2 (3a ) for some 2 6 a 6 7. There does not exist a uniserial module with socle
layers 4a , 1 and 4b , where 4a and 4b are any of the simple modules of dimension 4.
As a consequence, if a 6= 2 and M is a module with composition factors of dimension 42i−α , 1i for some
i > 0 and α > 0, then M has a trivial submodule or quotient.
30
Proof: For the first part, the only modules with 1-cohomology are 4i,i+1 = 2i ⊗ 2i+1 , so we may assume that
the socle of our uniserial module is 41,2 . To prove this we simply use a computer to compare Ext1 (4′ , 1/41,2 )
and Ext1 (4′ , 41,2 ), and note that they coincide for all 4-dimensional modules 4′ .
For the second statement, we firstly use induction: quotienting out by the socle, which consists of n
modules each of dimension 4, we get at most n trivial modules dropping into the socle, which we quotient
out to get another module of the same form as in the lemma but with fewer trivial composition factors.
Hence we may assume that we have removed all trivial modules in this process, i.e., that there are n modules
of dimension 4 in the socle, and at most n modules of dimension 1 in the second socle layer. Since we need a
4-dimensional to lie above each trivial in the third socle layer, the bounds on composition factors means that
our module has three socle layers, each with n composition factors, of dimensions 4, 1 and 4 respectively. We
now simply quotient out by all but one factor in the socle: this drops all but one trivial into the socle since
a 6= 2, which we also remove. The resulting module has a single trivial and so any more than two 4s must be
summands, so we remove them as well, constructing a uniserial module of the form 4/1/4, a contradiction
that completes the proof.
We end with a small lemma, needed at one point in the text.
Lemma 5.13 Let pa = 27. The projective cover of 41,2 is
41,2
1 32 42,3 41,3 122,3,1
1 31 33 42,3 41,3 41,2 41,2 41,2 9
1,3
121,3,2
1 1 32 32 42,3 42,3 41,3 41,3 41,2 121,2,3 122,3,1 122,3,1
1 31 33 42,3 41,3 41,2 41,2 41,2 91,3 121,3,2
1 32 42,3 41,3 122,3,1
41,2
Consequently, if V is a self-dual module of pressure 1 with at least five trivial composition factors then H
fixes a line or hyperplane of V .
Proof: The description of the projective is produced by a Magma calculation, and to see the consequence,
since V has pressure 1 is cannot be the whole projective, hence we can remove all quotients that are trivial,
41,3 or 42,3 , but clearly the top two 1s can be taken off.
5.3
Modules for SL2 (p)
Since H = SL2 (p) has a cyclic Sylow p-subgroup, there are only finitely many indecomposable modules for
it over a field of characteristic p. In this section we describe how to construct all indecomposable modules
for H in characteristic p, using the projective indecomposable modules as a starting point.
The Green correspondence [2] shows that the number of indecomposable modules of dimension congruent
to i modulo p for H is the same as that of the normalizer NH (P ) of a Sylow p-subgroup P of H, a soluble
group of order p(p − 1) with a centre of order 2. However, for this group, it is easy to construct the
indecomposable modules: the projective modules are of dimension p, and look like truncated polynomial
rings k[X]/(X p − 1), hence uniserial. Every indecomposable module is a quotient of such a module, and
as every simple module for NH (P ) is 1-dimensional, we see that there are exactly p − 1 indecomposable
modules of dimension i for each 1 6 i 6 p, with half of these faithful modules for SL2 (p) and half modules
for PSL2 (p).
31
In particular, we see that once we have constructed p(p − 1) non-projective, indecomposable modules for
H then we must have found them all. Thus we start with the simple and projective modules for H: letting
M = L(1) be the natural module for H, we construct all simple modules using symmetric powers
L(i) = S i (M )
0 6 i 6 p − 1,
with L(i) being of dimension i + 1. As with the case of SL2 (2a ), we will normally write the single number i to
refer to the simple module of dimension i, and so a module 3/5 for SL2 (7), for example, is an 8-dimensional
module with 5-dimensional socle L(4) and 3-dimensional top L(2). The odd-dimensional simple modules are
modules for PSL2 (p), and the even-dimensional ones are faithful modules for SL2 (p).
Having defined the simple modules, we consider the projectives: the Steinberg module L(p − 1) of
dimension p is already projective, and for each simple module i with 1 6 i 6 p − 1, the projective module
P (i) has structure
i/(p + 1 − i) , (p − 1 − i)/i,
except when i = 1, in which case p + 1 − i would have dimension p, and we have 1/(p − 1)/1, and when
i = p − 1, so p − 1 − i would have dimension 0, and we have (p − 1)/2/(p − 1).
We represent these in diagrams, with lines linking two composition factors A and B if there is a non-split
extension A/B as a subquotient of the module. For example, here are P (3) and P (5) for PSL2 (11).
3
5
7
9
5
7
3
5
Using these we construct indecomposable modules as follows: we have modules of the form i/(p + 1 − i)
and i/(p − 1 − i), and also two modules of the form i/(p − 1 − i), (p + 1 − i) and (i + 2)/(p − 1 − i). These
two indecomposables can be summed together, then quotiented by a diagonal submodule p − 1 − i to make
a new module with four composition factors.
It is easier to visualize using diagrams. In the example above, we can remove the socles of the two
projectives to get modules 3/7, 9 and 5/5, 7, take their direct product, and then quotient out by a diagonal
7.
5
3
5
5
7
3
→
⊕
7
9
5
7
9
This process certainly produces a module, with quotients both of our original summands, and so this module
must be indecomposable. Note that if one tries to do this with say two copies of 3/7, 9 then the fact that
Ext1 (3, 7) is 1-dimensional means that this module splits, so one needs the modules at the top (in this case
3 and 5) to be different.
One can continue this process until one constructs an indecomposable module M with all (non-projective)
simple modules appearing in the top and the socle of M exactly once. As an example, the diagrams of the
two such modules for p = 11 (one for PSL2 (11), one for faithful modules for SL2 (11)) are as follows:
9
1
7
3
5
5
3
7
32
1
9
2
4
10
8
6
6
8
4
10
2
We can take subquotients of these modules and construct new indecomposable modules, and we claim that
this constructs all non-projective, indecomposable modules for SL2 (p).
Firstly, the non-simple indecomposable subquotients of the module M are in one-to-one correspondence
with connected subdiagrams of the diagram with at least one edge, since one notes that no two distinct
subdiagrams of the diagram above have the same first and second rows. In the case of simple modules, of
course each appears twice.
The number of connected subdiagrams of each diagram with at least one edge is (p − 1)(p − 2)/2 (i.e.,
we choose the start and end points), and add in the p − 1 simple modules (other than the Steinberg), and
the p − 1 non-simple projective modules, to get
(p − 1)(p − 2) + 2(p − 1) = p(p − 1).
This is the number of indecomposable modules for the normalizer, and so we must have constructed all
indecomposable modules for SL2 (p).
It is clear from this ‘zigzag’ structure, that for any indecomposable module, the socle is a collection of
modules, and if A and B lie in the socle so does any module with dimension between dim(A) and dim(B).
We have proved the following proposition.
Proposition 5.14 Let H = SL2 (p), and let M be an indecomposable module for H.
(i) If M has one socle layer then M is simple, and there are p such modules, one of each dimension.
(ii) If M has three socle layers then M = P (i) for some 1 6 i 6 p − 1.
(iii) If M has two socle layers then the socle of M consists of modules of dimension i, i + 2, . . . , j (i 6 j),
and the top consists of modules p − j + ǫ, p − j + ǫ + 2, . . . , p − i + δ, where ǫ, δ = ±1. There are
(p − 1)(p − 2) such modules.
The indecomposable modules for PSL2 (7) other than the Steinberg are below, ordered so that the modules
in column i have dimension congruent to i modulo 7.
1
3, 5/3, 5
3
1, 3, 5/1, 3, 5
5
3/3
P (1)
3/5
1, 3/5
1, 3, 5/3, 5
3/3, 5
1, 3/3, 5
1/5
P (3)
5/3
5/1, 3
3, 5/1, 3, 5
3, 5/3
3, 5/1, 3
5/1
P (5)
As another example, the indecomposable modules for PSL2 (5) are
1, 3, 1/3, 3/1, P (1), P (3), 3, 1/3, 3/1, 3, 3/3, 1, 3/1, 3.
Since each of these modules M is in Green correspondence with an indecomposable module V of dimension
at most p − 1, and Green correspondence means that the restriction of M to NH (P ) is a sum of V and
projective modules, this means that a unipotent element u of order p in H acts on M with at most one
Jordan block of size not equal to p (if M is projective then all blocks have size p, of course). Notice that
the only indecomposable modules with no Jordan blocks of size p in the action of u are either simple or of
dimension p − 1.
The next lemma is an easy consequence of these facts.
33
Lemma 5.15 Let H = PSL2 (p), and let M be a module for H over a field of characteristic p. Let u ∈ H
have order p. If ni denotes the number of Jordan blocks of size i in the action of u on M , then
np >
X
n2i .
16i<(p−1)/2
Proof: The modules of dimension at most p are either simple, so u acts on then with a block of odd size, or
are i/(p − 1 − i), so u acts with a single block of size p − 1. Therefore if there is a block of even size i < p − 1
then it must come from an indecomposable module of dimension greater than p, and so we get at least one
block of size p, as needed.
For G = E7 we also must consider H = SL2 (p) with Z(H) = Z(G). In this case we want a similar result
to the above but for faithful modules.
Lemma 5.16 Let H = SL2 (p), and let M be a module for H over a field of characteristic p on which the
central involution z acts as the scalar −1. Let u ∈ H have order p. If ni denotes the number of Jordan
blocks of size i in the action of u on M , then
np >
X
n2i−1 .
16i<(p−1)/2
Proof: Similar to Lemma 5.15, and omitted.
We often want to understand self-dual modules for H, since the minimal modules Vmin are self-dual for
F4 and E7 , and the adjoint module L(G) is always self-dual. Using the statements above, if ni is odd,
where again ni is the number of blocks of size i in the action of u, there must be a self-dual indecomposable
summands of dimension congruent to i modulo p.
The next lemma follows from Proposition 5.14 and classifies self-dual indecomposable modules for SL2 (p).
From our zigzag diagrams above, it is clear which the self-dual modules are: choose the same simple module
as the start and end points of the subdiagram.
Lemma 5.17 Let H = SL2 (p), and let M be a self-dual indecomposable module for H. If M is not simple
or projective, then M has socle (and top) consisting of pairwise non-isomorphic modules N1 , N2 , . . . , Nr ,
where dim(Ni ) − dim(Ni−1 ) = 2 and dim(N1 ) + dim(Nr ) = p ± 1. In particular, there are exactly p − 1 nonprojective, indecomposable self-dual modules for PSL2 (p), and exactly p − 1 non-projective, indecomposable
and faithful self-dual modules for SL2 (p).
Therefore, if p ≡ 1 mod 4, all non-projective indecomposable self-dual modules for PSL2 (p) have dimension congruent to an odd number modulo p, and all non-projective, indecomposable, faithful self-dual
modules for SL2 (p) have dimension congruent to an even number modulo p.
We can use this to get a better handle on which possible Jordan block structures a given unipotent
element u can have, given it lies inside a copy of PSL2 (p) for p ≡ 1 mod 4. We split the result into two
corollaries depending on whether one has modules for PSL2 (p) or SL2 (p).
Corollary 5.18 Let H = PSL2 (p) with p ≡ 1 mod 4, and let M be a self-dual module for L. Let u be an
element of order p in H. The action of u has an even number of blocks of a given even size i, and there are
at least as many blocks of size p as there are blocks of size all even numbers less than p − 1.
34
Corollary 5.19 Let H = SL2 (p) with p ≡ 1 mod 4, and let M be a self-dual module for H on which the
central involution z acts as the scalar −1. Let u be an element of order p in H. The action of u has an even
number of blocks of a given odd size i, and there are at least as many blocks of size p as there are blocks of
size all odd numbers less than p.
We now turn to tensor products. By Steinberg’s tensor product theorem, simple modules for SL2 (pa ) are
tensor products of Frobenius twists of p-restricted modules, i.e., L(i) for i 6 p − 1. These restrict to SL2 (p)
as tensor products of simple modules, so it will come in handy to understand the tensor products of simple
modules for SL2 (p).
The next result gives the tensor product of any two simple modules for L, and will be of great use when
computing the restriction of simple SL2 (pa )-modules to SL2 (p).
Proposition 5.20 Let H = SL2 (p). If 0 6 µ 6 λ 6 p − 1 then L(λ) ⊗ L(µ) is given by one of the following:
(i) If λ + µ < p then
L(λ) ⊗ L(µ) = L(λ − µ) ⊕ L(λ − µ + 2) ⊕ · · · ⊕ L(λ + µ − 2) ⊕ L(λ + µ).
(ii) If λ + µ > p and λ < p − 1 then
L(λ) ⊗ L(µ) = L(λ − µ) ⊕ L(λ − µ + 2) ⊕ · · · ⊕ L(a)
P (λ + µ) ⊕ P (λ + µ − 2) ⊕ · · · ⊕ P (p + 1) ⊕ L(p − 1) µ even
⊕
P (λ + µ) ⊕ P (λ + µ − 2) ⊕ · · · ⊕ P (p)
µ odd
where a = 2p − (λ + µ + 4).
(iii) L(p − 1) ⊗ L(p − 1) = P (1) ⊕ P (3) ⊕ · · · ⊕ P (p − 1).
This result can be found, for example, in [12] and explicitly in [11, Lemma 3.1].
5.4
Modules for SL2 (pa ) for p > 5 and a > 1
As with modules for SL2 (3a ) we need a notation system for the simple modules, and as in that section, we
let 21 denote the natural module, i1 = S i−1 (21 ) for 2 6 i 6 p be the symmetric powers (the p-restricted
modules) and let ij+1 denote the application of the Frobenius morphism to ij . We then write, for module of
dimension n formed as the tensor product of m twisted fundamental modules, na1 ,...,am , in order of increasing
dimension of factor; for example, the module 21 ⊗ 32 will be denoted 61,2 , and 31 ⊗ 32 ⊗ 23 will be denoted
183,1,2 . If the integer n has a unique decomposition as a product of exactly m integers greater than 1 such
that the module would be for the correct group (i.e., PSL2 (pa ) or SL2 (pa )) then we simply write that, so
that 61 and 61,2 for PSL2 (49) are unambiguous. Sometimes there are modules that could be either for SL2
or PSL2 , such as 121,2 for p > 7, which is either 21 ⊗ 62 or 31 ⊗ 42 , but context will tell us which. When
there genuinely is ambiguity, for example, 181,2 when p > 11, as it could be 21 ⊗ 92 or 31 ⊗ 62 , we label them
(1)
(2)
with subscripts 181,2 and 181,2 according to the lexicographic ordering on the partitions of 18, but in these
rare cases we remind the reader which is which.
We start with some restrictions of simple PSL2 (q)-modules to PSL2 (p). This is needed because we often
understand the action of PSL2 (p) on the minimal or adjoint modules completely. We consider modules of
dimension at most 56 to include the minimal modules of F4 , E6 and E7 . We use Proposition 5.20 to compute
35
the restrictions of modules for SL2 (pa ) to SL2 (p). Since v(E7 ) = 75 we only list restrictions for pa 6 150
and dimension up to 56.
Lemma 5.21 Let H = PSL2 (5a ) for a = 2, 3, and let M be a simple module of dimension at most 56. The
restriction of M to L = PSL2 (5) is as below.
Module
Restriction
Composition factors of restriction
1
1
1
3
3
3
4=2⊗2
3⊕1
3, 1
5
5
5
8=2⊗4
5⊕3
5, 3
9=3⊗3
1⊕3⊕5
5, 3, 1
12 = 2 ⊗ 2 ⊗ 3
5 ⊕ 3⊕2 ⊕ 1
5, 32 , 1
15 = 3 ⊗ 5
5 ⊕ P (3)
5, 33 , 1
16 = 4 ⊗ 4
5 ⊕ P (3) ⊕ 1
5, 33 , 12
20 = 2 ⊗ 2 ⊗ 5
5⊕2 ⊕ P (3)
52 , 33 , 1
24 = 2 ⊗ 4 ⊗ 3
5⊕2 ⊕ P (3) ⊕ 3 ⊕ 1
52 , 34 , 12
25 = 5 ⊗ 5
5⊕2 ⊕ P (3) ⊕ P (1)
52 , 34 , 13
27 = 3 ⊗ 3 ⊗ 3
5⊕2 ⊕ P (3) ⊕ 3⊕2 ⊕ 1
52 , 35 , 12
40 = 2 ⊗ 4 ⊗ 5
5⊕3 ⊕ P (3)⊕2 ⊕ P (1)
53 , 37 , 14
5⊕4 ⊕ P (3)⊕2 ⊕ P (1)
54 , 37 , 14
45 = 3 ⊗ 3 ⊗ 5
48 = 4 ⊗ 4 ⊗ 3
5
⊕3
⊕2
⊕ P (3)
⊕ P (1) ⊕ 3
54 , 38 , 14
Consequently, if V is a module for H of dimension at most 56 such that V ↓L has more trivial than 3dimensional composition factors, then H stabilizes a line on V .
Proof: We prove the last statement: from the table above we see that only the trivial has more 1s than 3s
in its restriction to L. Suppose that the composition factors of V ↓L are 5i , 3j , 1k , with k > j. If there are α
trivial factors and β 8-dimensional factors in V , then α > k − (j − β) > β, so V has negative pressure and
so H fixes a line on V , as needed.
For PSL2 (25), we will need the eigenvalues of an element of order 12 on the simple modules, so we list
them here. These are of course easy to compute.
Lemma 5.22 Let H = PSL2 (25), and let x be a semisimple element of order 12 in H. Let ξ denote a
primitive 12th root of unity. Choosing ξ so that x acts on the symmetric square of the natural module for
SL2 (25) with eigenvalues ξ ±1 , the eigenvalues of x on the various simple modules for H are as follows.
36
Dimension
Eigenvalues
1
1
3
1, (ξ, ξ 11 )/(ξ 5 , ξ 7 )
4
(ξ 2 , ξ 10 ), (ξ 3 , ξ 9 )
5
1, (ξ 2 , ξ 10 ), (ξ, ξ 11 )/(ξ 5 , ξ 7 )
8
(ξ 2 , ξ 10 ), (ξ 3 , ξ 9 ), (ξ 4 , ξ 8 ), (ξ, ξ 11 )/(ξ 5 , ξ 7 )
9
1, (−1)2 , (ξ, ξ 11 ), (ξ 4 , ξ 8 ), (ξ 5 , ξ 7 )
15
1, (−1)2 , (ξ, ξ 11 ), (ξ 2 , ξ 10 ), (ξ 3 , ξ 9 ), (ξ 4 , ξ 8 ), (ξ 5 , ξ 7 ), (ξ, ξ 11 )/(ξ 5 , ξ 7 )
16
(−1)2 , (ξ, ξ 11 ), (ξ 2 , ξ 10 ), (ξ 3 , ξ 9 )2 , (ξ 4 , ξ 8 )2 , (ξ 5 , ξ 7 )
25
13 , (−1)2 , (ξ, ξ 11 )2 , (ξ 2 , ξ 10 )2 , (ξ 3 , ξ 9 )2 , (ξ 4 , ξ 8 )2 , (ξ 5 , ξ 7 )2
Here, (ξ, ξ 11 )/(ξ 5 , ξ 7 ) means either (ξ, ξ 11 ) or (ξ 5 , ξ 7 ), depending on the isomorphism type of the module.
We now give the analogue of Lemma 5.21 for p = 7. Again, we only need go up to pa = 150, so just 49
in this case.
Lemma 5.23 Let H = PSL2 (49), and let M be a simple module for H. The restriction of M to L = PSL2 (7)
is as below.
Module
Restriction
Composition factors of restriction
1
1
1
3
3
3
4=2⊗2
3⊕1
3, 1
5
5
5
7
7
7
8=2⊗4
5⊕3
5, 3
9=3⊗3
5⊕3⊕1
5, 3, 1
12 = 2 ⊗ 6
7⊕5
7, 5
15 = 3 ⊗ 5
7⊕5⊕3
7, 5, 3
16 = 4 ⊗ 4
7⊕5⊕3⊕1
7, 5, 3, 1
21 = 3 ⊗ 7
7 ⊕ P (5)
7, 52 , 3, 1
24 = 4 ⊗ 6
7 ⊕ P (5) ⊕ 3
7, 52 , 32 , 1
25 = 5 ⊗ 5
7⊕5
⊕3⊕1
7, 52 , 32 , 12
35 = 5 ⊗ 7
7 ⊕ P (5) ⊕ P (3)
7, 53 , 34 , 1
36 = 6 ⊗ 6
7 ⊕ P (5) ⊕ P (3) ⊕ 1
7, 53 , 34 , 12
49 = 7 ⊗ 7
7⊕2 ⊕ P (5) ⊕ P (3) ⊕ P (1)
72 , 54 , 34 , 13
⊕2
Having given restrictions of modules, we now need to understand Ext1 between simple modules. These
were completely determined in [3], but the information is not so easy to extract, and so we give a few special
cases that are necessary for us.
Of particular interest is which modules have non-trivial 1-cohomology, since we will often want to prove
that we stabilize a line. The next lemma gives this completely.
Lemma 5.24 Let p be a prime, a > 1 be an integer, and let M be a simple module for H = SL2 (pa ) with
non-trivial 1-cohomology. One of the following holds.
(i) pa = 2, M is the trivial module, with dim(H 1 (H, M )) = 1.
37
(ii) p is odd and a = 1, dim(M ) = p − 2, with dim(H 1 (H, M )) = 1.
(iii) pa = 9, dim(M ) = 4, with dim(H 1 (H, M )) = 2.
(iv) pa 6= 9 with a > 2, M is up to application of a Frobenius map L(p − 2) ⊗ L(1)σ , where σ is the
Frobenius twist (so that dim(M ) = 2(p − 1) and M = 2(p − 1)2,1 ), with dim(H 1 (H, M )) = 1.
Just knowing that modules have 1-cohomology is not going to be enough information. We need more
specific information about extensions between simple modules of low dimension, for p = 5, 7 and a > 1. The
final two lemmas of this section furnish us with this information.
Lemma 5.25 Let H = PSL2 (5a ) for a = 2, 3. The extensions between simple modules of dimension at most
8 are:
(i) 1 with 8i,i−1 ;
(ii) 3i with 4i,i+1 , 8i,i−1 ;
(iii) 4i,i+1 with 3i , 8i+1,i−1 (the latter only for a = 3);
(iv) 5i with nothing;
(v) for a = 2, 8i,i+1 with 3i ;
(vi) for a = 3, 8i,i+1 with 4i−1,i ;
(vii) for a = 3, 8i,i−1 with 1, 3i .
Lemma 5.26 Let H = PSL2 (49). The extensions between simple modules of dimension at most 9 are:
(i) 1 with nothing;
(ii) 3i with 8i+1,i ;
(iii) 41,2 with 51 , 52 ;
(iv) 5i with 41,2 ;
(v) 7i with nothing;
(vi) 8i,i+1 with 3i+1 , 91,2 ;
(vii) 91,2 with 81,2 , 82,1 .
38
Some PSL2s inside E6 in characteristic 3
6
In this short section we lay the groundwork for studying copies of H = PSL2 (3a ) (for a > 2) inside Ĝ = F4 (k)
by embedding Ĝ inside G = E6 (k), and attempting to construct many subgroups isomorphic to H inside
members of X σ of G other than Ĝ. Notice that field automorphisms of Ĝ lift to field automorphisms of G,
so we can embed an almost simple group with socle Ĝ into an almost simple group with socle G/Z(G). Let
Ḡ denote such an almost simple group, and note that Vmin is Ḡ-stable.
Suppose that NḠ (H) is contained inside both a σ-stable member X of X and inside Ĝ. The dimensions of
G and Ĝ are 78 and 52 respectively, so if X has dimension greater than 26, then X ∩ Ĝ is positive dimensional
(and of course still σ-stable), so that NḠ (H) lies inside a σ-stable, positive-dimensional subgroup of Ḡ, as
desired. Since the Borel subgroup of G has dimension 42, if NḠ (H) is contained in any parabolic subgroup
of G then we are done.
This also works if NḠ (H) is contained inside a conjugate of Ĝ, inside C4 – which has dimension 36 – and
A1 A5 – which has dimension 38. We are left with the irreducible G2 , the A2 G2 subgroup, and the A2 A2 A2
maximal-rank subgroup. In these cases we will show that H is a blueprint for Vmin .
If H is contained in X = G2 (k) then H acts on the natural module (up to Frobenius) as either 3⊕2
1 ⊕ 1,
or lies in a diagonal subgroup of A1 Ã1 acting as 41,i ⊕ 31 for any i > 1. In both cases H is contained in an
algebraic A1 subgroup Y .
The subgroup X acts irreducibly as L(20) on Vmin , and these two copies of H act on Vmin as
⊕3
3⊕3
,
1 ⊕ (1/41,2 /1)
91,i ⊕ 1/41,2 /1 ⊕ 41,i /(22 ⊗ 2i )/41,i ,
where of course 22 ⊗ 2i is 42,i if i > 2 and 1 ⊕ 32 if i = 2. The subgroups Y containing them act in the same
way, and stabilize the same subspaces as H, so that H is a blueprint for Vmin .
If X = A2 G2 , then X acts on Vmin as (10⊗10)⊕(02⊗00): if H lies inside the G2 factor then it centralizes
a 6-space on Vmin , so definitely lies inside a line stabilizer of F4 , and hence we may assume that H projects
along the A2 factor as 31 . Along the G2 it can act as either 3⊕2
i ⊕ 1 or 4i,j ⊕ 3i , for any i, j > 1 with i 6= j.
In the first case we get
⊕2
31 ⊕ 91,i
⊕ 1/41,2 /1
and
⊕3
3⊕3
,
1 ⊕ (1/41,2 /1)
for i > 1 and i = 1 respectively, and in the second case we get
31 ⊕1/41,2/1⊕41,i/(22 ⊗2i )/41,i ⊕1/41,2/1,
91,i ⊕41,j /(22 ⊗2j )/41,j ⊕1/41,2/1,
91,i ⊕12i,j,1 ⊕1/41,2/1,
for i = 1, j = 1 and i, j 6= 1 respectively. Again, in all cases, the algebraic subgroup containing H stabilizes
the same subspaces as H, so again H is a blueprint for Vmin .
Finally, we have X = A2 A2 A2 , which acts on Vmin as (up to duality) the sum of the three possible
configurations of natural times natural times trivial. If we act trivially on one or two or them we get, up to
twist,
⊕9
,
3⊕6
1 ⊕1
3⊕7
1 ⊕ 1/41,2 /1,
⊕3
91,i ⊕ 3⊕3
1 ⊕ 3i ,
according as one non-trivial, two non-trivial and the same, and two non-trivial and different. Thus we may
assume that H acts along the first factor as 31 , the second as 3i and the third as 3j , then we have, up to
twist, one of
⊕3
,
3⊕3
1 ⊕ (1/41,2 /1)
31 ⊕ 1/41,2 /1 ⊕ 9⊕2
1,j ,
271,i,j ,
according as whether i = j = 1, i = 1 6= j, and 1 6= i 6= j 6= 1. As with the other cases, each of these is
contained in an algebraic A1 stabilizing the same subspaces of Vmin , so again H is a blueprint for Vmin .
39
In particular, in all cases, either NḠ (H) is contained in a member of X σ or H is a blueprint for the
Ḡ-stable module Vmin , in which case NḠ (H) is contained in a member of X σ again.
We therefore have the following result.
Proposition 6.1 Let H = PSL2 (3a ), and let Ḡ be an almost simple group with socle G = F4 (k). If H
stabilizes a 3-space on the 25-dimensional minimal module for G, or the image of H in E6 centralizes a
2-space on the 27-dimensional minimal module for E6 , then NḠ (H) is contained in a member of X σ
40
7
Proof of the theorems: strategy
In this section we discuss the techniques that we will use in proving that a given copy of PSL2 (pa ) is contained
inside a member of X σ .
We have two cases to consider: when k, the field for the ambient group G, is not a splitting field for H,
and when k is. In the first case, we have to be more careful, as for example if 21 is a submodule of Vmin ↓H ,
H will not fix a 2-space over k, since 21 is not defined over k. Often we will deal with these small fields
separately, so we can use more uniform arguments in general, but H fixes a line over k if and only if it does
so over an extension field of k, so if that is what we prove we need no restriction on k.
The first step is usually to use the dimensions of modules and the traces of semisimple elements to produce
a list of potential sets of composition factors for the action of H on Vmin , which we call conspicuous sets of
composition factors. For many groups this list is small, but as the sizes of G and H grow the number grows
larger and we need more efficient methods that cut this number down, for example only considering possible
multisets of dimensions that have either no modules of dimension 1 or more modules of dimension 2(p − 1)
than modules of dimension 1, at least for pa odd and not equal to 9, i.e., modules of positive pressure.
Having done this, we can assume we know the composition factors of Vmin ↓H , and we have a few ways
to proceed.
• We can use the traces of semisimple elements to deduce a list, often a list with one element, of
potential sets of composition factors for L(G) ↓H . Sometimes this cannot exist, of course only if there
is no embedding of H with these composition factors. Other times L(G) ↓H has non-positive pressure,
so we are again done. Otherwise, we may analyse both Vmin ↓H and L(G) ↓H using the techniques
below. We will occasionally employ Lemma 2.1 in this regard.
• We can easily compute Ext1 between the composition factors of Vmin ↓H and determine if Vmin ↓H is
semisimple or not. If it is, the action of a unipotent element u must match one of the unipotent classes
of G, whose actions on Vmin and L(G) are tabulated in [13]. If it does not appear, or is generic, then
we are done.
• If V ↓H is not semisimple, and V is self-dual (i.e., all cases except when G = E6 and V = Vmin ) then
in order for a composition factor to appear in the socle and not be a summand, it must occur with
multiplicity at least 2. This allows us to cut down the possibilities for the socle of V ↓H .
• If the socle of V ↓H is W , then V is a submodule of P (W ), where P (W ) denotes the projective cover
of W . In particular, it is a submodule of the cf(V )-radical of P (W ), where we recall that cf(V ) is the
set of composition factors of V . This needs to contain at least as many copies of each composition
factor as there are in V , and further analysis of this radical can eliminate more cases.
• We can use Lemma 4.12: suppose that H = PSL2 (pa ) embeds in G, and an algebraic A1 -subgroup
X embeds in the algebraic version of G, such that for some module V , the highest weights of the
composition factors of both H and X on V are the same. Assume furthermore that the composition
factors of V ↓X satisfy the hypotheses of Lemma 4.12. We wish to conclude that H is a blueprint for
Vmin . In order to do this, an element x in H of order (pa ± 1)/2 must be guaranteed to come from a
class intersecting X. If the semisimple class containing x is determined by its eigenvalues on V then
this is true, but this is not true for every semisimple class, so we will have to check when we use the
lemma.
41
• In a similar vein, we can look for elements of G that do not lie in H and yet stabilize some eigenspaces
of an element of H on a module V : if ζ1 , . . . , ζr are roots of unity and y acts as a scalar on each
ζi -eigenspace of an element x ∈ H (i.e., preserves all subspaces of the eigenspace), then y stabilizes any
subspace of V on which x acts with eigenvalues some of the ζi . In particular, if there is a submodule W
of H with this property then hH, yi stabilizes W . Of course, it might be that hy, Hi is almost simple,
say PGL2 (pa ) for example, so we need to exclude this case by finding other such elements, proving
that the index of H in this group is not 2, or applying Corollary 4.14.
• If G = E6 , E7 and p = h − 1 where h is the Coxeter number of G, then in one case we prove that
H and NḠ (H) stabilize an sl2 -subalgebra of L(G). We can then apply Corollary 4.19 on positivedimensionality of such a stabilizer.
Some combination of these ideas is usually enough to solve any case we will see here.
42
F4
8
In this section, k is a field of characteristic p > 2 and G = F4 (k). Let Ḡ be an almost simple group with
socle G. From Section 4.2 we see that v(F4 ) = 18, so if H is any subgroup of G with a semisimple element
of order at least 19, then H is a blueprint for Vmin . The same holds for Ḡ except possibly if p = 2 and Ḡ
induces a graph involution on G. In [10] we proved that, when p = 2, SL2 (2a ) cannot be a maximal subgroup
of Ḡ if a > 5 regardless of this potential problem with graph automorphisms. In addition, in [9] we proved
that SL2 (4) cannot be a maximal subgroup of Ḡ either, so here we let H = PSL2 (pa ) with a = 3, 4 if p = 2
and pa 6 36 = 2 · v(F4 ) if p is odd. Let L = PSL2 (p) 6 H and let u denote a unipotent element of L of
order p.
8.1
Characteristic 2
Let p = 2. Since all semisimple elements of G lie inside D4 , which centralizes a 2-space on Vmin , and an
element of order 2a + 1 in H has a fixed point only on the trivial simple module, we see that Vmin ↓H has at
least two trivial composition factors. In particular, Lemma 5.2 applies in this situation, and so the pressure
of Vmin ↓H has to be at least 2 for H not to fix a line on Vmin .
We start by eliminating the possibility that k is not a splitting field for H, or where the composition
factors of Vmin ↓H are invariant under some outer automorphism of H, i.e., a field automorphism of H.
Proposition 8.1 Suppose that p = 2 and a = 3, 4, and suppose that k does not contain a splitting field for
H, or that the composition factors of Vmin ↓H are invariant under a field automorphism of H. Then either
H or its image under the graph automorphism fixes a line on Vmin .
Proof: Firstly let a = 3. The quickest way to do this is to use the traces of semisimple elements to see that
the composition factors of Vmin ↓H must be
82 , (21 , 22 , 23 ), 14 ,
(41,2 , 41,3 , 42,3 ), (21 , 22 , 23 )2 , 12 ,
83 , 12 .
The trace of an element of order 7 on these modules is 5, −2 and 5 respectively, and the trace of an element
of order 9 is 2, −1 and −1 respectively. Using Lemma 4.16, we therefore see that the first and second sets
of composition factors are swapped by the graph automorphism and the third does not exist. The first has
negative pressure, so fixes a line on Vmin , as needed.
Now let a = 4. We assume that k contains F4 but not F16 , as this includes the case where k ∩ F16 = F2 .
There are ten conspicuous sets of composition factors for Vmin ↓H , two of which are definable over F2 . They
fall into three orbits under the graph automorphism (recall that the graph automorphism squares to a field
automorphism), and well-chosen representatives of the three orbits are
41,3 , (21 , 23 )4 , 16 ,
(81,2,4 , 82,3,4 ), 41,3 , (21 , 23 ), 12 ,
41,2 , 42,3 , 43,4 , 41,4 , 41,3 , 42,4 , 12 .
The pressures of these modules are 2, 0 and −2 respectively, so the second and third definitely fix lines on
Vmin . For the first, if 41,3 lies in the socle of Vmin ↓H then since Vmin is self-dual, it must be a summand,
so we can assume that it is not; hence the socle consists of 21 s and 23 s, else H fixes a line on Vmin . The
{1, 21 , 23 , 41,3 }-radical of P (21 ) is
1/21 , 23 /1, 41,3 /21 ,
so there must be at least three 2i in the socle of Vmin ↓H , but then soc(Vmin ↓H ) has pressure 3 but Vmin ↓H
has pressure 2, contradicting Lemma 2.2. This completes the proof of the result.
43
We now turn to the case where k contains a splitting field for H. We split a = 3 and a = 4 into two
propositions.
Proposition 8.2 Suppose that p = 2 and a = 3, and that k contains a splitting field for H. If Vmin ↓H or
its image under the graph automorphism does not fix a line on Vmin , then up to field automorphism of H
the composition factors of Vmin ↓H are
421,3 , 241 , 22 , 223 , 14 ,
H stabilizes a 2-space on Vmin , and NḠ (H) = HCḠ (H) lies inside a member of X σ .
Proof: We use the traces of semisimple elements of order at most 17 to find all conspicuous sets of composition factors for Vmin ↓H . There are sixty-three such sets, ten of which do not have corresponding sets
of composition factors for L(G) ↓H . The rest fall into orbits of length 2, 3 and 6 under the graph and field
automorphisms of G. (The orbit of length 2 arose in Proposition 8.1, as did one of the ten sets that do not
yield corresponding composition factors on L(G).)
Note that there are two orbits of length six that share three points, because there are rational elements
of order 9 whose image under the graph automorphism is non-rational, so there are three options for their
image. This does not affect things, as the points they share correspond to conspicuous sets of composition
factors with non-positive pressure so fix lines on Vmin .
Six of the ten orbits contain sets with pressure 0, and a further two contain sets of pressure 1, which
are forbidden by Lemma 5.2. We are left with two orbits, one of length six and one of length three, with
representatives
8, 41,3 , 42,3 , 221 , 22 , 23 , 12 ,
and
421,3 , 241 , 22 , 223 , 14 .
Suppose that the first set of factors does not fix a line on Vmin : since the only non-trivial simple module
appearing more than once is 21 , Vmin ↓H is the sum of an indecomposable module with socle 21 and a
semisimple module, which we can ignore. We know that Vmin ↓H is self-dual, so if there is a non-split
extension between two non-trivial simple modules A and B inside Vmin ↓H , we must also have one between
B and A: i.e., A (or B) must be 21 , and so the other must be 41,3 by Lemma 5.1. In particular, 42,3 must
be a summand of Vmin ↓H . The {1, 21 , 22 , 23 , 41,3 }-radical of P (21 ) has three trivial composition factors, but
we remove all simple quotients other than 21 to obtain the smaller module
21 /1/21 , 22 /1, 41,3 /21 ,
on which an involution acts projectively. However, an involution cannot act projectively on Vmin (see [13,
Table 3]) so H must fix a line on Vmin .
We come to the final set of composition factors. There are modules with these composition factors that
do not fix a line, for example
⊕2
41,3
⊕ 21 /1/22/1/21 ⊕ 21 /1/23 ⊕ 23 /1/21,
and this has an allowable action of an involution as well. We note that Vmin ↓H always has a 2-dimensional
submodule, and so lies inside a member of X σ , and furthermore, since Vmin ↓H is not stable under the field
automorphism of H, if H is normalized by any element of an almost simple group Ḡ with socle G then H is
centralized by it.
To see that H stabilizes a 2-space on Vmin , note that otherwise Vmin ↓H is a submodule of P (41,3 ), but
P (41,3 ) has structure
41,3 /21 /1/22/1/21/41,3 ,
44
and there are many reasons why this cannot work: the dimension is 16, 23 is not involved in it, there are
not enough 1s or 21 s, the involution u acts projectively on it, and so on.
There is a copy of SL2 (8) inside F4 that does indeed not fix a line on Vmin up to the graph automorphism,
inside Ã2 A2 , with H projecting along the Ã2 factor as 21 /1 and along the A2 factor as 1/23: the product
of these two modules is an indecomposable module 21 /1/23 /41,3 with dual 23 /1, 41,3/21 , and the product of
21 /1 and its dual is
21 /1/22/1/21 ⊕ 1,
yielding an embedding into F4 with the required property (remember that the trivial in the last decomposition
is removed when considering Vmin .)
We now turn to a = 4, which ends this section on F4 in characteristic 2. Almost exactly the same result
holds in this case.
Proposition 8.3 Suppose that p = 2 and a = 4, and that k contains a splitting field for H. If Vmin ↓H or
its image under the graph automorphism does not fix a line on Vmin , then up to field automorphism of H
and graph automorphism of G the composition factors of Vmin ↓H are
421,3 , 241 , 22 , 223 , 14 ,
H stabilizes a 2-space on Vmin , and NḠ (H) = HCḠ (H) lies inside a member of X σ .
Proof: We proceed as in Proposition 8.2, starting by assembling all conspicuous sets of composition factors
using the traces of semisimple elements of order up to 17, this time finding 146 conspicuous sets of composition
factors, 16 of which have no corresponding set of composition factors on L(G), falling into eighteen orbits,
fifteen of length 8, two of length 4 and one of length 2.
Eleven of these orbits contain a conspicuous set of composition factors with negative pressure, and a
further two with factors with pressure 0. Using Lemma 5.2 we can exclude factors with pressure 1 as well,
eliminating a further two orbits. There remain three orbits, each of length 8, so not stable under Out(H) or
the graph automorphism of G.
The three orbits have representatives
81,2,3 , 41,3 , 42,3 , 221 , 22 , 23 , 12 ,
81,2,4 , 41,3 , 42,4 , 221 , 22 , 23 , 12 ,
421,3 , 241 , 22 , 223 , 14 .
For the first orbit we argue as in Proposition 8.2, to see that since the only non-trivial simple module
appearing more than once is 21 , so Vmin ↓H is a sum of a semisimple module, which can be ignored, and a
self-dual submodule of P (21 ) that has top 21 . The {1, 22, 23 , 41,3 , 42,3 , 81,2,3 }-radical of P (21 )/21 , lifted back
to P (21 ), is
23 /1, 42,3 /22 , 23 /1, 41,3/21 ,
but on top of some submodule of this must go a 21 , and the module must be self-dual. We therefore see that
we can remove all simple quotients other than 1 and 41,3 , as some submodule of this must be the second
radical layer (as 1 ⊕ 41,3 is the second socle layer). Doing this yields the smaller module
1/22/1, 41,3 /21 ,
and an involution u acts projectively on this, a contradiction.
45
The second orbit is almost identical, and the exact same proof works with the same smaller submodule
of P (21 ), and a slightly different original module of
1/22 , 23 /1, 41,3 /21 .
For the third orbit, since we constructed an example of this embedding not fixing a line on Vmin inside the
Ã2 A2 subgroup just after Proposition 8.2, we will not be able to prove that it fixes a line on Vmin . However,
it does stabilize a 2-space, as for a = 3: the {1, 21 , 22 , 23 , 41,3 }-radical of P (41,3 ) is
22 /1/21, 23 /41,3 ,
so soc(Vmin ↓H ) cannot be just 41,3 , and H either fixes a line or a 2-space on Vmin , as needed.
8.2
Characteristic 3
In characteristic 3, since v(F4 ) = 18 we only need consider a = 2, 3. We begin with the case a = 3, since we
have a complete result for that. In fact, we show that PSL2 (27) is always a blueprint for Vmin .
Proposition 8.4 Suppose that p = 3 and a = 3. Then H is a blueprint for Vmin .
Proof: There are forty conspicuous sets of composition factors for Vmin ↓H , but for only seven of these do
the elements of order 13 come from semisimple classes that are not blueprints for Vmin , and only three up
to field automorphism of H, which are
122,3,1 , 91,2 , 41,2 ,
122,3,1 , 91,3 , 41,2 ,
421,2 , 421,3 , 422,3 , 1.
Let ζ denote a primitive 13th root of unity, and let θ denote a primitive 26th root of unity with θ2 = ζ. Let
x ∈ H denote an element of order 13 that acts on 41,2 with eigenvalues ζ ±1 and ζ ±2 . In the first case, x acts
on Vmin with eigenvalues
12 , (ζ ±1 )2 , (ζ ±2 )3 , (ζ ±3 )2 , (ζ ±4 )2 , ζ ±5 , (ζ ±6 )2 .
There is an element x̂ of order 26 in G that squares to x and has eigenvalues
12 , (θ±1 )2 , (θ±2 )3 , (θ±3 )2 , (θ±4 )2 , θ±5 , θ±6 , (−θ±6 ).
This does not stabilize all the eigenspaces of x, but it only splits the ζ ±6 -eigenspaces, which are contained
inside the 12 factor. Hence x̂ stabilizes all subspaces stabilized by H and, since x̂ has order 26 > v(F4 ), this
means that H is a blueprint for Vmin .
In the second and third cases, the trace of x on Vmin is 0, and there is an element x̂ in G, of order 26,
such that x̂2 = x and x̂ acts on Vmin with eigenvalues
1, (θ±1 )2 , (θ±2 )2 , (θ±3 )2 , (θ±4 )2 , θ±5 , (−θ±5 ), θ±6 , (−θ±6 ),
so x̂ stabilizes any 41,2 in the socle of the second and third cases. This means that in the second case H is
contained inside a positive-dimensional subgroup stabilizing the 4-space.
Examining the list of maximal positive-dimensional subgroups of G, if H acts on Vmin with factors 12, 9, 4
then the only member of X σ in which H can lie is A1 C3 , which acts with factors of dimension 12 and 13, so
any positive-dimensional subgroup containing H must also stabilize the 12; hence H is a blueprint for Vmin
in this case as well.
46
We are left with the case of 421,2 , 421,3 , 422,3 , 1. We easily show that H lies inside a member of X , (although
not necessarily X σ yet). If H fixes a line on Vmin then we are done, so the socle consists of 4s, so by relabelling
we get 41,2 as a submodule of Vmin ↓H . But now the element x̂ above must stabilize any 41,2 in the socle
of Vmin ↓H , and since x̂ is a blueprint for Vmin , there is an infinite subgroup of G stabilizing a 4-space that
hH, x̂i stabilizes. Thus H is contained in an element of X , as claimed.
We now run through the elements of X , proving in fact that H does fix a line and u lies in the generic
class A2 , thus H is a blueprint for Vmin .
We cannot embed H in a maximal parabolic or A2 Ã2 as the dimensions are not compatible. If H 6 B4 ,
which acts as 9 ⊕ 16, then by Lemma 5.12 H must act semisimply on the 9 (as the 9 is self-dual), so the
action of u of order 3 has at least 36 , 13 on the 25-dimensional Vmin . Checking Table 4.1 we see that u
belongs to class A2 , generic.
If H embeds in X = A1 C3 then we may assume that H acts along A1 as 21 . Since (1, 100) is a summand of
Vmin ↓X , we need 6-dimensional modules whose tensor product with 21 only have 4-dimensional composition
⊕2
factors, each appearing at most twice, and there are three of these: 2⊕2
2 ⊕ 23 , 22 ⊕ 23 , and 63,1 , but none
of these has the correct exterior square, so H does not embed in A1 C3 .
We are left with A1 G2 , which acts on Vmin with composition factors (2, 10) of dimension 21 and (4, 00)
of dimension 4. This is impossible: H acts along the A1 factor as 31 , and so the action of H on the minimal
module for G2 cannot have a trivial or 3-dimensional composition factor, because the product with 31 is not
right. But then you cannot make a 7-dimensional module at all, a contradiction.
Thus if H embeds into G with these factors on Vmin then it is a blueprint, as needed.
We now consider a = 2, which we did not consider in [9] because we could not produce a complete answer
there, and we cannot here either. We will make substantial progress, pinning down precisely the action of
H on Vmin , but not enough to prove that it is always contained inside a positive-dimensional subgroup.
Proposition 8.5 Suppose that p = 3 and a = 2. One of the following holds:
(i) H fixes a line on Vmin or L(G);
(ii) H stabilizes a unique 3-space on Vmin ;
(iii) the action of H on the minimal module for E6 has two trivial submodules;
(iv) up to field automorphism of H, the action of H on Vmin is
9 ⊕ 4/1, 31/4 ⊕ 4,
where 4/1, 31/4 = 4 ⊗ 32 .
If (i), (ii) or (iii) hold, then NḠ (H) is contained in a member of X σ .
Proof: Using the traces of semisimple elements of orders 2, 4 and 5, one finds, up to field automorphism of
H, eight conspicuous sets of composition factors, namely
361 , 17 ,
43 , 331 , 14 ,
9, 43 , 31 , 1,
44 , 31 , 32 , 13 ,
4, 371 ,
4, 361 , 32 ,
9, 42 , 31 , 32 , 12
92 , 4, 31 .
The first case is semisimple and u lies in the generic class A2 . The sixth and seventh have the trace of an
involution being −7 on Vmin , whence it has trace 20 on L(G), so that L = PSL2 (3) acts with composition
47
factors 38 , 128 . By Lemma 5.8 any non-trivial simple module for H has at most two 3s for every three
1s on restriction to L, and so H always has at least sixteen trivial composition factors and at most eight
non-trivial composition factors. Since H 1 (H, 4) has dimension 2 by Lemma 5.9, L(G) ↓H has non-positive
pressure and hence has a trivial submodule. The second conspicuous set of composition factors must yield
a trivial submodule on Vmin by Lemma 5.11. This leaves the third, fourth, fifth and eighth conspicuous sets
of composition factors.
Suppose that we show that H stabilizes a unique 3-space on Vmin . The same must be true of NḠ (H), and
so if the stabilizer of that 3-space is positive-dimensional then NḠ (H) lies inside a member of X σ , and in
particular it is not maximal. However, Proposition 6.1 shows that if H stabilizes a 3-space on Vmin then the
stabilizer of that 3-space is positive dimensional, and so we are done whenever this is the case. In particular,
the semisimple eighth case must yield a unique 3-space being stabilized.
Also, if Vmin ↓H has no 1-cohomology then H stabilizes a line other than the F4 line, and Proposition
6.1 again states that NḠ (H) lies inside a member of X σ .
If the composition factors are 9, 42 , 31 , 32 , 12 , and we fix neither a line nor a unique 3-space on Vmin , then
the structure of Vmin ↓H is either
9 ⊕ 4/1, 1, 31, 32 /4
or
9 ⊕ 31 ⊕ 32 ⊕ 4/1, 1/4.
We claim that, in either case, Vmin ↓H has no 1-cohomology. The second module is uniquely determined and
so is an easy computer calculation. In the first case, the quotient by the unique 4 in the socle is also uniquely
determined, and this module has no 1-cohomology; adding a 4 on the bottom that already has two trivial
modules above it cannot add to the 1-cohomology, and so the claim holds. This proves, from the remarks
above, that NḠ (H) lies inside a member of X σ .
For 44 , 31 , 32 , 13 , we cannot have a single 3 in the socle of Vmin ↓H as we saw above. We cannot have two
3s in the socle, for then the module has 31 ⊕ 32 as a summand and 4, 4/1, 1, 1/4, 4, but since the self-dual
module 4/1, 1/4 is the {1, 4}-radical of P (4), such a module cannot exist.
Thus Vmin ↓H must have the form 4, 4/1, 1, 1, 31, 32 /4, 4. We now look at the image of H inside E6 , and
its action on the 27-dimensional module V27 . If soc(V27 ↓H ) = 1 then, since P (1) has dimension 27, we have
that P (1) = V27 ↓H . However, the action of u of order 3 on V27 is clearly now 39 , so acts on Vmin as 38 , 1,
from Table 4.1. But if we remove the top and socle from P (1) we get a 25-dimensional module on which u
acts as 37 , 22 , a contradiction.
Thus there exists an H-submodule 1 ⊕ 4 of the minimal module for E6 . Notice that P (4) has dimension
36 and has five socle layers, and P (1) has five socle layers, so since neither of these is contained in the module
V27 ↓H (where V27 is the minimal module for E6 ), we must have that V27 ↓H has at most four socle layers.
In particular, since V27 is self-dual, we cannot have a uniserial module 3i /4/1 as a subquotient of V27 ↓H ,
since 1/4/3i would also have to be a subquotient, and there is a unique 3i in V27 ↓H , hence V27 ↓H needs at
least five socle layers, not allowed.
Consider the preimage W of soc2 (Vmin ↓H ) in V27 , and in particular the {1, 4}-radical of W . This is the
preimage of a module 1, 1/4 ⊕ 1/4, and since 1, 1/4 has no extensions with 1, the {1, 4}-radical of W must
be a module 1, 1/4 ⊕ 1/4/1 (the uniserial module 1/4/1 is not uniquely determined). We need to place both
a 31 and a 32 on this module, but without constructing a uniserial 3i /4/1: this must yield a module
1, 1, 31, 32 /4 ⊕ 1/4/1,
the only way to exclude the uniserial 3i /4/1. Since there is no uniserial module 4/1/4 by Lemma 5.12, no 4
48
placed on top of this module W can cover the 1, and so Vmin ↓H has a trivial quotient, not allowed. Thus
(without loss of generality) 31 is diagonally placed across the 1, 1/4 and the 1/4/1.
Since Ext1 ((1, 1/4), 31) = 0, at most one (and in fact exactly one) of the modules 1/4/1 can have an
extension with 31 , and by replacing the summands by diagonal summands if necessary, we end up with a
module
1, 1/4 ⊕ 1, 31 /4/1,
and we have our uniserial 31 /4/1, not allowed. We therefore see that H cannot have non-trivial 1-cohomology,
as needed for the proposition.
The final case to consider is 9, 43 , 31 , 1. Again we cannot have the 31 in the socle, and not a trivial either,
and so Vmin ↓H must be
9 ⊕ 4/1, 31/4 ⊕ 4.
There is a unique self-dual module 4/1, 31/4, and it is 4 ⊗ 32 . To see that it is unique, notice that above
1, 1, 31 /4 one may place two copies of 4, one of which lies above 1, 1 only to make the module 4/1, 1/4, and
so any diagonal 4 between this and the one above 4 ⊗ 32 must cover both 1s, so we cannot peel a 1 off as a
quotient to make a different 4/1, 31/4. This proves (iv).
We can give a bit more information about case (iv) now. The action of such an H on V27 is unique as well:
although 4/1, 31/4 has non-trivial 1-cohomology, if the trivial does not lie only underneath the summand 4,
there must be five socle layers to V27 ↓H , which as we saw in the previous case leads to a contradiction.
This set of composition factors corresponds to 9, 44 , 351 , 332 , 13 on L(G), but even this does not help as H
can act on L(G) without fixing a line. We cannot say much about the 3i s, so remove them from the top and
bottom, as well as the 9, to leave a module W . With four 4s and three 1s we cannot have P (4) or P (1), so
from their structure above W has (at most) three socle layers. The two unipotent classes that act on Vmin as
38 , 1 are Ã2 and Ã2 + A1 : the former acts on L(G) with seven blocks of size 1, and so we end up fixing a line,
but the other acts as 316 , 22 , and need not. With two 4s in the socle of W , and three 1s above, we get as in
the previous analysis 1, 1/4 ⊕ 1/4, on which we can place 3i s and then two 4s. We must place 31 or 32 on top
of the 1/4 to avoid fixing a hyperplane, and then on top of the 1, 1/4 we need a 31 and a 32 , since we need
both classes of elements of order 3 to act as 316 , 22 and modules of the form 4/1, 1/4 and 4/1, 1, 3i/4 do not
allow this. However there is, unique up to isomorphism, a self-dual indecomposable module 4/1, 1, 31, 32 /4
on which both unipotent classes act with two blocks of size 2.
Inside C3 A1 , there is a copy of H whose projections along each factor act irreducibly on the respective
natural modules, and it acts on Vmin as stated, and on L(G) as
31⊕3 ⊕ 9 ⊕ 32 /4/1, 31/4/32 ⊕ 4/1, 1, 31, 32 /4.
We cannot push the analysis far enough to get uniqueness of this subgroup: it certainly exists, as we have
seen.
8.3
Characteristic at least 5
Let p > 5, and recall that H = PSL2 (pa ) for some a > 1, with pa 6 36 = 2 · v(F4 ), with u ∈ H of order
p. The possible actions of u on Vmin are given in [13, Table 3]; by Lemma 4.6 we may assume that our
unipotent class is not generic: this leaves us with the following three unipotent classes:
(i) C3 , p = 7, acting as 72 , 62 ;
49
(ii) F4 (a2 ), p = 7, acting as 73 , 5;
(iii) F4 , p = 13, acting as 132 .
This proves the following result immediately.
Proposition 8.6 If pa 6= 7, 13 then H is a blueprint for Vmin .
For p = 7 we have the following result.
Proposition 8.7 If pa = 7 then H fixes a line on either Vmin or L(G).
Proof: We use the traces of elements of orders 2, 3 and 4 to produce the possible composition factors of
Vmin ↓H , namely
36 , 18 ,
5, 37 ,
53 , 33 , 12 ,
7, 53 , 3, 1,
73 , 15 .
We saw in Section 5.3 that the only indecomposable module with the trivial composition factor but no trivial
submodule or quotient is P (3) = 5/1, 3/5. This immediately tells us that the first, third and fifth cases fix
lines on Vmin . (Indeed, the first and fifth cases have that all trivial factors are summands.)
The case 7, 53 , 3, 1 yields traces of elements of orders 2, 3 and 4 of 2, −1 and −2 respectively. This yields
traces on L(G) of −4 for an involution, −2 or 7 for an element of order 3, and finally 4 for an element of
order 4. There is no set of composition factors that are compatible with this, so this case cannot occur.
If the composition factors are 5, 37 , then the traces of the elements of orders 2 and 3 fix uniquely the
composition factors on L(G), and these become 57 , 3, 114 , and so H fixes lines on L(G), completing the proof.
For p = 13 we are left with one open possibility, which we will prove yields a Serre embedding (see
Definition 4.7).
Proposition 8.8 Suppose that pa = 13. Either H is a blueprint for Vmin , or u is a regular unipotent element
and Vmin ↓H and L(G) ↓H are given by
P (9) = 9/3, 5/9
and
P (3) ⊕ P (11) = 3/9, 11/3 ⊕ 11/1, 3/11
respectively. In particular, H is a Serre embedding.
Proof: From the list above, the regular class is the only non-generic one for p = 13, so if H is not a blueprint
for Vmin then u is regular and in particular acts projectively on Vmin and L(G), hence both modules must
restrict to H as projectives. The projective indecomposable modules for H are
1/11/1,
3/9, 11/3,
5/7, 9/5,
7/5, 7/7,
9/3, 5/9,
11/1, 3/11,
13.
Thus there are eight possible projective modules of dimension 26, two of which yield conspicuous sets of
composition factors on Vmin , namely P (5) and P (9). The first of these does not have corresponding factors
on L(G), and the second has factors 113 , 9, 33 , 1, which yield the projective module P (3) ⊕ P (11), as claimed.
We have therefore completed the proof of Theorem 1.1.
50
E6
9
In this section, k is a field of characteristic p > 2 and G = E6 (k), by which we mean the simply connected
form, i.e., |Z(G)| = gcd(3, |k × |) (if k is finite) and G′ = G. Let Ḡ be an almost simple group with socle
G/Z(G). From [10] we see that for real semisimple elements (and the semisimple elements of PSL2 (pa ) are
real), v(E6 ) = 18, so if H is any subgroup of G with a real semisimple element of order at least 19, then
H a blueprint for Vmin . The same holds for Ḡ, even when Ḡ involves the graph automorphism, by using
∗
Vmin ⊕ Vmin
instead of Vmin , which is Ḡ-stable, and applying Lemma 4.9. In addition, in [9] we prove that
almost simple groups with socles SL2 (4) and PSL2 (9) cannot be maximal subgroups of Ḡ either, so here we
let H = PSL2 (pa ) with a = 3, 4 if p = 2 and pa 6 36 = 2·v(F4 ) with pa 6= 9 if p is odd. Let L = PSL2 (p) 6 H
and let u denote a unipotent element of L of order p, as in Section 8.
9.1
Characteristic 2
Let p = 2. Unlike G = F4 the case of p = 2 is easy, since the graph automorphism, which could cause the
only problem, simply induces duality. As we see above, we just have to deal with a = 3, 4, and when a = 3
the group Out(H) has order 3, hence a graph automorphism must centralize H (hence H is an element of
X σ ), not merely normalize it. We therefore see that if H = SL2 (8) stabilizes a 2-space on Vmin then NḠ (H)
lies inside a positive-dimensional subgroup, even if Ḡ induces a graph automorphism on G.
For a = 4 we use the fact that, while not every semisimple element of order 17 in F4 is a blueprint for
the minimal module, almost every one is. This statement passes through to Vmin , since our real semisimple
elements lie in F4 , via Lemma 4.9.
We start with a = 3.
∗
Proposition 9.1 Suppose that p = 2 and a = 3. Then H fixes a line or 2-space on Vmin or Vmin
.
∗
Proof: Suppose that soc(Vmin ↓H ) and soc(Vmin
↓H ) have neither 1s nor 2s, so Vmin ↓H is a submodule of
P (4)s and 8s. The projective cover of 4i,i+1 is
4i,i+1 /2i+1 /1/2i−1 /1/2i+1/4i,i+1 ,
and thus Vmin ↓H is a sum of projectives P (4i,i+1 ) and 8s, but this has even dimension, not right.
Now we move on to a = 4, where we use semisimple elements of order 17 that are blueprints for Vmin , as
suggested earlier.
∗
Proposition 9.2 Suppose that p = 2 and a = 4. The subgroup H is always a blueprint for Vmin ⊕ Vmin
.
∗
Proof: Of the 230 semisimple classes in F4 of elements of order 17, all but two are blueprints for Vmin ⊕Vmin
,
as we saw in Section 8, with representatives x and x3 , where x has eigenvalues
±1 2
±2 2
±3
±4 2
±5
±6
±7
±8 2
13 , (ζ17
) , (ζ17
) , (ζ17
), (ζ17
) , (ζ17
), (ζ17
), (ζ17
), (ζ17
)
on Vmin . We thus may assume that every element of H of order 17 is conjugate to either x or x3 .
However, although there are 107766 possible sets of composition factors for a module of dimension 27,
none of them has the eigenvalues above, up to algebraic conjugacy. Thus a semisimple element of H of order
∗
17 is always a blueprint for Vmin ⊕ Vmin
, and so the result holds by Lemma 4.9.
51
9.2
Characteristic 3
Let p = 3. From the remarks at the start of this section we need only consider a = 3, i.e., H = PSL2 (27).
In the previous section we exploited the fact that most semisimple elements of order 17 are blueprints for
Vmin . We will do the same here with order 13 elements. Of the 104 semisimple classes of elements of order
13 in F4 , all but seven are blueprints, since there are elements of order 26 that square to them and preserve
the number of eigenvalues.
∗
Proposition 9.3 Suppose that p = 3 and a = 3. Either H is a blueprint for Vmin ⊕ Vmin
or it fixes a line
∗
on Vmin or Vmin
.
Proof: This is easier than the case of F4 , but will start in exactly the same way. There are fifty conspicuous
sets of composition factors for Vmin ↓H , but for only seven of these do the elements of order 13 come from
∗
semisimple classes that are not blueprints for Vmin ⊕ Vmin
, three up to field automorphism of H, which are
122,3,1 , 91,2 , 41,2 , 12 ,
122,3,1 , 91,3 , 41,2 , 12 ,
421,2 , 421,3 , 422,3 , 13 .
The first two have pressure −1 and so must fix a line on Vmin . For the third, since there is no module 4/1/4
for H by Lemma 5.12, Vmin ↓H cannot have the form 4, 4, 4/1, 1, 1/4, 4, 4, and so if H does not fix a line or
hyperplane on Vmin then there can be either one or two 4s in the socle.
The {1, 4}-radical of P (41,2 ) has three trivial factors but has a trivial quotient, so we may assume that
the socle of Vmin ↓H is the sum of two 4s. Since the pressure of Vmin ↓H is 3, we cannot have a submodule
with three 4s, but we need two 1s above this socle (else we could quotient out by one of them and get a
module with a simple socle), so the socle is 4, 4 and the second socle layer is 1, 1, so we must have a 4/1/4
∗
subquotient, not allowed. Thus H fixes a line on Vmin or Vmin
, as needed.
9.3
Characteristic at least 5
Let p > 5, and recall that H = PSL2 (pa ) for some a > 1, with pa 6 36, with u ∈ L 6 H of order p, where
L = PSL2 (p). The possible actions of u on Vmin are given in [13, Table 5]; by Lemma 4.6 we may assume
that our unipotent class is not generic, leaving us with the following seven unipotent classes:
(i) A4 , p = 5, acting as 55 , 12 ;
(ii) A4 + A1 , p = 5, acting as 55 , 2;
(iii) A5 , p = 7, acting as 72 , 62 , 1;
(iv) D5 (a1 ), p = 7, acting as 73 , 3, 2, 1;
(v) E6 (a3 ), p = 7, acting as 73 , 5, 1;
(vi) E6 (a1 ), p = 11, acting as 112 , 5;
(vii) E6 , p = 13, acting as 132 , 1.
We now go prime by prime, starting with p = 5.
Proposition 9.4 Suppose that p = 5. If a = 1 then H fixes a line on either Vmin or L(G). If a = 2 then H
either fixes a line or hyperplane on Vmin , or a line on L(G).
52
Proof: Suppose that a = 1. The conspicuous sets of composition factors of Vmin ↓H are
36 , 19 ,
5, 37 , 1,
53 , 33 , 13 .
The first set of composition factors has pressure −3, so fixes a line on Vmin by Lemma 2.2. In the second
case we switch to L(G), on which H acts with composition factors
58 , 38 , 114
or
511 , 35 , 18 .
In either case, we see that H fixes a line on L(G), as needed. The third set of composition factors has
pressure 0, so might only fix a hyperplane on Vmin . However, the only indecomposable modules with a trivial
composition factor but no trivial submodule are submodules of P (3) = 3/1, 3/3, so in order not to fix a line,
Vmin ↓H must be
5⊕3 ⊕ (1/3)⊕3 ,
on which u acts as 53 , 43 , but this does not appear on [13, Table 5], so H does indeed fix a line (and
hyperplane) on Vmin .
Now suppose that a = 2. By Lemma 5.21, if Vmin ↓L has more trivials than 3-dimensionals then H fixes
a line on Vmin . Thus if Vmin ↓L is the first set of composition factors then H fixes a line on Vmin , and if
Vmin ↓L is the second set of composition factors then H fixes a line on L(G).
We therefore assume that Vmin ↓L has factors 53 , 33 , 13 . At this point it seems easiest to use the traces of
semisimple elements of order at most 13, finding eighteen conspicuous sets of composition factors, each with
at least one trivial factor and with non-positive pressure, so fix either a line or a hyperplane on Vmin .
For p = 7 we do not need to go past a = 1, which makes this easier than the previous case.
Proposition 9.5 Suppose that p = 7 and a = 1. Then H fixes a line or hyperplane on Vmin .
Proof: The conspicuous sets of composition factors are, as for p = 5, the same as for F4 but with an extra
trivial factor, namely
36 , 19 ,
5, 37 , 1,
53 , 33 , 13 ,
7, 53 , 3, 12 ,
73 , 16 .
The only indecomposable module that has a trivial composition factor but no trivial submodule or quotient
is P (5) = 5/1, 3/5, thus all of these sets of composition factors fix either a line or hyperplane on Vmin .
Since any indecomposable module with a trivial composition factor but no trivial submodule has 5/1 as
a submodule, the first and fifth conspicuous sets of composition factors in the proof definitely fix lines on
Vmin . If H fixes a hyperplane but not a line then it cannot lie in F4 and must lie inside a D5 -parabolic, with
composition factors of dimensions 1, 16 and 10. These are incompatible with the second and fifth sets of
composition factors, hence H also lies inside F4 in this case.
For p = 11, we see the first use of the idea of fixing an sl2 -subalgebra.
Proposition 9.6 Suppose that p = 11. Either H is a blueprint for both Vmin and L(G), or H has a trivial
summand on Vmin , or H acts on Vmin and L(G) as
P (9) ⊕ 5
and
11⊕2 ⊕ P (7) ⊕ P (5) ⊕ 9 ⊕ 3,
and fixes an sl2 -subalgebra of L(G).
53
Proof: Examining [13, Tables 5 and 6], we see that there are only two unipotent classes of elements of order
11 that are not generic for both Vmin and L(G), namely D5 (generic for Vmin ) and E6 (a1 ) (not generic for
either). If u belongs to class D5 , then it acts on Vmin with Jordan blocks 11, 9, 5, 12, and since there are two
Jordan blocks of size 1 and only one of size 11, Vmin ↓H must have a trivial summand as each non-trivial
indecomposable summand of dimension congruent to 1 modulo 11 has dimension 12 and uses up a block of
size 11.
We therefore assume that u belongs to class E6 (a1 ), so acts as 112 , 5 on Vmin and as 116 , 9, 3 on L(G).
There are five indecomposable modules of dimension congruent to 5 modulo 11, which up to duality are
5,
7, 5, 3/5, 3,
9, 7, 5, 3/1, 3, 5, 7, 9,
with the last one having dimension 49, not allowed, and the second one having dimension 27, with trace of
an involution −1, so not allowed. Thus Vmin ↓H is the sum of 5 and a 22-dimensional projective module.
We now use traces of semisimple elements of orders at most 6 to see which sums of projectives and a 5
are conspicuous, finding two, namely
11 ⊕ P (1) ⊕ 5
and
P (9) ⊕ 5.
The first fixes a line on Vmin but does not have a trivial summand, hence lies inside a D5 -parabolic, acting
on Vmin uniserially as 10/16/1, and the image of H inside the D5 -Levi must act as 1/9 on the 10, not allowed
since this is a self-dual module. Thus the first case does not exist, and H must be the second.
The corresponding sets of composition factors on L(G) are
11, 93 , 74 , 33 , 13
and
112 , 9, 73 , 54 , 32 .
Since L(G) is self-dual and there is a unique self-dual module congruent to each dimension modulo 11 by
Lemma 5.17, we have that 9 and 3 must be summands of L(G) ↓H . The first set of factors cannot form a
projective and these summands, but the second case can, yielding
11⊕2 ⊕ P (7) ⊕ P (5) ⊕ 9 ⊕ 3.
By Proposition 4.17, the 3-dimensional summand is an sl2 -subalgebra, as claimed.
When p = 13, the only non-generic class is the regular unipotent class. We will show more generally that
if H contains a regular unipotent element then H either lies in F4 , or p = 13 and H is a non-G-cr subgroup
in a D5 -parabolic subgroup of G.
Proposition 9.7 Suppose that p > 13. If H contains a regular unipotent element then H is contained in a
conjugate of F4 , or p = 13 and H is a non-G-cr subgroup of the D5 -parabolic acting on Vmin as
1/11/1 ⊕ 9/5.
or its dual.
∗
If H does not contain a regular unipotent element, then H is a blueprint for Vmin ⊕ Vmin
.
Proof: Suppose that p > 17: the action of a unipotent element on Vmin is 17, 9, 1, and for p > 19 we
must have that Vmin ↓H = 17 ⊕ 9 ⊕ 1, and so H lies inside either F4 , as desired, or a D5 -parabolic, but this
has composition factors 10, 16, 1, incompatible. For p = 17 then 17, 1 could come from an 18-dimensional
indecomposable module, but the 9 is a summand, so in particular H has three composition factors on Vmin .
54
However, u is contained in the regular class, which is generic for p = 17, hence H is a blueprint for Vmin ,
in particular an element X of X . Since X contains a regular unipotent element (eliminating all reductive
maximal subgroups except for F4 from [14]) and must have at most three composition factors on Vmin , and
if it does have three then one has dimension 9 (eliminating all parabolic subgroups), we must have H 6 F4 ,
as claimed.
We therefore have that p = 13, and u acts on Vmin with factors 132 , 1. Suppose that the 1 in the action of
u arises from a trivial summand in Vmin ↓H . From the proof of Proposition 8.8 we see that the conspicuous
sets of composition factors are
5/7, 9/5 ⊕ 1
and
9/3, 5/9 ⊕ 1.
Since there is no 10-dimensional quotient not including the trivial summand, these structures are incompatible with coming from a D5 -parabolic, and so H 6 F4 , as needed.
We thus assume that Vmin ↓H has no trivial summand. We therefore have a projective of dimension 13
(either P (1) or 13, both with a trace of 1 for the involution) and a module i/(p + 1 − i), with a trace of ±2.
As the trace of an involution on Vmin is either 3 or −5, we see that it has a trace of +2 on i/(p + 1 − i), and
hence i = 5, 9. This means that, up to duality, Vmin ↓H is either
13 ⊕ 5/9
or
1/11/1 ⊕ 9/5.
The second case is as claimed in the proposition, so we are left to eliminate the first case. Here we take the
Borel subgroup B of H: the exact structure of B on the 27-dimensional module Vmin is up to duality as
follows, where ζ is a cube root of unity.
1
ζ
−ζ
−ζ 2
ζ2
1
−1
−ζ
ζ
ζ2
−ζ
2
−1
1
ζ
−ζ
−ζ 2
ζ2
1
−1
−ζ
ζ
ζ2
−ζ 2
−1
1
ζ
ζ2
Since F4 acts on Vmin as 26 ⊕ 1, the point that H fixes is either a D5 -parabolic point or a B4 point, but
either way H lies inside a D5 -parabolic, either one stabilizing a line or one stabilizing a hyperplane.
Let v be a unipotent element of D5 contained in the image of L inside the D5 -Levi. Thus v acts on the
10 and 16 as subquotients of the action of u on Vmin , namely 132 , 1. Therefore v acts on both the 10 and
the 16 with at most three Jordan blocks, and if it has three then one is of size 1.
We can read off the unipotent classes of D5 from the table for D6 , [14, Table 6], which shows that there
are only three unipotent classes, A4 , D5 (a1 ) and D5 , that act with at most three blocks on the 10. From
the embedding of the D5 -Levi into E6 we can easily deduce the actions of these on the 16, as we just consult
[13, Table 5] which lists the block sizes for the classes for E6 , and look for the unipotent classes with these
names. This gives us the list below.
55
Class
A4
2
Action on 10
5
Action on 16
7, 5, 3, 1
D5 (a1 )
D5
7, 3
9, 1
72 , 2
9, 7
We therefore see that v comes from the regular class D5 , and so the image B̄ of B in D5 , which contains v,
must act on the self-dual module 10 as
1 ⊕ (ζ 2 / − 1/ζ/ − ζ 2 /1/ − ζ/ζ 2 / − 1/ζ).
This is a submodule of the action of B above, and we therefore see that B̄ acts on the 16 with eigenvalues
(1, −1)2 , (ζ, ζ 2 , −ζ, −ζ 2 )3 ;
these cannot form modules of dimension 9 and 7, since a module of dimension 9 needs exactly three ±1
eigenvalues, and a module of dimension 7 needs at least two ±1s.
This proves that H cannot embed with these composition factors, and completes the proof of the proposition.
We will construct this non-G-cr subgroup of the D5 -parabolic when p = 13; the same construction works
for the E6 -parabolic subgroup of E7 and p = 19.
Let (G, p, X, Y ) be one of (E6 , 13, D5 , B4 ) and (E7 , 19, E6 , F4 ), and let Vmin denote the minimal module
for G. One of the stabilizers of a point on Vmin is a subgroup that is the extension of a unipotent group
by Y , so let H be a copy of PSL2 (p) inside Y that covers the regular unipotent element, the fixed points
of a principal PSL2 subgroup of Y . This copy of H embeds in X, of course, and the action of X on the
unipotent radical of the X-parabolic is as a single simple module, so that the 1-cohomology is easy to
compute. We see that the restriction of this simple module to H contains a summand of dimension p − 2,
hence the 1-cohomology of H on the unipotent radical is 1-dimensional. There is an action of the torus of
the X-parabolic outside of X on this cohomology group, and this yields two conjugacy classes of subgroups
H in the X-parabolic, one inside X and another class of complements. Given the composition factors of
H on Vmin , together with the table from [13], there is a unique possible module structure for Vmin ↓H if H
does not lie inside X but merely the X-parabolic subgroup of G, and this intersects non-trivially the regular
unipotent class of G.
56
10
E7 in characteristic 2
In this section, let k be a field of characteristic 2 and let G = E7 (k). Let a > 1 be a positive integer and
H = SL2 (2a ). The case of characteristic 2 is very different from odd characteristic because if p is odd then
a copy of PSL2 (pa ) inside the simple group of type E7 can lift in the simply connected group to either
PSL2 (pa ) × 2 or SL2 (pa ), and the two possibilities require very different strategies. In characteristic 2 there
is no such bifurcation.
The case a = 2 is done in [9], and since v(E7 ) = 75 for semisimple elements of odd order, if a > 7 then
H is a blueprint for Vmin ; so we may assume that 3 6 a 6 6. Furthermore, if Vmin ↓H has at least six trivial
composition factors then by Proposition 4.10 we can assume that H has no semisimple elements of order
more than 30, so a 6 4 in this case.
We can use a computer to find which semisimple elements are blueprints for Vmin even when they have
order smaller than 77, or 30 when they centralize a 6-space. For example, of the 2430 classes of elements
of order 17, 1892 of them are blueprints for Vmin , which helps reduce the number of conspicuous sets of
composition factors that need to be considered when a = 4.
Suppose firstly that there are no 1- or 2-dimensional composition factors in Vmin ↓H . In this case if H
is not a blueprint for Vmin then we have to switch to the Lie algebra L(G)′ , which we recall has dimension
132, not 133 in the case p = 2. We address this situation now.
Proposition 10.1 Suppose that p = 2 and a = 3, 4, 5, 6. Suppose that there are no 1- or 2-dimensional
composition factors in Vmin ↓H .
(i) We cannot have a = 3, 4.
(ii) If a = 5, 6 then H is a blueprint for Vmin .
Proof: If H acts on Vmin with no composition factors of dimension 1 or 2, then the trace of an element of
order 3 on Vmin is one of −25, −7, 2, 20, and so the dimensions of the composition factors are one of seven
possibilities:
32, 16, 42,
32, 8, 44 ,
163 , 8,
162 , 82 , 42 ,
16, 83, 44 ,
87 ,
84 , 46 .
For these, no arguments about unipotent classes or stabilizing subspaces will work if H is not a blueprint for
Vmin , and so we will just have to deal with them case by case, switching to the Lie algebra where we need
to.
Let a = 3: the only conspicuous set of composition factors for Vmin ↓H is
84 , 421,2 , 421,3 , 422,3 ,
which does not have a corresponding set of composition factors on L(G)′ . For a = 4 we get no conspicuous
sets of composition factors for Vmin ↓H at all.
For a = 5, up to field automorphism of H, there are two conspicuous sets of composition factors for
Vmin ↓H , namely
161234 , 8135 , 8235 , 8345 , 4213 , 423 , 434
and
82123 , 8124 , 8135 , 412 , 413 , 4214 , 4215 .
It is easy to check that an element of order 31 is a blueprint for Vmin in both cases by finding an element of
order 93 that cubes to it and has the same number of eigenvalues on Vmin .
57
For a = 6, we do not have lists of the traces of elements of orders 63 and 65, but we can check whether
a given matrix is the trace of a semisimple element of order 63 on Vmin by using the preimage trick from
Section 4.2. Doing this to the seven possible sets of dimensions yields the following table. In this, the number
of sets of composition factors up to field automorphism is given in the second column, and those that are
conspicuous using elements of order up to 21 and 63 are given in the third and fourth columns respectively.
Dimensions
32, 16, 4
32, 8, 4
2
4
Number of modules
Conspicuous up to 21
Conspicuous for 63
1800
1
0
61200
5
0
3
16 , 8
2270
1
0
162 , 82 , 42
504240
32
0
11781000
159
2
109660
1
0
57206136
934
9
3
16, 8 , 4
8
4
7
8 ,4
6
4
We thus simply need to check whether for a given conspicuous sets of composition factors, that any conjugacy
class of elements of order 63 with the correct eigenvalues on Vmin is a blueprint for Vmin . This can easily be
done with a computer, and so we prove the result.
We have now dealt with the case where Vmin ↓H has no 1- or 2-dimensional composition factors. We
generally cannot prove that H fixes a line on Vmin , and often want to prove that H fixes a 2-space on Vmin .
For this to be enough to show that H is contained in a positive-dimensional subgroup, we need that k
contains a splitting field for H, and for NḠ (H) we need to also consider when Vmin ↓H is stable under a
field automorphism of H. In general, we therefore consider conspicuous sets of composition factors that are
stable under some field automorphism.
Proposition 10.2 Suppose that 3 6 a 6 6 and that the composition factors of Vmin ↓H are stable under a
non-trivial field automorphism of H.
(i) If a = 3 then either H fixes a line on Vmin or L(G)′ , or the composition factors of Vmin ↓H are
8, (41,2 , 41,3 , 42,3 )2 , (21 , 22 , 23 )3 , 16 .
(ii) If a = 4 then H is a blueprint for Vmin or the composition factors of Vmin ↓H are
(41,4 , 42,3 )2 , 421,2 , (21 , 23 )4 , (22 , 24 )2 , 18
and the stabilizer of any simple submodule of Vmin ↓NḠ (H) is positive dimensional.
(iii) If a > 5 then H is a blueprint for Vmin .
Proof: Let a = 3. Over F2 there are eight conspicuous sets of composition factors for Vmin ↓H , two of which
have no corresponding set of composition factors on L(G)′ . One is as in the proposition, with corresponding
factors
82 , (41,2 , 41,3 , 42,3 )4 , (21 , 22 , 23 )9 , 114
on L(G)′ .
The other five are
8, (41,2 , 41,3 , 42,3 )3 , (21 , 22 , 23 )2 ,
(41,2 , 41,3 , 42,3 )2 , (21 , 22 , 23 )4 , 18
58
86 , 18 ,
84 , (21 , 22 , 23 )2 , 112 ,
8, (21 , 22 , 23 )8 ,
with corresponding sets of factors on L(G)′ given by
82 , (41,2 , 41,3 , 42,3 )2 , (21 , 22 , 23 )11 , 126 ,
82 , (41,2 , 41,3 , 42,3 )4 , (21 , 22 , 23 )9 , 114 ,
811 , (41,2 , 41,3 , 42,3 ), (21 , 22 , 23 )3 , 114 ,
88 , (41,2 , 41,3 , 42,3 ), (21 , 22 , 23 )6 , 120 ,
(41,2 , 41,3 , 42,3 )8 , (21 , 22 , 23 ), 130 .
By Lemma 5.4 we need three 2s for every two 1s in order not to fix a line or hyperplane, and so the third
and fourth cases fix lines on Vmin and the first, third, fourth and fifth all fix lines on L(G)′ , so we need to
consider the second case. We assume that Vmin ↓H has no trivial submodules.
Remove any 4s from the top and bottom of Vmin ↓H , and any summands of dimension 2 to yield a module
W . Since Vmin ↓H has pressure 4, there are at most four 2s in the socle of W . We cannot have a summand
P (2i ) for any i because P (2i ) has dimension 56 and only four trivial factors, so as in the proof of Lemma
5.4 we see that W has at most five socle layers, and in fact exactly five socle layers. As it has pressure 4 we
cannot have a subquotient of pressure greater than 4 or less than −4 by Lemma 2.2, so the first, third and
fifth socle layers each have four 2s, and the second and fourth socle layers each have four 1s.
It is clear that, since there are four 2s in the socle, two of them must be isomorphic, say 21 . The other
⊕2
⊕2
two 21 s must be in the top, and so the socle is either M1 = 2⊕2
1 ⊕ 22 or M2 = 21 ⊕ 22 ⊕ 23 . In the first
case, we construct the {21 , 22 }′ -submodule of P (M1 )/M1 , and note that this has only six trivial composition
factors, so it cannot be the right socle.
In the other case we construct the {21 }′ -submodule of P (M2 )/M2 , which has exactly eight trivial composition factors, so they must all lie in W . Therefore we take the {1}′-residual of this, and it has structure
1, 1, 1, 1/22, 22 , 22 , 23 , 23 /1, 1, 1, 1, 42,3/21 , 21 , 22 , 23 .
This must be a submodule of W , and yet it has five 2s in the third socle layer, too many. This contradiction
proves that H fixes a line on Vmin , as needed.
Thus, except for the set of composition factors explicitly stated, H must fix a line on either Vmin or
L(G)′ .
For a = 4 there are, up to field automorphism of H, six conspicuous sets of composition factors over
k = F4 , namely
421,3 , (21 , 23 )8 , 116 ,
482,4 , (21 , 23 )2 , 116 ,
(41,4 , 42,3 )2 , 421,2 , (21 , 23 )4 , (22 , 24 )2 , 18
(81,2,4 , 82,3,4 )2 , 421,3 , (21 , 23 )2 , 18 ,
(41,2 , 41,3 , 41,4 , 42,3 , 42,4 , 43,4 )2 , 18 ,
162 , (21 , 22 , 23 , 24 )2 , 18 .
In the proof of Proposition 9.2 we saw that there are only two semisimple classes of elements of order 17
in F4 that are not blueprints for the 26-dimensional minimal module, and hence for Vmin . The semisimple
elements of order 17 in the conspicuous sets of composition factors above are always conjugate into F4 , and
all but the fourth one are in fact blueprints for Vmin (as they are for the minimal module for F4 ), so we get
the first part of the result.
We are left with the fourth set of composition factors. Let ζ denote a primitive 17th root of unity, and
choose x ∈ H of order 17 such that the eigenvalues of x on 21 are ±ζ, and therefore the eigenvalues of x on
Vmin are
18 , (ζ ±1 )4 , (ζ ±2 )4 , (ζ ±3 )2 , (ζ ±4 )4 , (ζ ±5 )2 , (ζ ±6 )2 , (ζ ±7 )2 , (ζ ±8 )4 .
59
There exists an element y1 in the algebraic group F4 , cubing to x, with eigenvalues
18 , (θ±1 )4 , (θ±2 )4 , (θ±3 )2 , (θ±4 )4 , (θ±5 )2 , (θ±6 )2 , (θ±7 )2 , (θ±8 )2 , (θ±9 )2
on Vmin , where θ is a primitive 51st root of unity with θ3 = ζ. This preserves all eigenspaces except for
the ζ ±8 ones, which lie in 24 and 41,4 . There also exists an element y2 in F4 , again cubing to x, and with
eigenvalues
18 , (θ±1 )4 , (θ±2 )4 , (θ±3 )2 , (θ±4 )2 , (θ±21 )2 , (θ±22 )2 , (θ±23 )2 , (θ±24 )2 , (θ±25 )4 ,
with now the ζ ±4 -eigenspaces being split and all others being preserved. Only the module 23 has ζ 4 as
an eigenvalue, so the stabilizer of any simple submodule of Vmin ↓H contains either y1 or y2 , and hence is
positive dimensional, either using the fact that v(F4 ) = 18 or Proposition 4.10.
This proves that, regardless of the field k, or whether NḠ (H) induces a field automorphism on H, NḠ (H)
is contained in a positive-dimensional subgroup stabilizing a σ-stable subspace of Vmin , as needed.
For a = 5, we use the traces of semisimple elements and get a unique set of composition factors,
32, (21 , 22 , 23 , 24 , 25 )2 , 14 .
An element of order 31 in H has 1-eigenspace of dimension 6 on these factors (since it has a 1-eigenspace of
dimension 2 on the 32), hence is a blueprint for Vmin by Proposition 4.10. Thus H is a blueprint for Vmin ,
as claimed.
When a = 6, the group of field automorphisms has order 6, so we need to consider those factors that
are stable under field automorphisms of orders 2 and 3. We start with those definable over F4 , i.e., stable
under the automorphism of order 3. There are sixteen conspicuous sets of composition factors for elements
of order up to 63, six of which have no factors of dimension 1 or 2, hence dealt with in Proposition 10.1,
with the rest given by
842,4,6 , (21 , 23 , 25 )2 , 112 ,
(41,3 , 43,5 , 41,5 )2 , (21 , 23 , 25 )4 , 18 ,
81,3,5 , (41,3 , 43,5 , 41,5 )2 , 41,4 , 42,5 , 43,6 , (21 , 23 , 25 )2 ,
81,3,5 , (21 , 23 , 25 )8 ,
81,3,5 , (41,3 , 43,5 , 41,5 )2 , 41,6 , 42,3 , 44,5 , (21 , 23 , 25 )2 ,
up to field automorphism of H. The first two of these have more than six trivial composition factors, hence
are blueprints for Vmin by Proposition 4.10; for the other three it can be checked manually that the elements
of order 63 are blueprints for Vmin , by finding elements of order 315 = 5 · 63 that power to them and have
the same number of eigenvalues on Vmin .
We also have to check sets of composition factors defined over F8 , where up to field automorphism there
are ten conspicuous sets of composition factors for semisimple elements of order 21, three of which lose their
conspicuousness on elements of order 63. The remaining seven all have at least eight trivial composition
factors, hence are blueprints for Vmin by Proposition 4.10. These are
421,4 , (21 , 24 )8 , 116 ,
483,6 , (21 , 24 )2 , 116 ,
(41,5 , 42,4 )2 , 421,4 , (21 , 24 )4 , (22 , 25 )2 , 18 ,
1621,3,4,6 , (21 , 24 )2 , (22 , 25 )2 , 18 ,
1622,3,5,6 , (21 , 24 )2 , (23 , 26 )2 , 18 ,
(82,3,6 , 82,4,5 )2 , 422,5 , (21 , 24 )2 , 18 ,
This completes the proof of the proposition.
60
(81,3,6 , 83,4,6 )2 , 421,4 , (21 , 24 )2 , 18
We now can assume that k contains a splitting field for H and, moreover, that NḠ (H) = HCḠ (H),
as the composition factors of Vmin ↓H , which is stable under Out(G), are not compatible with an outer
automorphism of H.
Proposition 10.3 Suppose that Vmin ↓H has at least one 2-dimensional composition factor and no trivial
composition factors.
(i) If a = 3 then H fixes a 2-space on Vmin .
(ii) If a = 4, 5, 6 then H fixes a 2-space on Vmin or is a blueprint for Vmin .
Proof: Suppose that a = 3. Any 8s split off, so we just consider the 4s and 2s. The projective cover of
4i,i+1 is
P (4i,i+1 ) = 4i,i+1 /2i+1 /1/2i−1 /1/2i+1 /4i,i+1 ,
and from this we see that no module can have a 2-dimensional composition factor, no trivial composition
factor, and no have a 2-dimensional submodule or quotient. This proves (i).
For a = 4, we first compute the conspicuous sets of composition factors, finding eighty-one sets up to
field automorphism of H. We have a list of those classes of elements of order 17 that are blueprints for Vmin ,
and all but fifteen of these sets appear on that list. We can also compute which have positive 2i -pressure (or
no 2i ) for every i, and find that only eighteen of these sets do. The intersection of these two short lists has
just two sets of composition factors on it, and so we consider these two:
81,2,3 , 421,2 , 431,3 , 432,3 , 42,4 , 221 , 222 , 223 ,
81,2,4 , 82,3,4 , 421,2 , 41,3 , 421,4 , 432,3 , 221 , 223 .
The first of these must stabilize a 2-space on Vmin , as notice that otherwise the socle can consist of summands
of Vmin ↓H and a submodule of 41,2 ⊕ 41,3 ⊕ 42,3 , but the largest submodules of those projectives with
composition factors those in Vmin ↓H are
42,4 /22 /41,2 ,
41,3 /23 /42,3 /21 , 23 /41,3 ,
23 /41,3 , 42,3 /23 , 81,2,3 /42,3 ,
so at most a single 21 can lie in Vmin ↓H , a contradiction.
The second case is even easier, given that the corresponding submodules are
41,2 /81,2,4 /41,2 ,
41,3 /21 /41,4 ,
41,3 /23 /42,3 .
This completes the proof for a = 4.
We now let a = 5. There are thirty possible multisets of dimensions for the composition factors of Vmin ↓H
that have at least one 2, no 1s, and have the right trace of an element of order 3. If H does not fix a 2-space,
then we need two 4s in the dimensions, removing twenty multisets of dimensions from the list. We can also
apply Lemma 5.6, which shows that if there are no 8s in Vmin ↓H then we need at least as many 4s as 2s,
removing another three. Since any 4 has 2-pressure at most 2, there needs to be more than half as many 4s
as 2s in all cases; this brings us down to ten. These are
410 , 28 ,
162 , 44 , 24 ,
8, 49 , 26 ,
83 , 47 , 22 ,
16, 47 , 26 ,
82 , 48 , 24 ,
16, 82 , 45 , 22 ,
16, 8, 46, 24 ,
32, 45 , 22 ,
162 , 8, 43 , 22 .
In these cases we switch to proving that H is a blueprint for Vmin . (This could be done for the other cases
but the amount of extra work is significant and so this has not been done.)
61
We give a table listing the total number of possible sets of composition factors for Vmin ↓H , then those
that are conspicuous, and finally those for which an element of order 31 in H is a blueprint for Vmin . These
numbers are all up to a field automorphism of Vmin .
Case
10
4 ,2
8
8, 49 , 26
7
6
8
4
16, 4 , 2
2
8 ,4 ,2
6
16, 8, 4 , 2
2
4
16 , 4 , 2
3
7
8 ,4 ,2
2
4
4
2
5
16, 8 , 4 , 2
5
32, 4 , 2
2
3
2
2
16 , 8, 4 , 2
2
Number
Conspicuous
31 is blueprint
9145422
23
23
20420400
32
32
2402400
3
2
18718700
52
51
3503500
12
8
150150
2
2
7550400
22
21
1651650
20
19
6006
0
0
99000
4
4
We now focus on these remaining eight conspicuous sets of composition factors:
161,2,3,4, 42,5 , 41,4 , 422,4 , 41,3 , 41,5 , 42,3 , 223 , 222 , 221 ,
161,3,4,5 , 82,4,5 , 42,5 , 43,5 , 41,5 , 424,5 , 42,3 , 23 , 231 ,
82,3,5 , 81,2,5 , 41,4 , 423,5 , 41,3 , 421,5 , 422,3 , 223 , 221 ,
161,3,4,5 , 82,4,5 , 41,4 , 43,5 , 41,3 , 424,5 , 42,3 , 223 , 221 ,
161,2,3,4 , 81,4,5 , 42,5 , 43,5 , 42,4 , 41,5 , 43,4 , 42,3 , 223 , 221 ,
82,4,5 , 81,3,4 , 81,2,5 , 42,4 , 41,5 , 424,5 , 43,4 , 42,3 , 41,2 , 23 , 21 ,
161,2,3,5 , 82,3,4 , 41,4 , 43,5 , 421,3 , 44,5 , 42,3 , 24 , 23 , 221 ,
161,2,4,5, 81,2,4 , 81,2,5 , 42,4 , 41,5 , 43,4 , 42,3 , 41,2 , 22 , 21 .
Recall from Lemma 5.1 that 2i has extensions only with 4i,j for j 6= i, i + 1: from this we see that the third,
fourth, fifth, seventh and eighth have non-positive 21 -pressure so have 21 s as submodules; the first case has
23 -pressure 0 so fixes a 23 submodule; the sixth case has 24 -pressure 0 so fixes a 24 submodule. This leaves
the second case, which will fix a 21 submodule: we quotient out by any simple submodule other than 41,5 ,
41,4 and 41,3 , and take any submodule whose quotient is simple and not one of these, to leave a module
W containing both 21 s and having only 41,5 in the socle, except possibly for 41,4 or 41,3 appearing as a
summand. We therefore consider the largest submodule of P (41,5 ) with composition factors from the factors
of Vmin ↓H , and this is
41,4 /21 /41,5 .
Thus Vmin ↓H must have 21 as a submodule, and we are done for the case a = 5.
Finally, let a = 6. We have the same ten multisets of dimensions of composition factors for Vmin ↓H as
the case of a = 5, and we perform the same analysis as before.
62
Case
10
4 ,2
Number
8
Conspicuous up to 21
Conspicuous up to 63
63 is blueprint
420696342
68
41
41
8, 49 , 26
1258472670
369
76
76
16, 47 , 26
134306100
121
9
9
1410195600
1068
104
104
244188000
750
38
38
7712064
89
12
12
626749200
983
90
90
128200860
1097
80
80
244188
15
1
1
5712000
208
24
24
2
8
8 ,4 ,2
4
6
16, 8, 4 , 2
2
4
16 , 4 , 2
3
7
8 ,4 ,2
2
4
4
2
5
16, 8 , 4 , 2
5
32, 4 , 2
2
2
2
3
16 , 8, 4 , 2
2
As every conspicuous set of composition factors for Vmin ↓H has an element of order 63 that is a blueprint
for Vmin , H is always a blueprint for Vmin , as needed.
Proposition 10.4 Suppose that Vmin ↓H has either two or four trivial composition factors.
(i) If a = 3 then Vmin ↓H has a submodule of dimension at most 2.
(ii) If a = 4 then either H fixes a line on L(G)′ or NḠ (H) fixes a σ-stable, proper, non-zero subspace of
Vmin whose stabilizer is positive dimensional.
(iii) If a = 5 then NḠ (H) fixes a σ-stable, proper, non-zero subspace of Vmin whose stabilizer is positive
dimensional.
(iv) If a = 6 then NḠ (H) is either a blueprint for Vmin or fixes a subspace of dimension at most 2 of Vmin .
Proof: Let a = 3. As we have seen before, the projective cover of 4i,i+1 is
P (4i,i+1 ) = 4i,i+1 /2i+1 /1/2i−1 /1/2i+1 /4i,i+1 ,
whence if Vmin ↓H has no 1- or 2-dimensional submodules or quotients, it is a sum of 8s and P (4i,i+1 )s for
various i. In particular, since dim(P (4i,i+1 ) = 16, we have one of P (4), 85 and P (4)2 , 83 , as there are between
two and four trivial factors. Thus the composition factors of Vmin ↓H are either 85 , 42 , 23 , 12 or 83 , 44 , 26 , 14 ,
on which an element of order 3 acts with trace −4 and −1 respectively, not a trace of an element of order 3
on Vmin , which is one of −25, −7, 2, 20. This completes the proof for a = 3.
Let a = 4. Using all semisimple elements, there are (up to field automorphism) 113 conspicuous sets
of composition factors for Vmin ↓H with exactly two trivial composition factors. Only seventy-nine of these
have corresponding factors on L(G), and of these only thirty-eight have either no 2i or positive 2i -pressure
for every i. One can eliminate three more as an element of order 17 is a blueprint for Vmin , leaving us with
thirty-five. Two more of these have no 4-dimensional factors appearing with multiplicity greater than 1, so
must stabilize either a line or 2-space as Vmin is self-dual.
We are left with thirty-three conspicuous sets of composition factors for Vmin ↓H , still too many to list.
Let W be the subquotient obtained from Vmin ↓H by quotienting out by the {8, 16}-radical and taking the
{8, 16}-residual, and remove any 4-dimensional simple summands. Since H can be assumed not to fix a line
or 2-space on Vmin , the socle of W consists of 4-dimensional modules, and the factors of soc(W ) consist
of 4-dimensional simple modules that occur with multiplicity at least 2 in Vmin ↓H , and hence W . Let
63
S1 , . . . , Sr be the 4-dimensional simple modules that appear in Vmin ↓H with multiplicity at least 2. (If a
module appears more than twice, we take the floor of half of its multiplicity, since this is the maximum
number of times it may appear in the socle.)
We construct the largest submodule W ′ of P (S1 ⊕ · · · ⊕ Sr ) that consists solely of composition factors
from Vmin ↓H ; certainly W 6 W ′ . Thus W ′ must have at least two trivial factors, and all the requisite
2-dimensional factors.
In fact, only twenty-two out of the thirty-three cases yield modules W ′ with any trivial factors, with
five even being the zero module (as there are no such Si ). Another six can be removed for not having the
correct 2-dimensional factors, leaving sixteen sets of composition factors. Five of these have corresponding
set of composition factors on L(G)′ having pressure less than 6, hence H fixes a line on L(G)′ by Lemma
5.3 and the fact that an involution has at least six trivial Jordan blocks on L(G)′ . The remaining eleven are
as follows:
821,3,4 , 41,2 , 421,3 , 441,4 , 241 , 22 , 12 ,
81,3,4 , 421,3 , 431,4 , 422,3 , 42,4 , 231 , 22 , 223 , 24 , 12 ,
81,3,4 , 41,2 , 41,3 , 41,4 , 42,3 , 422,4 , 423,4 , 231 , 22 , 23 , 224 , 12 ,
81,2,4 , 421,2 , 421,3 , 41,4 , 42,3 , 42,4 , 43,4 , 221 , 222 , 223 , 24 , 12 ,
82,3,4 , 41,2 , 421,3 , 41,4 , 422,3 , 422,4 , 221 , 222 , 223 , 24 , 12 ,
16, 41,2, 41,3 , 41,4 , 422,4 , 43,4 , 221 , 222 , 223 , 24 , 12 ,
81,2,3 , 81,2,4 , 421,2 , 421,3 , 421,4 , 42,3 , 221 , 22 , 223 , 12 ,
821,3,4 , 41,2 , 421,3 , 431,4 , 42,3 , 221 , 22 , 223 , 12 ,
81,2,4 , 81,3,4 , 41,2 , 41,3 , 421,4 , 422,3 , 42,4 , 221 , 22 , 23 , 24 , 12 ,
821,2,4 , 81,3,4 , 431,2 , 41,3 , 421,4 , 221 , 22 , 12 ,
16, 81,2,3, 81,3,4 , 41,2 , 41,3 , 421,4 , 221 , 22 , 12 .
We can eliminate some more using module structures: in the second case, suppose that 41,4 lies in the
socle. If it is a summand, we quotient it out and ignore it, so suppose it is a submodule but not a summand,
and quotient this out, also quotienting out any 2- and 1-dimensional factors that become submodules to
produce a module U . This is a submodule of P (41,4 ) and so we use Lemma 5.7, seeing that U is a submodule
of
21 /1/22/1/21 /41,4 ;
if both trivials are in U then Vmin ↓H /U has 23 -pressure 0, so has 23 as a submodule, a contradiction from the
definition of U . If there is a single trivial in U then firstly replace U by the 7-dimensional submodule 1/21 /41,4
of U , and since Vmin is self-dual, there is a (unique) corresponding submodule U ′ such that Vmin ↓H /U ′ ∼
= U ∗.
If U 6 U ′ then U ′ /U has no trivials and again it has 23 -pressure 0 and so we get a contradiction. Thus
U ′ does not contain U , and we claim that in this case an involution u must act with exactly two trivial
Jordan blocks, not allowed by [13, Table 7]. To see this, firstly let M denote the {1}′ -residual modulo the
{1}′ -radical of Vmin ↓H , so it is a submodule of P (1), as otherwise it is simply 1⊕2 , with this impossible by
[13, Table 7]. The submodule U of Vmin ↓H has image inside M which is just soc(M ), and the image of U ′
has image inside M which is simply rad(M ). It is therefore clear that the image of U ′ contains the image of
U and, since U is uniserial, U ′ contains U .
Thus we can remove any 41,4 in the socle and top, perhaps remove two 21 s that are now in the socle and
top, and assume that the resulting module V ′ is a self-dual submodule of P (41,3 ) ⊕ P (42,3 ).
We now give the three modules obtained from the following procedure, given a socle S that is a submodule
of 41,3 ⊕ 42,3 :
(i) Take the preimage S1 in P (S) of the radical of P (S)/S corresponding to all composition factors of
Vmin ↓H other than those in S;
64
(ii) Take the preimage S2 in P (S) of the cf(S)-radical of the quotient P (S)/S1 ;
(iii) Take the cf(S)′ -residual S3 of S2 .
This must contain the module V ′ , so we examine the composition factors of the modules S3 for the choices
of S, which are
41,3 , 41,3 /21 , 23 /41,4 , 42,3 /21 , 23 /41,3 ,
42,3 /23 /1/24 /1/23/42,3 ,
42,3 /23 /1, 41,3/21 , 24 /1, 41,3 , 41,4 , 42,3 /21 , 23 , 23 /41,3 , 42,3 .
None of these has a 22 as a composition factor, and this yields a contradiction. Thus in the second case H
must fix a line or 2-space on Vmin .
In the third case, W ′ might have enough 2-dimensional factors, but in order to have three 21 s in W ′ we
need both 42,4 and 43,4 in the socle, whence they cannot appear elsewhere in the module (which they can
do in our construction of W ′ ). With this restriction, that 42,4 and 43,4 can only appear in the socle and top
of Vmin ↓H , we get the analogue of W ′ to be
43,4 /23 , 24 /1, 1, 41,3, 42,4 /21 , 21 , 22 , 23 /1, 1, 41,2, 42,4 , 43,4 /22 , 24 , 24 /42,4 , 43,4 ,
which still does not have three 21 s in it, a contradiction.
In the fourth case, the socle of W cannot simply be 41,2 as there are no 21 s in its contribution to W ′ . If
it is 41,2 ⊕ 41,3 then, arguing as in the previous case, we get
41,2 /22 /1, 41,3 , 41,3 /21 , 23 , 23 /1, 41,2, 41,4 , 42,3 /21 , 22 , 23 , 81,2,4 /41,2 , 41,3 ,
and if it is just 41,3 then we get
41,3 , 41,3 /21 , 23 /41,4 , 42,3 /21 , 23 /41,3 ,
so in neither case can we contain all of W .
In the sixth case, the socle of W must be 42,4 , and if so no 42,4 can appear outside of the socle and top
of W . Taking the radical of P (42,4 )/42,4 with factors all other composition factors of Vmin ↓H , then adding
on as many 42,4 s on top of that, then taking the {42,4 }′ -residual of this (since the socle of W must be 42,4 ),
we end up with
42,4 , 42,4 /22 , 24 /41,2 , 43,4 /22 , 24 /42,4 ,
clearly wrong. Thus in the sixth case H fixes a 1- or 2-space on Vmin .
In the ninth case, the possible factors of soc(W ) are 41,4 and 42,3 , with both required for all of the
2-dimensional factors to be present, as an examination of W ′ proves. In this case, we do as above to find
that W is a submodule of
41,4 /21 /1, 41,4 /21 , 22 /1, 1, 41,3, 41,4 /21 , 23 , 81,3,4 /41,4 , 42,3 ,
which does not have a 24 , a contradiction.
For the other cases, the existence of the uniserial modules M1 = 41,4 /21 /1/22/1/21/41,4 and M2 =
41,3 /21 /41,4 /21 /41,3 prove directly that the first (M1 plus M2 ), fifth (M1 twisted by the square of the field
automorphism, M2 untwisted, and M2 twisted by the field), seventh (M1 twisted by the field, plus M2 ),
eighth (M1 plus M2 twisted by the field squared), tenth and eleventh cases (both a single M1 ) cannot be
solved in the same way.
65
For these we will show that H fixes a (σ-stable) subspace whose stabilizer is positive dimensional, since
it will contain an element of order 85 > v(E7 ). In fact, five of the six satisfy an extra property that we will
use later on: an element x of order 17 has seventeen distinct eigenvalues on Vmin , and has a preimage x̂ of
order 85 that has eighteen eigenvalues on Vmin . This means that x̂ must have the same eigenspaces as x,
except that the 1-eigenspace of x must split into two. Since only the trivial module has a 1-eigenspace for
the action of x, this means that every submodule of Vmin ↓H that does not contain a trivial module must be
stabilized by x̂ and hence by a positive-dimensional subgroup of G.
We are left with a single case to consider, which is
821,3,4 , 41,2 , 421,3 , 431,4 , 42,3 , 221 , 22 , 223 , 12 ;
if we choose x of order 17 and ζ a primitive 17th root of unity so that x acts on 21 with eigenvalues ζ ±1 ,
then x acts on Vmin with eigenvalues
12 , (ζ ±1 )3 , (ζ ±2 )2 , (ζ ±3 )5 , (ζ ±4 )4 , (ζ ±5 )4 , (ζ ±6 )3 , (ζ ±7 )3 , (ζ ±8 )3 ,
and there is an element x̂ of order 85 in G that powers to x and has the same eigenspaces, except it splits
the ζ ±1 and 1-eigenspaces, so has twenty eigenvalues on Vmin . An easy calculation shows that the only
composition factors of Vmin on which x has 1 or ζ ±1 as an eigenvalue are 1, 21 and 41,2 : if 1 or 21 is a
submodule of Vmin ↓H then its stabilizer is positive dimensional anyway, and if 41,2 is a submodule then
it is a summand, so there must be another factor in the socle, which therefore has a positive-dimensional
stabilizer. This completes the proof of the proposition in the case where there are exactly two trivial factors
in Vmin ↓H .
Moving to exactly four trivial composition factors, up to field automorphism there are 114 conspicuous
sets of composition factors for Vmin ↓H . Twenty-two of these contain an element of order 17 that is a blueprint
for Vmin , and twenty-five have no corresponding sets of composition factors for L(G). Taken together, this
leaves seventy sets of composition factors. Removing those with non-positive 1- or 2i -pressure and those
without two isomorphic 4-dimensional composition factors brings us down to fifty-one sets of factors.
As with the previous case, we construct the modules W and W ′ and apply the same test, reducing us to
twenty-three sets of composition factors. Another three have pressure less than 6 on L(G)′ , so fix a line on
L(G)′ and can be discarded.
As we saw when considering two trivial factors, construction of the module W ′ does not take into account
that if a 4-dimensional factor lies in the socle of W and has multiplicity exactly 2 in Vmin ↓H then it cannot
appear anywhere other than the socle or the top of W . Including this, and ranging over all possible socles
rather than just the largest one, yields a collection of modules for each case, all smaller than the original
W ′ , and another twelve that no longer have enough 1- or 2-dimensional factors, bringing us down to eight.
The last eight cases are as follows:
431,3 , 41,4 , 432,3 , 42,4 , 241 , 222 , 233 , 24 , 14 ,
41,2 , 431,3 , 421,4 , 42,3 , 42,4 , 241 , 222 , 233 , 24 , 14 ,
81,3,4 , 421,3 , 421,4 , 42,3 , 423,4 , 231 , 22 , 223 , 224 , 14 ,
81,3,4 , 82,3,4 , 431,4 , 42,4 , 423,4 , 231 , 22 , 224 , 14 ,
81,2,3 , 81,3,4 , 41,2 , 41,3 , 421,4 , 422,3 , 221 , 22 , 223 , 24 , 14 ,
66
81,3,4 , 41,2 , 421,3 , 431,4 , 42,3 , 241 , 222 , 223 , 14 ,
821,3,4 , 431,4 , 42,4 , 423,4 , 231 , 22 , 224 , 14 ,
821,3,4 , 41,3 , 431,4 , 42,3 , 43,4 , 231 , 222 , 23 , 14 .
In the first case, the socle of W can be either 41,3 or 41,3 ⊕ 42,3. If the socle of W is 41,3 then the module
′
W in which W can be found is
41,3
21 23
1 41,4 42,3
21 22 23 24
1 1 41,3 41,3 42,4
21 22 23 24
1 41,4 42,3
21 23
41,3
This is self-dual, so has a simple top, and since it is 64-dimensional, W must be contained in rad(W ′ ), and
indeed in the {41,3 , 42,3 }′ -residual of this, which is
42,3 /23 /1, 41,3 , 41,3 /21 , 23 , 24 /1, 41,4, 42,3 /21 , 23 /41,3 ,
which has no 22 , so 41,3 cannot be the socle. If 41,3 ⊕ 42,3 is the socle, then the module W ′ is the sum of the
one above and
41,3 , 42,3 /21 , 23 /1, 41,4 /21 , 24 /1, 41,3/23 /42,3 ,
which also has no 22 . The same statement about the top 41,3 not appearing in W remains true, and so we
take the same residual (this is why we took the {41,3 , 42,3 }′ -residual rather than the {41,3 }′ -residual above)
and see no 22 again. Thus H must fix a 1- or 2-space on Vmin .
In the second case, the socle of W ′ must be 41,3 , and indeed W ′ is the same module as in the previous
case, so the same method works there.
In the third case, the socle of W ′ must be 41,3 ⊕ 41,4 ⊕ 43,4 . We can construct such a module, namely
41,4
43,4
41,3
21
24
23
1
1
42,3 ⊕
⊕
22
21
23
1
1
41,3
21
24
41,4
43,4
⊕ 81,3,4 ,
so we will need to look at elements of G to solve this case.
In the fourth case, the module W ′ is the self-dual module
41,4
21 81,3,4
1 41,3 41,4
21 22 23
1 1 42,3
21 22 23
1 41,3 41,4
21 81,3,4
41,4
67
which has two 81,3,4 s, so we can as in the first two cases take the {41,4 }′ -residual of rad(W ′ ) to get a module
41,4 /21 /1/22/1, 41,4/21 , 81,3,4 /41,4 ,
which cannot work for several reasons, so that H fixes a line or 2-space on Vmin . The exact same module
appears as W ′ in the eighth case as well, so this method works there.
In the fifth and sixth cases, W ′ must have socle 41,4 ⊕ 43,4 and we can construct a module with the right
composition factors, namely
41,4
43,4
21
24
1
1
22
⊕
21
1
1
21
24
41,4
43,4
with the remaining factors being summands.
In the seventh case, W ′ must have 41,4 ⊕42,3 in the socle, and a module with the right composition factors
has the first summand above and the second summand twisted by three iterations of the field automorphism,
so
41,4
42,3
21
23
1
1
22
⊕
24
1
1
21
23
41,4
42,3
We have therefore eliminated the first, second, fourth and eighth cases, and will look at semisimple
elements in the third, fifth, sixth and seventh cases.
In the third case, the element x acts on Vmin with eigenvalues
14 , (ζ ±1 )3 , (ζ ±2 )2 , (ζ ±3 )3 , (ζ ±4 )5 , (ζ ±5 )5 , (ζ ±6 )2 , (ζ ±7 )2 , (ζ ±8 )4 ,
and there exists an element of order 85 that powers to x and has nineteen distinct eigenvalues on Vmin , only
splitting the ζ ±2 -eigenspaces. In Vmin ↓H , these lie in the 22 and 42,3 , the latter of which can only lie in
the socle if it is a summand. Therefore every other simple submodule is preserved by an element of order
85 > v(E7 ) and therefore a positive-dimensional subgroup of G.
We do the same thing in the fifth case, finding an element of order 85 that powers to x and only disturbs
the 1- and ζ ±1 -eigenspaces. Since these only lie in the trivial and 21 , every simple submodule of Vmin ↓H is
stabilized by a positive-dimensional subgroup of G.
For the sixth case, there are eight elements of order 85 that power to x and have nineteen eigenvalues
on Vmin : four split the ζ ±6 -eigenspace and the other four split the ζ ±8 -eigenspace. The ζ ±6 -eigenspace is
contributed to by 42,4 and 81,3,4 from Vmin ↓H , and the ζ ±8 -eigenspace is contributed to by 24 and 41,4 .
Thus any simple submodule of Vmin ↓H is stabilized by at least one element of order 85, and so the stabilizer
is positive dimensional, as claimed.
Finally, we have the seventh case. Here, the smallest number of eigenvalues that an element of order 85
powering to x has on Vmin is twenty-three, but there is one that splits the ζ ±1 , ζ ±2 and ζ ±5 -eigenspaces. The
68
element x has eigenvalues ζ ±7 , ζ ±8 on 41,4 , which is in the socle of Vmin ↓H as we saw above. (We actually
saw that it was in the socle of W , but all modules that appear with multiplicity greater than 1 are either in
the socle of W or have dimension at most 2, so that soc(W ) 6 soc(Vmin ↓H ) or H fixes a 1- or 2-space on
Vmin .) Thus we see that an element of order 85, and hence a positive-dimensional subgroup of G, stabilizes
the (σ-stable) submodule 41,4 of Vmin ↓H , completing the case for a = 4.
Let a = 5, and firstly suppose that there are two trivial composition factors, so we need at least three
2s and two 4s, to avoid fixing a line or 2-space on Vmin . There are seventeen possible sets of dimensions of
composition factors with these properties that also have the correct trace of an element of order 3: if there
are exactly two 4s in Vmin ↓H then we can use Lemma 5.7 to see that we can have exactly three 2s, thus
eliminating two of these cases, and if there are three 4s we can have at most eight 2s, eliminating two more.
We now give a table listing the possible sets of dimensions, together with the number of sets of composition
factors (up to field automorphism) with those dimensions, those that are conspicuous, those for which the
element x of order 31 is a blueprint for Vmin , and those for which there exists an element x̂ of order 93,
cubing to x, and such that x̂ has one more distinct eigenvalue than x. This last condition does not ensure
that H is a blueprint for Vmin , but does show that H lies inside a positive-dimensional subgroup stabilizing
every simple submodule of soc(Vmin ↓H ) not of dimension 1 or 32.
To see this, if x̂ has one more eigenvalue than x then, since x̂ must be real as it lies in E7 , all eigenspaces
are preserved except for the 1-eigenspace. As only the trivial and 32-dimensional have 1 as an eigenvalue
for x, this means that x̂, and a positive-dimensional subgroup of G stabilizing the same subspaces as x̂, fix
any simple submodule of Vmin ↓H not of dimension 1 or 32. In particular, this proves that H lies inside a
member of X σ .
Case
46 , 215 , 12
5
13
8, 4 , 2 , 1
2
4
2
11
8 ,4 ,2 ,1
9
9
4 ,2 ,1
8
2
7
8, 4 , 2 , 1
2
6
7
2
7
5
2
16, 4 , 2 , 1
2
2
8 ,4 ,2 ,1
5
5
16, 8, 4 , 2 , 1
2
3
5
16 , 4 , 2 , 1
2
2
Number
Conspicuous
31 is blueprint
One more eigenvalue
3879876
5
4
0
9529520
2
2
0
10735725
13
12
1
6952660
16
12
0
16044600
30
23
0
1651650
9
3
1
15855840
54
29
12
2522520
24
10
3
83160
6
6
0
83 , 46 , 23 , 12
7707700
22
5
9
16, 82 , 44 , 23 , 12
1376375
19
14
3
5005
1
0
0
57750
3
1
1
4
3
32, 4 , 2 , 1
2
2
3
2
16 , 8, 4 , 2 , 1
2
Excluding both those that are blueprints and where there is an element with one more eigenvalue, we are left
with forty-eight conspicuous sets of composition factors. Twenty of these forty-eight have no corresponding
set of composition factors on L(G)′ , so cannot yield embeddings of H into G.
We are left with twenty-eight conspicuous sets of composition factors for Vmin ↓H , still too many to list.
Let W be the subquotient obtained from Vmin ↓H by quotienting out by the {8, 16, 32}-radical and taking
the {8, 16, 32}-residual, and remove any 4-dimensional simple summands. Since H can be assumed not to
fix a line or 2-space on Vmin , the socle of W consists of 4-dimensional modules, and the factors of soc(W )
69
consist of 4-dimensional simple modules that occur with multiplicity at least 2 in Vmin ↓H , and hence W .
Let S1 , . . . , Sr be the 4-dimensional simple modules that appear in Vmin ↓H with multiplicity at least 2.
(Note that no composition factor of Vmin ↓H , in the twenty-eight remaining sets of factors, appears with
multiplicity greater than 3, so we need only one copy of each Si .)
We construct the largest submodule W ′ of P (S1 ⊕ · · · ⊕ Sr ) that consists solely of composition factors
from Vmin ↓H ; certainly W 6 W ′ . Thus W ′ must have at least two trivial factors, and all the requisite
2-dimensional factors. In fact, only nine out of the twenty-eight cases yield modules W ′ with any trivial
factors, with ten even being the zero module (as there are no such Si ). Another seven can be removed for
not having the correct 2-dimensional factors, leaving the following two sets of factors:
161,3,4,5 , 81,3,4 , 81,4,5 , 41,4 , 421,5 , 41,2 , 22 , 221 , 12 .
81,3,5 , 81,4,5 , 41,4 , 421,3 , 421,5 , 42,3 , 41,2 , 223 , 22 , 221 , 12 ,
In these final two cases we need to consider a preimage x̂ that does not stabilize all eigenspaces, but does
stabilize those that make up some submodule of Vmin ↓H . In the first case, x has thirty-one eigenvalues on
Vmin , and the fewest number of eigenvalues for a preimage x̂ of order 93 is thirty-five (two such preimages,
each a power of the other), with the four eigenvalues of x not being stabilized being ζ ±14 , ζ ±15 where ζ is a
primitive 31st root of unity.
In the second case, x again has thirty-one eigenvalues and the fewest number of eigenvalues for x̂ is thirtyfour (four preimages, yielding two subgroups of order 93), with the three eigenvalues of x not stabilized being
either 1, ζ ±2 or 1, ζ ±3 , depending on the choice of preimage.
The eigenvalues of x on 41,2 are ζ ±1 , ζ ±3 , so if this is a submodule (hence summand) of Vmin ↓H then
that submodule is stabilized by a positive-dimensional subgroup of G, as needed. There are no extensions
between 41,2 and any of 41,3 , 41,5 , 21 or 23 , and so since all other composition factors are multiplicity free,
and 41,2 is multiplicity free, there can be no extensions between 41,2 and any other composition factor, as
Vmin is self-dual. Thus 41,2 splits off in both cases, and our result is proved for two trivial factors.
If there are four trivial composition factors in Vmin ↓H , then there are twenty-eight possible sets of
dimensions for the factors of Vmin ↓H that have a good trace of an element of order 3, and we exclude those
that do not have at least five 2s – bringing us down to sixteen sets – and those that do not have three 4s as
needed by Proposition 5.7. We apply this lemma again to see that we cannot have too many 2s per 4, and
this brings us down to six possible sets of dimensions, given in the table below.
Case
5
16
4
8
10
4
4 ,2 ,1
4 ,2 ,1
Number
Conspicuous
31 is blueprint
1939938
3
2
4866862
14
14
8, 47 , 28 , 14
11325600
30
28
16, 45 , 28 , 14
990990
7
3
11561550
45
45
1501500
19
18
2
6
6
8 ,4 ,2 ,1
4
6
4
16, 8, 4 , 2 , 1
4
This leaves just six sets of composition factors that are not guaranteed to be blueprints for Vmin ↓H .
These are
81,4,5 , 423,5 , 421,3 , 421,5 , 42,3 , 224 , 223 , 222 , 221 , 14 ,
161,2,3,4 , 41,4 , 41,3 , 41,5 , 423,4 , 25 , 224 , 23 , 222 , 221 , 14 ,
161,2,4,5 , 42,5 , 43,5 , 422,4 , 43,4 , 225 , 224 , 22 , 231 , 14 ,
161,2,3,5, 422,5 , 421,3 , 42,3 , 224 , 223 , 222 , 221 , 14 ,
161,3,4,5 , 41,4 , 41,5 , 424,5 , 42,3 , 225 , 223 , 22 , 231 , 14 ,
161,2,3,4, 81,3,4 , 41,4 , 42,4 , 43,4 , 42,3 , 25 , 23 , 222 , 221 , 14 .
70
They all have corresponding sets of composition factors on L(G), but the easiest way to eliminate them is
to consider the modules W and W ′ from the case of two trivial factors: in each of the six cases, we have at
most two trivial factors in W ′ , and so H must always fix a line or 2-space on Vmin , as needed.
When a = 6, we have exactly the same possible dimensions for composition factors for Vmin ↓H as for
a = 5. The traces of semisimple elements of order up to 21 are known, but not 63 or 65, so we can check
if a set of composition factors are conspicuous up to 21. Letting x be an element of order 63 in H, we note
that we have a list of all semisimple elements of order 21 in G, but not 63, so we use the preimage trick
from Section 4.2 firstly to see if the composition factors are conspicuous up to 63, and then use the preimage
trick again to see if there exists an element x̂ of order 63 · 5 = 195 with the same eigenspaces as x and with
x̂5 = x. In every case, we find that the element of order 63 is a blueprint for Vmin .
Case
6
15
4 ,2 ,1
5
2
13
8, 4 , 2 , 1
2
4
2
11
8 ,4 ,2 ,1
2
49 , 29 , 12
8, 48 , 27 , 12
6
7
2
7
5
2
16, 4 , 2 , 1
2
8 ,4 ,2 ,1
5
5
16, 8, 4 , 2 , 1
2
3
5
16 , 4 , 2 , 1
3
6
3
8 ,4 ,2 ,1
2
4
2
2
2
3
16, 8 , 4 , 2 , 1
2
32, 44 , 23 , 12
2
2
3
16 , 8, 4 , 2 , 1
2
Case
5
16
4
8
10
4
4 ,2 ,1
4 ,2 ,1
Number
Conspicuous up to 21
Conspicuous up to 63
63 is blueprint
100155870
6
6
6
332095680
22
3
3
467812800
60
18
18
272669110
164
21
21
844192800
1201
40
40
76744800
254
16
16
1025589600
3079
93
93
146512800
1203
59
59
3427200
53
20
20
557110500
2665
54
54
89964000
996
63
63
171360
14
5
5
2688000
58
18
18
Number
Conspicuous up to 21
Conspicuous up to 63
63 is blueprint
39437442
4
4
4
160048350
170
19
19
8, 47 , 28 , 14
498841200
792
47
47
16, 45 , 28 , 14
37414170
61
12
12
626754246
1484
85
85
70686000
146
37
37
2
6
6
8 ,4 ,2 ,1
4
6
4
16, 8, 4 , 2 , 1
4
This completes the proof for 3 6 a 6 6, as needed.
We are left with H having at least six trivial composition factors, where by the remarks at the start of
this subsection we noted that if a = 5, 6 then H is always a blueprint for Vmin .
Proposition 10.5 Suppose that a > 3 and Vmin ↓H has at least six trivial composition factors.
(i) If a = 3 then Vmin ↓H has a 1- or 2-dimensional submodule or Vmin ↓H is
8 ⊕ P (41,2 ) ⊕ P (42,3 ) ⊕ P (41,3 ).
(ii) If a = 4 then H is a blueprint for Vmin or H fixes a subspace of dimension at most 2 on Vmin .
71
(iii) If a > 5 then H is a blueprint for Vmin .
Proof: (iii) follows from the remarks above. For (i) we use the proof of the previous proposition to note
that the only possibility is that Vmin ↓H is the sum of three P (4)s and an 8, so we consider the ten possible
such modules, and note that only one has a conspicuous set of composition factors for Vmin ↓H , the one
mentioned.
We are left with a = 4. Here we use only non-blueprint elements of order 17 to restrict the number of
possibilities. We also assume that Vmin ↓H has positive pressure, else H fixes a line on Vmin , and has at least
two 4s, else it would fix a 2-space on Vmin .
Remove any 8s and 16s in the top and socle of Vmin ↓H , together with any simple summands of dimension
4, leaving a self-dual module W whose top and socle consist of 4-dimensional modules, with W having all
trivial factors in Vmin ↓H .
The projectives P (41,2 ) and P (41,3 ) both have exactly four trivial composition factors, and have dimension
64. Therefore we cannot have the whole projective, so remove the simple top, then any 1-, 2- and 8dimensional modules from the top to find the following modules:
41,2 , 42,4 /22 , 24 /1, 43,4 /23 , 24 /1, 41,2, 42,4 /22 , 81,2,4 /41,2 ;
41,4 , 42,3 /21 , 23 /1, 1, 41,3, 41,3 , 42,4 /21 , 22 , 23 , 24 /1, 41,4, 42,3 /21 , 23 /41,3 .
From this we see that we need at least two 4s in the socle, and can have two only if they are both 41,3 or
42,4 . This means that we need either three different 4s appearing in Vmin ↓H with multiplicity at least two,
or 441,3 , 442,4 or 421,3 , 422,4 .
Using the traces of non-blueprint semisimple elements of order 17, and traces of all elements of order 3, 5
and 15, we end up with, up to field automorphism, ten conspicuous sets of composition factors with at least
six trivials, positive pressure, and at least two 4s. These are
421,3 , 421,4 , 422,3 , 241 , 222 , 243 , 224 , 18 ,
431,3 , 41,4 , 432,3 , 241 , 222 , 243 , 24 , 16 ,
421,3 , 421,4 , 42,3 , 423,4 , 241 , 232 , 223 , 224 , 16 ,
81,3,4 , 431,3 , 421,4 , 42,4 , 241 , 22 , 223 , 224 , 16 ,
16, 41,3, 41,4 , 423,4 , 231 , 232 , 23 , 224 , 16 ,
82,3,4 , 421,3 , 421,4 , 423,4 , 241 , 232 , 224 , 16 ,
81,2,4 , 82,3,4 , 431,3 , 422,4 , 241 , 22 , 223 , 16 ,
82,3,4 , 41,2 , 41,3 , 422,3 , 423,4 , 231 , 222 , 223 , 224 , 16 ,
821,3,4 , 421,3 , 41,4 , 42,3 , 42,4 , 231 , 223 , 224 , 16 ,
821,2,3 , 41,2 , 41,3 , 41,4 , 422,3 , 221 , 222 , 223 , 24 , 16 .
By our previous remarks, in all but the first three cases H must fix either a 1-space or a 2-space on Vmin , as
needed. In those three cases, all 4s that appear with multiplicity greater than 1 must appear in the socle of
Vmin ↓H .
The first case we saw before in Proposition 10.2, but we will come back to it. In the second case we
take the preimages of the {1, 2, 42,3}-radicals of P (41,3 )/41,3 , P (41,4 )/41,4 and P (43,4 )/43,4 to produce three
modules in whose direct sum rad(W ) is a submodule, but these are
1/22 , 23 , 24 /1, 42,3 /21 , 23 /41,3 ,
21 /1/22/1/21/41,4 ,
24 /1/21/1/24 /43,4 ,
and rad(W ) cannot have a trivial quotient, so there are only five 1s in W , a contradiction. In the third case
we do the same thing, but with the {1, 21, 22 , 24 , 82,3,4 }-radicals, to get
21 /41,3 ,
21 /1/22/1/21 /41,4 ,
72
24 /1/21 /1/24, 82,3,4 /43,4 ,
and clearly we have a contradiction here. Back to the first case, the appropriate modules here are
22 , 24 /1/21 , 23 /41,3 ,
21 /1/22/1/21 /41,4 ,
so we again must stabilize a 2-space on Vmin .
73
23 /1/24 /1/23/42,3 ,
11
E7 in odd characteristic: PSL2 embedding
In this section, k is a field of characteristic p > 3 and G = E7 (k), by which we mean the simply connected
form, i.e., |Z(G)| = 2 and G′ = G. Let Ḡ be an almost simple group with socle G/Z(G). From [10] we see
that v(E7 ) = 75 for odd integers, so if H is any subgroup of G with a semisimple element of odd order 77
or more, then H is a blueprint for Vmin . In addition, in [9] we prove that PSL2 (9) cannot be a maximal
subgroup of Ḡ either, so here we let H = PSL2 (pa ) with a = 3, 4 if p = 3 and pa 6 150 = 2 · v(E7 ) if p > 5.
Let L = PSL2 (p) 6 H and let u denote a unipotent element of L of order p.
By Proposition 4.10, if a semisimple element x has order at least 31 in G and centralizes a 6-space on
Vmin , then x is a blueprint for Vmin . Since any semisimple element in H has a 1-dimensional 1-eigenspace
on every odd-dimensional simple module, if H has at least six odd-dimensional composition factors on Vmin
and then pa > 60 then H is a blueprint for Vmin . This normally ends up being the case.
11.1
Characteristic 3
Now let H = PSL2 (3a ) for some a > 1. Since v(E7 ) = 75, we assume that a 6 4, and from [9] we assume that
a 6= 2, so a = 3, 4. If a = 4 then we may assume that there are fewer than six odd-dimensional composition
factors in Vmin ↓H , by the discussion at the start of this section.
We begin by computing the composition factors of Vmin ↓L , which depends only on the trace of an
involution, ±8. This means that there are eight more of one factor than the other, so 312 , 120 and 316 , 18 .
From Lemma 5.8 we can see the possible dimensions of composition factors for Vmin ↓H : if Vmin ↓L has
factors 316 , 18 then we must have at least eight 3-dimensional factors in Vmin ↓H , and if the factors are
312 , 120 then as only 9 and 1 for H have more 1s than 3s on restriction to L, we need at least eight of these
in Vmin ↓H , and again have at least eight odd-dimensional composition factors in Vmin ↓H . This gives us the
first proposition.
Proposition 11.1 Let p = 3 and a = 4. A semisimple element of order 41 in H is always a blueprint for
Vmin , and hence H is always a blueprint for Vmin .
We turn to a = 3, where we cannot quite get the same result, but we come close.
Proposition 11.2 Let p = 3 and a = 3. Either H is a blueprint for Vmin or H fixes a line on either Vmin or
L(G).
Proof: As with F4 and E6 , we want to discount conspicuous sets of composition factors where a semisimple
element is a blueprint for Vmin . We already know that there are 97 classes of semisimple elements of order
13 that are blueprints for the minimal module for F4 , and there are 188 classes of semisimple elements of
order 13 in E7 whose 1-eigenspace is at least 8-dimensional, leaving 91 classes to which an element of order
13 in H can belong.
Using this, we find up to field automorphism eight conspicuous sets of composition factors, two of which
have negative pressure so will not be displayed. The other six are
91,3 , 491,3 , 31 , 18 ,
441,2 , 441,3 , 442,3 , 18 ,
932,3 , 451,2 , 41,3 , 15 ,
461,2 , 391 , 32 , 12 ,
41,2 , 451,3 , 42,3 , 391 , 1,
92,3 , 451,2 , 351 , 342 .
The first and third have pressure 1 so fix a line on Vmin by Lemma 5.13. The second case fixes a line on Vmin
by Lemma 5.12.
74
For the fourth, if H does not fix a line on Vmin then we may assume that the socle consists of 41,2 s by
quotienting out any 3s in the socle, and the {1, 31 , 32 , 41,2 }-radical of P (41,2 ) is
41,2 /1, 32 /41,2 ,
but since there is only one 32 in Vmin ↓H we cannot cover both trivials in this way, thus H fixes a line on
10 6 21
Vmin . (Alternatively, the factors of H on L(G) are 416
1,2 , 31 , 32 , 1 , so H fixes a line on L(G).)
The fifth case is 41,2 , 42,3 , 451,3 , 391 , 1, which yields a set of composition factors on L(G) of
10
16
451,2 , 452,3 , 411
1,3 , 31 , 33 , 1 ,
which has pressure 5, and so might not have a trivial submodule or quotient. However, the largest submodule
of P (4i,j ) with these composition factors has three trivial composition factors in all cases, and so we need
at least six 4s in the socle of L(G) ↓H (once we remove all 3s), contradicting the fact that the module has
pressure 5, so we fix a line.
The final case is 91,2 , 451,2 , 351 , 342 . The 9 splits off and we may quotient out by the {31 , 32 }-radical to get
a module with 41,2 s in the socle. On this we can only place 32 s, and so u would act on Vmin as 317 , 15 , not
a valid unipotent action in [13, Table 7], so H cannot embed with these factors.
11.2
Characteristic at least 5
We now let p > 5, let H = PSL2 (pa ) with a > 1, let L = PSL2 (p) 6 H and let u ∈ L have order p. We
begin by producing a list of all unipotent classes to which u can belong, excluding those that come from
generic classes (see Lemma 4.6) and those that fail Lemma 5.15. Moreover, we make a few remarks now
about indecomposable modules for L, which can cut down our list.
When p = 5, 13, 17, we use Corollary 5.18, so for these primes the number of blocks of each even size less
than p − 1 is even. For p = 7, 11, 19, 23, there exists a unique indecomposable module for L of dimension
congruent to a given even number modulo p.
For p = 11, the self-dual module of dimension congruent to 6 modulo p has socle structure
1, 3, 5, 7, 9/1, 3, 5, 7, 9
and has dimension 50. The trace of an involution on the module is 0, and since involutions have trace ±8
on Vmin , we would need a trace of ±8 from the remaining factors of Vmin ↓L , a module of dimension 6, so
not possible. Thus this is not a summand of Vmin ↓L .
For p = 19, the module of dimension congruent to 6 modulo p has socle structure
5, 7, 9, 11, 13, 15/5, 7, 9, 11, 13, 15
and has dimension 120, so cannot be a summand of Vmin ↓L .
For p = 23 the module of dimension congruent to 10 modulo p has socle structure
3, 5, 7, 9, 11, 13, 15, 17, 19, 21/3, 5, 7, 9, 11, 13, 15, 17, 19, 21
and has dimension 240, so cannot be a summand of Vmin ↓L .
We now list the unipotent classes of interest using [13, Table 7].
(i) A3 + A2 , p = 5, acting as 56 , 42 , 34 , 22 , 12 ;
75
(ii) A4 , p = 5, acting as 510 , 16 ;
(iii) A4 + A1 , p = 5, acting as 510 , 22 , 12 ;
(iv) A4 + A2 , p = 5, acting as 510 , 32 ;
(v) (A5 )′′ , p = 7, acting as 72 , 67 ;
(vi) D4 + A1 , p = 7, acting as 76 , 25 , 14 ;
(vii) D5 (a1 ), p = 7, acting as 76 , 32 , 22 , 14 ;
(viii) (A5 )′ , p = 7, acting as 74 , 64 , 14 ;
(ix) A5 + A1 , p = 7, acting as 74 , 63 , 52 ;
(x) D5 (a1 ) + A1 , p = 7, acting as 76 , 4, 25 ;
(xi) D6 (a2 ), p = 7, acting as 76 , 52 , 4;
(xii) E6 (a3 ), p = 7, acting as 76 , 52 , 14 ;
(xiii) E7 (a5 ), p = 7, acting as 76 , 6, 42 ;
(xiv) A6 , p = 7, acting as 78 ;
(xv) D6 , p = 11, acting as 114 , 10, 12;
(xvi) E6 (a1 ), p = 11, acting as 114 , 52 , 12 ;
(xvii) E7 (a3 ), p = 11, acting as 114 , 10, 2;
(xviii) E6 , p = 13, acting as 134 , 14 ;
(xix) E7 , p = 19, acting as 192 , 18.
We start with p = 5, proving that there are always at least six odd-dimensional summands so that we
can assume that a = 1, 2 when doing the hard work.
Proposition 11.3 Let p = 5 and a > 1.
(i) If a = 1, 2 then H fixes a line on either Vmin or L(G).
(ii) If a > 3 then H is a blueprint for Vmin .
Proof: The traces on elements of orders 2 and 3 yield conspicuous sets of composition factors for L =
PSL2 (5) of
312 , 120 ,
52 , 314 , 14 ,
59 , 33 , 12 ,
56 , 36 , 18 .
As we have seen, only P (3) = 3/1, 3/3 has a trivial composition factor and no trivial submodule or quotient,
and so the first, third and fourth cases all fix lines on Vmin . However, in the third case this means that L
cannot embed with these factors: as it fixes a line on Vmin , from Lemma 2.5 we see that L lies in either an
E6 -parabolic, with factors 1, 1, 27, 27∗ or a B5 -subgroup, factors 1, 1, 112, 32, neither of which is compatible
with 59 , 33 , 12 , so this case cannot occur.
76
For the remaining case of 52 , 314 , 14 , we switch to the Lie algebra. There are two possibilities for the
corresponding composition factors of L(G) (since an element of order 3 with trace 2 on Vmin can have trace
either −2 or 7 on L(G)) are
510 , 322 , 117 ,
and
513 , 319 , 111 ,
both of which must have trivial submodules as again we can only cover a 1 by 3/1, 3/3. This proves (i).
Recall from Lemma 5.24 that the only simple modules for H with non-trivial 1-cohomology when a > 2
have dimension 8 and restrict to L as 5 ⊕ 3 by Lemma 5.21. When a = 2, if Vmin ↓L has composition factors
312 , 120 then Vmin ↓H has at least eight trivial composition factors and no factors of dimension 8, so Vmin ↓H
has eight trivial summands. Similarly, if the composition factors of Vmin ↓L are 56 , 36 , 18 then Vmin ↓H must
have at least two trivial composition factors by Lemma 5.21, and for every composition factor of dimension
8 we must have another trivial factor, so Vmin ↓H always has pressure at most −2, so (ii) holds.
We thus may assume that Vmin ↓L has factors 52 , 314 , 14 . We have at most two 8s in Vmin ↓H since there
are only two 5s in Vmin ↓L , and hence there can be at most a single trivial composition factor in Vmin ↓H ,
else H fixes a line on Vmin . We thus get two cases: there is a trivial composition factor and there is not.
If there is a trivial factor we have 82 , 1 in Vmin ↓H , and the remaining factors of Vmin ↓H restrict to L as
312 , 13 , so we need factors 82 , 43 , 39 , 1. For a = 2, there is no such set of composition factors, so we cannot
have a trivial composition factor in Vmin ↓H .
For a = 2 there are up to field automorphism five possible sets of composition factors, which are
521 , 44 , 310
1 ,
82,1 , 51 , 44 , 391 ,
822,1 , 44 , 371 , 32 ,
822,1 , 44 , 361 , 322 ,
1522,1 , 42 , 341 , 322 .
The fact that 4 has an extension with 31 and 32 (see Lemma 5.25) makes deducing the module structure
difficult, and so we turn to the Lie algebra in all cases. These are
5
11 6
832,1 , 510
1 , 41,2 , 31 , 1 ,
11
852,1 , 551 , 461,2 , 310
1 , 32 , 1 ,
93 , 842,1 , 531 , 48 , 361 , 32 , 16 ,
9, 872,1 , 531 , 351 , 332 , 17 ,
1522,1 , 941,2 , 821,2 , 82,1 , 531 , 52 , 431,2 , 321 , 32 , 12 :
each of these has non-positive pressure, as needed. (Remember that 82,1 has 1-cohomology but 81,2 does not
by Lemma 5.24.)
Finally, suppose that a > 3. From Lemma 5.21 we see the following facts: firstly, in any even-dimensional
composition factor of Vmin ↓H there are the same number of 3s as 5s and 1s combined on restriction to L, and
secondly, in any odd-dimensional factor of Vmin ↓H there is at most one more 5 and 1 combined than 3 on
restriction to L. This means that if Vmin ↓L has factors 56 , 36 , 18 , there must be at least six odd-dimensional
composition factors. Lemma 5.21 easily shows that if the factors of Vmin ↓L are 312 , 120 then there must be
at least eight trivial factors in Vmin ↓H , and if we have 52 , 314 , 14 then we have at least six 3s in Vmin ↓H , so
in all cases we have at least six odd-dimensional composition factors, so an element of order (pa ± 1)/2 > 30
has a 1-eigenspace of dimension at least 6. This means that H is a blueprint by Proposition 4.10, as needed
for the proposition.
Having completed p = 5, we now move on to p = 7. This time pa = 7, 49 will need to be considered, but
pa = 343 is above 2 · v(E7 ) = 150.
Proposition 11.4 Suppose that p = 7.
77
(i) If a = 1 then either H fixes a line on Vmin or L(G), or the actions of H on Vmin and L(G) are
7⊕4 ⊕ P (3)⊕2
7⊕5 ⊕ P (5)⊕6 ⊕ P (3).
and
(ii) If a = 2 then either H is a blueprint for Vmin or H fixes a line on Vmin .
Proof: We first compute the possible composition factors of Vmin ↓H when a = 1, using the traces of
elements of orders 2, 3 and 4. There are seven of these, given by
312 , 120 ,
52 , 314 , 14 ,
56 , 36 , 18 ,
7, 59 , 3, 1,
72 , 56 , 32 , 16 ,
74 , 52 , 36 ,
76 , 114 .
As in the case of p = 5, the only indecomposable module with a trivial composition factor but no trivial
submodule or quotient is P (5) = 5/1, 3/5, so we need twice as many 5s as 1s and as many 3s as 1s. Thus all
but the fourth and sixth cases must fix lines on Vmin .
Case 4: If the factors are 7, 59 , 3, 1, then H cannot fix a line or hyperplane on Vmin , since by Lemma 2.5
the line stabilizers for Vmin are contained in either an E6 -parabolic – composition factors 27, 27∗, 12 – or a
subgroup q 1+32 B5 (q) · (q − 1), – composition factors 32, 112, 12 – neither of which can work. As there are no
self-extensions of the 5, Vmin ↓H is
7 ⊕ P (5) ⊕ 5⊕7 ,
with u acting as 73 , 57 , not in [13, Table 7], so there does not exist an embedding of H into G with these
factors.
Case 6: We are left with 74 , 52 , 36 . Here, the lack of trivials means that the possible summands are
3/5 ⊕ 5/3,
3/3
3/3, 5/3,
3/3, 5 ⊕ 3, 5/3,
3, 5/3, 5.
Therefore we need an even number of 1s, 4s and 7s in the Jordan block structure of u, with at least four
more 7s than the 1s, 2s and 4s combined. Examining the list above, we see only two examples of this, namely
(xiii) and (xiv). this yields the two possible embeddings Vmin ↓H to be
7⊕4 ⊕ 3/3 ⊕ 3/3, 5 ⊕ 3, 5/3,
and
7⊕4 ⊕ P (3)⊕2 .
The traces of semisimple elements of orders 3 and 4 on Vmin yield two possibilities each for the semisimple
class, and so we get four possible sets of composition factors for L(G) ↓H , namely
76 , 510 , 310 , 111 ,
78 , 510 , 36 , 19 ,
73 , 513 , 313 , 18 ,
75 , 513 , 39 , 16 .
Apart from the last one, each of these has enough trivials and not enough 5s to ensure that H fixes a line
on L(G), since P (5) = 5/1, 3/5 is the only indecomposable module for H with a trivial factor but no trivial
submodule or quotient.
We now remove the first possible action on Vmin , using the simple fact that for p > 5, the symmetric
square of Vmin is the sum of L(G) and the 1463-dimensional module L(2λ1 ), Lemma 2.1. The symmetric
square of the first module is a sum of projectives and
(3, 5/1, 3, 5 ⊕ 1, 3, 5/3, 5)⊕2 ⊕ 5 ⊕ 3 ⊕ 1.
Since u comes from class E7 (a5 ), and this acts on L(G) as 717 , 5, 33 , we must have two of the summands of
dimension 17 in L(G) ↓H , hence H fixes a line on L(G). This completes the proof for a = 1.
78
Now let a = 2, so that H = PSL2 (49), and recall that L 6 H is a copy of PSL2 (7). At the start of this
proof we gave the conspicuous sets of composition factors for Vmin ↓L , and from Lemmas 5.23 and 5.24 we
see that the only simple modules for H with non-trivial 1-cohomology have dimension 12 and restrict to L
as 7 ⊕ 5, and only the trivial module for H restricts to L with more 1s than 3s. These two facts mean that if
Vmin ↓L has factors the first, third, fifth and seventh cases then Vmin ↓H has trivial composition factors, and
these are summands except for the fifth case, and there we have at least four trivials and at most two 12s, so
pressure at most −2. The fourth case cannot occur, as we proved, so Vmin ↓L has factors either 52 , 314 , 14 ,
and Vmin ↓H can have no trivial factors else it has a trivial summand, or 74 , 52 , 36 .
In the case of 52 , 314 , 14 , from Lemma 5.23, apart from 3, there are no simple modules for H whose
restriction to L has more 3s than other factors, and the composition factors of Vmin ↓H have dimensions 1,
3, 4, 5, 8 and 9. In particular, this means that Vmin ↓H has at least eight 3-dimensional composition factors.
By Lemma 5.26, of these modules only 8s can have an extension with 3s, with there being at most two of
those, so the 3-pressure is at least 6. This means that Vmin ↓H has at least four 3-dimensional summands,
so the action of the unipotent element u on Vmin has at least four Jordan blocks of size 3. There are no
non-generic unipotent classes with this property, as we saw in the list at the start of this section, and so H
is a blueprint for Vmin .
Thus we end with Vmin ↓L being 74 , 52 , 36 , and Lemma 5.23 implies that H has at least two 7s and
four 3s on Vmin . The remaining composition factors have dimension 3, 5, 7, 8, 12 or 15: using the traces of
semisimple elements, we find exactly four conspicuous sets of composition factors for Vmin ↓H with the correct
restriction to L, and for each of these the eigenvalues of an element of order 24 determine its conjugacy class,
and this lies inside F4 , hence a blueprint by Lemma 4.9. Thus H is a blueprint for Vmin , as needed.
Proposition 11.5 Suppose that p = 11.
(i) If a = 1 then either H is a blueprint for Vmin or H fixes a line on Vmin .
(ii) If a = 2 then H is a blueprint for Vmin .
Proof: Let a = 1 firstly, and suppose that H is not a blueprint for Vmin , so in particular the class to which
the unipotent u belongs is non-generic, thus cases (xv) to (xvii) from the start of this section. In each case
we either have 52 or 10 in the action of u. A single block of size 10 (as Vmin is self-dual) must come from 5/5,
and for the 52 the indecomposable modules of dimension congruent to 5 modulo 11 are 5 itself, 5, 7/3, 5, 7
and its dual of dimension 27, and 3, 5, 7, 9/1, 3, 5, 7, 9 and its dual, of dimension 49. Thus if we are in case
(xvi), so u belongs to class E6 (a1 ) acting as 114 , 52 , 12 , we therefore have 5⊕2 as a summand of Vmin ↓H , or
Vmin ↓H = 5, 7/3, 5, 7 ⊕ 3, 5, 7/5, 7 ⊕ 1⊕2 ;
an involution x ∈ H acts with trace 0 on this module, so it is not allowed. Therefore in all cases Vmin ↓H
has either 5/5 or 5⊕2 as a summand, contributing 2 to the trace of x.
If u comes from class E7 (a3 ), acting as 114 , 10, 2, the self-dual module that can contribute 2 to the
action of u is 5, 7/5, 7, so that Vmin ↓H has composition factors at least 72 , 54 . The only conspicuous sets of
composition factors with this many 7s and 5s are
92 , 72 , 54 , 14
and
72 , 56 , 32 , 16 ,
with the latter being incompatible with the unipotent action and the former implying that Vmin ↓H is
P (1)⊕2 ⊕ 5/5 ⊕ 5, 7/5, 7. Clearly therefore H fixes a line on Vmin .
79
For D6 and E6 (a1 ) we already have the composition factors of 52 , and the Jordan blocks 12 in the action
of u come either from two trivial summands, so we are done, or come from 5/7 or 3/9 and their duals, on
each of which x acts with trace 0. From the previous case we know that 5/7 yields
P (1)⊕2 ⊕ 5/7 ⊕ 7/5 ⊕ M,
where M is either 5⊕2 or 5/5, and again we fix two lines on Vmin . If we have 3/9 ⊕ 9/3, we again have a
unique conspicuous set of composition factors, and this time it is the very similar
P (1)⊕2 ⊕ 3/9 ⊕ 9/3 ⊕ M,
where M is as before, proving the result.
Now let a = 2, and suppose that H is not a blueprint for Vmin . In particular, by Proposition 4.10 there
are at most four odd-dimensional composition factors in Vmin ↓H . When the composition factors of Vmin ↓L
are 92 , 72 , 54 , 14 , as they are in two cases above, all of the 1s become trivial composition factors for H,
as there are no 3s in Vmin ↓L . Also, because there are no 11s either, there can be no simple modules in
Vmin ↓H with non-trivial 1-cohomology by Lemma 5.24, as a 20-dimensional module restricts as 11 ⊕ 9. Thus
the potential embeddings of L into G with these factors, where L has a trivial submodule but no trivial
summand, cannot be extended to embeddings of H into G. The same statement holds when the composition
factors are 94 , 52 , 32 , 14 , as we have at least two trivial composition factors.
Thus we may assume that Vmin ↓L has two trivial summands, along with the two 5s, with the rest being
projective. The conspicuous such modules are
11⊕2 ⊕ P (1)⊕2 ⊕ 1⊕2 ⊕ M
and P (9)⊕2 ⊕ 1⊕2 ⊕ M.
The first case has six trivial factors and no 3s, so Vmin ↓H has six trivial factors and hence H is a blueprint
for Vmin . In the second case there are two more trivials than 3s so Vmin ↓H has two trivial factors, and
Vmin ↓L has four 9s and no 7s or 11s, so Vmin ↓H has four 9s, thus H has six odd-dimensional factors on Vmin
and so is a blueprint for Vmin , as needed.
For p = 13, since 169 > 150 = 2 · v(E7 ), we need only consider a = 1.
Lemma 11.6 Suppose that p = 13 and a = 1. Then either H is a blueprint for Vmin or H fixes a line on
either Vmin or L(G).
Proof: If H is not a blueprint for Vmin then u acts non-generically, and so we are in case (xviii) of the list of
possible unipotent actions at the start of this subsection. Suppose that H has no trivial summand on Vmin ,
so that the 14 in the action of u all comes from indecomposables of dimension 14; this means that there
are eight composition factors of dimensions between 3 and 11, as indecomposables of dimension 14 have the
form i/(14 − i). Since the trace of an involution x is ±8, and the trace of x on i/(14 − i) is ±2, each of them
has the same trace. Adding in the traces of elements of orders 3 and 4, the unique such set of composition
factors is 112 , 74 , 32 . This gives Vmin ↓H as
11/3 ⊕ 3/11 ⊕ 7/7 ⊕ 7/7.
Using these traces, we can determine two possibilities for the action of L on the adjoint module L(G): these
are 133 , 11, 93 , 75 , 53 , 3, 13 and 132 , 113 , 94 , 7, 54 , 33 , 12 . Either way, H fixes a line on L(G), since the only
module with a trivial factor but not a trivial submodule or quotient is P (11) = 11/1, 3/11. The proof is
complete.
80
The last case is p = 19, where again we only have a = 1.
Proposition 11.7 Suppose that p = 19 and a = 1. If H is not a blueprint for Vmin then H centralizes a
2-space on Vmin and is a non-G-cr subgroup of the E6 -parabolic acting on Vmin as
P (1)⊕2 ⊕ 9/9.
Proof: If H is not a blueprint for Vmin then in particular u is non-generic, and so we are in case (xix) from
the list at the start of the section, i.e., u is regular and acts as 192 , 18. We need a self-dual indecomposable
module of dimension congruent to 18 modulo 19, and there is only one of these by Lemma 5.17, namely 9/9,
and the remainder of the module is projective. If x denotes an involution in H then x has trace ±8 on Vmin ,
and has trace 2 on 9/9, leaving a trace of 6 or −10 on the remaining projective summand. The trace of x
on P (i) for 3 6 i 6 17 is ±2, the trace of x on 19 is −1, and on P (1) it is 3. Thus Vmin ↓H is P (1)⊕2 ⊕ 9/9,
as needed.
This non-G-cr subgroup was constructed at the end of Section 9.
81
12
E7 in odd characteristic: SL2 embedding
In this section, k is a field of characteristic p > 3 and G = E7 (k), by which we mean the simply connected
form, i.e., |Z(G)| = 2 and G′ = G. Let Ḡ be an almost simple group with socle G/Z(G). From [10] we see
that v(E7 ) = 75 for odd integers, so if H is any subgroup of G with a semisimple element of odd order 77
or more, then H is a blueprint for Vmin , with the same holding for NḠ (H). In addition, in [9] we prove that
SL2 (9) cannot yield a maximal subgroup of Ḡ either, so here we let H = SL2 (pa ) with a = 3, 4 if p = 3 and
pa 6 150 = 2 · v(E7 ) if p > 5. In order for H not to be contained in a centralizer of a non-central involution,
we require that Z(G) = Z(H). Let L = SL2 (p) 6 H and let u denote a unipotent element of L of order p.
On Vmin , since we consider SL2 (pa ) rather than PSL2 (pa ), there can be no trivial composition factors
in Vmin ↓H , but rather 2-dimensional factors. We will thus normally aim to show that H is a blueprint for
Vmin , that H fixes a line on L(G), or that H fixes a 2-space on Vmin .
12.1
Characteristic 3
We let p = 3 and a = 3, 4. We begin with the case where the action of H = SL2 (3a ) on Vmin is definable
over a subfield of F3a .
Proposition 12.1 Suppose that p = 3, that a = 3, 4, and that the composition factors of Vmin ↓H are
invariant under a non-trivial field automorphism of H.
(i) If a = 3 then the composition factors of Vmin ↓H are either
8, (21 , 22 , 23 )8 ,
or
84 , (21 , 22 , 23 )4 ;
in the first case H fixes a line on L(G), and in the second case NḠ (H) is contained in a member of
X σ.
(ii) The case a = 4 cannot occur.
Proof: For a = 4, it turns out that the traces of semisimple elements of orders 5 and 8 are enough to
eliminate all possible sets of composition factors for Vmin ↓H , so we concentrate on the case a = 3.
The traces of semisimple elements of H are enough to confirm the first part of this statement. If the
composition factors of Vmin ↓H are 8, (21 , 22 , 23 )8 then the composition factors of H on L(G) are
(41,2 , 42,3 , 41,3 )8 , (31 , 32 , 33 ), 128 ,
which has pressure −4 and so H fixes a line on L(G). Suppose therefore that the composition factors of
Vmin ↓H are 84 , (21 , 22 , 23 )4 , and to start that there is an 8 in the socle of Vmin ↓H . Let y be the diagonal
matrix with entries ζ and ζ −1 , where ζ is a primitive 26th root of unity. The eigenvalues of y 2 on 8 are
12 , ζ ±4 , ζ ±10 , ζ ±12 ,
and so we look for elements of order 26 in G that square to y 2 and stabilize these eigenspaces of the action of
y 2 on Vmin : in fact, there are elements of G lying in seven distinct conjugacy classes of semisimple elements,
one of which contains y itself, and all of which have no (−1)-eigenspace on L(G). Any of these can be used
with Corollary 4.14, but taken together they show that the stabilizer of an 8 that is a submodule of Vmin ↓H
must be very large, and in particular positive dimensional.
82
Thus 8 is not a submodule of Vmin ↓H . If a 2i submodule is stabilized by NḠ (H) and Ḡ then H is
contained inside a member of X σ , so we may assume that this is not the case, either because NḠ (H)
induces the field automorphism on H or that no 2i is σ-stable (because k does not contain F27 for example).
In this case there must be an NḠ (H)-stable H-submodule W isomorphic to 21 ⊕ 22 ⊕ 23 , and the eigenvalues
of y 2 on W are ζ ±2 , ζ ±6 and ζ ±8 . In this case we do the same as for 8, finding again seven distinct classes
containing elements that square to y 2 and preserve these eigenspaces. Thus NḠ (H) is always contained in a
member of X σ as needed.
We may now attack the case of a = 3. We will not prove anything specific about the embeddings of H
into G, except that they are always contained in positive-dimensional subgroups.
Proposition 12.2 Suppose that p = 3 and a = 3. The subgroup NḠ (H) is contained inside a member of
X σ.
Proof: There are 284 conspicuous sets of composition factors for Vmin ↓H , but only 137 of these have
corresponding sets of factors on L(G), each of these being unique. Of these, seventy-seven have either no
2i or positive 2i -pressure for each i = 1, 2, 3, and of these only sixty-seven have either no trivial or positive
pressure on L(G). One of these is invariant under the field automorphism, which we saw earlier, so we are
left with twenty-two sets of composition factors, up to field automorphism.
The module 2i has non-split extensions only with 2i±1 , 6i−1,i and 8, so if there are 2s but no 6i,i−1 or 8
appearing with multiplicity 2 or above, then H fixes a 2-space on Vmin : two sets of composition factors (up
to field automorphism) satisfy this, so we are down to twenty sets of composition factors.
There are eighteen (six up to field automorphism) sets of composition factors with no 8s, and since 2i
only has extensions with the 8, 2i±1 and 6i−1,i , and of course there can be no 2s in the socle of Vmin ↓H ,
the socle must consist of 6i−1,i s, plus modules we can remove without exposing 2s. In each case there is a
unique i such that 6i−1,i appears with multiplicity at least 2, so this must be the socle and all 2s must be
stacked on top of it in some way. In each case we cannot place enough 2s on top of each 6i−1,i , and so we
must fix a 2-space in all these cases. This reduces us to fourteen sets of composition factors.
Thus we have at least one 8 in Vmin ↓H . There are five conspicuous sets of composition factors with a
single 8, up to field automorphism, namely
8, 622,1 , 623,1 , 261 , 242 , 223 ,
8, 621,3 , 62,3 , 623,1 , 251 , 222 , 223 ,
8, 621,2 , 622,1 , 63,1 , 63,2 , 221 , 222 , 223 ,
8, 653,1 , 251 , 222 , 223 ,
182,3,1 , 8, 62,1 , 623,1 , 231 , 22 , 223 ,
Since there is a single 8, we must still have the 6i−1,i s in the socle, and so again we can eliminate these cases
as follows: each of these only has a single 6i−1,i appearing with a non-unital multiplicity, and it appears with
multiplicity 2, so this appears once in the socle and all 2s lie above this, so we consider the {2, 6, 8}-radical
of P (61,2 ), and then remove all quotients that are not 61,2 , since Vmin is self-dual. This module is
61,2 /22 /23 /22 /61,2 ,
and so we can only support three 2s, but there are clearly far too many. This is enough to complete the
first four cases, but the fifth has an 18. We take the appropriate radical of P (63,1 ), and again remove all
composition factors from the top that are not 63,1 , to get the module
63,1 /21 /22 , 63,1 /21 , 62,1 , 182,3,1 /63,1 ,
83
which does not have a 23 in it. Thus all these conspicuous sets of composition factors yield a stabilized
2-space.
Thus we are down to nine, which are below.
82 , 622,1 , 61,2 , 63,1 , 241 , 232 , 23 ,
82 , 623,2 , 62,1 , 623,1 , 221 , 222 , 23 ,
82 , 62,1 , 61,3 , 623,1 , 241 , 222 , 223 ,
82 , 61,3 , 63,1 , 622,3 , 61,2 , 221 , 222 , 23 ,
181,2,3 , 82 , 62,1 , 63,1 , 221 , 222 , 23 ,
181,2,3 , 83 , 61,3 , 231 , 22 ,
82 , 623,2 , 61,2 , 62,1 , 63,1 , 61,3 , 21 , 22 ,
82 , 63,2 , 62,1 , 62,3 , 621,3 , 221 , 222 , 23 ,
181,2,3 , 82 , 62,1 , 61,3 , 231 , 222 .
Recall that 2i has extensions with 2i±1 , 6i−1,i and 8, and no other simple modules.
Cases 1,6,7,8,9: In the first, sixth, seventh, eighth and ninth cases, 6i−1,i does not occur with multiplicity
greater than 1 for all i, and so if it occurs in the socle then it is a summand. Therefore, for these six sets
of composition factors, we can remove all quotients and submodules from Vmin ↓H other than 8, and yield a
submodule W of P (8), which contains all composition factors of dimension 2 in Vmin ↓H .
The {2i , 6i,i+1 , 6i+1,i , 181,2,3 }-radical of P (8)/ soc(P (8)) is
21 , 22 , 23 , 61,2 , 62,3 , 63,1 /21 , 22 , 23 , 62,1 , 63,2 , 61,3 /8,
and so if W has two 8s and three 2i s then H must stabilize a 2-space on Vmin , eliminating the first and ninth
cases immediately. It also means that W has at most four socle layers except for the eighth case, which we
eliminate easily, since the {21 , 22 , 61,3 , 8, 181,2,3 }-radical of P (8) is
8/21 , 22 , 61,3 /8,
which doesn’t have enough 21 s, so H fixes a 2-space of Vmin .
Since W has at most four socle layers and is self-dual, none of the 6i,i+1 can occur in W . Also, any
2i that occurs with multiplicity 1 in Vmin ↓H must occur in the second socle layer, and cannot have any
extensions with a 2i±1 in the third socle layer. However, both 2i±1 in the third layer have an extension with
the 2i in the second layer, so cannot exist in W . In other words, we cannot have two 2i in W , eliminating
the sixth and seventh cases.
Case 5: Remove all composition factors from the top and bottom of Vmin ↓H to yield a module W with
socle a submodule of 8 ⊕ 62,3 and all 2s in it.
The preimages of the {21 , 22 , 23 , 61,2 , 61,3 , 62,3 }-radical of the module P (8)/ soc(P (8)) and {21 , 22 , 23 , 61,2 , 61,3 , 8}radical of P (62,3 )/ soc(P (62,3 )) are
21 , 22 , 23 , 62,3 /21 , 22 , 23 , 61,3 /8
22 , 23 /21 , 8/23, 61,3 /62,3 ,
and the fact that there is a single 23 means it must lie in the second socle, and has no extensions with the
other 2i composition factors. But then all other 2i s lie in the second socle layer as well, and that means we
cannot fit enough 2s in, so H fixes a 2-space on Vmin .
Cases 2,3,4: Here we wish to apply Corollary 4.14, so let y be the diagonal matrix with entries ζ 2 , ζ −2 for
ζ a primitive 26th root of unity, so that y has order 13 and acts with eigenvalues ζ ±2 on 21 . The eigenvalues
of y 2 on 63,1 and 8 are
ζ ±4 , ζ ±8 , ζ ±12
and
12 , ζ ±4 , ζ ±10 , ζ ±12 .
If we can find an element ŷ of order 26 in G \ H that has no (−1)-eigenspace on L(G) and that stabilizes
a submodule W of Vmin ↓H , then by Corollary 4.14 the subgroup hH, ŷi, which cannot be all of G, is not
84
almost simple modulo Z(G) either, and so by Proposition 3.3 the stabilizer of W , and any submodule of
Vmin ↓H isomorphic to W , is contained in a member of X σ , and so H and indeed NḠ (H) are contained
inside a member of X σ .
In the third case, the composition factor 63,1 only has extensions with 21 and 62,1 from the composition
factors of Vmin ↓H , so this must split off as a summand. In the second and fourth cases, if 8 is not a
submodule of Vmin ↓H and Vmin ↓H does not a 2-dimensional submodule then Vmin ↓H is a submodule of
P (63,1 ) and P (63,2 ⊕ 63,1 ) respectively. The cf(Vmin ↓H )-radical of P (63,1 ) in case 2 is
63,1 /21 , 23 /22 , 8/21 , 62,1 /63,1 ,
so 8 must be a submodule of Vmin ↓H . Similarly, the cf(Vmin ↓H )-radicals of P (63,1 ) and P (63,2 ) in case 4 is
62,1 /63,1 , 8/21 , 23 , 63,2 /22 , 8/21 , 62,1 /63,1
and
63,1 /21 , 62,1 /8/63,2 ,
so since there is no 222 in this, 8 must be in the socle of Vmin ↓H again.
We have therefore proved that in cases 2 and 4, 8 must be a submodule, and in case 3, 63,1 must be a
submodule. We therefore find, in each case, an element ŷ in G \ H of order 26, squaring to y 2 , and stabilizing
the eigenspaces of the particular stabilized submodule. On Vmin these have eigenvalues
14 , (−ζ ±1 )6 , (ζ ±2 )4 , (ζ ±3 ), (−ζ ±3 )2 , (ζ ±4 )2 , (−ζ ±4 )3 , (ζ ±5 )3 , (−ζ ±6 )5 ,
14 , (ζ ±1 )2 , (−ζ ±1 )2 , (−ζ ±2 )5 , (ζ ±3 )3 , (−ζ ±3 ), (ζ ±4 )4 , (ζ ±5 ), (−ζ ±5 )3 , (−ζ ±6 )5 ,
14 , (ζ ±1 )2 , (−ζ ±1 ), (−ζ ±2 )6 , (ζ ±3 )4 , (−ζ ±3 ), (ζ ±4 )5 , (−ζ ±5 )3 , (−ζ ±6 )4 .
Thus the result holds.
Proposition 12.3 Suppose that p = 3 and a = 4. Either H is a blueprint for Vmin or the composition
factors of Vmin ↓H are
184,2,3 , 821,2,3 , 81,2,4 , 61,3 , 64,1 , 21 ,
and NḠ (H) lies inside an element of X σ .
Proof: Using semisimple elements of order up to 41, one whittles down the 55 million or so possible sets
of composition factors for a module of dimension 56 to just 190 up to field automorphism, of which two fail
the trace of an element of order 80, leaving 188.
Of these, we consider an element y of order 41 in H, and whether there exists an element of order 123
in G cubing to y and stabilizing the same subspaces of Vmin . If this is true then y is a blueprint for Vmin
since 123 > v(E7 ). Indeed, a computer check shows that this is true for 187 of the 188 semisimple elements
involved. The remaining one comes from the conspicuous set of composition factors in the statement of the
proposition,
184,2,3 , 821,2,3 , 81,2,4 , 61,3 , 64,1 , 21 ,
where y has 38 distinct eigenvalues on Vmin , and there exist elements of order 123 cubing to y and with 40
distinct eigenvalues, but none with 38. If ζ is a primitive root of unity then y can be chosen to have the
following eigenvalues.
85
Module
Eigenvalues
21
ζ ±1
61,3
ζ ±1 , ζ ±17 , ζ ±19
64,1
ζ ±12 , ζ ±14 , ζ ±16
81,2,3
ζ ±5 , ζ ±7 , ζ ±11 , ζ ±13
81,2,4
ζ ±10 , ζ ±12 , ζ ±16 , ζ ±18
184,2,3
ζ ±2 , ζ ±3 , ζ ±4 , ζ ±8 , ζ ±9 , ζ ±10 , ζ ±14 , ζ ±15 , ζ ±20
There exists an element ŷ of order 123 cubing to y and stabilizing all eigenspaces except for the ζ ±1 eigenspaces.
Since 81,2,3 is the only composition factor to occur with multiplicity greater than 1, any other factor
in the socle must be a summand. The module 81,2,3 only has extensions with 21 , 61,3 and 81,2,4 from the
composition factors of Vmin ↓H , and so the structure of Vmin ↓H must be W ⊕ 64,1 ⊕ 184,2,3, where W consists
of the remaining factors.
Since ŷ stabilizes all but the ζ ±1 -eigenspaces, it fixes the 32 ⊕ 6 ⊕ 18 decomposition above, and if Ŷ
denotes an infinite subgroup of G containing ŷ and stabilizing the same subspaces of Vmin as ŷ (which exists
since 123 > v(E7 ) = 75), then X = hŶ , Hi certainly stabilizes the 6- and 18-dimensional summands of
Vmin ↓H . Thus H is contained inside an element of X σ .
12.2
Characteristic at least 5
We now let p > 5: since there are elements of orders (pa + 1)/2 or (pa − 1)/2, and one of these is odd, we
assume that pa 6 150. As all unipotent classes are generic for Vmin for all p > 29, we only need consider
pa = 5, 7, 11, 13, 17, 19, 23, 25, 49, 121, 125.
As for PSL2 (pa ), there are some restrictions we can place on the possible actions of a unipotent element
u, above and beyond appearing on [13, Table 7], given by Lemma 5.16. This yields twenty-nine possible
non-generic classes for various primes, as given below.
(i) (A3 + A1 )′′ , p = 5, acting as 52 , 48 , 27 ;
(ii) D4 (a1 ) + A1 , p = 5, acting as 56 , 4, 34 , 25 ;
(iii) A3 + A2 , p = 5, acting as 56 , 42 , 34 , 22 , 12 ;
(iv) A4 , p = 5, acting as 510 , 16 ;
(v) A3 + A2 + A1 , p = 5, acting as 56 , 44 , 25 ;
(vi) A4 + A1 , p = 5, acting as 510 , 22 , 12 ;
(vii) A4 + A2 , p = 5, acting as 510 , 32 ;
(viii) (A5 )′′ , p = 7, acting as 72 , 67 ;
(ix) D4 + A1 , p = 7, acting as 76 , 25 , 14 ;
(x) D5 (a1 ), p = 7, acting as 76 , 32 , 22 , 14 ;
(xi) (A5 )′ , p = 7, acting as 74 , 64 , 14 ;
86
(xii) A5 + A1 , p = 7, acting as 74 , 63 , 52 ;
(xiii) D5 (a1 ) + A1 , p = 7, acting as 76 , 4, 25 ;
(xiv) D6 (a2 ), p = 7, acting as 76 , 52 , 4;
(xv) E6 (a3 ), p = 7, acting as 76 , 52 , 14 ;
(xvi) E7 (a5 ), p = 7, acting as 76 , 6, 42 ;
(xvii) A6 , p = 7, acting as 78 ;
(xviii) E7 (a4 ), p = 11, acting as 112 , 10, 8, 6, 42, 2;
(xix) D6 , p = 11, acting as 114 , 10, 12;
(xx) E6 (a1 ), p = 11, acting as 114 , 52 , 12 ;
(xxi) E7 (a3 ), p = 11, acting as 114 , 10, 2;
(xxii) E6 , p = 13, acting as 134 , 14 ;
(xxiii) E7 (a3 ), p = 13, acting as 132 , 12, 10, 6, 2;
(xxiv) E7 (a2 ), p = 13, acting as 134 , 4;
(xxv) E7 (a2 ), p = 17, acting as 172 , 10, 8, 4;
(xxvi) E7 (a1 ), p = 17, acting as 172 , 16, 6;
(xxvii) E7 (a1 ), p = 19, acting as 192 , 12, 6;
(xxviii) E7 , p = 19, acting as 192 , 18;
(xxix) E7 , p = 23, acting as 232 , 10.
Thus we need to consider p = 5, 7, 11, 13, 17, 19, 23, and we will examine each in turn.
Proposition 12.4 Suppose that p = 5.
(i) If a = 1 then H fixes a 2-space on Vmin .
(ii) For a = 2, one of the following holds: H and NḠ (H) fix a 2-space on Vmin ; H fixes a line on L(G) or H
stabilizes a 4-space on Vmin that has a positive-dimensional stabilizer; H stabilizes an sl2 -subalgebra
of L(G).
(iii) If a = 3 then an element of order 63 in H is a blueprint for Vmin , and hence H is a blueprint for Vmin .
Proof: Only the element of order 3 is important here, and it has trace one of −25, −7, 2, 20, with the last
case not possible, and so we get
4, 226 ,
47 , 214 ,
410 , 28 .
As P (4) = 4/2/4, each of these must fix a 2-space on Vmin , as claimed.
When a = 2, there are 106 conspicuous sets of composition factors for Vmin ↓H , but fifty of these have
no corresponding set of composition factors for L(G), so can be ignored. None of the 106 sets is definable
87
over F5 , so if H stabilizes a line on L(G) or a 2-space on Vmin then we are done. Removing those sets of
factors with non-positive 2i -pressure (and at least one 2i ) leaves eighteen, nine up to field automorphism
of H. Exactly one of these has two possible sets of composition factors on L(G), one of which has a single
trivial and a single 8-dimensional, so that second option will be ignored as having pressure 0. Of the other
eight, which all have a unique corresponding set of composition factors on L(G), three have non-positive
pressure and trivial factors, so that leaves us with six conspicuous sets of composition factors on Vmin . Up
to field automorphism of H, these are
1221,2 , 101,2 , 422 , 421 , 231 ,
121,2 , 101,2 , 62,1 , 621,2 , 432 , 221 ,
121,2 , 122,1 , 102,1 , 62,1 , 61,2 , 42 , 41 , 21 ,
121,2 , 1021,2 , 62,1 , 61,2 , 42 , 41 , 22 , 21 ,
1221,2 , 102,1 , 101,2 , 61,2 , 41 , 21 ,
1222,1 , 101,2 , 631,2 , 41 .
The simple modules with extensions with 21 are 42 , 62,1 and 121,2 .
Case 1: This has pressure 1, and we may assume that we have a submodule of P (121,2 ) or P (42 ) with
three 21 s. The {21 , 41 , 42 , 101,2 , 121,2 }-radicals of these two modules are 101,2 /121,2 /21 , 41 , 101,2 /121,2 and
42 /21 /42 , so H fixes a 2-space on Vmin .
Case 2: The corresponding submodule of P (42 ) in the second case is also 42 /21 /42 , so again this fixes a
2-space on Vmin .
Cases 3, 4: The 21 -pressure is 1, but the only simple module with multiplicity more than 1 is 101,2 , so 21
(and 22 ) split off, and the result holds again, and of course the same idea works for the fourth case.
Case 6: There are no extensions between the simple modules involved so Vmin ↓H is semisimple. The
corresponding composition factors for L(G) ↓H have no extensions between them either and so the restriction
is also semisimple, acting as
⊕3
⊕3
⊕3
151,2
⊕ 15⊕2
⊕ 51 ⊕ 52 ⊕ 3⊕4
2,1 ⊕ 9
1 ⊕ 32 .
Write x for an element of order 13. Choosing ζ a primitive 13th root of unity appropriately (so that x acts
on 21 with eigenvalues ζ ±1 ) the eigenvalues of x on Vmin are
14 , (ζ ±1 )7 , (ζ ±2 )6 , (ζ ±3 )3 , (ζ ±4 )6 , ζ ±5 , (ζ ±6 )3 .
Looking through the elements of order 26 in E7 , we find one x̂ that squares to x, and if θ is a primitive
26th root of 1 with θ2 = ζ, we have that the eigenvalues of x̂ are
14 , (θ±1 )7 , (θ±2 )6 , (θ±3 )3 , (θ±4 )6 , θ±5 , (θ±6 )2 , (−θ±6 ).
This stabilizes the 4-space of Vmin stabilized by H (as well as the 6-spaces and the sum of the 12- and
10-spaces), so the stabilizer Y of this 4-space is either an almost simple group PGL2 (25) modulo Z(G) or H
is not the socle of an almost simple maximal subgroup. The eigenvalues of x̂ on L(G) are
117 , (θ±1 )10 , (θ±2 )13 , (θ±3 )12 , (θ±4 )6 , (θ±5 )8 , (−θ±5 )3 , (θ±6 )4 , (−θ±6 )2 ,
so hH, x̂i is not PGL2 (25) by Corollary 4.14. By Proposition 3.3, either H is contained in a member of X σ
or H is contained in a copy of the Rudvalis group Ru (which is really 2 · Ru) acting on Vmin as 28 ⊕ 28∗
by the table in [23]. Since this is incompatible with the action of H, we get that H is contained inside a
member of X σ , and this 4-space stabilizer must itself be positive dimensional.
88
Case 5: The 102,1 must split off as it has no extensions with the 121,2 , but the rest of the composition
factors of Vmin ↓H can lie above it, and there is a unique module
102,1 ⊕ 121,2 /21 , 41 , 61,2 , 101,2 /121,2 ,
with u acting as 510 , 32 , class A4 + A2 . The action of u on the direct sum of the composition factors of
Vmin ↓H has block structure 58 , 42 , 24 , and examining [13, Table 7], we see that there are only two possible
actions for u: 510 , 32 and 510 , 22 , 12 . Thus the 41 and 61,2 cannot be summands (as they both have a 4 in
the action of u), and we assume that the 21 is not a submodule, so if Vmin ↓H is not the module above then
only the 101,2 can be removed from the non-simple summand. However, the {21 , 41 , 61,2 , 121,2 }-radical of
P (121,2 ) is
121,2 /21 , 41 , 61,2 /121,2,
but with a 61,2 quotient, not allowed. Thus Vmin ↓H is as above, and in particular the symmetric square of
this has L(G) ↓H as a summand (since S 2 (Vmin ) = L(G) ⊕ L(2λ1 )).
The composition factors of L(G) ↓H are
16, 1521,2, 1522,1 , 821,2 , 832,1 , 321 , 332 , 12 ,
and u must act on L(G) as 526 , 3. There are only six isomorphism types of indecomposable module appearing
as a summand of S 2 (Vmin ↓H ) whose composition factors appear on the list above, and these have structures
31 ,
152,1 ,
82,1 /1, 32/82,1 ,
152,1 /82,1 /1, 32/82,1 /152,1 ,
151,2 /81,2 /1, 31 /81,2 /151,2,
32 /82,1 , 16/32,
with the second module appearing only once. There is only one way to assemble these summands into a
module with the right unipotent action and composition factors, and this is
152,1 /82,1 /1, 32/82,1 /152,1 ⊕ 151,2 /81,2 /1, 31/81,2 /151,2 ⊕ 32 /82,1 , 16/32 ⊕ 31 .
In particular, this has 31 as a summand and so this is an sl2 -subalgebra by Proposition 4.17, and also 32 as
a submodule and subalgebra, but not necessarily a copy of sl2 . This completes the proof for a = 2.
Finally, let a = 3. Using the traces of semisimple elements of order up to 31 there are 434 conspicuous
sets of composition factors, 146 up to field automorphism. Checking the traces of elements of order 63,
we find that eleven of these are not conspicuous for elements of order 63, and the remaining 135 all have
preimages of order 5 · 63 = 315 that have the same number of eigenspaces on Vmin . Since 315 > v(E7 ) = 75
(for odd-order elements) we have that elements of order 63 in H are always blueprints for Vmin , and hence
H is as well.
In the introduction we claimed that this potential SL2 (25) is maximal if it exists, so we must show
that it is not contained in any positive-dimensional subgroup. By consideration of composition factors and
summand dimensions, it can only lie inside a D6 -parabolic; then one can proceed either by showing that
the 12-dimensional factor cannot support a symmetric bilinear form, or by noting that if the subgroup lies
inside the D6 -parabolic then another lies inside the D6 -Levi, hence acting semisimply, but the action of the
unipotent element would be 58 , 42 , 24 , not appearing in [13, Table 7].
The next case is p = 7, where we cannot prove that there are no maximal SL2 (7)s in all cases.
Proposition 12.5 Suppose that p = 7.
89
(i) Let a = 1. If H does not fix either a 2-space on Vmin , or a 1-space on L(G), then the action of H on
Vmin and L(G) are
P (6)⊕2 ⊕ P (4) ⊕ 6 ⊕ 4⊕2
and 7⊕5 ⊕ P (5)⊕3 ⊕ P (3)⊕3 ⊕ 5 ⊕ 3⊕3
respectively.
(ii) Let a = 2. If H is not a blueprint for Vmin then either NḠ (H) fixes a 2-space on Vmin or a line on
L(G).
Proof: We start with a = 1. The conspicuous sets of composition factors are
4, 226 ,
47 , 214 ,
6, 49 , 27 ,
63 , 47 , 25 ,
64 , 43 , 210 ,
65 , 45 , 23 ,
66 , 4, 28 ,
67 , 43 , 2.
As the projective indecomposable modules are
P (2) = 2/4, 6/2,
P (4) = 4/2, 4/4,
P (6) = 6/2/6,
we look through the list above, checking to see whether we have enough 4s and 6s (three of the first or two
of the second) to cover all 2s; this leaves the sixth and eighth cases of 65 , 45 , 23 and 67 , 43 , 2 to deal with.
Case 8: We switch to L(G), and there is only one corresponding set of composition factors on L(G), namely
7, 515 , 310 , 121 , which means that H fixes a line on L(G).
Case 6: The only possible structure that does not fix a 2-space on Vmin and also yields a unipotent action
from the list at the start of this section is P (6)⊕2 ⊕ P (4) ⊕ 6 ⊕ 4⊕2 , with u lying in class E7 (a5 ). The factors
of L(G) ↓H aren’t uniquely determined, and can be any one of
75 , 515 , 32 , 117 ,
72 , 518 , 35 , 114 ,
78 , 57 , 312 , 16 ,
75 , 510 , 315 , 13 .
The first three of these must fix a line on L(G), but the last one could in theory not, with module action
7⊕5 ⊕ P (5)⊕3 ⊕ P (3)⊕3 ⊕ 5 ⊕ 3⊕3 ,
this action compatible with the action of u on Vmin .
Now let a = 2, and recall that L = SL2 (7). The eigenvalues of an element y of order 25 on Vmin are
enough to determine the semisimple class of E7 to which y belongs. This allows us to apply Lemma 4.12 to
see that if there is an A1 subgroup with 24-restricted composition factors, then any subgroup H of G whose
composition factors on Vmin match the restriction of this A1 to SL2 (49) is a blueprint for Vmin .
There are 150 conspicuous sets of composition factors for Vmin ↓H , and this is too many to analyse one
at a time, although none of these is definable over F7 , so we can eliminate those of negative 2i -pressure. Of
the 150, only 92 of them have a corresponding set of composition factors for L(G), and only 42 of these
have either no trivial factors or positive pressure. Of these 42, only 26 have either no 2i or positive 2i pressure for i = 1, 2, so we have thirteen conspicuous sets of composition factors left to deal with, up to field
automorphism. Six of these have a 21 composition factor:
652 , 61,2 , 442 , 221 ,
1821,2 , 102,1 , 41 , 42 , 21 ,
141,2 , 101,2 , 622 , 61,2 , 432 , 21 ,
281,2 , 102,1 , 622 , 41 , 21 ,
182,1 , 1022,1 , 61 , 61,2 , 42 , 21 ,
90
142,1 , 1032,1 , 61,2 , 42 , 21 .
Case 1: Consider the diagonal A1 inside the A1 A1 maximal subgroup of G, acting along each factor as
L(1): this acts on Vmin as
(L(6) ⊗ L(3)) ⊕ (L(2) ⊗ L(5)) ⊕ (L(4) ⊗ L(1)),
which has factors L(5)5 , L(3)4 , L(7)2 , L(9), up to field automorphism the same as the first case above.
Case 4: Inside C3 G2 , consider an A1 subgroup X acting along the first factor as L(5) and along the second
as L(2) ⊕ L(8). The action of X on Vmin is
L(5)/L(7)/L(5) ⊕ L(3) ⊕ L(11) ⊕ L(13) ⊕ L(3)/L(9)/L(3),
so the restriction to SL2 (49) has composition factors 142,1 , 102,1 , 621 , 62,1 , 431 , 22 , up to field automorphism a
match for the fourth case.
Case 5: For the fifth case, we note that the composition factors are the same as
22 ⊗ ((61 ⊗ 22 ) ⊕ 51 /41,2 /51 ⊕) ⊕ 42 :
inside A1 F4 , let X be a copy of A1 acting along the first factor as L(7) and along the second factor as
L(12) ⊕ L(4)/L(8)/L(4). This subgroup of F4 exists inside the A1 C3 subgroup, acting irreducibly on the
minimal modules of both subgroups as L(7) and L(5). The composition factors of Vmin ↓X match the fifth
case.
Case 6: For the sixth case, as in the fifth, we note that the composition factors are the same as
22 ⊗ (71 ⊕ 51 /41,2 /51 ⊕ 51 ) ⊕ 42 :
inside A1 F4 , let X be a copy of A1 acting along the first factor as L(7) and along the second factor as
L(6) ⊕ L(4)/L(8)/L(4) ⊕ L(4). This subgroup of F4 exists inside the A1 G2 subgroup, acting irreducibly on
the minimal modules of both subgroups as L(1) and L(6). The composition factors of Vmin ↓X match the
sixth case.
Since these embeddings have factors up to L(21), in all cases H is a blueprint for Vmin by Lemma 4.12.
Case 3: We now address the third case. Inside C3 G2 , let X be an A1 subgroup acting along the C3 as
L(1) ⊕ L(21) and acting along G2 as L(6). The action of X on Vmin is
L(27) ⊕ L(5)/L(7)/L(5) ⊕ L(3) ⊕ L(29).
Up to field automorphism, the composition factors match the third case. While these are not 24-restricted,
they are close: checking the weight spaces against the eigenvalues of the SL2 (49) contained within it, all weight
spaces that have the same eigenvalues when restricted to SL2 (49) are contained within the L(27) ⊕ L(29).
Thus if H is not a blueprint then H is contained in a positive-dimensional subgroup Y with composition
factors of dimension 38, 6, 6, 4, 2. Notice that therefore Y , and hence H, must lie in either C3 G2 itself – and
we know from above that in that case H is a blueprint for Vmin – or inside A1 F4 , but it is easily seen to not
be possible to place H inside this subgroup by the action of H on Vmin . Thus H is indeed a blueprint for
Vmin .
Case 2: Here if H is not semisimple – and hence stabilizes a 2-space of Vmin – then the action of H on Vmin
is
181,2 /21 , 41 /181,2 ⊕ 102,1 ⊕ 42 ;
we claim that such a subgroup H does not lie in a positive-dimensional subgroup of G. To see this, firstly
the dimensions of the composition factors are not compatible with coming from any maximal parabolic, so
91
that H must be contained in a reductive maximal subgroup, where the dimensions and multiplicities exclude
A7 and A2 , and it is easy to see that it doesn’t lie in the A1 A1 .
For H to lie in A1 D6 , 181,2 would have to lie in the product of a module for SL2 of dimension 2 and
a module for PSL2 of dimension 12: this is possible, but only with 21 being tensored by 21 ⊗ 62 , and this
yields 62 as well, not in Vmin ↓H .
If H lies in A2 A5 then we must have that H acts on the natural modules along each factor as 31 and 62 ,
whence the module (00, λ3 ) is 42 /61,2 /42 ⊕ 62 , obviously not correct.
If H lies in G2 C3 , the the tensor product of the two minimal modules for these groups must be 1821,2 , 4i
for some i, and this is obviously impossible.
For H to lie in A1 G2 , 181,2 would have to be a composition factor of the tensor product of a module for
SL2 (49) of dimension 4 and a module for PSL2 (49) of dimension 7, not possible.
We finally have H in A1 F4 , where H must act on the natural module as 2i for some i, yielding 4i as a
composition factor of Vmin ↓H , and the rest of the module must be 2i ⊗ M for some 26-dimensional module
M for PSL2 (49), and this cannot yield 1821,2 , so H cannot embed in this subgroup either.
Having proved this, we now show that the stabilizer of the 42 submodule is positive dimensional, a
contradiction. There are eight elements of order 50 in a maximal torus of G squaring to y of order 25 and
preserving the eigenspaces making up the 42 : these eight elements generate a subgroup Z50 × Z2 × Z2 of
order 200, and so the stabilizer of the 4-space contains H as a subgroup of index at least 4, ruling out the
possibility that it is almost simple with socle H. Thus H is contained in a member of X σ by Proposition
3.3, a contradiction and we are done.
We now give the seven conspicuous sets of composition factors with no 2i in them.
1822,1 , 61 , 62,1 , 421 ,
182,1 , 142,1 , 102,1 , 62,1 , 421 ,
1222,1 , 101,2 , 631,2 , 41 ,
421,2 , 62,1 , 421 ,
282,1 , 122,1 , 622,1 , 41 ,
281,2 , 141,2 , 61 , 421
281,2 , 182,1 , 101,2 .
Cases 1 and 2: Inside A2 A5 , let a subgroup X of type A1 act along the two factors as L(14) and L(5)
respectively. The composition factors of Vmin ↓X are L(21)2 , L(5), L(3)2 , L(9). This is the first case.
Inside C3 G2 , consider an A1 subgroup X acting along the first factor as L(5) and along the second as
L(14) ⊕ L(8). The action of X on Vmin is
L(19) ⊕ L(13) ⊕ L(11) ⊕ L(3)/L(9)/L(3).
The composition factors of X on Vmin match the second case.
Case 3: Inside C3 G2 , consider an A1 subgroup X acting along the first factor as L(5) and along the second
as L(42). The action of X on Vmin is
L(47) ⊕ L(3)/L(9)/L(3),
a match for the third case, but of course these are not 24-restricted, but do satisfy the second condition of
Lemma 4.12, so that H is a blueprint for Vmin
Case 4: Inside F4 A1 , let X denote an A1 subgroup acting along the second factor as L(1), and along the
first factor as L(4) ⊕ L(44), which exists inside the A1 G2 subgroup of F4 . The action of X on Vmin is
L(43) ⊕ L(45) ⊕ L(5) ⊕ L(3)⊕2 ,
92
so this is the fourth case. Of course, these are not 24-restricted, so we proceed as in the second case of
the previous set of composition factors, looking for elements of order 50 squaring to y and stabilizing the
eigenspaces the comprise the 41 in the socle. Again, we find a subgroup Z50 × Z2 × Z2 , and we conclude as
before that the stabilizer of the 41 is a positive-dimensional subgroup of G.
We claim that H is a blueprint for Vmin . With the dimensions and multiplicities of the composition
factors, the only maximal positive-dimensional subgroups it can lie in are D6 A1 , C3 G2 , A1 G2 , A1 F4 and
A1 A1 , with the last one clearly impossible.
If H 6 D6 A1 then 141,2 ⊕ 61 ⊕ 41 is a tensor product of a 12-dimensional and a 2-dimensional module,
so must be 21 ⊗ (72 ⊕ 51 ). Thus H lies inside the product of the A1 and a product of two orthogonal groups,
Spin7 × Spin5 , and there is a unique action of an A1 subgroup inside these of acting as L(42) ⊕ L(0) and
L(42) on the two relevant modules of Spin7 , and as L(3) and L(4) on the two modules of Spin5 . This A1
fixes the same subspaces of Vmin as H, so H is a blueprint for Vmin . The same statement holds from above
for A1 F4 .
If H 6 C3 G2 then a similar analysis shows that H acts on the minimal modules of the two factors as
21 ⊕ 41 and 72 respectively, and the A1 acting along each factor as L(1) ⊕ L(3) and L(42) again stabilizes
the same subspaces of Vmin as H, so H is a blueprint for Vmin .
We are left with A1 G2 , where in order to find the 281,2 we must have that H acts on the two natural
modules as 21 and 72 respectively, so that the factor 3 ⊗ 10 in Vmin ↓A1 G2 yields 281,2 , but then the other
factor of 1 ⊗ 01 yields two copies of 61,2 , not correct. (Indeed, this is how we get the sixth case above.)
Thus, whenever H is a subgroup of a positive-dimensional subgroup of G, it is a blueprint for Vmin , as
needed.
Case 5: Inside the maximal subgroup A1 G2 , let X be an A1 subgroup acting along the first factor as L(7)
and along the second as L(2)⊕2 ⊕ L(0): the composition factors of X on Vmin are L(9)3 , L(11), L(21), L(23)2,
matching up with the fifth case, up to field automorphism. Since they are all 24-restricted, H is a blueprint
for Vmin by Lemma 4.12.
Case 6: Inside the same subgroup A1 G2 , let X instead be an A1 subgroup acting along the first factor
as L(7) and along the second as L(6): the composition factors of X on Vmin are L(3), L(9)2 , L(17), L(27),
matching up with the sixth case, but no longer 24-restricted, but H is still a blueprint for Vmin by Lemma
4.12(ii).
Case 7: Inside the maximal A1 A1 subgroup, take a diagonal A1 as we have done before, but this time acting
as L(1) and L(7) along the two factors. The composition factors of this on Vmin are L(27), L(13), L(37), of
course not 24-restricted, and even Lemma 4.12(ii) doesn’t work in this case. If H is contained in a positivedimensional subgroup other than A1 A1 then the dimensions of the composition factors and multiplicities
show that it can only come from A1 G2 , and in order to get 281,2 appearing, H must act along A1 as 21 and
G2 as 72 . However, the other factor of dimension 28 must have 61,2 as a composition factor, not allowed.
Thus H 6 A1 A1 , and so H is a blueprint for Vmin .
It remains to show that H is always contained inside a member of X . Of course, since Vmin ↓H is
multiplicity free, it is semisimple, and so the 101,2 is a submodule. As with previous cases, we find more
than one element of order 50 squaring to y and stabilizing the eigenspaces that comprise 101,2 . The subgroup
generated by these is Z50 × Z2 , and we wish to apply Corollary 4.14, so we need to find an element ŷ of order
50 in this subgroup whose action on L(G) has no (−1)-eigenspace, but this is easy: its eigenspaces are
17 , (θ±2 )6 , (θ±4 )4 , (θ±5 ), (θ±6 )2 , (θ±7 )3 , (θ±8 ), (θ±9 )5 , (θ±11 )6 , (θ±13 )6 ,
(θ±15 )5 , (θ±16 ), (θ±17 )3 , (θ±18 )2 , (θ±19 )2 , (θ±20 )4 , (θ±21 ), (θ±22 )5 , (θ±24 )6 ,
93
where θ is a primitive 50th root of unity and y = ŷ 2 acts on 21 with eigenvalues θ±2 .
We have thus shown that all cases are blueprints for Vmin , fix a 2-space on Vmin , or fix a line on L(G), as
needed.
Proposition 12.6 Suppose that p = 11.
(i) Let a = 1. If H is not a blueprint for Vmin , then NḠ (H) fixes a 2-space on Vmin or a line on L(G).
(ii) Let a = 2. If H is not a blueprint for Vmin , then NḠ (H) fixes a 2-space on Vmin .
Proof: Let a = 1. As p = 11 the action of u is one of cases (xviii) to (xxi) in the list above. In the first
unipotent class, E7 (a4 ) acting as 112 , 10, 8, 6, 42, 2, there are single blocks of size 10, 8, 6 and 2, which must
come from simple summands of those dimensions. Since we cannot have a faithful indecomposable module of
dimension 11 + 4 = 15, the 4s must also come from simple summands, and so Vmin ↓H is a single projective
plus a semisimple module. The conspicuous such sets of composition factors yield
P (10) ⊕ 10 ⊕ 8 ⊕ 6 ⊕ 4⊕2 ⊕ 2
P (4) ⊕ 10 ⊕ 8 ⊕ 6 ⊕ 4⊕2 ⊕ 2.
and
The second of these has corresponding set of factors on L(G) given by 98 , 72 , 57 , 112 , and the action of u on
L(G) is 118 , 9, 72 , 52 , 34 . Blocks of size 3 come from, up to duality,
3,
7, 9/1, 3, 5,
3, 5, 7/5, 7, 9,
and so H cannot embed with these factors and this action of u. (The other set of composition factors yields
11⊕4 ⊕ P (9) ⊕ P (7) ⊕ 9 ⊕ 7⊕2 ⊕ 5⊕2 ⊕ 3⊕4 .)
If u comes from class D6 , acting as 114 , 10, 12 , then the 10 must come from a simple summand, and
the 12 comes from (up to duality) 6/6, 4/8 or 2/10. The sum of one of these plus its dual, a single
projective indecomposable module, and the 10, must be conspicuous. There are five such conspicuous sets
of composition factors, all of which have corresponding sets of factors on L(G), three of them having two
different sets. The action of u on L(G) is 1110 , 102 , 13 , so certainly H has a trivial summand on L(G), with
the 102 coming from (5/5)⊕2 , 3/7 ⊕ 7/3 or 1/9 ⊕ 9/1, the other 12 being either semisimple, 9/3 ⊕ 3/9 or
5/7 ⊕ 7/5, with the rest of the module being projective. There are 937 such sets of composition factors, and
when taking the intersection of that list with those of the corresponding sets of composition factors to our
list for Vmin ↓H , we find two members:
P (4) ⊕ 10/2 ⊕ 2/10 ⊕ 10
and
P (6) ⊕ (6/6)⊕2 ⊕ 10,
with corresponding embeddings
11⊕4 ⊕ P (5) ⊕ P (3) ⊕ 3/9 ⊕ 9/3 ⊕ 3/7 ⊕ 7/3 ⊕ 1
and
11⊕6 ⊕ P (5) ⊕ P (3) ⊕ 3/7 ⊕ 7/3 ⊕ 1⊕3 .
Of course, these fix a line on L(G), as needed.
If u comes from class E6 (a1 ) acting as 114 , 52 , 12 , then the two blocks of size 5 must come from summands
of dimension 16, so 4, 6/6 ⊕ 6/4, 6, and the two 1s must come from summands of dimension 12. The
conspicuous such sets of composition factors yield
4, 6/6 ⊕ 6/4, 6 ⊕ 10/2 ⊕ 2/10,
4, 6/6 ⊕ 6/4, 6 ⊕ 4/8 ⊕ 8/4,
94
4, 6/6 ⊕ 6/4, 6 ⊕ (6/6)⊕2 .
Of these, only the last has a corresponding set of composition factors on L(G), and this is 112 , 98 , 55 , 114 ;
however, u acts as 1110 , 9, 52 , 3, 1 on L(G), so we need a 3 as a summand of L(G) ↓H , not possible. Thus H
does not embed with u from this class.
Finally, if u acts as 114 , 10, 2, coming from E7 (a3 ), then the 2 and 10 must come from summands of the
same dimension, so we need two projectives plus 10⊕2. There are three such conspicuous sets of composition
factors, yielding
P (4)⊕2 ⊕ 10 ⊕ 2,
P (4) ⊕ P (10) ⊕ 10 ⊕ 2,
P (6) ⊕ P (10) ⊕ 10 ⊕ 2.
Each of these has corresponding sets of composition factors on L(G), with the third having two. However,
u acts on L(G) with blocks 1111 , 9, 3, so L(G) ↓H is the sum of 9 ⊕ 3 and projectives, and since we have
1111 in the action of u, the number of summands of L(G) ↓H that are either 11s or P (1)s must be odd. In
particular, any trivial composition factors lie either in P (1)s or P (9)s. The four corresponding sets of factors
are
98 , 73 , 57 , 112 ,
114 , 93 , 74 , 53 , 36 , 1,
113 , 92 , 76 , 55 , 35 ,
112 , 94 , 77 , 5, 36 , 13 ;
the first and second cases need an odd number of P (1)s, and the second case can have no P (9)s as it has no
3s, leading to a contradiction. The fourth case cannot work with the unipotent class either, but the third
case yields
11⊕3 ⊕ P (7)⊕2 ⊕ P (5) ⊕ P (3) ⊕ 9 ⊕ 3.
Thus P (6) ⊕ P (10) ⊕ 10 ⊕ 2 is the only acceptable embedding of H into Vmin with this unipotent action.
Now let a = 2, and recall that L is a copy of SL2 (11) inside H. We have traces of semisimple elements of
order up to 40, and can use the preimage trick from Section 4.2 to find traces of elements of orders 60 and
120 as well. We would like that the eigenvalues of an element y of order 120 on Vmin uniquely determine the
semisimple class of E7 to which y belongs. This is not true in general for all classes, but will be true for the
particular classes that arise from conspicuous sets of composition factors for Vmin ↓H , by a check using the
preimage trick.
Suppose that u comes from class E7 (a4 ), so that the composition factors of Vmin ↓L are 103 , 8, 6, 42 , 2.
There are, up to field automorphism, twenty conspicuous sets of composition factors for Vmin ↓H , using
semisimple elements of order up to 40; using the preimage trick, we can eliminate seven of these from
contention, as they fail either the trace of an element of order 60 or 120, leaving thirteen. Note that, for
these thirteen remaining sets of composition factors, the eigenvalues of an element y of order 120 on Vmin
determine the semisimple class to which y belongs.
Nine of the thirteen have a 2-dimensional composition factor, and are
141,2 , 1032 , 422 , 221 ,
(1)
1031 , 81 , 61 , 421 , 21 , 22 ,
301,2 , 102 , 101,2 , 41 , 21 ,
(1)
(1)
222,1 , 182,1 , 61 , 421 , 21 ,
(1)
221,2 , 181,2 , 101,2 , 41 , 21 ,
(2)
(1)
222,1 , 142,1 , 101 , 421 , 21 ,
181,2 , 1022 , 62 , 61,2 , 42 , 21 ,
1031 , 81 , 61 , 62,1 , 41 , 21 ,
(2)
221,2 , 181,2 , 102 , 42 , 21 .
(1)
(Recall that 18i,j = 2i ⊗ 9j , 18i,j = 3i ⊗ 6j and 30i,j = 3i ⊗ 10j .) The simple modules with non-trivial
(1)
(1)
extensions with 21 are 102 , 182,1 and 301,2 , and in order for H not to stabilize a 2-space, one of these must
occur with multiplicity 2: thus all cases must fix a 2-space on Vmin except for the first and eighth. However,
even in the first case the {21 , 42 , 102 , 141,2 }-radical of P (102 ) is simply 102 /21 /102 , so there must be a 21
submodule of Vmin ↓H , so that case also fixes a 2-space.
For the eighth case, up to field automorphism we find this inside D6 A1 , by taking an A1 subgroup
acting on the natural modules for the two factors as 91 ⊕ 31 = L(8) ⊕ L(2) and 22 = L(11) respectively.
95
(1)
This yields 182,1 ⊕ 62,1 = L(19) ⊕ L(13), and for the spin module for the D6 term, we need a module
with unipotent action 112 , 6, 4 (one sees this from the entry for D6 (a1 ) in [13, Table 7]) and composition
factors 1021 , 61 , 41 , 22 = L(9)2 , L(5), L(3), L(11) (obtained from the traces of semisimple elements), and so
the restriction of the spin module to this A1 subgroup is
L(9)/L(11)/L(9) ⊕ L(5) ⊕ L(3),
so we apply Lemma 4.12 to see that H is a blueprint for Vmin , as needed.
The remaining four conspicuous sets of composition factors, which have no 2-dimensional composition
factor, are
362,1 , 101 , 61,2 , 41 ,
421,2 , 101 , 41 ,
222,1 , 101 , 102,1 , 81 , 61,2 ,
282,1 , 222,1 , 62,1 .
For the last of these, consider a diagonal A1 inside A1 G2 , acting as 22 = L(11) along A1 and as 71 = L(6)
along the G2 factor. The composition factors on Vmin are L(39), L(21), L(13), matching up with the fourth
case above, so we satisfy the conditions of Lemma 4.12. Consider an A1 subgroup of G2 C3 , acting as
72 = L(66) along the G2 factor and as 61 = L(5) along the C3 factor: the composition factors on Vmin
are L(71), L(9), L(3), matching up with the second case above. Again we apply Lemma 4.12, this time the
second statement, and therefore H is a blueprint for Vmin , as claimed.
For the third case, we find an A1 subgroup Y inside D6 A1 that works: consider Y acting as 21 = L(1)
along the A1 factor, and as 91 ⊕ 32 = L(8) ⊕ L(22) along the second factor. This second A1 is contained
diagonally as an irreducible subgroup inside the product of orthogonal groups O9 × O3 , i.e., B4 A1 , so its
action on the 32-dimensional half-spin module is as the tensor product of the spins. The action on Spin3
must be as 22 = L(11), and the action on the Spin9 has unipotent factors 11, 5 and can be seen to be
111 ⊕ 51 = L(10) ⊕ L(4). Thus the subgroup Y has composition factors
L(21), L(15), L(9), L(7), L(23),
and hence satisfies Lemma 4.12, with the SL2 (121) inside Y having the same factors on Vmin as the third
case.
We are left with 362,1 , 101 , 61,2 , 41 . We firstly note that if there is such a subgroup H then NḠ (H) is
not contained in a positive-dimensional subgroup of Ḡ: to see this, notice that since Vmin ↓H is multiplicityfree and contains a 36-dimensional composition factor, the only positive-dimensional subgroups that could
contain it are G2 C3 and A1 F4 . If H 6 G2 C3 , the module 10 ⊗ 100 has dimension 42, so must restrict to
362,1 ⊕ 61,2 , but 362,1 = 91 ⊗ 42 is not a composition factor of any tensor product of a 6-dimensional module
and a 7-dimensional module, a contradiction. Since A1 F4 acts on Vmin with factors 1 ⊗ 0001 and 3 ⊗ 0000,
we see that if H 6 A1 F4 then the projection of H along A1 must act on the natural module as 21 , so that
L(3) restricts to H as 41 . However, again 362,1 is not a composition factor of any tensor product of 21 and a
module of dimension at most 26, so H cannot lie in A1 F4 either. Thus H, and by extension NḠ (H), cannot
lie in a positive-dimensional subgroup of G.
We therefore must have that NḠ (H) contains SL2 (121) with index at most 2 by Proposition 3.3. Recalling
that y is an element of order 120, chosen so that the eigenvalues of y on 21 are ζ ±1 for ζ a primitive 120th
root of unity, we consider the elements of order 120 in a maximal torus of G squaring to y 2 , noting that the
ζ 2 - and ζ 6 -eigenspaces of y 2 on Vmin are both 3-dimensional and coincide with the ζ- and ζ 3 -eigenspaces of
y respectively. We thus look for elements that square to y 2 and preserve the ζ 2 - and ζ 6 -eigenspaces of y 2 ;
of course, y is one of these elements, and we find four elements with 3-dimensional ζ- and ζ 3 -eigenspaces
(and therefore four with 3-dimensional (−ζ)- and (−ζ 3 )-eigenspaces), which together generate a subgroup
96
Z120 × Z2 × Z2 of the torus. Thus the stabilizer in G of the 4-dimensional submodule 41 of Vmin contains H
with index at least 4, a contradiction, and so H doesn’t exist by the above proof. This completes the proof
of the proposition when u comes from class E7 (a4 ).
Suppose that u comes from class D6 , and that the composition factors of Vmin ↓L are 10, 67 , 4: the trace
of an element of order 5 is 6, and this is enough to seriously restrict the possibilities. Using other semisimple
elements of order up to 20, we find up to field automorphism a single conspicuous set of composition factors,
namely 101 , 671 , 41 , and this must be semisimple as there are no non-trivial extensions between the factors.
However, this is incompatible with the action of u, so H cannot embed with this restriction to L. The other
set of composition factors are the same as for E7 (a4 ), which we have already considered above.
Suppose that u comes from E7 (a3 ), so that Vmin ↓L has composition factors 103 , 63 , 4, 22 . Checking traces
of elements of order up to 40 yields, up to field automorphism, only five conspicuous sets of composition
factors, which are
1031 , 631 , 41 , 221 ,
1032 , 632 , 42 , 221 ,
1031 , 102,1 , 621 , 21 , 22 ,
101 , 1022 , 61 , 622 , 61,2 , 21 ,
221,2 , 102 , 101,2 , 622 , 21 .
The last two of these fail the traces of elements of order 60, so do not exist. The 21 -pressures of the remaining
three are −2, 1 and −1 respectively, as only 102 from these simple modules has an extension with 21 , so
only the second need not fix a 2-space. In this case the {21 , 42 , 62 , 102 }-radical of P (102 ) is 102 /21 /102 , so
Vmin ↓H has a 2-dimensional submodule.
We now have that p > 13, for which we only need consider a = 1. For p = 17, 19, 23, if H is not in
a member of X σ , we will prove that H fixes an sl2 -subalgebra, and for p = 19 we will also get a Serre
embedding (see Definition 4.7).
We begin with p = 13.
Proposition 12.7 Suppose that p = 13 and a = 1. If H is not a blueprint for Vmin then NḠ (H) either fixes
a 2-space or 4-space, and in both cases lies in a member of X σ .
Proof: There are three possibilities for the action of u on Vmin , namely cases (xxii), (xxiii) and (xxiv) from
the list above, with the last of these being semiregular.
In the first case, u acts as 134 , 14 , and so Vmin ↓H is the sum of four modules of dimension 14, which are
2/12,
4/10,
6/8,
8/6,
10/4,
12/2.
There are only two conspicuous sets of composition factors consisting of dual pairs of these, and they are
2/12 ⊕ 12/2 ⊕ 4/10 ⊕ 10/4
and
4/10 ⊕ 10/4 ⊕ 6/8 ⊕ 8/6.
Neither of these has a corresponding set of composition factors on L(G), so H does not embed in G with u
coming from class E6 .
If u comes from class E7 (a3 ) then it acts as 132 , 12, 10, 6, 2. The single block of size 2 must come from a
self-dual indecomposable module of dimension congruent to 2 modulo 13, and the two of these are 2 itself
– so H fixes a 2-space on Vmin – or a 28-dimensional module 6, 8/6, 8; from here, the blocks of sizes 6 and
10 must come from simple summands and the 12 comes either from a 12 or a 6/6, yielding two possible
sets of composition factors, neither of which is conspicuous, having trace −1 for an element of order 3. This
completes the proof for the second action of u.
97
The final unipotent class to consider is the semiregular E7 (a2 ), acting as 134 , 4 on the minimal module.
The single block of size 4 comes either from a summand 4 or from the indecomposable module
4, 6, 8, 10/4, 6, 8, 10, which is conspicuous, but we saw above that it has no corresponding set of factors
for L(G). Hence Vmin ↓H has a 4 as a summand, with two projective indecomposable summands. The
conspicuous such sets of composition factors yield
P (2) ⊕ P (4) ⊕ 4,
P (12) ⊕ P (10) ⊕ 4,
P (10) ⊕ P (8) ⊕ 4,
P (6) ⊕ P (4) ⊕ 4.
The first and second of these cannot occur because they do not have corresponding factors on L(G).
In the fourth case we again switch to L(G), and find two corresponding sets of composition factors,
namely
13, 11, 92, 79 , 53 , 34 , 1,
113 , 93 , 75 , 54 , 36 .
The action of u on L(G) must be 1310 , 3, and the single 3 in this action comes from a summand isomorphic
to either 3 or 5, 7, 9/5, 7, 9. The first case cannot occur as the single 1 must lie in a P (11), but this cannot
occur. In the second case, no trivials means there can be no P (11)s, so we must have P (3)⊕3 in L(G) ↓H ,
but then there are no 3s or 9s remaining, so the summand contributing the 3 to the action of u cannot occur,
a contradiction. Thus H cannot embed with these factors either.
In the third case, there are again two corresponding sets of composition factors on L(G), namely
13, 113, 92 , 77 , 5, 34 , 13 ,
115 , 93 , 73 , 52 , 36 , 12 .
In the first of these, the single 5 means one has no P (5) and at most one P (7), but these are the only two
projectives containing 7, so we cannot use up the seven 7s. The second case does have a unique possibility,
however, of
P (11)⊕2 ⊕ P (9) ⊕ P (7) ⊕ P (3) ⊕ 3.
Although it has 3 as a summand, the presence of a P (3) means that we cannot guarantee that it is an
sl2 -subalgebra of L(G), although the 3 ⊕ 3 in the socle of L(G) ↓H does form a subalgebra.
Let x denote an element of order 14 in H and let ζ be a primitive 14th root of unity, arranged so that
the eigenvalues of x on 4 are ζ ±1 , ζ ±3 . The eigenvalues of x on Vmin are
(−1)8 , (ζ ±1 )9 , (ζ ±3 )8 , (ζ ±5 )7 .
Let θ denote a primitive 28th root of unity with θ2 = ζ, and we find x̂ ∈ G such that x̂2 = x and x̂ has
eigenvalues
(±i)4 , (θ±1 )9 , (θ±3 )8 , (θ±5 )5 , (−θ±5 )2
on Vmin . This stabilizes the eigenspaces intersecting the 4, and so x̂ stabilizes the 4-space stabilized by H.
The action of x̂ on L(G) has a 21-dimensional 1-eigenspace but −1 is not an eigenvalue of x̂: if H̄ =
hH, x̂i/Z(G) ∼
= PSL2 (13) then all of the composition factors of L(G) ↓ must be the plus-type factors, in the
H̄
notation of Section 4.4. Thus we extend the action of H/Z(G) ∼
= PSL2 (13) on L(G) uniquely to an action
of PGL2 (13), and doing so yields a trace of 11 on L(G) for an element of order 4 in PGL2 (13) \ PSL2 (13).
However, 11 is not the trace of an element of order either 4 or 8 (in case it powers to the central involution)
in G on L(G), a contradiction, so that hH, x̄i, which stabilizes the 4-space, is not almost simple modulo
Z(G). Thus NḠ (H) is contained inside a member of X σ by Proposition 3.3, as needed.
Proposition 12.8 Suppose that p = 17 and a = 1. If H is not a blueprint for Vmin then either NḠ (H)
stabilizes a 4-space on Vmin whose stabilizer not NḠ (H), and hence NḠ (H) is contained in a member of X σ ,
or H stabilizes an sl2 -subalgebra of L(G).
98
Proof: There are two non-generic unipotent classes, cases (xxv) and (xxvi) above, where there are single
Jordan blocks of sizes 4, 6, 8, 10, 16. Apart from the simple modules of dimension congruent to 4, 6, 8, 10, 16
modulo 17, the self-dual indecomposable modules congruent to those dimensions have dimensions 72, 108,
144, 114 and 16 respectively, so only the 16 might not come from a simple summand.
For u belonging to class E7 (a2 ), so acting as 172 , 10, 8, 4, we therefore have a single projective plus
10 ⊕ 8 ⊕ 4. Applying the traces of semisimple elements yields two possibilities:
P (16) ⊕ 10 ⊕ 8 ⊕ 4
and
P (10) ⊕ 10 ⊕ 8 ⊕ 4.
The second of these does not yield an action on L(G) as the traces do not match up, but the first of these
has a unique set of composition factors on L(G) which yields
17 ⊕ P (15) ⊕ P (11) ⊕ 15 ⊕ 11 ⊕ 9 ⊕ 7 ⊕ 3⊕2 .
The subspace 3⊕2 is an H-invariant Lie subalgebra of L(G), but we proceed as in the case of p = 13, finding
an element of order 36 that preserves the 4-dimensional submodule.
Let x be an element of H of order 18 and ζ be a primitive 18th root of unity, arranging our choices so
that the eigenvalues of x on 4 are ζ ±1 and ζ ±3 . The eigenvalues of x on Vmin are
(−1)6 , (ζ ±1 )6 , (ζ ±3 )7 , (ζ ±5 )6 , (ζ ±7 )6 .
Letting θ denote a primitive 36th root of unity squaring to ζ, we find an element x̂ of order 36 in G with
x̂2 = x and with eigenvalues on Vmin given by
(±i)3 , (θ±1 )6 , (θ±3 )7 , (θ±5 )5 , −θ±5 , (θ±7 )4 , (−θ±7 )2 .
We see immediately that x̂ preserves the 4-space stabilized by H, and we proceed as in Proposition 12.7,
noting that −1 is not an eigenvalue of x̂ on L(G), and so there is a unique extension of PSL2 (17) to PGL2 (17).
However, an element of order 6 in PGL2 (17) \ PSL2 (17) has trace 8 on L(G). While 8 is the trace of an
element of order 6 in G (but not an element of order 12 powering to the central involution), the square of
this element needs trace 34 on L(G), which is not a trace of an element of order 3 in H, a contradiction. We
then proceed as in Proposition 12.7.
For u belonging to class E7 (a1 ), so acting as 172 , 16, 6, the 6 must come from a simple summand, but
the 16 comes from either a simple summand or 8/8. Thus our embedding of H is either a single projective
plus 8/8 ⊕ 6 or a single projective plus 16 ⊕ 6. Using traces, the two options are
P (12) ⊕ 16 ⊕ 6
and
P (4) ⊕ 16 ⊕ 6.
The second of these has no corresponding set of composition factors on L(G), but the first has a unique set,
which implies that L(G) ↓H is
17 ⊕ P (15) ⊕ P (11) ⊕ P (7) ⊕ 11 ⊕ 3,
and so the 3 is an sl2 -subalgebra by Proposition 4.17 and we complete the proof.
Proposition 12.9 Suppose that p = 19 and a = 1. If H is not a blueprint for Vmin then H stabilizes an
sl2 -subalgebra of L(G), or H is a Serre embedding.
99
Proof: When p = 19, there are two non-generic unipotent classes, E7 (a1 ) and E7 , cases (xxvii) and (xxviii)
above. As p ≡ 3 mod 4 there is a unique self-dual indecomposable module congruent to any given integer
modulo p. For E7 (a1 ) the 12 and 6 in the action of u must therefore come from simple summands, leaving a
single projective module of dimension 38. Only two possibilities yield conspicuous sets of composition factors,
namely P (16) ⊕ 12 ⊕ 6 and P (4) ⊕ 12 ⊕ 6. The second of these has no corresponding set of composition
factors on L(G), with the first of these yielding the unique action
P (15) ⊕ P (11) ⊕ 19 ⊕ 17 ⊕ 11 ⊕ 7 ⊕ 3,
with the 3 being an sl2 -subalgebra of L(G) by Proposition 4.17.
The remaining case is u coming from the regular class, where as with the E7 (a1 ) case the 10 from the
action of u must yield a simple summand, with the rest projective. There are again two conspicuous sets of
composition factors for Vmin ↓H , coming from P (4) ⊕ 18 and P (10) ⊕ 18. The first has no corresponding set
of composition factors on L(G), and the second has a single set, which since u is projective on L(G), must
be arranged so that L(G) ↓H is
19 ⊕ P (15) ⊕ P (11) ⊕ P (3).
While the 3-dimensional submodule is a subalgebra of L(G), it is not obviously an sl2 -subalgebra because
we cannot apply Proposition 4.17. This is a Serre embedding as defined in Definition 4.7, as needed.
The last case is p = 23 and the regular unipotent class, to conclude this section.
Proposition 12.10 Suppose that p = 23 and a = 1. If H is not a blueprint for Vmin then H stabilizes an
sl2 -subalgebra of L(G).
Proof: The only non-generic class for p = 23 and Vmin is the regular class, with blocks 232 , 10, case (xxix)
above. The 10 in the action of u must come from a simple summand, leaving a single projective module of
dimension 46. Only two possibilities yield conspicuous sets of composition factors, namely 20, 18, 10, 42 and
182 , 10, 6, 4. The first of these has no corresponding set of composition factors on L(G), with the second of
these yielding the unique action
P (19) ⊕ P (11) ⊕ 23 ⊕ 15 ⊕ 3,
with the 3 being an sl2 -subalgebra of L(G) by Proposition 4.17.
100
A
Actions of maximal positive-dimensional subgroups on minimal
and adjoint modules
In this appendix we collate information on the actions of the reductive and parabolic maximal subgroups of
positive dimension on the minimal and adjoint modules for the algebraic groups F4 , E6 and E7 that we have
used in the text, other than those in Lemmas 2.4 and 2.5. These have been documented in many places, but
we give them here as well for ease of reference.
We need information for F4 and E6 in characteristic 3, and for E6 in characteristics 7 and 11. We
list the composition factors of every maximal closed, connected subgroup of positive dimension in these
characteristics, taken from [20], on Vmin , and L(G) for G = F4 . We list the reductive subgroups first, and
then the parabolics. Write M ± to mean a module M and its dual M ∗ .
We begin with the table for F4 in characteristic 3.
Subgroup of F4 for p = 3
Factors on Vmin
Factors on L(G)
B4
1000, 0001
0100, 0001
Ã1 C3
(1, 100), (0, 010)
(2, 000), (0, 200), (1, 001)
A2 Ã2
(10, 10), (01, 01), (00, 11)
(11, 00), (00, 11), (10, 02), (01, 20), (00, 00)2
A1 G2
(2, 10), (4, 00)
(2, 00), (0, 01), (0, 10)2, (4, 10)
B3
100, 0012, 0002
1002 , 010, 0012, 000
C3
1002 , 010
200, 0012, 0003
(10, 1)± , (10, 0)± ,
(11, 0), (10, 2)±, (10, 1)± , (10, 0)± ,
(00, 2), (00, 1)2
(00, 2), (00, 1)2, (00, 0)2
A2 Ã1
(10, 1)± , (10, 0)± , (11, 0)
Ã2 A1
(20, 1)± , (20, 0)± , (11, 0),
(00, 2), (00, 1)2, (00, 0)2
Next, the subgroups of E6 in characteristic 3.
Subgroup of E6 for p = 3
Factors on Vmin
A5 A1
(λ4 , 0), (λ1 , 1)
A2 A2 A2
(10, 01, 00), (00, 10, 01), (01, 00, 10)
F4
0001, 00002
C4
0100
G2 A2
(10, 10), (00, 02)
G2 (2 classes)
20
D5
λ1 , λ4 , 0
A5
λ21 , λ4
A4 A1
(1000, 1), (0001, 0), (0010, 0), (1, 0000)
A2 A2 A1
(10, 01, 1), (01, 00, 1), (00, 10, 1), (01, 00, 0), (00, 10, 0)
Finally, the subgroups of E7 in characteristics 7 and 11.
101
Subgroup of E7 for p = 7, 11
Factors on Vmin
D6 A1
(λ1 , 1), (λ5 , 0)
A7
λ±
1
A5 A2
(λ1 , 10)± , (λ3 , 00)
C3 G2
(001, 00), (100, 10)
G2 A1
(01, 1), (10, 3)
F4 A1
(0001, 1), (3, 0000)
A2
06±
A1 A1
(6, 3), (4, 1), (2, 5)
E6
2
λ±
1 ,0
D6
λ21 , λ5
A6
±
λ±
1 , λ2
A5 A1
(λ1 , 10)± , (λ3 , 00)
A4 A2
(10, 1000)±, (10, 0000)±, (00, 0100)±
A3 A2 A1
(000, 10, 1)±, (010, 00, 1), (100, 10, 0)±, (100, 00, 0)±
102
References
[1] Jonathan Alperin, Projective modules for SL(2, 2n ), J. Pure Appl. Algebra 15 (1979), 219–234.
[2]
, Local representation theory, Cambridge Studies in Advanced Mathematics, vol. 11, Cambridge
University Press, Cambridge, 1996.
[3] Henning Haahr Andersen, Jens Jørgensen, and Peter Landrock, The projective indecomposable modules
of SL(2, pn ), Proc. London Math. Soc. 46 (1983), 38–52.
[4] Michael Aschbacher, The maximal subgroups of E6 , preprint, 170pp.
[5] Michael Aschbacher and Leonard Scott, Maximal subgroups of finite groups, J. Algebra 92 (1985), 44–80.
[6] Richard Block, Trace forms on Lie algebras, Canad. J. Math. 14 (1962), 553–564.
[7] Alexandre Borovik, Structure of finite subgroups of simple algebraic groups, Algebra i Logika 28 (1989),
249–279 (Russian).
[8] Arjeh Cohen, Martin Liebeck, Jan Saxl, and Gary Seitz, The local maximal subgroups of exceptional
groups of Lie type, finite and algebraic, Proc. London Math. Soc. 64 (1992), 21–48.
[9] David A. Craven, Alternating subgroups of exceptional groups of Lie type, submitted.
[10]
, Subspace stabilizers and maximal subgroups of exceptional groups of Lie type, submitted.
[11]
, On tensor products of simple modules for simple groups, Algebr. Represent. Theory 126 (2013),
377–404.
[12] Stephen Doty and Anne Henke, Decomposition of tensor products of modular irreducibles for SL2 , Quart.
J. Math. 56 (2005), 189–207.
[13] Ross Lawther, Jordan block sizes of unipotent elements in exceptional algebraic groups, Comm. Algebra
23 (1995), 4125–4156.
[14]
, Unipotent classes in maximal subgroups of exceptional algebraic groups, J. Algebra 322 (2009),
270–293.
[15] Martin Liebeck and Jan Saxl, On the orders of maximal subgroups of the finite exceptional groups of
Lie type, Proc. London Math. Soc. 55 (1987), 299–330.
[16] Martin Liebeck, Jan Saxl, and Donna Testerman, Simple subgroups of large rank in groups of Lie type,
Proc. London Math. Soc. 72 (1996), 425–457.
[17] Martin W. Liebeck and Gary M. Seitz, Maximal subgroups of exceptional groups of Lie type, finite and
algebraic, Geom. Dedicata 35 (1990), 353–387.
[18]
, On the subgroup structure of exceptional groups of Lie type, Trans. Amer. Math. Soc. 350
(1998), 3409–3482.
[19]
, On finite subgroups of exceptional algebraic groups, J. reine angew. Math. 515 (1999), 25–72.
[20]
, The maximal subgroups of positive dimension in exceptional algebraic groups, Mem. Amer.
Math. Soc. 169 (2004), no. 802, vi+227.
103
[21]
, Maximal subgroups of large rank in exceptional groups of Lie type, J. London. Math. Soc. 71
(2005), 345–361.
[22] Alastair Litterick, Finite simple subgroups of exceptional algebraic groups, Ph.D. thesis, Imperial College,
London, 2013.
[23]
, Finite simple subgroups of exceptional algebraic groups, Mem. Amer. Math. Soc., to appear.,
2015.
[24] Kay Magaard, The maximal subgroups of the Chevalley groups F4 (F ) where F is a finite or algebraically
closed field of characteristic not equal to 2 or 3, Ph.D. thesis, California Institute of Technology, 1990.
[25] Alexander Ryba, Short proofs of embeddings into exceptional groups of Lie type, J. Algebra 249 (2002),
402–418.
[26] Gary M. Seitz, Unipotent elements, titling modules, and saturation, Invent. Math. 141 (2000), 467–502.
[27] Jean-Pierre Serre, Exemples de plongements des groupes PSL2 (Fp ) dans des groupes de Lie simples,
Invent. Math. 124 (1996), 525–562.
[28] David Stewart and Adam Thomas, The Jacobson–Morozov theorem and complete reducibility of Lie
subalgebras, submitted.
104
| 4math.GR
|
Whodunnit? Crime Drama as a Case for Natural Language Understanding
arXiv:1710.11601v1 [cs.CL] 31 Oct 2017
Lea Frermann Shay B. Cohen Mirella Lapata
Institute for Language, Cognition and Computation
School of Informatics, University of Edinburgh
10 Crichton Street, Edinburgh EH8 9AB
[email protected]
[email protected]
[email protected]
Abstract
In this paper we argue that crime drama exemplified in television programs such as CSI:
Crime Scene Investigation is an ideal testbed
for approximating real-world natural language
understanding and the complex inferences associated with it. We propose to treat crime
drama as a new inference task, capitalizing on
the fact that each episode poses the same basic question (i.e., who committed the crime)
and naturally provides the answer when the
perpetrator is revealed. We develop a new
dataset1 based on CSI episodes, formalize perpetrator identification as a sequence labeling
problem, and develop an LSTM-based model
which learns from multi-modal data. Experimental results show that an incremental inference strategy is key to making accurate
guesses as well as learning from representations fusing textual, visual, and acoustic input.
1
Introduction
The success of neural networks in a variety of applications (Sutskever et al., 2014; Vinyals et al.,
2015) and the creation of large-scale datasets have
played a critical role in advancing machine understanding of natural language on its own or together
with other modalities. The problem has assumed
several guises in the literature such as reading comprehension (Richardson et al., 2013; Rajpurkar et
al., 2016), recognizing textual entailment (Bowman
et al., 2015; Rocktäschel et al., 2016), and notably
question answering based on text (Hermann et al.,
1
Our dataset is available at https://github.com/
EdinburghNLP/csi-corpus.
2015; Weston et al., 2015), images (Antol et al.,
2015), or video (Tapaswi et al., 2016).
In order to make the problem tractable and
amenable to computational modeling, existing approaches study isolated aspects of natural language
understanding. For example, it is assumed that understanding is an offline process, models are expected to digest large amounts of data before being
able to answer a question, or make inferences. They
are typically exposed to non-conversational texts or
still images when focusing on the visual modality,
ignoring the fact that understanding is situated in
time and space and involves interactions between
speakers. In this work we relax some of these simplifications by advocating a new task for natural language understanding which is multi-modal, exhibits
spoken conversation, and is incremental, i.e., unfolds sequentially in time.
Specifically, we argue that crime drama exemplified in television programs such as CSI: Crime Scene
Investigation can be used to approximate real-world
natural language understanding and the complex inferences associated with it. CSI revolves around a
team of forensic investigators trained to solve criminal cases by scouring the crime scene, collecting
irrefutable evidence, and finding the missing pieces
that solve the mystery. Each episode poses the same
“whodunnit” question and naturally provides the answer when the perpetrator is revealed. Speculation
about the identity of the perpetrator is an integral
part of watching CSI and an incremental process:
viewers revise their hypotheses based on new evidence gathered around the suspect/s or on new inferences which they make as the episode evolves.
We formalize the task of identifying the perpetrator in a crime series as a sequence labeling problem.
Like humans watching an episode, we assume the
model is presented with a sequence of inputs comprising information from different modalities such
as text, video, or audio (see Section 4 for details).
The model predicts for each input whether the perpetrator is mentioned or not. Our formulation generalizes over episodes and crime series. It is not specific to the identity and number of persons committing the crime as well as the type of police drama
under consideration. Advantageously, it is incremental, we can track model predictions from the
beginning of the episode and examine its behavior,
e.g., how often it changes its mind, whether it is consistent in its predictions, and when the perpetrator is
identified.
We develop a new dataset based on 39 CSI
episodes which contains goldstandard perpetrator
mentions as well as viewers’ guesses about the
perpetrator while each episode unfolds. The sequential nature of the inference task lends itself naturally to recurrent network modeling. We
adopt a generic architecture which combines a
one-directional long-short term memory network
(Hochreiter and Schmidhuber, 1997) with a softmax
output layer over binary labels indicating whether
the perpetrator is mentioned. Based on this architecture, we investigate the following questions:
1. What type of knowledge is necessary for performing the perpetrator inference task? Is the
textual modality sufficient or do other modalities (i.e., visual and auditory input) also play a
role?
2. What type of inference strategy is appropriate?
In other words, does access to past information
matter for making accurate inferences?
3. To what extent does model behavior simulate humans? Does performance improve over
time and how much of an episode does the
model need to process in order to make accurate guesses?
Experimental results on our new dataset reveal
that multi-modal representations are essential for the
task at hand boding well with real-world natural language understanding. We also show that an incremental inference strategy is key to guessing the perpetrator accurately although the model tends to be
less consistent compared to humans. In the remainder, we first discuss related work (Section 2), then
present our dataset (Section 3) and formalize the
modeling problem (Section 4). We describe our experiments in Section 5.
2
Related Work
Our research has connections to several lines of
work in natural language processing, computer vision, and more generally multi-modal learning. We
review related literature in these areas below.
Language Grounding Recent years have seen increased interest in the problem of grounding language in the physical world. Various semantic space
models have been proposed which learn the meaning
of words based on linguistic and visual or acoustic input (Bruni et al., 2014; Silberer et al., 2016;
Lazaridou et al., 2015; Kiela and Bottou, 2014).
A variety of cross-modal methods which fuse techniques from image and text processing have also
been applied to the tasks of generating image descriptions and retrieving images given a natural language query (Vinyals et al., 2015; Xu et al., 2015;
Karpathy and Fei-Fei, 2015). Another strand of research focuses on how to explicitly encode the underlying semantics of images making use of structural representations (Ortiz et al., 2015; Elliott and
Keller, 2013; Yatskar et al., 2016; Johnson et al.,
2015). Our work shares the common goal of grounding language in additional modalities. Our model is,
however, not static, it learns representations which
evolve over time.
Video Understanding Work on video understanding has assumed several guises such as generating descriptions for video clips (Venugopalan et al.,
2015a; Venugopalan et al., 2015b), retrieving video
clips with natural language queries (Lin et al., 2014),
learning actions in video (Bojanowski et al., 2013),
and tracking characters (Sivic et al., 2009). Movies
have also been aligned to screenplays (Cour et al.,
2008), plot synopses (Tapaswi et al., 2015), and
books (Zhu et al., 2015) with the aim of improving scene prediction and semantic browsing. Other
work uses low-level features (e.g., based on face detection) to establish social networks of main characters in order to summarize movies or perform genre
Peter Berglund:
Grissom doesn't look
You're still going to have to worried.
convince a jury that I killed He takes his gloves off and
two strangers for no reason. puts them on the table.
Grissom:
You ever been to the
theater Peter?
There 's a play called six
degrees of separation.
It 's about how all the
Camera holds on Peter
people in the world are
Berglund's worried look.
connected to each other by
no more than six people.
All it takes to connect you to
the victims is one degree.
Figure 1: Excerpt from a CSI script (Episode 03, Season 03: “Let the Seller Beware”). Speakers are shown in bold,
spoken dialog in normal font, and scene descriptions in italics. Gold-standard entity mention annotations are in color.
Perpetrator mentions (e.g., Peter Berglund) are in green, while words referring to other entities are in red.
classification (Rasheed et al., 2005; Sang and Xu,
2010; Dimitrova et al., 2000). Although visual features are used mostly in isolation, in some cases they
are combined with audio in order to perform video
segmentation (Boreczky and Wilcox, 1998) or semantic movie indexing (Naphide and Huang, 2001).
A few datasets have been released recently
which include movies and textual data. MovieQA
(Tapaswi et al., 2016) is a large-scale dataset
which contains 408 movies and 14,944 questions,
each accompanied with five candidate answers,
one of which is correct. For some movies, the
dataset also contains subtitles, video clips, scripts,
plots, and text from the Described Video Service
(DVS), a narration service for the visually impaired.
MovieDescription (Rohrbach et al., 2017) is a related dataset which contains sentences aligned to
video clips from 200 movies. Scriptbase (Gorinski
and Lapata, 2015) is another movie database which
consists of movie screenplays (without video) and
has been used to generate script summaries.
In contrast to the story comprehension tasks envisaged in MovieQA and MovieDescription, we focus on a single cinematic genre (i.e., crime series),
and have access to entire episodes (and their corresponding screenplays) as opposed to video-clips or
DVSs for some of the data. Rather than answering
multiple factoid questions, we aim to solve a single
problem, albeit one that is inherently challenging to
both humans and machines.
Question Answering A variety of question answering tasks (and datasets) have risen in popularity
in recent years. Examples include reading compre-
hension, i.e., reading text and answering questions
about it (Richardson et al., 2013; Rajpurkar et al.,
2016), open-domain question answering, i.e., finding the answer to a question from a large collection
of documents (Voorhees and Tice, 2000; Yang et al.,
2015), and cloze question completion, i.e., predicting a blanked-out word of a sentence (Hill et al.,
2015; Hermann et al., 2015). Visual question answering (VQA; Antol et al. (2015)) is a another related task where the aim is to provide a natural language answer to a question about an image.
Our inference task can be viewed as a form of
question answering over multi-modal data, focusing on one type of question. Compared to previous
work on machine reading or visual question answering, we are interested in the temporal characteristics
of the inference process, and study how understanding evolves incrementally with the contribution of
various modalities (text, audio, video). Importantly,
our formulation of the inference task as a sequence
labeling problem departs from conventional question answering allowing us to study how humans and
models alike make decisions over time.
3
The CSI Dataset
In this work, we make use of episodes of the U.S.
TV show “Crime Scene Investigation Las Vegas”
(henceforth CSI), one of the most successful crime
series ever made. Fifteen seasons with a total of 337
episodes were produced over the course of fifteen
years. CSI is a procedural crime series, it follows
a team of investigators employed by the Las Vegas
Police Department as they collect and evaluate ev-
per case
episodes with one case
episodes with two cases
total number of cases
sentences
sentences with perpetrator
scene descriptions
spoken utterances
characters
murder
accident
type of crime
suicide
other
19
20
59
min
228
0
64
144
8
51
4
2
2
max
1209
267
538
778
38
avg
689
89
245
444
20
Table 1: Statistics on the CSI data set. The type of crime
was identified by our annotators via a multiple-choice
questionnaire (which included the option “other”). Note
that accidents may also involve perpetrators.
idence to solve murders, combining forensic police
work with the investigation of suspects.
We paired official CSI videos (from seasons 1–5)
with screenplays which we downloaded from a website hosting TV show transcripts.2 Our dataset comprises 39 CSI episodes, each approximately 43 minutes long. Episodes follow a regular plot, they begin
with the display of a crime (typically without revealing the perpetrator) or a crime scene. A team of five
recurring police investigators attempt to reconstruct
the crime and find the perpetrator. During the investigation, multiple (innocent) suspects emerge, while
the crime is often committed by a single person, who
is eventually identified and convicted. Some CSI
episodes may feature two or more unrelated cases.
At the beginning of the episode the CSI team is
split and each investigator is assigned a single case.
The episode then alternates between scenes covering each case, and the stories typically do not overlap. Figure 1 displays a small excerpt from a CSI
screenplay. Readers unfamiliar with script writing
conventions should note that scripts typically consist
of scenes, which have headings indicating where the
scene is shot (e.g., inside someone’s house). Character cues preface the lines the actors speak (see
boldface in Figure 1), and scene descriptions explain
what the camera sees (see second and fifth panel in
Figure 1).
Screenplays were further synchronized with the
2
http://transcripts.foreverdreaming.org/
video using closed captions which are time-stamped
and provided in the form of subtitles as part of
the video data. The alignment between screenplay
and closed captions is non-trivial, since the latter
only contain dialogue, omitting speaker information
or scene descriptions. We first used dynamic time
warping (DTW; Myers and Rabiner (1981)) to approximately align closed captions with the dialogue
in the scripts. And then heuristically time-stamped
remaining elements of the screenplay (e.g., scene
descriptions), allocating them to time spans between
spoken utterances. Table 1 shows some descriptive statistics on our dataset, featuring the number
of cases per episode, its length (in terms of number
of sentences), the type of crime, among other information.
The data was further annotated, with two goals
in mind. Firstly, in order to capture the characteristics of the human inference process, we recorded
how participants incrementally update their beliefs
about the perpetrator. Secondly, we collected goldstandard labels indicating whether the perpetrator is
mentioned. Specifically, while a participant watches
an episode, we record their guesses about who the
perpetrator is (Section 3.1). Once the episode is finished and the perpetrator is revealed, the same participant annotates entities in the screenplay referring
to the true perpetrator (Section 3.2).
3.1
Eliciting Behavioral Data
All annotations were collected through a webinterface. We recruited three annotators, all postgraduate students and proficient in English, none of
them regular CSI viewers. We obtained annotations
for 39 episodes (comprising 59 cases).
A snapshot of the annotation interface is presented in Figure 2. The top of the interface provides
a short description of the episode, i.e., in the form of
a one-sentence summary (carefully designed to not
give away any clues about the perpetrator). Summaries were adapted from the CSI season summaries
available in Wikipedia.3 The annotator watches the
episode (i.e., the video without closed captions) as
a sequence of three minute intervals. Every three
minutes, the video halts, and the annotator is pre3
See e.g., https://en.wikipedia.org/wiki/
CSI:_Crime_Scene_Investigation_(season_1).
Number of cases: 2
Case 1: Grissom, Catherine, Nick and Warrick investigate
when a wealthy couple is murdered at their house.
Case 2: Meanwhile Sara is sent to a local high school where a
cheerleader was found eviscerated on the football field.
(
It
’s
a shell casing
.
)
Perpetrator
Suspect
Other
GRISSOM moves his light to
Perpetrator
Suspect
Other
the canopy below
Figure 3: Annotation interface (second pass): after
watching the episode, the annotator indicates for each
word whether it refers to the perpetrator.
Screenplay
Perpetrator
mentioned?
Relates
to case
1/2/none?
(Nick cuts the canopy around
MONICA NEWMAN.)
Nick okay, Warrick, hit it
(WARRICK starts the crane
support under the awning to remove the body and the canopy
area that NICK cut.)
Nick white female, multiple
bruising . . . bullet hole to the
temple doesn’t help
Nick .380 auto on the side
Warrick yeah, somebody manhandled her pretty good before
they killed her
Figure 2: Annotation interface (first pass): after watching three minutes of the episode, the annotator indicates
whether she believes the perpetrator has been mentioned.
sented with the screenplay corresponding to the part
of the episode they have just watched. While reading through the screenplay, they must indicate for
every sentence whether they believe the perpetrator
is mentioned. This way, we are able to monitor how
humans create and discard hypotheses about perpetrators incrementally. As mentioned earlier, some
episodes may feature more than one case. Annotators signal for each sentence, which case it belongs
to or whether it is irrelevant (see the radio buttons
in Figure 2). In order to obtain a more fine-grained
picture of the human guesses, annotators are additionally asked to press a large red button (below the
video screen) as soon as they “think they know who
the perpetrator is”, i.e., at any time while they are
watching the video. They are allowed to press the
button multiple times throughout the episode in case
they change their mind.
Even though the annotation task just described
reflects individual rather than gold-standard behavior, we report inter-annotator agreement (IAA) as
a means of estimating variance amongst participants. We computed IAA using Cohen’s (1960)
Kappa based on three episodes annotated by two
participants. Overall agreement on this task (second column in Figure 2) is 0.74. We also measured
percent agreement on the minority class (i.e., sentences tagged as “perpetrator mentioned”) and found
it to be reasonably good at 0.62, indicating that despite individual differences, the process of guessing
the perpetrator is broadly comparable across participants. Finally, annotators had no trouble distinguishing which utterances refer to which case (when
the episode revolves around several), achieving an
IAA of κ = 0.96.
3.2
Gold Standard Mention Annotation
After watching the entire episode, the annotator
reads through the screenplay for a second time, and
tags entity mentions, now knowing the perpetrator.
Each word in the script has three radio buttons attached to it, and the annotator selects one only if a
word refers to a perpetrator, a suspect, or a character
who falls into neither of these classes (e.g., a police investigator or a victim). For the majority of
words, no button will be selected. A snapshot of
our interface for this second layer of annotations is
shown in Figure 3. To ensure consistency, annotators were given detailed guidelines about what constitutes an entity. Examples include proper names
and their titles (e.g., Mr Collins, Sgt. O’ Reilly),
pronouns (e.g., he, we), and other referring expressions including nominal mentions (e.g., let’s arrest
the guy with the black hat).
Inter-annotator agreement based on three
episodes and two annotators was κ = 0.90 on
the perpetrator class and κ = 0.89 on other entity annotations (grouping together suspects with
other entities). Percent agreement was 0.824 for
perpetrators and 0.823 for other entities. The high
agreement indicates that the task is well-defined and
the elicited annotations reliable. After the second
pass, various entities in the script are disambiguated
in terms of whether they refer to the perpetrator or
other individuals.
Note that in this work we do not use the tokenlevel gold standard annotations directly. Our model
is trained on sentence-level annotations which we
obtain from token-level ones, under the assumption
that a sentence mentions the perpetrator if it contains
a token that does.
4
Figure 4: Overview of the perpetrator prediction task.
The model receives input in the form of text, images, and
audio. Each modality is mapped to a feature representation. Feature representations are fused and passed to an
LSTM which predicts whether a perpetrator is mentioned
(label l = 1) or not (l = 0).
Model Description
We formalize the problem of identifying the perpetrator in a crime series episode as a sequence labeling task. Like humans watching an episode, our
model is presented with a sequence of (possibly
multi-modal) inputs, each corresponding to a sentence in the script, and assigns a label l indicating
whether the perpetrator is mentioned in the sentence
(l = 1) or not (l = 0). The model is fully incremental, each labeling decision is based solely on information derived from previously seen inputs.
We could have formalized our inference task as a
multi-label classification problem where labels correspond to characters in the script. Although perhaps more intuitive, the multi-class framework results in an output label space different for each
episode which renders comparison of model performance across episodes problematic. In contrast, our
formulation has the advantage of being directly applicable to any episode or indeed any crime series.
A sketch of our inference task is shown in Figure 4. The core of our model (see Figure 5)
is a one-directional long-short term memory network (LSTM; Hochreiter and Schmidhuber (1997;
Zaremba et al. (2014)). LSTM cells are a variant
of recurrent neural networks with a more complex
Figure 5: Illustration of input/output structure of our
LSTM model for two time steps.
computational unit which have emerged as a popular
architecture due to their representational power and
effectiveness at capturing long-term dependencies.
LSTMs provide ways to selectively store and forget
aspects of previously seen inputs, and as a consequence can memorize information over longer time
periods. Through input, output and forget gates, they
can flexibly regulate the extent to which inputs are
stored, used, and forgotten.
The LSTM processes a sequence of (possibly
multi-modal) inputs s = {xh1 , xh2 , ..., xhN }. It utilizes
a memory slot ct and a hidden state ht which are incrementally updated at each time step t. Given input
xt , the previous latent state ht−1 and previous memory state ct−1 , the latent state ht for time t and the
updated memory state ct , are computed as follows:
it
σ
ft σ
=
W ht−1
ot σ
xt
cˆt
tanh
ct = ft
ct−1 + it
ht = o t
tanh(ct ).
cˆt
The weight matrix W is estimated during inference,
and i, o, and f are memory gates.
As mentioned earlier, the input to our model consists of a sequence of sentences, either spoken utterances or scene descriptions (we do not use speaker
information). We further augment textual input with
multi-modal information obtained from the alignment of screenplays to video (see Section 3).
Textual modality Words in each sentence are
mapped to 50-dimensional GloVe embeddings, pretrained on Wikipedia and Gigaword (Pennington
et al., 2014). Word embeddings are subsequently
concatenated and padded to the maximum sentence
length observed in our data set in order to obtain
fixed-length input vectors. The resulting vector
is passed through a convolutional layer with maxpooling to obtain a sentence-level representation xs .
Word embeddings are fine-tuned during training.
Visual modality We obtain the video corresponding to the time span covered by each sentence and
sample one frame per sentence from the center of
the associated period.4 We then map each frame to
a 1,536-dimensional visual feature vector xv using
the final hidden layer of a pre-trained convolutional
network which was optimized for object classification (inception-v4; Szegedy et al. (2016)).
Acoustic modality For each sentence, we extract
the audio track from the video which includes all
sounds and background music but no spoken dialog. We then obtain Mel-frequency cepstral coefficient (MFCC) features from the continuous signal. MFCC features were originally developed in
the context of speech recognition (Davis and Mermelstein, 1990; Sahidullah and Saha, 2012), but
4
We also experimented with multiple frames per sentence
but did not observe any improvement in performance.
have also been shown to work well for more general sound classification (Chachada and Kuo, 2014).
We extract a 13-dimensional MFCC feature vector
for every five milliseconds in the video. For each
input sentence, we sample five MFCC feature vectors from its associated time interval, and concatenate them in chronological order into the acoustic
input xa .5
Modality Fusion Our model learns to fuse multimodal input as part of its overall architecture. We
use a general method to obtain any combination
of input modalities (i.e., not necessarily all three).
Single modality inputs are concatenated into an
m-dimensional vector (where m is the sum of dimensionalities of all the input modalities). We then
multiply this vector with a weight matrix W h of dimension m × n, add an m-dimensional bias bh , and
pass the result through a rectified linear unit (ReLU):
xh = ReLU([xs ; xv ; xa ]W h + bh )
The resulting multi-modal representation xh is of dimension n and passed to the LSTM (see Figure 5).
5
Evaluation
In our experiments we investigate what type of
knowledge and strategy are necessary for identifying the perpetrator in a CSI episode. In order to shed
light on the former question we compare variants of
our model with access to information from different
modalities. We examine different inference strategies by comparing the LSTM to three baselines. The
first one lacks the ability to flexibly fuse multi-modal
information (a CRF), while the second one does not
have a notion of history, classifying inputs independently (a multilayer perceptron). Our third baseline
is a rule-base system that neither uses multi-modal
inputs nor has a notion of history. We also compare
the LSTM to humans watching CSI. Before we report our results, we describe our setup and comparison models in more detail.
5.1
Experimental Settings
Our CSI data consists of 39 episodes giving rise to
59 cases (see Table 1). The model was trained on
5
Preliminary experiments showed that concatenation outperforms averaging or relying on a single feature vector.
53 cases using cross-validation (five splits with 47/6
training/test cases). The remaining 6 cases were
used as truly held-out test data for final evaluation.
We trained our model using ADAM with
stochastic gradient-descent and mini-batches of six
episodes. Weights were initialized randomly, except
for word embeddings which were initialized with
pre-trained 50-dimensional GloVe vectors (Pennington et al., 2014), and fine-tuned during training. We
trained our networks for 100 epochs and report the
best result obtained during training. All results are
averages of five runs of the network. Parameters
were optimized using two cross-validation splits.
The sentence convolution layer has three filters of
sizes 3, 4, 5 each of which after convolution returns
75-dimensional output. The final sentence representation xs is obtained by concatenating the output of
the three filters and is of dimension 225. We set the
size of the hidden representation of merged crossmodal inputs xh to 300. The LSTM has one layer
with 128 nodes. We set the learning rate to 0.001
and apply dropout with probability of 0.5.
We compared model output against the gold standard of perpetrator mentions which we collected as
part of our annotation effort (second pass).
5.2
Model Modality
Cross-val
T V A pr re f1
PRO + – – 19.3 76.3 31.6
CRF + – – 33.1 15.4 20.5
+ – – 36.7 32.5 33.7
+ + – 37.4 35.1 35.1
MLP
+ – + 39.6 34.2 35.7
+ + + 38.4 34.6 35.7
+ – – 39.2 45.7 41.3
+ + – 39.9 48.3 43.1
LSTM
+ – + 39.2 52.0 44.0
+ + + 40.6 49.7 44.1
Humans
74.1 49.4 58.2
Held-out
pr re f1
19.5 77.2 31.1
30.2 16.1 21.0
35.9 36.8 36.3
38.0 41.0 39.3
38.7 36.5 37.5
38.5 42.3 40.2
36.9 50.4 42.3
40.9 54.9 46.8
36.8 56.3 44.5
42.8 51.2 46.6
76.3 60.2 67.3
Table 2: Precision (pr) recall (re) and f1 for detecting the
minority class (perpetrator mentioned) for humans (bottom) and various systems. We report results with crossvalidation (center) and on a held-out data set (right) using
the textual (T) visual (V), and auditory (A) modalities.
sheds light on the importance of sequential information for perpetrator identification task. All results are
best checkpoints over 100 training epochs, averaged
over five runs.
Model Comparison
CRF Conditional Random Fields (Lafferty et al.,
2001) are probabilistic graphical models for sequence labeling. The comparison allows us to examine whether the LSTM’s use of long-term memory
and (non-linear) feature integration is beneficial for
sequence prediction. We experimented with a variety of features for the CRF, and obtained best results
when the input sentence is represented by concatenated word embeddings.
MLP We also compared the LSTM against a
multi-layer perceptron with two hidden layers, and
a softmax output layer. We replaced the LSTM in
our overall network structure with the MLP, keeping the methodology for sentence convolution and
modality fusion and all associated parameters fixed
to the values described in Section 5.1. The hidden
layers of the MLP have ReLU activations and layersize of 128, as in the LSTM. We set the learning rate
to 0.0001. The MLP makes independent predictions
for each element in the sequence. This comparison
PRO Aside from the supervised models described
so far, we developed a simple rule-based system
which does not require access to labeled data. The
system defaults to the perpetrator class for any sentence containing a personal (e.g., you), possessive
(e.g., mine) or reflexive pronoun (e.g., ourselves).
In other words, it assumes that every pronoun refers
to the perpetrator. Pronoun mentions were identified using string-matching and a precompiled list
of 31 pronouns. This system cannot incorporate any
acoustic or visual data.
Human Upper Bound Finally, we compared
model performance against humans. In our annotation task (Section 3.1), participants annotate sentences incrementally, while watching an episode for
the first time. The annotations express their belief as
to whether the perpetrator is mentioned. We evaluate these first-pass guesses against the gold standard
(obtained in the second-pass annotation).
Human
LSTM
LSTM avg
Human avg
precision in final 10% of the episode
1
0.8
0.6
0.4
0.2
0
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
test episode ID
Figure 6: Precision in the final 10% of an episode, for 30 test episodes from five cross-validation splits. We show
scores per episode and global averages (horizontal bars). Episodes are ordered by increasing model precision.
5.3
Which Model Is the Best Detective?
We report precision, recall and f1 on the minority
class, focusing on how accurately the models identify perpetrator mentions. Table 2 summarizes our
results, averaged across five cross-validation splits
(left) and on the truly held-out test episodes (right).
Overall, we observe that humans outperform all
comparison models. In particular, human precision
is superior, whereas recall is comparable, with the
exception of PRO which has high recall (at the expense of precision) since it assumes that all pronouns refer to perpetrators. We analyze the differences between model and human behavior in more
detail in Section 5.5. With regard to the LSTM, both
visual and acoustic modalities bring improvements
over the textual modality, however, their contribution appears to be complementary. We also experimented with acoustic and visual features on their
own, but without high-level textual information, the
LSTM converges towards predicting the majority
class only. Results on the held-out test set reveal that
our model generalizes well to unseen episodes, despite being trained on a relatively small data sample
compared to standards in deep learning.
The LSTM consistently outperforms the nonincremental MLP. This shows that the ability to utilize information from previous inputs is essential for
this task. This is intuitively plausible; in order to
identify the perpetrator, viewers must be aware of
the plot’s development and make inferences while
the episode evolves. The CRF is outperformed by
all other systems, including rule-based PRO. In contrast to the MLP and PRO, the CRF utilizes sequential information, but cannot flexibly fuse information from different modalities or exploit non-linear
mappings like neural models. The only type of
input which enabled the CRF to predict perpetrator mentions were concatenated word embeddings
(see Table 2). We trained CRFs on audio or visual
features, together with word embeddings, however
these models converged to only predicting the majority class. This suggests that CRFs do not have the
capacity to model complex long sequences and draw
meaningful inferences based on them. PRO achieves
a reasonable f1 score but does so because it achieves
high recall at the expense of very low precision. The
precision-recall tradeoff is much more balanced for
the neural systems.
5.4
Can the Model Identify the Perpetrator?
In this section we assess more directly how the
LSTM compares against humans when asked to
identify the perpetrator by the end of a CSI episode.
Specifically, we measure precision in the final 10%
of an episode, and compare human performance
(first-pass guesses) and an LSTM model which uses
all three modalities. Figure 6 shows precision results
for 30 test episodes (across five cross-validation
splits) and average precision as horizontal bars.
Perhaps unsurprisingly, human performance is superior, however, the model achieves an average precision of 60% which is encouraging (compared to
Episode 12 (Season 03): “Got Murder?”
0.8
LSTM f1
Human f1
0.4
0.2
0.4
0.2
0
0
0
100
200
300
400
500
600
0
180
60
40
120
400
500
200
300
400
500
200
300
#sentences observed
400
500
30
0
0
100
200
300
400
500
600
0
10
LSTM tp
Human tp
Gold tp
4
2
100
LSTM tp
Human tp
Gold tp
8
count
count
300
90
0
6
200
60
20
8
100
LSTM tp
Human tp
Gold tp
150
count
count
100
LSTM tp
Human tp
Gold tp
80
10
LSTM f1
Human f1
0.6
score
0.6
score
Episode 19 (Season 03): “A Night at the Movies”
0.8
6
4
2
0
0
0
100
200
300
400
#sentences observed
500
600
0
100
Figure 7: Human and LSTM behavior over the course of two episodes (left and right). Top plots show cumulative
f1; true positives (tp) are shown cumulatively (center) and as individual counts for each interval (bottom). Statistics
relating to gold perpetrator mentions are shown in black. Red vertical bars show when humans press the red button to
indicate that they (think they) have identified the perpetrator.
85% achieved by humans). Our results also show
a moderate correlation between model and humans:
episodes which are difficult for the LSTM (see left
side of the plot in Figure 6) also result in lower human precision. Two episodes on the very left of
the plot have 0% precision and are special cases.
The first one revolves around a suicide, which is not
strictly speaking a crime, while the second one does
not mention the perpetrator in the final 10%.
5.5
How Is the Model Guessing?
We next analyze how the model’s guessing ability compares to humans. Figure 7 tracks model
behavior over the course of two episodes, across
100 equally sized intervals. We show the cumulative development of f1 (top plot), cumulative true
positive counts (center plot), and true positive counts
within each interval (bottom plot). Red bars indicate
times at which annotators pressed the red button.
Figure 7 (right) shows that humans may outperform the LSTM in precision (but not necessarily in
recall). Humans are more cautious at guessing the
perpetrator: the first human guess appears around
sentence 300 (see the leftmost red vertical bars in
Figure 7 right), the first model guess around sentence 190, and the first true mention around sentence
30. Once humans guess the perpetrator, however,
they are very precise and consistent. Interestingly,
model guesses at the start of the episode closely follow the pattern of gold-perpetrator mentions (bottom
plots in Figure 7). This indicates that early model
guesses are not noise, but meaningful predictions.
Further analysis of human responses is illustrated
in Figure 8. For each of our three annotators we
plot the points in each episode where they press the
red button to indicate that they know the perpetrator (bottom). We also show the number of times (all
three) annotators pressed the red button individually
for each interval and cumulatively over the course of
the episode. Our analysis reveals that viewers tend
to press the red button more towards the end, which
is not unexpected since episodes are inherently designed to obfuscate the identification of the perpetrator. Moreover, Figure 8 suggests that there are
two types of viewers: eager viewers who like our
model guess early on, change their mind often, and
therefore press the red button frequently (annotator
1 pressed the red button 6.1 times on average per
annotator 1
annotator 2
annotator 3
all annotators frequency
all annotators cumulative
0
0.2
First correct perpetrator prediction
min max
avg
LSTM
2
554
141
Human
12 1014
423
0.4
0.6
0.8
1
Table 3: Sentence ID in the script where
the LSTM and Humans predict the true
Figure 8: Number of times the red button is pressed by each anno- perpetrator for the first time. We show
tator individually (bottom) and by all three within each time interval the earliest (min) latest (max) and avand cumulatively (top). Times are normalized with respect to length. erage (avg) prediction time over 30 test
episodes (five cross-validation splits).
Statistics are averaged across 18/12/9 cases per annotator 1/2/3.
portion of episode lapsed
s1
Grissom
pulls
out
a
small
evidence
bag
with
the filling
s1
Grissom
What’s so
amusing?
s2
He puts
it on the
table
Episode 03 (Season 03): “Let the Seller Beware”
s3
s4
s5
s6
Tooth fill- 10-7-02
Brass We Peter B. Look
ing 0857
also found I’m sure you’ll
your finger- find me all
prints and over the house
your hair
s2
Adam Trent
So let’s say
you find out
who did it and
maybe
it’s
me.
s7
Peter B.
I wanted
to buy it
Episode 21 (Season 05): “Committed”
s3
s4
s5
s6
Adam Trent Adam Trent Adam
Grissom
What are you Are you going smirks
Is it you?
going to do?
to convict me and starts
of murder and biting his
put me in a nails.
bad place?
s8
Peter B.
I was everywhere
s7
Adam Trent
Check
the
files sir.
s9
Brass well
you made
sure
you
were everywhere too
didn’t you?
s8
Adam Trent
I’m a rapist
not a murderer.
Table 4: Excerpts of CSI episodes together with model predictions. Model confidence (p(l = 1)) is illustrated in red,
with darker shades corresponding to higher confidence. True perpetrator mentions are highlighted in blue. Top: a
conversation involving the true perpetrator. Bottom: a conversation with a suspect who is not the perpetrator.
episode) and conservative viewers who guess only
late and press the red button less frequently (on average annotator 2 pressed the red button 2.9 times
per episode, and annotator 3 and 3.7 times). Notice
that statistics in Figure 8 are averages across several
episodes each annotator watched and thus viewer
behavior is unlikely to be an artifact of individual
episodes (e.g., featuring more or less suspects). Table 3 provides further evidence that the LSTM behaves more like an eager viewer. It presents the time
in the episode (by sentence count) where the model
correctly identifies the perpetrator for the first time.
As can be seen, the minimum and average identification times are lower for the LSTM compared to
human viewers.
Table 4 shows model predictions on two CSI
screenplay excerpts. We illustrate the degree of the
model’s belief in a perpetrator being mentioned by
color intensity. True perpetrator mentions are highlighted in blue. In the first example, the model
mostly identifies perpetrator mentions correctly. In
the second example, it identifies seemingly plausible
sentences which, however, refer to a suspect and not
the true perpetrator.
5.6
What if There Is No Perpetrator?
In our experiments, we trained our model on CSI
episodes which typically involve a crime, committed by a perpetrator, who is ultimately identified.
How does the LSTM generalize to episodes without
60
LSTM fp
Human fp
50
count
40
30
20
10
0
0
50
100
150
200
250
300
#sentences observed
350
400
450
Figure 8: Cumulative counts of false positives (fp) for
the LSTM and a human viewer for an episode with no
perpetrator (the victim committed suicide). Red vertical
bars show the times at which the viewer pressed the red
button indicating that they (think they) have identified the
perpetrator.
a crime, e.g., because the “victim” turns out to have
committed suicide? To investigate how model and
humans alike respond to atypical input we present
both with an episode featuring a suicide, i.e., an
episode which did not have any true positive perpetrator mentions.
Figure 8 tracks the incremental behavior of a human viewer and the model while watching the suicide episode. Both are primed by their experience with CSI episodes to identify characters in the
plot as potential perpetrators, and predict false positive perpetrator mentions. The human realizes after roughly two thirds of the episode that there is
no perpetrator involved (he does not annotate any
subsequent sentences as “perpetrator mentioned”),
whereas the LSTM continues to make perpetrator predictions until the end of the episode. The
LSTM’s behavior is presumably an artifact of the
recurring pattern of discussing the perpetrator in the
very end of an episode.
6
Conclusions
In this paper we argued that crime drama is an ideal
testbed for models of natural language understanding and their ability to draw inferences from complex, multi-modal data. The inference task is welldefined and relatively constrained: every episode
poses and answers the same “whodunnit” question. We have formalized perpetrator identification as a sequence labeling problem and developed
an LSTM-based model which learns incrementally
from complex naturalistic data. We showed that
multi-modal input is essential for our task as well
an incremental inference strategy with flexible access to previously observed information. Compared
to our model, humans guess cautiously in the beginning, but are consistent in their predictions once they
have a strong suspicion. The LSTM starts guessing
earlier, leading to superior initial true-positive rates,
however, at the cost of consistency.
There are many directions for future work. Beyond perpetrators, we may consider how suspects
emerge and disappear in the course of an episode.
Note that we have obtained suspect annotations but
did not used them in our experiments. It should also
be interesting to examine how the model behaves
out-of-domain, i.e., when tested on other crime series, e.g., “Law and Order”. Finally, more detailed
analysis of what happens in an episode (e.g., what
actions are performed, by who, when, and where)
will give rise to deeper understanding enabling applications like video summarization and skimming.
Acknowledgments The authors gratefully acknowledge the support of the European Research
Council (award number 681760; Frermann, Lapata) and H2020 EU project SUMMA (award number 688139/H2020-ICT-2015; Cohen). We also
thank our annotators, the anonymous TACL reviewers whose feedback helped improve the present paper, and members of EdinburghNLP for helpful discussions and suggestions.
References
Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu,
M̃argaret Mitchell, Dhruv Batra, C. Lawrence Zitnick,
and Devi Parikh. 2015. VQA: Visual Question Answering. In Proceedings of the IEEE International
Conference on Computer Vision (ICCV), pages 2425–
2433, Santiago, Chile.
Piotr Bojanowski, Francis Bach, Ivan Laptev, Jean Ponce,
Cordelia Schmid, and Josef Sivic. 2013. Finding actors and actions in movies. In The IEEE International
Conference on Computer Vision (ICCV), pages 2280–
2287, Sydney, Australia.
John S. Boreczky and Lynn D. Wilcox. 1998. A hidden Markov model framework for video segmentation
using audio and image features. In Proceedings of
the 1998 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), pages 3741–
3744, Seattle, Washington, USA.
Samuel R. Bowman, Gabor Angeli, Christopher Potts,
and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference.
In Proceedings of the 2015 Conference on Empirical
Methods in Natural Language Processing, pages 632–
642, Lisbon, Portugal.
Elia Bruni, Nam Khanh Tran, and Marco Baroni. 2014.
Multimodal distributional semantics. J. Artif. Int. Res.,
49(1):1–47, January.
Sachin Chachada and C.-C. Jay Kuo. 2014. Environmental sound recognition: A survey. APSIPA Transactions
on Signal and Information Processing, 3.
Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1):37–46.
Timothee Cour, Chris Jordan, Eleni Miltsakaki, and Ben
Taskar. 2008. Movie/script: Alignment and parsing
of video and text transcription. In Proceedings of the
10th European Conference on Computer Vision, pages
158–171, Marseille, France.
Steven B. Davis and Paul Mermelstein. 1990. Comparison of parametric representations for monosyllabic
word recognition in continuously spoken sentences.
In Alex Waibel and Kai-Fu Lee, editors, Readings in
Speech Recognition, pages 65–74. Morgan Kaufmann
Publishers Inc., San Francisco, California, USA.
Nevenka Dimitrova, Lalitha Agnihotri, and Gang Wei.
2000. Video classification based on HMM using text
and faces. In Proceedings of the 10th European Signal
Processing Conference (EUSIPCO), pages 1–4. IEEE.
Desmond Elliott and Frank Keller. 2013. Image description using visual dependency representations. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1292–
1302, Seattle, Washington, USA.
Philip John Gorinski and Mirella Lapata. 2015. Movie
script summarization as graph-based scene extraction.
In Proceedings of the 2015 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language Technologies,
pages 1066–1076, Denver, Colorado, USA.
Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman,
and Phil Blunsom. 2015. Teaching machines to read
and comprehend. In C. Cortes, N. D. Lawrence, D. D.
Lee, M. Sugiyama, and R. Garnett, editors, Advances
in Neural Information Processing Systems 28, pages
1693–1701. Curran Associates, Inc.
Felix Hill, Anoine Bordes, Sumit Chopra, and Jason Weston. 2015. The Goldilocks principle: Reading children’s books with explicit memory representations. In
Proceedings of the 3rd International Conference on
Learning Representations (ICLR), San Diego, California, USA.
Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long
short-term memory. Neural Computation, 9(8):1735–
1780, November.
Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li,
David A Shamma, Michael S Bernstein, and Li FeiFei. 2015. Image retrieval using scene graphs. In
Proceedings of the 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages
3668–3678, Boston, Massachusetts, USA.
Andrej Karpathy and Li Fei-Fei. 2015. Deep visualsemantic alignments for generating image descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3128–
3137, Boston, Massachusetts.
Douwe Kiela and Léon Bottou. 2014. Learning image
embeddings using convolutional neural networks for
improved multi-modal semantics. In Proceedings of
the 2014 Conference on Empirical Methods in Natural
Language Processing (EMNLP), pages 36–45, Doha,
Qatar.
John D. Lafferty, Andrew McCallum, and Fernando C. N.
Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data.
In Proceedings of the 18th International Conference
on Machine Learning, pages 282–289, San Francisco,
CA, USA. Morgan Kaufmann Publishers Inc.
Angeliki Lazaridou, Nghia The Pham, and Marco Baroni. 2015. Combining language and vision with a
multimodal skip-gram model. In Proceedings of the
2015 Conference of the North American Chapter of
the Association for Computational Linguistics: Human Language Technologies, pages 153–163, Denver,
Colorado, USA.
Dahua Lin, Sanja Fidler, Chen Kong, and Raquel Urtasun. 2014. Visual semantic search: Retrieving videos
via complex textual queries. In IEEE Conference
on Computer Vision and Pattern Recognition, pages
2657–2664, Columbus, Ohio, USA.
Cory S. Myers and Lawrence R. Rabiner. 1981. A comparative study of several dynamic time-warping algorithms for connected word recognition. The Bell System Technical Journal, 60(7):1389–1409.
Milind R. Naphide and Thomas S. Huang. 2001. A probabilistic framework for semantic video indexing, filtering, and retrieval. IEEE Transactions on Multimedia,
3(1):141–151.
Luis Gilberto Mateos Ortiz, Clemens Wolff, and Mirella
Lapata. 2015. Learning to interpret and describe abstract scenes. In Proceedings of the 2015 NAACL: Human Language Technologies, pages 1505–1515, Denver, Colorado, USA.
Jeffrey Pennington, Richard Socher, and Christopher D.
Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference
on Empirical Methods in Natural Language Processing (EMNLP), pages 1532–1543, Doha, Qatar.
Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and
Percy Liang. 2016. SQuAD: 100,000+ questions for
machine comprehension of text. In Proceedings of
the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392, Austin,
Texas, USA.
Zeeshan Rasheed, Yaser Sheikh, and Mubarak Shah.
2005. On the use of computable features for film classification. IEEE Transactions on Circuits and Systems
for Video Technology, 15(1):52–64.
Matthew Richardson, Christopher J.C. Burges, and Erin
Renshaw. 2013. MCTest: A challenge dataset for the
open-domain machine comprehension of text. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 193–203,
Seattle, Washington, USA.
Tim Rocktäschel, Edward Grefenstette, Karl Moritz Hermann, Tomas Kocisky, and Phil Blunsom. 2016. Reasoning about entailment with neural attention. In
Proceedings of the 4th International Conference on
Learning Representations (ICLR), San Juan, Puerto
Rico.
Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket
Tandon, Christopher Pal, Hugo Larochelle, Aaron
Courville, and Bernt Schiele. 2017. Movie description. International Journal of Computer Vision,
123(1):94–120.
Md Sahidullah and Goutam Saha. 2012. Design, analysis and experimental evaluation of block based transformation in MFCC computation for speaker recognition. Speech Communication, 54(4):543–565.
Jitao Sang and Changsheng Xu. 2010. Character-based
movie summarization. In Proceedings of the 18th
ACM International Conference on Multimedia, pages
855–858, Firenze, Italy.
Carina Silberer, Vittorio Ferrari, and Mirella Lapata.
2016. Visually grounded meaning representations.
IEEE Transactions on Pattern Analysis and Machine
Intelligence, 99.
Josef Sivic, Mark Everingham, and Andrew Zisserman.
2009. “Who are you?” – Learning person specific
classifiers from video. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1145–
1152, Miami, Florida, USA.
Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014.
Sequence to sequence learning with neural networks.
In Proceedings of the 27th International Conference
on Neural Information Processing Systems, NIPS’14,
pages 3104–3112, Cambridge, MA, USA. MIT Press.
Christian Szegedy, Sergey Ioffe, and Vincent Vanhoucke.
2016. Inception-v4, inception-ResNet and the im-
pact of residual connections on learning. CoRR,
abs/1602.07261.
Makarand Tapaswi, Martin Bäuml, and Rainer Stiefelhagen. 2015. Aligning plot synopses to videos for storybased retrieval. International Journal of Multimedia
Information Retrieval, (4):3–26.
Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen,
Antonio Torralba, Raquel Urtasun, and Sanja Fidler.
2016. MovieQA: Understanding stories in movies
through question-answering. In The IEEE Conference
on Computer Vision and Pattern Recognition (CVPR),
pages 4631–4640, Las Vegas, Nevada.
Subhashini Venugopalan, Marcus Rohrbach, Jeff Donahue, Raymond J. Mooney, Trevor Darrell, and Kate
Saenko. 2015a. Sequence to sequence – Video to text.
In Proceedings of the 2015 International Conference
on Computer Vision (ICCV), pages 4534–4542, Santiago, Chile.
Subhashini Venugopalan, Huijuan Xu, Jeff Donahue,
Marcus Rohrbach, Raymond Mooney, and Kate
Saenko. 2015b. Translating videos to natural language using deep recurrent neural networks. In Proceedings the 2015 Conference of the North American
Chapter of the Association for Computational Linguistics – Human Language Technologies (NAACL HLT
2015), pages 1494–1504, Denver, Colorado, June.
Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. 2015. Show and tell: A neural image
caption generator. Proceedings of the 2015 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), pages 3156–3164.
Ellen M. Voorhees and Dawn M. Tice. 2000. Building a
question answering test collection. In ACM Special Interest Group on Information Retrieval (SIGIR), pages
200–207, Athens, Greece.
Jason Weston, Antoine Bordes, Sumit Chopra, and
Tomas Mikolov. 2015. Towards AI-complete question answering: A set of prerequisite toy tasks. CoRR,
abs/1502.05698.
Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho,
Aaron Courville, Ruslan Salakhudinov, Rich Zemel,
and Yoshua Bengio. 2015. Show, attend and tell: Neural image caption generation with visual attention. In
Proceedings of the 32nd International Conference on
Machine Learning, pages 2048–2057, Boston, Massachusetts, USA.
Yi Yang, Wen-tau Yih, and Christopher Meek. 2015.
WikiQA: A challenge dataset for open-domain question answering. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2013–2018, Lisbon, Portugal.
Mark Yatskar, Luke Zettlemoyer, and Ali Farhadi. 2016.
Situation recognition: Visual semantic role labeling
for image understanding. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), pages 5534–5542, Zurich, Switzerland.
Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals.
2014.
Recurrent neural network regularization.
CoRR, abs/1409.2329.
Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja
Fidler. 2015. Aligning books and movies: Towards
story-like visual explanations by watching movies and
reading books. In The IEEE International Conference
on Computer Vision (ICCV), Santiago, Chile.
| 7cs.IT
|
Consensus-Based Torque Control of Deloaded Wind DFIGs for
Distributed and Fair Dynamic Dispatching
arXiv:1602.04234v1 [cs.SY] 12 Feb 2016
Stefanos Baros1
Abstract— In this paper we aim to address the problem of
dynamically dispatching a group of state-of-the-art deloaded
wind generators (WGs) in a fair-sharing manner. We use the
term dynamically since the WGs aim to dispatch themselves
according to a varying committed WF power output. We first
propose a leader-follower protocol whose execution guarantees
asymptotically, two control objectives. These are 1) reaching
asymptotic consensus on the utilization level of all WGs and
2) the total power output of the WGs asymptotically converges
to the reference value. Thereafter, we combine singular perturbation and Lyapunov theory to prove that, under certain
conditions, the proposed protocol will asymptotically converge
to its equilibrium. Finally, we derive a cooperative Control
Lyapunov Function-based (CLF) controller for the rotor side
converter (RSC) of each WG that realizes the protocol in
practice. We demonstrate the effectiveness of our proposed
protocol and the corresponding RSC controller design via
simulations on the modified IEEE 24-bus RT system.
I. I NTRODUCTION
A recent study conducted by the US Department Of
Energy [1] outlines the future of wind power in the US.
Specifically, it mentions that 10% of the US electricity
demand is expected to be produced by wind power by 2020,
20% of the US electricity demand by 2020 and 35% of
the US electricity demand by 2050. In a similar status is
Europe, where wind power integration is expected to increase
significantly in the next years [2]. These studies evidence a
recent world-wide tendency toward integrating a lot of wind
power into power systems. On the other hand, integrating
high levels of wind power into power systems raises an
important challenge for those systems. That is, to maintain
their stability, their reliability and their robustness [2].
It is worthwhile realizing that in high-wind-integration
settings, the WGs control will have a pronounced impact
on the stability and performance of the power grids that
accommodate the wind power. For this reason, the ongoing
regulations for the operation of WGs require the WGs to provide multiple advanced capabilities into the grid [2]. Between
them, frequency regulation, inertial response, power output
smoothing, Low Voltage Ride-Through (LVRT) capability
and voltage control [2]. Yet, by allowing communication
between WGs, capabilities that require coordination between
WGs can be achieved. Examples can be coordinated voltage control for regulation of the WF terminal voltage and
coordinated power control with load-sharing between WGs.
Such capabilities can be achieved efficiently with distributed
control methods that require a limited number of dedicated
communication links.
1 Stefanos Baros is with the ECE department of Carnegie Mellon University, Pittsburgh, PA, 15213 USA e-mail: [email protected]
In this paper, we focus on an important and advanced
capability that can be provided by a group of deloaded WGs.
The capability for regulating their total power extracted from
the wind such that it tracks a reference while they dynamically dispatch themselves in a fairly fashion. Deloaded WGs
are characterized by their flexibility to increase their power
output when they are commanded to. A WF that has its WGs
operating in a deloaded regime aiming to provide the above
capability has to address the following challenges. Firstly, to
timely compute the reference power points of its WGs while
taking into account the local varying wind speed conditions.
Secondly, to timely communicate these set-points to the WGs
in order for the WF total power output to match the required
reference.
In the literature, most of the studied methods relied on
centralized control schemes. Centralized schemes presume
that information about the wind speed, which is scattered
throughout the WF, is obtained from the central controller.
The central controller, having this information together with
information regarding the total WF power output can then
compute the power set-point for each WG individually.
Then, it can dispatch the WGs accordingly. Finally, the
power output of each WG can be regulated by its local
controller such that the WG generates the reference power.
An approach belonging to the above category can be found
in [3]. Despite the fact that centralized-based approaches can
address the problem discussed above they come with several
drawbacks, rather critical to be neglected. Among others are,
single-point failures, increased computational cost, extensive
communication network costs and delays [4]. The delays can
hamper a fast-responding control action from the WF and can
also compromise its tracking performance when (in a given
set-up) the dispatching of the WGs has to happen fast e.g
for maintaining power balance in a microgrid.
The literature on distributed methods for addressing the
problem of dispatching distributively a group of WGs given
a varying WF committed power output is not very broad.
As far as the authors are aware, the only references related
to the above problem are [5] and [6]. In [6], the authors
proposed a distributed WF controller for regulating the power
references of multiple WGs. At the same time, the controller
ensured that the fatigue experienced by the WGs was reduced
while the total power of the WF was reaching a pre-assigned
value. In a similar line of research, the authors in [5]
proposed a multi-agent-based strategy for addressing the
same problem in a microgrid setting. The global information
of the total demand and total available wind power were
retrieved via a consensus protocol that was executed by the
agents. Subsequently, this information was used by each
agent to define the set point of its corresponding WG.
In this work we make several contributions toward addressing the problem of distributively and dynamically dispatching a group of WGs for the purpose of having the
WF power output tracking a reference. To this end, we first
propose a distributed leader-follower consensus protocol that
realizes two basic control objectives 1) asymptotic consensus
of the utilization levels of all WGs i.e consensus on the
ratio defined by the available (from the wind) mechanical
power over the maximum mechanical power of each WG
2) asymptotic tracking of a varying reference by the WF
total power output. We prove that the proposed protocol
asymptotically converges to its equilibrium point under specific conditions. Our proof relies on results from singular
perturbation theory [7]. In the last part of our approach, we
develop a Control Lyapunov Function-based (CLF) [8] RSC
controller that realizes the proposed protocol in practice.
The rest of the paper is outlined in the following way.
Section II describes the problem of distributed dynamic
dispatching of the WGs for WF power output tracking.
Section III, presents the relevant WF model. Section IV
and V provide the main results of the paper. Section IV,
introduces the proposed protocol and Section V presents the
stability analysis. Section VI gives the derivation of the CLFbased torque RSC controller. In Section VII, the effectiveness
of the proposed approach is evaluated via simulations on
the modified IEEE-RTS 24-bus system. Finally, Section VIII
concludes the paper.
II. P ROBLEM F ORMULATION
A. Notation
With G being a set we use |G| to denote its cardinality.
We denote by R the set of reals and by C the set of
complex numbers. Also, we denote by R+ the set of nonnegative real numbers and with R++ the set of positive
reals. We denote the m-dimensional Euclidean space by Rm .
We denote vectors and matrices with bold characters. Let
A ∈ Rm×n be a m × n matrix of reals. With A> we denote
the transpose of A and with [a]ij the (i, j)-entry of the matrix
A. Let A 0 (A 0) denote that the matrix A is positive
definite (semi-definite). The spectrum of the matrix A (set of
eigenvalues) is denoted by σ(A). A n×n diagonal matrix B
is denoted by B = diag[bi ]ni=1 . The maximum value of the
vector a is denoted by ā. Similarly the maximum value of a
scalar quantity z is given by z̄. With In we denote the n × n
identity matrix and with 0n×1 and 1n×1 a n × 1 column
vector of zeros and ones respectively. With ẋ we denote the
time derivative of a variable x with respect to t, dx
dt . With
2
ẍ we denote the second derivative ddt2x The operator Re(·)
returns the real part of an imaginary number (·) ∈ C. With
C 2 we denote the space of functions with continuous second
derivatives.
B. Fair Dynamic Dispatching of WGs While the WF Power
Output is Tracking a Reference
To formulate the main problem, we consider a set-up
where we have a WF comprised with n wind generators.
We denote these generators by the set G , {1, ..., n} and
index each WG by i where i ∈ G. The available mechanical
power that can be extracted from the wind by each WG is
given by [9]:
Pm,i ,
1
3
ρCp,i Ai vw,i
,
2
∀i ∈ G
(1)
kg
where ρ ∈ R++ is the air density [ m
3 ], Cp,i ∈ R+ is the
2
power coefficient, Ai = πRi ∈ R++ is the area (swept by
the blades) and vw,i ∈ R++ , the local wind speed in [ m
s ].
Notice that, the only controllable variable in (1) is Cp,i which
can be regulated by the rotor speed of the WG, ωr,i . The
standard functionality provided by DFIGs wind turbines is
Maximum Power Point Tracking (MPPT). Achieving MPPT
presumes that the WG is controlled such that Cp,i = C̄p,i ,
where C̄p,i is the maximum value of Cp,i . In that case, (1)
can be recasted to:
1
3
,
∀i ∈ G
(2)
P̄m,i , ρC̄p,i Ai vw,i
2
The total power that a WF is required to extract from the
wind at any given moment can be described by a reference
Pd that equals the WF committed power output to the grid.
The latter is true under the mild assumption that the WF
power losses are negligible. In our case, we consider a setting
where the WGs are operating in a deloading regime and
can
P always meet the demanded power reference i.e Pd ≤
i∈G P̄m,i . For that setting, the problem we aim to address
can be formulated as follows.
Problem 1: To develop a fully distributed control scheme
for the RSC of the WGs that guarantees meeting the next two
conditions.
P
Condition 1: lim
Pm,i = Pd
t→∞ i∈G
Pm,i
P
Condition 2: lim P̄m,i
= lim P̄m,j
, ∀i, j ∈ G
m,j
t→∞
t→∞
The first condition ensures that the total WF power extracted
from the wind is tracking the reference while the second
condition that the WGs are dynamically dispatched in a fairly
manner i.e the ratio of the mechanical power to the maximum
mechanical power of each WG is the same. We suppose
that all WGs have identical technical characteristics such that
Ai = Aj , ∀i, j ∈ G. The following remark holds.
Cp,i
P
Remark 1: P̄m,i
=
,
∀i ∈ G
C̄p,i
m,i
Remark 1 directly appears when dividing (1) over (2). With
this, the Condition 2 becomes:
Condition 3:
C
C
p,i
p,j
= lim
,
∀i, j ∈ G
lim
t→∞ C̄p,i
t→∞ C̄p,j
Observe that:
Condition 2⇔ Condition 3
In the sequel, we use this observation to introduce an
approach that addresses Problem 1.
Power Coefficient
h
Cp,i (λi , θi ),0.22 116(
III. M ATHEMATICAL M ODELING
We present the WF-related models for providing the
ground of the forthcoming analysis. Specifically, we present
the wind-speed stochastic model, the rotor-voltage dynamical
model including the RSC control input and the rotor-speed
dynamical model.
The effective wind speed vw,i ∈ R+ can be modeled by
integrating two basic components, the slowly-varying mean
wind-speed, vm,i ∈ R+ , and the fast turbulence, vs,i ∈ R+
[10],[11]. Therefore, vw,i appears as:
i∈G
(3)
We note that, the turbulent component can be modeled in the standard state-space form parameterized by the
mean wind-speed vm,i with p1,i , p1,i (vm,i ), p2,i ,
p2,i (vm,i ), ki , ki (vm,i ) ∈ R as:
!
0
1
0
vs,i
v̇s,i
e
+
,
p +p2,i
ki
− p1,i1p2,i − p1,i
v̇s,i
v̈s,i
p1,i p2,i
1,i p2,i
(4)
With e ∈ N (0, 1), we denote a white noise process [10],[11].
B. Wind Generator Model
Since the problem we aim to address involves only the
rotor of the WG, we only present the rotor-side dynamics of
the WG. These, can be fully described by the electromagnetic
state-variables (rotor-voltages) dynamics and the electromechanical state-variable (rotor-speed) dynamics [9],[12] as:
Rotor-voltage Dynamics
0
0
0
1 h
Ėd,i , 0 − (Ed,i − (Xs,i − Xs,i )Iqs,i )
T0,i
i
0
0
Xm,i
+ T0,i (−ωs
Vqr,i + (ωs − ωr,i )Eq,i ) , ∀i∈G
Xr,i
(5a)
0
0
0
1 h
Ėq,i , 0 − (Eq,i + (Xs,i − Xs,i )Ids,i )
T0,i
i
0
0
Xm,i
+ T0,i (ωs
Vdr,i − (ωs − ωr,i )Ed,i ) , ∀i∈G
Xr,i
(5b)
Rotor-speed Dynamics
ω̇r,i ,
ωs
(Tm,i − Te,i ), ∀i∈G
2Hi
(5c)
Mechanical Torque
Tm,i ,
1
i +0.08θi
)
− 0.035
3
θ +1
i
, ∀i∈G
(5e)
Tip-speed Ratio
2k ω R
i
r i
λi ,
, ∀i∈G
pi
vw,i
(5f)
All the variables are explained in the Appendix.
A. Wind Speed Model
vw,i = vm,i + vs,i ,
·e
−12.5( λ
0.035 i
1
)
− 3
λi + 0.08θi
θi + 1
1 ρπRi2 ωs
3
Cp (λi , θi )vw,i
, ∀i∈G
2 Sb,i ωr,i
(5d)
IV. L EADER -F OLLOWER C ONSENSUS P ROTOCOL
By establishing the equivalence between Condition 2
and 3, we readily observe that having WGs achieving fair
sharing is the same as ensuring that Condition 3 is met.
Consequently, we can pose this problem as a consensus
agreement problem among all WGs on the utilization levels
Cp,i
, ∀i ∈ G. For this reason, we introduce an appropriate
C̄p,i
Leader-follower Consensus Protocol which we prove that
it converges to an equilibrium point that solves the exact
consensus agreement problem. Let WG 1 be the group leader
l , 1 and Ḡ , {2, ..., n} where Ḡ ⊂ G be the set of WGs
without the leader. Then, we propose the following protocol.
Protocol P1
Leader WG
X
dξh
, (Pd − Pm,l −
Pm,i )
dt
i∈Ḡ
dzl
, −kα,l (zl − ξh ) ,
dt
zl , z1
ξh ∈ R
(6a)
zl ∈ R
(6b)
WG i
dzi
, −kα,i (zi − zi−1 ) , i ∈ Ḡ
zi ∈ R
(6c)
dt
C
Cp,i
where the consensus states are zl , C̄p,l
and
z
,
i
C̄
p,i
p,l
respectively and the auxiliary state-variable of the leader is
ξh . The allowable communication links for implementing
this protocol can be seen in Fig. 3. We briefly describe
the mechanism by which the protocol is executed. The WF
supervisory controller obtains the WF reference Pd from
the system operator and passes its value to the leader WG.
Next, the leader WG computes its auxiliary state ξh and its
consensus state zl , using the reference and information from
all WGs. The leader communicates its consensus state with
the time derivative of its consensus state to the neighboring
WG. The same process is executed by all WGs i.e they obtain
their consensus state and communicate it to a neighboring
WG, concurrently. An assumption that has to be valid in
order for the protocol to be realizable
is that the leader
P
WG can retrieve the information i∈Ḡ Pm,i . The following
methods can be used for this purpose.
1) information passing from each WG to the leader (indirectly, via intermediate WGs).
2) average consensus protocol [13] P
with consensus statevariable the mechanical power i.e i∈Ḡ (Pm,i +Pm,l ) =
(Pm,avg · n).
V. S TABILITY A NALYSIS OF THE P ROPOSED P ROTOCOL
In this section, we study the asymptotic behavior of the
proposed protocol and the properties of its equilibrium point.
We begin by defining the following coefficients:
1
3
ρC̄p,l Al vw,l
, α l ∈ R+
(7)
2
1
3
αi , ρC̄p,i Ai vw,i
, αi ∈ R+ , i ∈ Ḡ
(8)
2
where in vector form are written as α
=
[αl , α2 , ..., αn ]> , α ∈ Rn . We define the consensusstates vector compactly as z = [z1 , ..., zi , ..., zn ]> , z ∈ Rn .
With these, Eq. (6a) become:
X
dξh
= (Pd − αl zl −
α i zi )
(9a)
dt
αl ,
i∈Ḡ
The equilibrium of the consensus protocol as obtained from
Eq. (9a),(6b),(6c) is:
ξh0 =
(αl +
P
Pd
zl0 = ξh0
i∈Ḡ
αi )
zi0 = ξh0 , ∀i ∈ Ḡ
(10)
(11)
(12)
Without loss of generality we take kα,i , kα,l , ∀i ∈ Ḡ.
we have the next Theorem.
Defining ε ∈ R+ as ε = kᾱ
a,i
Theorem 1: ∃ε∗ > 0 s.t ∀ε < ε∗ the equilibrium point
(ξh0 , z0 ) is asymptotically stable.
Proof: First, we define a new time-scale τ = ᾱ t with
dτ = ᾱ dt. Using this, equations (6b),(6c), (9a) become:
X αi
dξh
Pd
αl
,(
− zl −
zi )
(13a)
dτ
ᾱ
ᾱ
ᾱ
i∈Ḡ
ᾱ dz
l
, −(zl − ξh )
(13b)
kα,l dτ
ᾱ dz
i
, −(zi − zi−1 ),
i ∈ Ḡ
(13c)
kα,i dτ
ā
ā
Letting kα,l
= kα,l
= ε, we can write the above equations
more compactly as:
Slow quasi-steady system
dξh
, gh ,
dτ
gh ∈ R
(14a)
Fast boundary-layer system
ε
dz
, g,
dτ
g ∈ Rn
(14b)
where
X αi
Pd
αl
gh , (
− zl −
zi )
ᾱ
ᾱ
ᾱ
i∈Ḡ
g , [−(zl − ξh ) · · · − (zi − zi−1 ) · · · − (zn − zn−1 )]>
(14c)
Equations (14a),(14b) are in the standard singularly perturbed form [7] with ξh being the slow state-variable and z
being the fast state-variables. A system possessing a multitime-scale property enables a compartmental stability analysis of its system dynamics. By exploiting this property, we
first study the fast boundary-layer system dynamics (14b) in a
new time scale τ̃ = τε . Assuming that the slow state-variable
h
ξh is “frozen” i.e dξ
dτ̃ ≈ 0, and using the transformation
yi = (zi − ξh ) we can write the system equations (14b) as:
dy
, Af y,
dτ̃
y , [y1 , ..., yn ]>
−1 0 · · · 0
1 −1 · · · 0
Af , . .
.. . .
|
(15)
(16)
0
0
..
.
(17)
0 0 · · · 1 −1
{z
}
n×n
Since Af is a lower-triangular matrix, the diagonal terms
represent also the eigenvalues of Af . From that, we can
conclude that Af is a Hurwitz matrix and that the equilibrium y0 = 0n×1 is asymptotically stable. Equivalently that,
the equilibrium z0 = (ξh0 · 1n×1 ) is asymptotically stable
and attractive to the trajectories of the fast state-variables
z. Thus, all zi will converge toward the slow state-variable
ξh . We are left to show that the ξh converges toward ξh0 .
To do that, we focus on the behavior of the fast sub-system
equation (14b) when ε , 0, and observe that it degenerates
into the algrebraic equation:
0n×1 = g(ξh , z)
(18)
Solving for z results into the n−dimensional equilibrium
slow-manifold of (18), described by z = η(ξh ) = (ξh ·1n×1 ).
Direct substitution into (14a) yields the slow model:
dξh
= gh (ξh , η(ξh ))
dτ
P
X αi
αl
d
=
− ξh −
ξh
ᾱ
ᾱ
ᾱ
(19)
i∈Ḡ
The slow sub-system (19) has an asymptotically stable
equilibrium equal to:
ξh0 =
(αl +
P
Pd
i∈Ḡ
αi )
Having established that, the fast and the slow sub-systems
have asymptotically stable equilibria, from Theorem 11.4 (in
[7]) we conclude the next statement. That, ∃ε∗ > 0 such that
∀ε < ε∗ the equilibrium point of the full system (13a)-(13c)
is asymptotically stable. That, completes the proof.
VI. D ESIGN OF THE C ONSENSUS - BASED T ORQUE
C ONTROLLER
The previous Section was dedicated to establishing asymptotic convergence of the proposed consensus protocol to an
equilibrium point that realizes the desired control objectives,
described in Problem 1. For implementing the proposed
protocol, we design a cooperative torque controller for the
RSC of each WG that will force the system dynamics to
evolve as in (13a)-(13c). Writing equation (6c) analytically
leads to:
Cp,i
Cp,i−1
Ċp,i
= −kα,i (
−
),
C̄p,i
C̄p,i
C̄p,i−1
i∈G
(20)
i∈G
(21)
WG i
Gearbox
Noticing from (5e) that when θi = 0 we have
Cp,i , Cp,i (λi , 0),
The term
Ċp,i
C̄p,i
Te,i
i∈G
1
∗ 2
) ,
∀i ∈ G
(24)
Ve,i = (Te,i − Tei
2
∗
where Ve,i > 0, ∀Te,i ∈ De,i \ {Tei
}. Now, consider the
following proposition.
Proposition 1: Ve,i ∈ C 2 is a CLF and the equilibrium
∗
Te,i = Te,i
can be rendered asymptotically stable.
∗
Proof: Define the variable T̃e,i = (Te,i − Te,i
), ∀i ∈ G
and let the electrical torque expressed as:
0
Te,i
(25)
Computing the time-derivative of Ve,i along T̃˙e,i dynamics
gives us:
0
0
dVe,i
Vs,i 1 h
= T̃e,i 0
− (Eq,i + (Xs,i − Xs,i )Ids,i )
0
dt
Xs,i T0,i
i
0
0
Xm,i
∗
+ T0,i (ωs
Vdr,i − (ωs − ωr,i )Ed,i ) − T̃e,i Ṫe,i
Xr,i
(26)
This expression can be compactly expressed as:
dVe,i
∂Ve,i
=
[fi + hi Vdr,i ],
dt
∂ T̃e,i
∀i ∈ G
(27)
where
0
0
∂Ve,i
Vs,i 1 h
fi , T̃e,i 0
− (Eq,i + (Xs,i − Xs,i )Ids,i )
0
Xs,i T0,i
∂ T̃e,i
i
0
0
∗
+ T0,i (−(ωs − ωr,i )Ed,i ) − T̃e,i Ṫe,i
,
∀i ∈ G
(28)
∂Ve,i
Vs,i Xm,i
hi , T̃e,i 0 ωs
,
∀i ∈ G
(29)
Xr,i
Xs,i
∂ T̃e,i
-
k
+
,i
0
+
+
Xs,i +
Vs,i
⇤
Ṫe,i
⇤
Te,i
(22)
Letting (22) and (20) to be equal gives the electrical torque
∗
Te,i
∈ R as:
Cp,i Cp,i−1 i
1 ∂Cp,i ∂λi ωs −1 h
∗
−kα,i (
)
−
)
Te,i
= Tm,i −(
C̄p,i ∂λi ∂ωr,i 2Hi
C̄p,i C̄p,i−1
(23)
∗
To guarantee that limt→∞ (Te,i ) = Te,i
we consider the
candidate Control Lyapunov Function (CLF):
Eq,i Vs,i
=
, ∀i ∈ G
0
Xs,i
GSC
Vdr,i
in (20) can be expanded as:
1 ∂C ∂λ ω
Ċp,i
p,i
i
s
,
(Tm,i −Te,i ),
∂λi
∂ωr,i 2Hi
C̄p,i
C̄p,i
Capacitor
RSC
-
⇤
⇤
Computation of Te,i
, Ṫe,i
xi
Cp,i
C̄p,i
Ċp,i
,
1 C̄p,i
1
From WG (i
local
1
i
local
1
1)
Fig. 1: Cooperative torque controller of WG i
∂V
We observe that ∂ Te,i
˜ hi 6= 0 whenever T̃e,i 6= 0. That, means
e,i
the feedback control input Vdr,i can always guarantee that
∂Ve,i
fi < 0, ∀T̃e,i 6= 0. That, proves that Ve,i is indeed
∂ T˜
e,i
∗
a CLF and that the equilibrium Te,i = T̃e,i
can be rendered
exponentially stable. That, completes the proof.
We note that, Ve,i being a CLF is a necessary condition for
the existence of a stabilizing feedback control Vdr,i . Hence,
we proceed by designing a RSC stabilizing controller Vdr,i .
For having V̇e,i < 0 we take V̇e,i to be:
dVe,i
∗ 2
= −kβ,i (Te,i − Te,i
) < 0,
dt
∗
∀Te,i ∈ De,i \ {Te,i
}
(30)
dV
Equation (30) can be written as dte,i = −2kβ,i (Ve,i ) which
−2kβ,i t
has solution Ve,i = Ve,i0 e
i.e limt→∞ Ve,i = 0, For
having (30) the following equation has to hold:
∗
dTe,i
dTe,i
∗
=
− kβ,i (Te,i − Te,i
), ∀i ∈ G
dt
dt
(31)
dV
Furthermore, we assume that dts,i = 0, ∀i ∈ G
∂λi
∂
and ∂t
( ∂ω
) = 0, ∀i ∈ G hold. Combining equar,i
tions (31),(25),(5b) we derive the RSC controller as:
0
∗
i
Xs,i h dTe,i
0
1 h
∗
− kβ,i (Te,i − Tei
) − 0 − (Eq,i
Vdr,i =
Vs,i dt
T0,i
0
+ (Xs,i − Xs,i )Ids,i )
i X
0
0
r,i
+ T0,i (−(ωs − ωr,i )Ed,i )
,
Xm,i ωs
i∈G
(32)
This controller is depicted in Fig. 1 with all the variables
explained in the Appendix. The expressions of the appearing
2
∂λi ∂Cp,i ∂ Cp,i
terms ∂ω
, ∂λi , ∂λ2 are ommitted due to space limitar,i
i
18
21
22
23
17
16
20
19
15
14
13
11
24
3
12
9
10
6
4
5
8
1
2
7
W F Bus
Fig. 2: IEEE 24-bus RT system with a WF (with 10 WGs) at bus 22
2
1
n
3
W F Supervisory Controller
(a)
2
1
3
n
(b)
Fig. 3: a) Physical connectivity of the WF b) Communication structure
between WGs
tion. Nevertheless, they can be derived from (5e),(5f).
VII. P ERFORMANCE E VALUATION
The effectiveness of the proposed approach is explored
via simulations on the modified IEEE 24-bus RT system. In
this system, a WF comprised with 10 WGs lies at bus 22.
The physical connectivity and the allowable communication
links among the WGs (for n = 10) are shown in Fig. 3. The
control logic for each RSC follows equation (32) (Fig. 1)
whereas the group objective for the RSC controllers is to
dynamically self-dispatch their WGs in a fair-sharing way
and the WF power output to track a reference shown in
Fig. 4b. We studied the following critical scenarios.
Scenario 1 : At t = 0s, the WF power output reference is
0.38p.u and suddenly at t = 0.2s, the reference
changes step-wise to 0.42p.u as seen in Fig. 4b.
The response of the WF power output (blue) is tracking the
reference (red) closely with good performance given standard
metrics e.g overshoot, response time (Fig. 4b). The response
of the consensus state-variables is depicted in Fig. 4a. Notice
that, the response for all 10 WGs is completely identical.
That, verifies the “fair-dispatching” between the WGs i.e
each WG extracts mechanical power from the wind according to the local wind-speed conditions. In our setting, we
regard that all WGs experience the same local wind-speed
conditions. The mechanism by which the CLF-based RSC
controller carries out its objectives is understood as follows.
When the leader WG obtains the new power reference its
consensus state-variable is ordered to increase value. Since
all the WGs are trying to reach consensus with the leader
they increase their consensus state variables, leading all the
utilization factors to exceed 0.8 while starting from a value
of 0.73 (Fig. 4a). To achieve that, the various RSC torque
controllers slowed-down the WGs (Fig. 4c), enabling them
to increase the mechanical power that they extract from the
wind until their total power reached the pre-assigned reference value (Fig. 4b). In summary, the proposed protocol and
the developed RSC controllers effectively address Problem 1.
VIII. C ONCLUDING R EMARKS
This paper introduced a leader-follower consensus protocol that is able to dynamically dispatch a fleet of WGs
according to their local wind-speed conditions such that the
WF total power output reaches a new assigned value. By
employing singular perturbation theory [7], we provided theoretical guarantees in the form of asymptotic stability of desired equilibria, proving that the protocol will asymptotically
accomplish its aforementioned objectives. On the practical
side, we developed a cooperative CLF-based RSC controller
that implements the above protocol. We demonstrated the
performance of the proposed methodology via simulations
on the IEEE 24-bus RT system.
IX. A PPENDIX
The various variables related to WG i are explained below.
Variable
Te,i ∈ R+
Tm,i ∈ R+
0
T0,i ∈ R+
0
Xs,i ∈ R+
Xs,i ∈ R+
Xr,i ∈ R+
Xm,i ∈ R+
H i ∈ R+
0
0
Eq,i , Ed,i ∈ R+
Iqs,i , Ids,i ∈ R+
Vqr,i , Vdr,i ∈ R+
ωs
Sb,i ∈ R+
ωr,i ∈ R+
λi ∈ R+
θi ∈ R
ki ∈ R +
p i ∈ R+
Corresponds to
electrical torque
mechanical torque
transient open-circuit time constant
stator transient reactance
stator reactance
rotor reactance
mutual reactance of the stator-rotor
combined inertia of the WG
q, d axis rotor voltages
q, d axis stator current
q and d axis RSC control inputs
synchronous speed 2π · 60 [ rad
s ]
base power
electrical rotor speed of the WG
tip speed ratio
pitch angle
gearbox ratio
poles
0.81
R EFERENCES
0.8
9h
z1
z2
z3
z4
z5
z6
z7
z8
z9
z10
0.79
0.78
0.77
0.793
0.76
0.7925
0.75
0.792
0.74
0.73
6.5
0
6.6
5
6.7
10
15
time(s)
(a) Response of the Cp,i coefficients
0.43
0.42
P
0.41
i2G
Pm;i
(p:u)
Pd
0.4
0.39
0.38
0.37
0
3
6
9
12
15
time(s)
(b) WF total mechanical power tracking response
(reference Pd )
415
398.2
410
398.1
398
405
[ rs ]
5
5.2
400
395
390
0
5
!r;1
!r;2
!r;3
!r;4
!r;5
!r;6
!r;7
!r;8
!r;9
!r;10
10
15
time(s)
(c) Rotor speed response
Fig. 4: System response under Scenario 1
xi = [Tm,i , Ṫm,i , ωr,i , ω̇r,i ,
∂Cp,i ∂ 2 Cp,i Cp,i Ċp,i
,
,
,
,
∂λi
∂λ2i
C̄p,i C̄p,i
∂λi
, vw,i ]>
∂ωr,i
0
0
1 h
γi = 0 − (Eq,i + (Xs,i − Xs,i )Ids,i )
T0,i
i X
0
0
r,i
+ T0,i (−(ωs − ωr,i )Ed,i )
Xm,i ωs
ωs ∂λi −1 ∂Cp,i −2
∗
Ṫe,i
= Ṫm,i + kβ,i C̄p,i (
) (
)
2Hi ∂ωr,i
∂λi
h Ċ
Ċp,i−1 ∂Cp,i
p,i
· (
−
)
C̄p,i
C̄p,i−1 ∂λi
i
Cp,i
Cp,i−1 ∂ 2 Cp,i
−(
−
)
·
ω̇
, ∀i ∈ G
r,i
C̄p,i
C̄p,i−1 ∂λ2i
(33)
(34)
(35)
[1] U.S Deparment of Energy. Windvision: A new era for wind power in
the united states. Technical report, U.S Deparment of Energy, March
2015.
[2] Thomas Ackermann, editor. Wind Power in Power Systems. John
Wiley and Sons, second edition edition, 2012.
[3] Liyan Qu and Wei Qiao. Constant power control of dfig wind turbines
with supercapacitor energy storage. IEEE Transactions on Industry
Applications, 47(1):359–367, January/February 2011.
[4] Huanhai Xin, Zhihua Qu, John Seuss, and Ali Maknouninejad. A selforganizing strategy for power flow control of photovoltaic generators
in a distribution network. IEEE Transactions on Power Systems,
26(3):1462–1473, August 2011.
[5] Wei Zhang, Yinliang Xu, Wenxin Liu, Frank Ferrese, and Liming Liu.
Fully distributed coordination of multiple dfigs in a microgrid for load
sharing. IEEE Transactions on Smart Grid, 4(2):806–815, June 2013.
[6] Biegel B., Madjidian D., Spudic V., Rantzer A., and Stoustrup J.
Distributed low-complexity controller for wind power plant in derated
operation. In Proceeding of 2013 IEEE International Conference on
Control Applications (CCA), pages 146–151. IEEE, August 2013.
[7] H.K.Khalil. Nonlinear Systems. Prentice Hall, 3rd edition, 2002.
[8] E.D.Sontag. Mathematical Control Theory: Deterministic Finite Dimensional Systems. Number 6 in Textbooks in Applied Mathematics.
Springer, 2nd edition, 1998.
[9] H.A. Pulgar-Painemal and P.W. Sauer. Dynamic modeling of wind
power generation. In NAPS, pages 1–6. IEEE, October 2009.
[10] A.J Larsen and T.S Mogensen.
Individuel pitchregulering af
vindmølle. Master’s thesis, Technical University of Denmark., 2006.
[11] S.Thomsen. Nonlinear control of a wind turbine. Master’s thesis,
Technical University of Denmark, 2006.
[12] W.Qiao. Dynamic modeling and control of doubly fed induction
generators driven by wind turbines. In PSCE, 2009, pages 1–8. IEEE,
March 2009.
[13] Reza Olfati-Saber, J. Alex Fax, and Richard M. Murray. Consensus
and cooperation in networked multi-agent systems. Proceedings of the
IEEE, 95(1):215–233, January 2007.
[14] Soummya Kar and Jose M.F.Moura. Distributed consensus algorithms
in sensor networks with imperfect communication: Link failures and
channel noise. IEEE Transactions on Signal Processing, 57(1):355–
369, January 2009.
[15] Ali Bidram, Ali Davoudi, Frank L. Lewis, and Josep M. Guerrero. Distributed cooperative secondary control of microgrids using feedback
linearization. IEEE Transactions on Power Systems, 28(3):3462–3470,
March 2013.
[16] Alejandro D.Dominguez-Garcia and Christoforos N.Hadjicostis. Resilient networked control of distributed energy resources. IEEE
Journal on Selected Areas in Communications, 30(6):1137–1148, May
2012.
[17] Alejandro D.Dominguez-Garcia and Christoforos N.Hadjicostis. Coordination and control of distributed energy resources for provision of
ancillary services. In IEEE International Conference on Smart Grid
Communications (SmartGridComm), pages 537–542, October 2010.
[18] Soummya Kar and Gabriela Hug. Distributed robust economic
dispatch in power systems: A consensus + innovations approach. In
Power and Energy Society General Meeting, pages 1–8, July 2012.
[19] Roger A. Horn and Charles R. Johnson. Matrix Analysis. Cambridge
University Press, 2013.
[20] Stefanos Baros. A novel ectropy-based control scheme for a dfig driven
by a wind turbine with an integrated energy storage. In American
Control Conference (submitted), July 2015.
[21] S.Baros and M. Ilic. Robust ectropy-based cooperative control of a
wind dfig for transient stabilization and mppt (submitted). IEEE PES
General Meeting, July 2015.
[22] Chad Abbey and Geza Joos. Supercapacitor energy storage for wind
energy applications. IEEE Transactions on Industry Applications,
43(3):769–776, May/June 2007.
[23] Faruk A.Bhuiyan and Amirnaser Yazdani. Multimode control of a dfigbased wind-power unit for remote applications. IEEE Transactions on
Power Delivery, 24(4):2079–2089, October 2009.
[24] Chad Abbey and Geza Joos. Integration of energy storage with a
doubly-fed induction machine for wind power applications. In IEEE
35th Annual on Power Electronics Specialists Conference, volume 3,
pages 1964–1968. IEEE, June 2004.
| 3cs.SY
|
Association Rules Mining Based Clinical Observations
Mahmood A. Rashid1, Md Tamjidul Hoque2, Abdul Sattar1
1
Institute for Integrated and Intelligent Systems (IIIS),
Discovery Biology, Eskitis Institute for Cell & Molecular Therapies,
Griffith University
Nathan, QLD, Australia
{m.rashid, t.hoque, a.sattar}@griffith.edu.au
2
Abstract
Healthcare institutes enrich the repository of patients’
disease related information in an increasing manner
which could have been more useful by carrying out
relational analysis. Data mining algorithms are proven to
be quite useful in exploring useful correlations from
larger data repositories. In this paper we have
implemented Association Rules mining based a novel idea
for finding co-occurrences of diseases carried by a
patient using the healthcare repository. We have
developed a system-prototype for Clinical State
Correlation Prediction (CSCP) which extracts data from
patients’ healthcare database, transforms the OLTP data
into a Data Warehouse by generating association rules.
The CSCP system helps reveal relations among the
diseases. The CSCP system predicts the correlation(s)
among primary disease (the disease for which the patient
visits the doctor) and secondary disease/s (which is/are
other associated disease/s carried by the same patient
having the primary disease).
Key words
Disease Correlation, Association Mining, e-Health,
Healthcare, Medicare
1. Introduction
In the recent era, medical science has revealed that the
occurrence of one disease can lead to several associated
diseases [1]. For example, Heart-Block can lead to the
occurrences of other diseases like Hypertension, CardiacArrest and so on. It is, however, still an interesting
problem [1], to see how far the medical philosophy holds
from statistical point of view. Data mining based
techniques, like association rule mining, have gained
popularity [1-3] among contemporary scientists to gain
clearer understanding of different physical and scientific
phenomenon. In this paper, we apply association rule
mining to extract knowledge from clinical data for
predicting correlation of diseases carried by a patient.
From the viewpoint of scientific research, data mining is
relatively a new discipline that has been developed mainly
from studies carried out in various disciplines such as
computing, marketing, statistics and so on [4]. Data
mining problems and corresponding solutions have roots
in classical data analysis. Many of the methodologies
used in data mining has come from two branches of
research: i) one is developed in machine leaning (artificial
intelligence) community, and ii) the other is developed in
the statistical community particularly in multivariate and
computational statistics. Both have made great
contributions to the understanding and applications of
data mining techniques [5, 6].
The remainder of the paper is organized as follows.
Section 2 highlights previous research in related areas.
Section 3 introduces the Architecture of the CSCP
Prototype-system. It also briefly discusses the working
procedures and Algorithms used to compute the
correlations. Section 4 highlights the output of the system
with the hypothetical datasets. Concluding remarks and
future research are sketched in section 5.
2. Related Works
Automated healthcare systems are accumulating large
quantities of information about patients and their medical
conditions everyday. Unfortunately, few methodologies
have been developed and applied to discover this hidden
knowledge [7].
The cluster-analysis based model is suggested and
discussed [8] for assigning prostate cancer patients into
homogenous groups with the aim to support future
clinical treatment decisions as an illustration. To explore
association rules in noisy and high dimensional medical
data-repository an improved algorithm has been
introduced with several constraints [9]. A statistical
analysis of decision tree based classification approach on
diagnosing the Ovarian Cancer using Bio-marker Patterns
Software (BPS) has been applied [10]. A task [11] has
been accomplished on comparison of data mining
methods supporting diagnosis for Melanoma. Association
rule classifiers have been applied to diagnose breast
cancer using digital mammograms [12], Neural Network
based classification approach also used for the same
purpose [13]. Association Mining applied on
questionnaire responses related to human sleeping [14]
where questionnaire data and clinical summaries
comprised a total of 63 variables including gender, age,
body mass index, and Epworth and depression scores.
Many Clinical Decision Support Systems (CDSS) have
been developed. CHICA [15] is a CDSS, developed to
improve preventive paediatric primary care. Dynamic
forms are generated and tailored to patients’ needs based
on the Medical Logic Modules (MLMs). A knowledge
management framework for distributed healthcare
systems has been proposed [16] to integrate the
heterogeneous systems used by different departments
from clinical care to administration.
However, the aforementioned developments are
application specific and thus hard to apply in general.
Instead of developing an application limited to a specific
purpose such as prostate cancer [8], skin cancer [11], and
sleeping [14] and so on, we proposed for a more generic
version of CSCP system that can work for all diseases in
similar fashion and generate correlations depending on
the input dataset. In the next section, the architecture of
the application and its working procedures are stated.
3. Framework and Working Procedures
The CSCP system will extract data from an OLTP system.
Thus to implement the CSCP system, we require the
regular operational OLTP system to generate input data
for the CSCP system. The architecture of the system can
be described as follows.
OLTP Application: When a patient will visit a doctor
he/she must have to fill up a prescribed form and the
information from the form is proposed to be captured
through web enabled OLTP system.
CSCP Application: This application will import
transactional records from OLTP application for further
processing to generate correlations among diseases using
Association Rules data mining.
The OLTP system is a regular database application to
capture patient’s information and to preserve the records
into a database repository. It is simple and no analytical
process has been incorporated in this portion and thus the
CSCP system is our major focus in this paper and we
presume that the OLTP system has been implemented
successfully and data has been captured accordingly from
healthcare institutes.
3.2 Association Rules Mining and Apriori Algorithm
In data mining, association rule is used for discovering
interesting relations between variables in large databases.
The two key terms support and confidence are used in
computing correlations between variables which can be
defined as follows:
Support: In a fixed number of transactions the
occurrences of a particular event is the support of that
event. For example there are T transactions among which
Txy transaction s contain the itemset {X,Y} is Tx.
Confidence: confidence is a relative support. For example
there are T transactions among which Txy transactions
contain itemset {X, Y} and Tx transactions contain item X
and thus confidence of occurring X and Y together is Txy/
Tx.
The Apriori Algorithm [17] is used here for Association
Rule Mining to find out frequent dataset that satisfy the
predefined minimum support and confidence from a given
database. As computers are handy now-a-days, a viable
CSCP system can be setup to collect large volumes of
data and to stored in the database simultaneously. This
kind of data includes the transaction records of clinics,
hospitals, supermarkets, banks, stock markets, telephone
companies and so on. The next few sections we try to
discover some hidden information from a sample
transactional dataset.
3.3 Sample Dataset
Figure 1 - The Complete Architecture of the Proposed
System for Prediction Correlation among diseases using
association mining on healthcare data. The upper-portion
of the figure describes the architecture of the OLTP
system and the lower portion is the CSCP system.
3.1 System Architecture
The disease diagnosis application proposed here based on
two different software systems. One is online transaction
processing (OLTP) system and another is disease
correlation predicting (CSCP) application.
In a clinic various patients come but most of them come
for a particular disease, which mentioned here as primary
disease. When Doctors or their associates make an
interview with the patient and note down other problems
(diseases), which mentioned here as secondary or
associated diseases are inserted into a database.
The main objective of the paper is to find out relations
among the primary disease and other secondary diseases.
The following table represents sample dataset of a
Medicare database that contains the patient-wise diseases.
Table 1 - Transactional data sample
Patient Id
Disease
P000000001
P000000001
P000000001
P000000001
P000000002
P000000002
P000000002
…
P000001000
Bradycardia
Cardiac Arrest
Hypertension
Myocarditis
Bradycardia
Cardiac Arrest
Hypertension
…
Cardiac Arrest
3.4 Producing Itemsets
A set of diseases obtained by each patient presented here
along with the number of diseases counted by the
algorithm in Figure 3 and put the result into the column
headed CNT in figure 3
Algorithm 1 - CountDisease
1
PROCEDURE CountDisease
2
FOR each p in P
3
Ds ""
4
c0
5
find records for p
6
FOR each r in R
7
c c 1
8
Ds Ds r " , "
9
NEXT r
10
Ds Ds without comma at the end
11
INSERT p, c, Ds in to database table
12
NEXT p
13
END CountDisease
Figure 2 - Pseudo code to count diseases carried by any
patient.
Table 2: Patient records with multiple diseases
Patient
Count
P000000001
4
P000000002
…
3
…
P000001000
1
Diseases
Heart-Block, Hypertension, CardiacArrest, Bradycardia
Heart-Block, Hypertension, CardiacArrest
…
Hypertension
3.5 Counting Support of an Item (Disease) in the
sample dataset
The frequency of every item in all the transactions has
been calculated in the following table implementing the
following algorithm and for a 1-item itemsets for the first
pass.
Algorithm 2- FindSupport
PROCEDURE FindSupport
d disease
1
2
3
db database
in considarat
ion
4
r record
5
rs recordset
s support
6
7
BEGIN
8
FOR each d in db
9
s0
10
find records for d
11
FOR each r in rs
12
s s 1
13
NEXT r
14
UPDATE Database with s
15
NEXT d
16
END
17 END FindSupport
Figure 3 - Pseudo code to calculate support.
Disease
Table 3 - Distinct diseases
Support
Heart-Block
334
Hypertension
549
Myocarditis
532
Cardiac-Arrest
Bradycardia
536
305
Pass
1
1
1
1
1
3.6 Generating Candidate Itemset
The following procedure generates candidate itemsets
taking the transactional records as input and maximum
pass and minimum support as parameters.
Algorithm 3 - GenerateRule
PROCEDURE GenerateRule (maxpass, minsup)
// the method requires two user input
// maxpass- maximum number of items are consider in an
itemset
// minsup- minimum support for the candidate sets
BEGIN
1. Produce maxpass copies of transaction data
by aliasing the original one for generating all
possible combination of items in each pass by
producing Cartesian product and the filter the
meaning sets
2. Calculate the supports for the filtered sets
3. Eliminate the sets having support less than
minsup
END
Figure 4 - Generating candidate itemset from sample
dataset.
Table 4 - Candidate Sets for stop-level 2
Itemset
Count
{Bradycardia, Cardiac-Arrest}
26
{Bradycardia, Heart Block}
{Bradycardia, Hypertension}
{Bradycardia, Myocarditis}
9
28
21
{Cardiac Arrest, Heart-Block}
{Cardiac Arrest, Myocarditis}
13
32
{Heart Block, Hypertension}
{Heart Block, Myocarditis}
10
11
3.7 Association Rule Generation
Using the sample transactions of Table 1, after second
pass (maxpass=2), the CSCP system has generated the
rule data as listed in Table 5 below
Table 5- Candidate Sets for stop-level 2
Itemset
Supp (%) Conf (%)
Figure 5 represents the support and confidence of second
pass for the disease set{Hypertension, Heart-Block} for
the patients of different age groups.
The graph in the Figure 6 represents the supports and
confidences for different itemsets for the female patients
in the sample transactions.
In our work we implement the system to compute the
correlation of diseases for different age groups and the
output in Figure 5 is clearly illustrating that at the age of
45-50 the risk of Heart-Block is more than other ages if a
patient has already been living with Hypertension. Our
system has also been applied on the database for the
patent of different sexualities and the graphical form of
empirical output has been shown in Figure 6 which
illustrates that for the female patients the risk of
Hypertension is significant if the patient already carry
Bradycardia.
Suport
12.00
10.00
{Bradycardia, Cardiac-Arrest}
{Bradycardia, Heart-Block}
2.60
0.90
8.52
2.69
…………
…..
…..
4.00
2.00
{Heart Block, Hypertension}
1.00
1.82
0.00
{Heart Block, Myocarditis}
1.10
3.29
4. Discussion on the Empirical Results
In practical applications, a rule can be considered
statistically significant when it generates from a data
repository that contain hundred of thousands or millions
of trusted transactions. We have carried out our
experiments on a relatively small number of transactions1
due to the lack of availability of real transactional data
where the disease Heart-Block and the disease
Hypertension conjunctionally occurs once and thus its
support of {Heart Block , Hypertension} is 1.0 as shown
in Table 5.
Support
Confidence
40.00
35.00
30.00
25.00
20.00
15.00
Conf idence
14.00
8.00
6.00
}
}
}
n}
s}
n}
n}
a}
ia
st
ia
io
i ti
io
di
io
re
rd
rd
rd
ns
ar
ns
ns
Ar
ca
a
ca
e
c
e
e
t
y
c
t
y
t
c
d
d
a
dy
er
er
er
yo
ra
ra
di
ra
yp
yp
yp
,M
,B
ar
,B
,B
H
st
,H
is
st
ck
,C
t,H
k,
t
e
a
i
i
s
c
r
k
e
o
r
rd
lo
rd
re
Bl
Ar
oc
Ar
ca
rt
Ar
ca
Bl
tB
c
ac
yo
rt
ia
ar
dy
ea
di
ac
a
i
r
e
a
d
H
M
r
r
e
a
{
{
d
a
{H
{B
ar
{H
{C
{C
{C
Figure 6 - For the different itemsets in second pass, the
supports and confidences for patients of different
sexualities (here females only). X-axis represents the
disease sets and Y-axis represents the value of support and
confidence respectively.
The system can easily be enhanced for further crucial
aspects: such as patients of various professions, localities,
physiological conditions and so on. The nature of working
environment for IT professionals is remarkably different
from that of the construction professionals. IT
professionals work in static indoor environment where as
construction professionals work in dynamic and outdoor
environment. By implementing our system on the
patients’ database of IT profession and construction
profession, new trend of disease correlation can be
predicted.
10.00
5.00
5. Conclusions
0.00
0
20
40
60
80
100
Figure 5- For the itemset {Hypertension, Heart-Block} in
second pass, the support and confidence for patients of
different age groups. X-axis represents the ages and Yaxis represents the value of support and confidence
respectively.
1
Randomly generated 1000 transactions
In this paper, we have implemented system-prototype,
named CSCP system, using the association rules of data
mining technique applied to a patients’ (assumed)
database for discovering patterns of diseases that might be
carried by a patients. As a novel idea of mining the data
capturing process can further be modified in the clinics as
well as in the data-warehouses which should further be
involved to enhance the CSCP system we have proposed.
The recognised pattern by this implementation definitely
can improve the healthcare services along with medical
researchers for further exploring trends of diseases that
are correlated. To ensure strong national economy and
bio-security [18] by having healthier inhabitants, for
example, Medicare Australia [19] can use CSCP system
to ensure wellbeing further.
In this research-work, we have succeeded to investigate
correlation among diseases for patients of different age
and sex groups providing the outcome in statistical as
well as in graphical format.
The system we developed has been based on computer
generated data, since the real data were not handy.
However, it is our future target to enhance the system for
the aforementioned various cases, applying the system on
collected real-life data while enhancing the proposed
system rigorously.
6. Acknowledgement
This research is partially supported by the ARC
(Australian Research Council) grant no DP0557303.
7. References
[1] Lovell NH, Magrabi F, Celler BG, Huynh K, and
Garsden H, Web-based acquisition, storage, and
retrieval of biomedical signals, IEEE Eng. Medicine
and Biology, vol. 20(3), pp. 38-44, 2001.
[2] Mazzi C, Ganguly P, and Kidd M, Healthcare
application based on software agents, in Medinfo
2001 Proceedings, 2001, pp. 136-140.
[3] Takeuchi H, Kodama N, Hashiguchi T, and Hayashi
D, Automated Healthcare Data Mining Based on a
Personal Dynamic Healthcare System, the 28th IEEE
EMBS Annual International Conference, New York
City, USA, 2006.
[4] Giudici P, Applied Data Mining: Statistical Methods
for Business and Industry: Wiley Publications, 2003.
[5] Mannila H, Data mining: machine learning, statistics,
and databases, in Eighth International Conference on
Scientific and Statistical Database Management,
1996, pp. 1-8.
[6] Kantardzic M, Data Mining: Concepts, Models, and
Algorithms. NJ: Wiley Interscience, 2003.
[7] Prather J, Lobach D, Goodwin L, Hales J, Hage M,
and Hammond W, Medical data mining: knowledge
discovery in a clinical data warehouse, in Proc AMIA
Annu Fall Symp., 1997, pp. 101-5.
[8] Churilov L, Bagirov A, Schwartz D, Smith K, and
Dally M, Improving Risk Grouping Rules for
Prostate Cancer Patients with Optimization, the 37th
International Conference on System Sciences, Hawaii
2004.
[9] Ordonez C, Santana C, and Braal LD, Discovering
Interesting Association Rules in Medical Data, in
ACM SIGMOD Workshop on Research Issues on
Data Mining and Knowledge Discovery, 2000, pp.
78-85.
[10] Vlahou A, Schorge JO, Gregory BW, and. Coleman
RL, Diagnosis of Ovarian Cancer Using Decision
Tree Classification of Mass Spectral Data, Journal of
Biomedicine and Biotech-nology, vol. 5, pp. 308–
314, 2003.
[11] Grzymala-Busse JW and Hippe JS, Data Mining
Methods Supporting Diagnosis of Melanoma, the
18th IEEE Symposium on Computer-Based Medical
Systems (CBMS 2005), 2005, pp. 371–373.
[12] Za¨ıane OR, Antonie ML, and Coman A,
Mammography classification by an association rulebased classifier, in MDM/KDD: Inter-national
Workshop on Multimedia Data Mining (with ACM
SIGKDD 2002), 2002, pp. 62–69.
[13] Land WH, Masters JT, Lo JY, McKee DW, and
Anderson FR, New results in breast cancer
classification obtained from an evolutionary
computation/adaptive
boosting
hybrid
using
mammogram and history data in Industrial
Applications, Proceedings of the 2001 IEEE
Mountain Workshop on Soft Computing, Virginia
Tech, Blacksburg. Virginia, 2001, pp. 47–52.
[14] Laxminarayan P, Ruiz C, Alvarez SA, and Moonis
M, Mining associations over human sleep time series,
the 18th IEEE Symposium on Computer- Based
Medical Systems (CBMS), 2005, pp. 323–328.
[15] Anand A, Biondich PG, Liu G, Rosenman M, and
Downs SM, Child health improvement through
computer automation: The CHICA system, in
MEDINFO, 2004.
[16] Kazemzadeh RS and Sartipi K, Interoperability of
Data and Knowledge in Distributed Health care
Systems, in The 13th IEEE International Workshop
on Software Technology and Engineering Practice,
2005.
[17] Agrawal and Srikant, Fast algorithm for Mining
Association Rules, in The 20th VLDB conference
Santiago, Chile, 1994.
[18] A Healthier Future For All Australians. vol. ISBN: 174186-940-4 Publications Number: P3 -5499:
Commonwealth of Australia, 2009.
[19] Medicare Australia, www.medicareaustralia.gov.au,
last access: Oct, 2009.
Address for correspondence
Mahmood A. Rashid
Building N34, Room 1.45
Institute for Integrated and Intelligent Systems (IIIS)
Griffith University, Nathan, QLD 4111, Australia
Email: [email protected]
Phone: +61(0)7 373 53757, Fax: +61 (0)7 373 54066
| 5cs.CE
|
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
OF A REGULAR LOCAL RING
arXiv:1512.03848v2 [math.AC] 2 Jan 2016
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
Abstract. Let (R, m) be a d-dimensional regular local domain with d ≥ 2 and
let V be a valuation domain birationally dominating R such that the residue
field of V is algebraic over R/ m. Let v be a valuation associated to V . Associated to R and V there exists an infinite directed family {(Rn , mn )}n≥0 of
d-dimensional regular local rings dominated by V with
R = R0 and Rn+1 the
S
local quadratic transform of Rn along V . Let S := n≥0 Rn . Abhyankar proves
that S = V if d = 2. Shannon observes that often S is properly contained in V
if d ≥ 3, and Granja gives necessary and sufficient conditions for S to S
be equal
to V . The directed family {(Rn , mn )}n≥0 and the integral domain S = n≥0 Rn
may be defined without first prescribing a dominating valuation domain V . If
{(Rn , mn )}n≥0 switches strongly infinitely often, then S = V is a rank one valuaord n (f )
)
tion domain and for nonzero elements f and g in m, we have v(f
.
= lim ordR
v(g)
R (g)
n→∞
n
If {(Rn , mn )}n≥0 is a family of monomial local quadratic transforms, we give necessary and sufficient conditions for {(Rn , mn )}n≥0 to switch strongly infinitely
often. If these conditions hold, then S = V is a rank one valuation domain of
rational rank d and v is a monomial
P valuation. Assume that V is rank one and
birationally dominates S. Let s = ∞
i=0 v(mi ). Granja, Martinez and Rodriguez
show that s = ∞ implies S = V . We prove that s is finite if V has rational
rank at least 2. In the case where V has maximal rational rank, we give a sharp
upper bound for s and show that s attains this bound if and only if the sequence
switches strongly infinitely often.
1. Introduction
Let R = R0 be a d-dimensional regular local ring, and for each integer n ≥ 0, let
Rn+1 be a d-dimensional local quadratic transform of Rn . Thus {(Rn , mn )}n≥0 is a
S
directed family of d-dimensional regular local rings. Let S := n≥0 Rn . If d = 2,
Abhyankar proves in [A] that S is always a valuation domain. In the case where
Date: March 2, 2018.
1991 Mathematics Subject Classification. Primary: 13H05, 13A18, 13C05; Secondary: 13E05,
13H15.
Key words and phrases. local quadratic transform, infinite directed family, switches strongly
infinitely often, rank and rational rank of a valuation domain, transform of an ideal, monomial
ideal, valuation ideal.
1
This paper was supported by Faculty Research Fund, Sungkyunkwan University, 2013.
Correspondence with Alan Loper, Bruce Olberding and Hans Schoutens that motivated our
interest in the work of Shannon and Granja on infinite directed unions of local quadratic transformations is gratefully acknowledged.
1
2
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
d ≥ 3, Shannon in [S] and later Granja in [Gr] consider conditions in order that S
be a valuation domain. In this connection, Shannon gives the following definition
in [S, page 314].
Definition 1.1. Let {(Rn , mn )}n≥0 be an infinite directed family of local quadratic
transforms of a regular local ring (R, m). We say that {(Rn , mn )}n≥0 switches
strongly infinitely often if there does not exist an integer j and a height one
S
prime ideal pj of Rj with the property that ∞
n=0 Rn ⊂ (Rj )pj .
S
Assume that V is a rank one valuation domain that birationally dominates S :=
n≥0 Rn .
If V is non-discrete, Shannon proves in [S, Proposition 4.18] that S = V
if and only if {(Rn , mn )}n≥0 switches strongly infinitely often. It is observed in [GR,
Theorem 6] that the proof given by Shannon also holds if V is rank one discrete.
S
Granja [Gr, Proposition 7] shows that if S := n≥0 Rn is a valuation domain V ,
then V has real rank either one or two, and in [Gr, Theorem 13], he characterizes
the sequence of local quadratic transforms of R along V . If V has rank one, then
{(Rn , mn )}n≥0 switches strongly infinitely often. If V has rank two, then the value
group of V is Z ⊕ G, where G has rational rank one. In this case Granja proves
[Gr, Theorem 13] that the sequence {(Rn , mn )}n≥0 is height one directed as in
Definition 1.2.
Definition 1.2. Let {(Rn , mn )}n≥0 be an infinite directed family of local quadratic
transforms of a regular local ring (R, m). The sequence {(Rn , mn )}n≥0 is height
one directed if there exists a nonnegative integer j and a height one prime ideal p
S
of Rj such that ∞
n=0 Rn ⊂ (Rj )p , and if for some nonnegative integer k and some
S
height one prime ideal q of Rk we have ∞
n=0 Rn ⊂ (Rk )q , then (Rj )p = (Rk )q .
Let {(Rn , mn )}n≥0 be a directed family of Noetherian local domains. Assume that
S
ordRn defines a valuation for each n. If ∞
n=0 Rn = V is a rank one valuation domain,
we prove in Theorem 3.9 that the valuation v is related to the order valuations of
the Rn as follows:
ordRn (f )
v(f )
= lim
,
n→∞ ordRn (g)
v(g)
for all nonzero elements f, g in the maximal ideal of V .
Let (R, m) be a d-dimensional regular local ring with d ≥ 2 and fix d elements
x, y, . . . , z such that m = (x, y, . . . , z)R. In Section 4, we consider finite sequences
R = R0 ⊂ R1 ⊂ · · · ⊂ Rn of monomial local quadratic transforms. We prove in
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
3
Theorem 4.3 that the order in R of each nonzero f ∈ m is greater than the order of
the transform of f in Rn if and only if in passing from R0 to Rn we transform in
each monomial direction at least once.
In Section 5, we consider infinite sequences {(Rn , mn )}n≥0 of monomial local quaS
dratic transforms. In Theorem 5.4, we prove that ∞
n=0 Rn is a rank one valuation
domain if and only if the transform in each monomial direction occurs infinitely
many times.
In Section 6, we present examples of infinite sequences of local quadratic transforms {(Rn , mn )}n≥0 that switch strongly infinitely often and have the property that
S
n≥0 Rn = V is a rank one valuation ring having rational rank less than dim R.
Let {(Rn , mn )}n≥0 be a directed sequence of local quadratic transforms along
a zero-dimensional rank one valuation V . In Section 7, we consider the invariant
P
s= ∞
i=0 v(mi ). Granja, Martinez and Rodriguez prove in [GMR, Prop. 23] that
S
s = ∞ implies n≥0 Rn = V . We observe in Proposition 7.3 that s is finite if V has
rational rank at least 2 and that s = ∞ or s < ∞ are both possible if V has rational
rank one and is not a DVR. In the case where V has maximal rational rank, we give
in Theorem 7.2 a sharp upper bound for s, and show that s attains this bound if
and only if the sequence switches strongly infinitely often. In Theorem 7.5, we give
necessary and sufficient conditions for a sequence {(Rn , mn )}n≥0 to be height one
S
directed. This yields examples where S = n≥0 Rn is a rank 2 valuation domain
with value group Z ⊕ H such that H is rational rank one but not discrete.
We use µ(I) to denote the minimal number of generators of an ideal I, and λR (M )
to denote the length of an R-module M . We use the notation A ⊂ B to denote that
A is a subset of B that may be equal to B.
2. Preliminaries
Definition 2.1. Let R ⊆ T be unique factorization domains (UFDs) with R and T
having the same field of fractions. Let p be a height-one prime in R. If pT ∩ R = p,
then there exists a unique height-one prime q in T such that q ∩ R = p. We then
have Rp = Tq . On the other hand, if p ( pT ∩ R, then (R \ p)−1 T = Q(T ), the field
of fractions of T . The transform pT of p in T is the ideal q if pT ∩ R = p and is
the ring T if p ( pT ∩ R. Thus
pT =
(
q
T
if
if
p = pT ∩ R,
p ( pT ∩ R.
4
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
Let I be a nonzero ideal of R. Then I has a unique factorization I = pa11 · · · pann J,
where the pi are principal prime ideals, the ai are positive integers, and J is an ideal
with J −1 = R.
(1) The transform I T of I in T is the ideal
I T = q1a1 · · · qnan (JT )(JT )−1
where qi = pTi for each i with 1 ≤ i ≤ n.
(2) The complete transform of I in T is the completion I T of I T .
In the case where J is a nonzero principal ideal, the definition here of the transform
agrees with the definition given by Granja [Gr, page 701] for the strict transform.
For an ideal I of a local ring (R, m), the order of I, denoted ordR I, is r if I ⊆ mr
but I * mr+1 . If (R, m) is a regular local ring, the function that associates to
an element a ∈ R, the order of the principal ideal aR, defines a discrete rank-one
valuation, denoted ordR on the field of fractions of R. The associated valuation ring
(DVR) is called the order valuation ring of R.
Definition 2.2. Let V be a valuation domain corresponding to the valuation v.
The rank of v is defined to be the Krull dimension of V . The rational rank
of v, denoted rat. rank v, is the rank of the value group Γv of V over Q. Thus,
rat. rank v = dimQ (Γv ⊗Z Q).
Definition 2.3. Let R be an integral domain. An ideal I of R is said to be a
valuation ideal if there exists a valuation domain V such that R ⊆ V ⊆ Q(R)
and IV ∩ R = I. We then say that I is a V -ideal in R, and if v is a valuation
corresponding to V that I is a v-ideal in R.
If R is a subring of a valuation domain V and mV is the maximal ideal of V ,
then the prime ideal mV ∩R of R is called the center of V on R. Let (R, m) be a
Noetherian local domain with field of fractions Q(R). A valuation domain (V, mV )
is said to birationally dominate R if R ⊆ V ⊆ Q(R) and mV ∩R = m, that is, m
b.d.
is the center of V on R. We write R ⊂ V to denote that V birationally dominates
R. The valuation domain V is said to be a prime divisor of R if V birationally
dominates R and the transcendence degree of the field V / mV over R/ m is dim R−1.
If V is a prime divisor of R, then V is a DVR [A, p. 330].
Definition 2.4. Let R be a local domain with maximal ideal m and let V be a
valuation domain that birationally dominates R. Let v be the valuation associated
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
5
with V . The dimension of v on R, or the dimension of V over R, is defined to
be the transcendence degree of the residue field k(v) of v over the field R/ m.
The quadratic dilatation or blowup of m along V , cf. [N, page 141], is the
unique local ring on the blowup Blm (R) of m that is dominated by V . The ideal
m V is principal and is generated by an element of m. Let a ∈ m be such that
aV = m V . Then R[m /a] ⊂ V . Let Q := mV ∩R[m /a]. Then R[m /a]Q is the
quadratic transformation of R along V . In the special case where (R, m) is a
d-dimensional regular local domain we use the following terminology.
Definition 2.5. Let d be a positive integer and let (R, m, k) be a d-dimensional
regular local ring with maximal ideal m and residue field k. Let x ∈ m \ m2 and
let S1 := R[ mx ]. The ring S1 is a d-dimensional regular ring in the sense that each
localization of S1 at a prime ideal is a regular local ring. To see this, observe that
S1 /xS1 is isomorphic to a polynomial ring in d − 1 variables over the field k, cf. [SH,
Corollary 5.5.9], and S1 [1/x] = R[1/x] is a regular ring. Moreover, S1 is a UFD since
x is a prime element of S1 and S1 [1/x] = R[1/x] is a UFD, cf. [M, Theorem 20.2].
Let I be an m-primary ideal of R with r := ordR (I). Then one has in S1
IS1 = xr I1
for some ideal
I1
of
S1 .
It follows that either I1 = S1 or ht I1 ≥ 2. Thus I1 is the transform I S1 of I in S1
as in Definition 2.1.
Let p be a prime ideal of R[ mx ] with m ⊆ p. The local ring
m
R1 : = R[ ]p = (S1 )p
x
is called a local quadratic transform of R; the ideal I1 R1 is the transform of I
in R1 as in Definition 2.1.
Remark 2.6. Let (R, m) be a regular local ring and let (R1 , m1 ) be a local quadratic
transform of R as in Definition 2.5. If q is a nonzero prime ideal of R1 such that
q ∩R =: p is properly contained in m, then Rp = (R1 )q .
Proof. We have Rp ⊆ (R1 )q, and (R1 )q dominates Rp . Since R1 is a localization
/ p. Thus
of S1 = R[ mx ], we have xR1 ∩ R = m. Hence q ∩R 6= m implies that x ∈
R[ x1 ] ⊂ Rp , and so R ⊂ S1 ⊂ Rp ⊆ (R1 )q. Since (R1 )q is a localization of S1 and
q(R1 )q ∩ S1 = p Rp ∩ S1 , we have Rp = (R1 )q .
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
6
Definition 2.7. Let (R, m) be a d-dimensional regular local ring and consider a
sequence of local quadratic transforms along R,
R = R0 ⊂ R1 ⊂ . . . ⊂ Rn .
We say there is a change of direction from R0 to Rn if m0 ⊂ m2n .
Remark 2.8. Assume the notation of Definition 2.7. Let V be a valuation domain
birationally dominating Rn and let v be a valuation associated to V . There is a
change of direction from R0 to Rn if and only if v(m0 ) > v(mn−1 ).
Proof. It suffices to prove the case where n = 2. Assume that v(m0 ) = v(m1 ). We
show there is no change of direction from R0 to R2 . Let x ∈ m0 be such that
v(x) = v(m0 ). Then R2 is a localization of R1 [ mx1 ]. Hence x ∈ m2 \ m22 .
Conversely, assume that v(m0 ) > v(m1 ). We show there is a change of direction
from R0 to R2 . Let x ∈ m0 . Let w ∈ m1 be such that v(w) = v(m1 ). Then R2 is a
localization of R1 [ mw1 ]. Since v(x) ≥ v(m0 ) > v(m1 ) = v(w), we have
x=
w wx
∈
m22 .
We conclude that m0 ⊂
x
w
m22 .
∈ m2 . Thus
3. Directed unions of local quadratic transforms
Remark 3.1. If (R, m) is a 2-dimensional regular local ring, then a well-known
result of Abhyankar [A, Lemma12] states that an infinite directed union of local
quadratic transforms of R is always a valuation domain. Examples given by Shannon
in [S, Examples 4.7 and 4.17] show that this fails in general if d ≥ 3.
Remark 3.2. Let {(Rn , mn )}n≥0 be an infinite directed family of local quadratic
transforms of a regular local ring R. As shown by Granja in [Gr, Lemma 10], the
following are equivalent:
(1) The sequence {(Rn , mn )}n≥0 switches strongly infinitely often.
(2) For each integer j ≥ 0 and nonzero element f ∈ Rj , there exists a positive
integer n ≥ j such that the transform (f Rj )Rn = Rn .
Let V be a valuation domain birationally dominating a regular local ring (R, m)
and let {(Rn , mn )}n≥0 be the family of local quadratic transforms of R along V .
Granja in [Gr, Theorem 13] proves the following:
Remark 3.3. Let {(Rn , mn )}n≥0 be an infinite directed family of local quadratic
transforms of a regular local ring R. If {Rn }n≥0 switches strongly infinitely often,
S
then S := ∞
n=0 Rn is a rank one valuation domain.
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
7
Proof. To prove that S is a valuation domain, it suffices to show for nonzero elements
f, g ∈ R := R0 that either f /g or g/f is in Rn for some n > 0. Let I = (f, g)R.
We prove the assertion by induction on ordR (I). The case where ordR (I) = 0 is
clear. Let ordR (I) = r > 0, and assume the assertion holds for nonnegative integers
less than r. We may assume that r = ordR (f ) ≤ ordR (g). Let I R1 = (f1 , g1 )R1 ,
where
f1
g1
= fg , denote the transform of I in R1 . Since r = ordR (f ) = ordR (I), the
ideal f1 R1 is the transform of f R in R1 . If ordR1 (f1 ) < r or ordR1 (g1 ) < r, then
we are done by the induction hypothesis. Otherwise, ordR1 (I1 ) = ordR1 (f1 ) = r
and I1 := (f1 , g1 )R1 satisfies the same hypotheses as I. Let I R2 = (f2 , g2 )R2 ,
where
f2
g2
= fg , denote the transform of I in R2 , and continue in this way to define
In = (fn , gn )Rn . Since the sequence switches strongly infinitely often, there exists
a positive integer n such that either ordRn (gn ) < r or the strict transform fn of f
S
in Rn has ordRn (fn ) < r. This proves that S = ∞
n=0 Rn is a valuation domain.
Assume that dim S > 1 and let P be a nonzero nonmaximal prime ideal of S. For
S
S
each n ≥ 0, let Pn := P ∩ Rn . Then P = n≥0 Pn and SP = n≥0 (Rn )Pn . There
exists a positive integer t for which Pt is a nonzero, nonmaximal prime ideal of the
regular local ring Rt . For each positive integer i, we have Pt+i ∩ Rt = Pt . Therefore
by Remark 2.6 (Rt )Pt = (Rt+i )Pt+i . It follows that SP = (Rt )Pt , a contradiction to
our hypothesis that {Rn }n≥0 switches strongly infinitely often. We conclude that
dim S = 1 and S is a rank one valuation domain.
We use the following setting to observe in Theorem 3.5 a connection between
valuation ideals and an infinite directed family that switches strongly infinitely often.
Setting 3.4. Let {(Rn , mn )}n≥0 be an infinite directed family of local quadratic
transforms of a regular local ring R. Assume that {(Rn , mn )}n≥0 switches strongly
S
infinitely often and let V = n≥0 Rn . As noted in Remark 3.3, V is a rank one
valuation domain. Let v be a valuation associated to V . By [ZS2, Lemma 3, p.
T
343], the V -ideals of R form a descending sequence {In }n≥0 , and n≥0 In = (0).
Theorem 3.5. Assume notation as in Setting 3.4. Then there exists an ascending
sequence of integers {τn }n≥0 such that for each valuation domain W that birationally
dominates Rτn , the ideals I1 , . . . , In are valuation ideals for W .
Proof. We construct the sequence {τn }n≥0 by induction on n. Suppose we have
constructed the sequence up to length n. Since the sequence of local quadratic
transforms switches strongly infinitely often, there exists an integer τn+1 ≥ τn such
8
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
that In Rτn+1 is principal, say In Rτn+1 = gRτn+1 , where g ∈ In . Since V birationally
dominates Rτn+1 , it is clear that v(g) = v(In ). Notice that we may take g to be
any element in In such that v(g) = v(In ). Let h be another such element. Since
h ∈ gRτn+1 ,
element
h
g
h
g
∈ Rτn+1 . Since V birationally dominates Rτn+1 and v(g) = v(h), the
is a unit in Rτn+1 , so gRτn+1 = hRτn+1 .
Let W birationally dominate Rτn+1 . Then W birationally dominates Rτn . By
the induction hypothesis, I0 , . . . , In are W -ideals. To see that In+1 is a W -ideal,
consider J = In+1 W ∩ R0 . Since In+1 ( In and In is a W -ideal, we have In+1 ⊂
J ⊂ In . Suppose by way of contradiction that J 6= In+1 . Let f ∈ J \ In+1 . Since
In+1 = {a ∈ R0 | v(a) > v(In )}, it follows that v(f ) ≤ v(In ). Because f ∈ In ,
we have v(f ) = v(In ). Thus In Rτn+1 = f Rτn+1 , so J = In . Since In+1 ( In and
In Rτn+1 = f Rτn+1 , there exists a proper ideal L of Rτn+1 such that In+1 Rτn+1 = f L.
Since L is a proper ideal of Rτn+1 and W dominates Rτn+1 , we have w(L) > 0. Since
J = f LW ∩ R and f ∈ J, it follows that f ∈ f LW , which contradicts the fact that
w(L) > 0.
Remark 3.6. Let (R, m) be a Noetherian local domain and let V be a valuation
domain birationally dominating R such that the vector space dimension of V / mV
over R/ m denoted dimR/ m V / mV = e < ∞. Let {In }n≥0 be the sequence of V ideals of R. Then 1 ≤ λR (In /In+1 ) ≤ e for all n ≥ 0. To see this, fix an integer n,
and consider the natural embedding as R-modules,
In
In+1
∼
=
In V
In V ∩ R
V
∼
֒→
,
=
m V In V ∩ R
m V In V
mV
where the last isomorphism follows because In V = f V , for some element f ∈ In .
Thus there is an R-module embedding of In /In+1 into V / mV , and hence we have
0 < λR (In /In+1 ) ≤ λR (V / mV ) = e. In particular, if R/ m = Rn / mn for all n ≥ 0
and {In }n≥0 is the sequence of V -ideals in R, then λR (In /In+1 ) = 1 for all n ≥ 0.
Remark 3.7. Let R and the family {(Rn , mn )}n≥0 be as in Theorem 3.5. It is
natural to ask about conditions on the integer τn given in Theorem 3.5. The smaller
the integer τn , the sharper the assertion. We describe the situation in more detail
in Discussion 3.8
Discussion 3.8. Let the ring R, the family of local quadratic transforms {(Rn , mn )}n≥0 ,
and the descending chain of ideals {In }n≥0 be as in Theorem 3.5. Let S denote the
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
9
set of valuation domains that birationally dominate R. For each integer n ≥ 0, let
Sn = {W ∈ S | I0 , I1 , . . . , In are all W -ideals},
b.d.
Tn = {W ∈ S | Rn ⊂ W }.
Notice that S0 = S1 = S = T0 and the families {Sn } and {Tn } are both linearly
ordered with respect to inclusion and decreasing. Since I2 = PRR1 is the special
∗-simple complete ideal in R associated to R1 as defined by Lipman in [L, Proposition 2.1], we also have S2 = T1 . Associated with these sequences, we define:
sn = min{ j ∈ N0 | Tj ⊆ Sn }.
(
min{ j ∈ N0 | Sj ⊆ Tn },
tn =
∞ if Sj is not contained in Tn for all j.
We have s0 = 0 = t0 and s1 = 0, while t1 = 2 and s2 = 1. The sequences {sn } and
{tn } are increasing in the sense that sn ≤ sn+1 and tn ≤ tn+1 for each n ≥ 0.
Theorem 3.5 implies that for each positive integer n there exists a positive integer
τn such that
Tτn ⊆ Sn
that is,
W ∈ Tτn =⇒ W ∈ Sn .
Thus τn is any integer such that τn ≥ sn . The proof of Theorem 3.5 also proves
that:
(1)
sn ≤ min{ j ∈ N0 | Iµ Rj is principal for all µ with 0 ≤ µ ≤ n − 1}.
Theorem 3.9 applies to a directed family {(Rn , mn )}n≥0 of Noetherian local doS
mains such that ∞
n=0 Rn is a rank one valuation domain.
Theorem 3.9. Let {(Rn , mn )}n≥0 be an infinite family of Noetherian local domains
such that Rn is a subring of Rn+1 for each integer n ≥ 0. Assume that ordRn defines
S
a valuation for each integer n ≥ 0, and that ∞
n=0 Rn = V is a rank one valuation
domain. Let v denote a valuation associated to V , and let f, g be nonzero elements
in the maximal ideal of V . Then for n sufficiently large, f and g are in mn , and we
have
v(f )
ordRn (f )
= lim
.
n→∞ ordRn (g)
v(g)
Proof. Let L =
v(f )
v(g)
and let ǫ > 0. There exist positive integers p and q such that
L−ǫ <
Then
p+1
p
≤ L <
< L + ǫ.
q
q
p
v(f )
p+1
≤ L=
<
q
v(g)
q
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
10
implies that
pv(g) ≤ qv(f ) < (p + 1)v(g)
For n sufficiently large, we have
and
f q g p+1
gp , f q
v(g p ) ≤ v(f q ) < v(g p+1 ).
∈ Rn . Therefore
ordRn (g p ) ≤ ordRn (f q ) < ordRn (g p+1 ).
It follows that
p
ordRn (f )
p+1
≤
.
<
q
ordRn (g)
q
We conclude that
L−ǫ<
ordRn (f )
<L+ǫ
ordRn (g)
for all sufficiently large integers n.
Corollary 3.10 is immediate from Theorem 3.9.
Corollary 3.10. Let {(Rn , mn )}n≥0 be an infinite directed family of local quadratic
transforms of a regular local ring R. Assume that {(Rn , mn )}n≥0 switches strongly
S
infinitely often, so V = ∞
n=0 Rn is a rank 1 valuation. Let v be a valuation associated to V . Then for nonzero f, g ∈ m, we have
ordRn (f )
v(f )
= lim
.
n→∞ ordRn (g)
v(g)
4. Directed unions of monomial local quadratic transforms
Definition 4.1. Let (R, m) be a d-dimensional regular local ring with d ≥ 2, and
fix d elements x, y, . . . , z such that m := (x, y, . . . , z)R. An element p ∈ R is called a
monomial in x, y, . . . , z if there exists (α, β, . . . , γ) ∈ Nd0 such that p = xα y β · · · z γ .
An ideal I of R is said to be a monomial ideal if I is generated by monomials.
Let
x
A = R
y
m
z
= R , ...,
x
x
x
x1 := x,
y1 : =
y
,
x
. . . , z1 :=
z
.
x
If I is a monomial ideal in R, the transform of I in x A is generated by elements
of the form xa1 y1b · · · z1c with a, b, . . . , c ∈ N0 . This motivates us to define an ideal
J of x A to be a monomial ideal if J is generated by monomials in x1 , y1 , . . . , z1 .
We consider monomial quadratic transformations of R defined as follows: the ring
xR = R m
x (x, y , ..., z ) is a monomial local quadratic transformation of R in
x
x
the x-direction. An ideal J of x R is said to be a monomial ideal if J is generated
by monomials in x1 , y1 , . . . , z1 .
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
11
In a similar manner, we define y R, . . ., z R to be the local monomial quadratic
transformations of R in the y-direction, . . ., z-direction, respectively, where
m
m
z
y
R
=
R
.
R = R
z , ...,
x
x y
y ( y , y, ..., y )
z ( z , z , ..., z)
We define an ideal of y R, . . ., z R to be a monomial ideal if it is generated by
monomials in the respective rings. We refer to the elements in the fixed set of
minimal generators of the regular local ring as variables.
For a monomial ideal I of one of these rings, there exists a unique set of monomial
minimal generators of I, [KS, Corollary 4]. We let ∆(I) denote the set of monomial
minimal generators of I.
Setting 4.2. Let (R, m) and x, y, . . . , z be as in Definition 4.1. For n a positive
integer, consider a sequence of regular local rings
(2)
{(Ri , mi )}ni=0 ≡ R =: R0 ⊂ R1 ⊂ R2 ⊂ · · · ⊂ Rn ,
where Ri+1 is a monomial local quadratic transformation of Ri , for each i < n
mi = (xi , yi , . . . , zi )Ri
and
mi+1 = (xi+1 , yi+1 , . . . , zi+1 )Ri+1 ,
and where one of the following d-cases occur:
xi+1 := xi ,
yi+1 :=
yi
,
xi
···
, zi+1 :=
zi
,
xi
in which case we say the transform from Ri to Ri+1 is in the x-direction, or
xi+1 :=
xi
,
yi
yi+1 := yi ,
···
, zi+1 :=
zi
,
yi
in which case we say the transform from Ri to Ri+1 is in the y-direction, or
..
.
xi
yi
, yi+1 := , · · · , zi+1 := zi ,
zi
zi
in which case we say the transform from Ri to Ri+1 is in the z-direction.
xi+1 :=
Let Dx (respectively, Dy , . . . , Dz ) denote the set of integers i ∈ {0, 1, . . . , n − 1}
for which the transform from Ri to Ri+1 is in the x-direction (respectively, the
y-direction, . . . , the z-direction).
Theorem 4.3. Let notation be as in Setting 4.2. The following are equivalent:
(1) The sets Dx , Dy , . . . , Dz are all nonempty.
(2) For every nonzero f ∈ m, we have ordRn (f R)Rn < ordR (f ).
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
12
Proof. Assume that item 1 fails. Then at least one of the sets Dx , Dy , . . . , Dz is
empty. By relabeling, we may assume that Dx is empty. Then Rn ⊂ RxR and we
have ordRn (xR)Rn = 1 = ordR (x).
Assume that item 1 holds and fix a nonzero element f ∈ m. Let r := ordR (f ) > 0.
The order of the transform of f R never increases, cf [HKT, Lemma 3.6]. By [HubS,
Proposition 1.3], there exist monomials g1 , . . . , gt and units λ1 , . . . , λt in R such that
f = λ1 g1 + · · · + λt gt .
Let gi = xai y bi · · · z ci be a monomial with ordR (gi ) = r. By relabeling the variables,
we may assume ai > 0. Since the set Dx is nonempty, there exists s ∈ {0, . . . , n − 1}
such that the transform from Rs to Rs+1 is in the x-direction. Let s be minimal
with this property.
If ordRs (f R)Rs < r, we are done. Otherwise, we have ordRs (f R)Rs = r, and
there exists a monomial q in xs , ys , . . . , zs such that
g1
gt
f
(f R)Rs = ( )Rs = ( λ1 + . . . + λt )Rs .
q
q
q
g1
gt
The elements q , . . . , q are distinct monomials in xs , ys , . . . , zs . Since ordRs is a
monomial valuation and ordRs (f R)Rs = r, we have ordRs (gi R)Rs = r. Thus
gi
(gi R)Rs = ( )Rs = (xas i ysbi · · · zsci )Rs .
q
Since the transform from Rs to Rs+1 is in the x-direction, ordRs+1 (gi R)Rs+1 =
r − ai < r. Set q ′ = qxrs . Then
g1
gt
f
(f R)Rs+1 = ( ′ )Rs+1 = ( ′ λ1 + . . . + ′ λt )Rs+1 .
q
q
q
gi
Rs+1 . Since ord
Rs+1 is a monomial valuation and
q ′ Rs+1 = (gi R)
gt
g1
s+1 ≤ r − a < r. This
i
q ′ , . . . , q ′ are distinct, we have ordRs+1 (f R)
We have
the mono-
mials
completes
the proof of Theorem 4.3.
Remark 4.4. For the transform of an ideal as defined in Definition 2.1, we observe
some properties of the transform of a monomial ideal with respect to a monomial
local quadratic transform from R0 to R1 .
Let (R1 , m1 ) be the monomial local quadratic transformation of R0 in the wdirection, where w ∈ ∆(m0 ), and m1 := (x1 , y1 , . . . , w1 , . . . , z1 ), where
x
y
z
x1 :=
, y1 :=
, . . . , w1 := w, . . . , z1 :=
.
w
w
w
Let p := xα1 y β1 · · · wη1 · · · z γ1 and q := xα2 y β2 · · · wη2 · · · z γ2 be distinct monomials
in R and let I := (p, q)R. Assume that r := ordR (p) ≤ ordR (q) := s. Thus
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
13
r = ordR (I). In R1 , the monomials p and q can be written as follows:
p : = xα1 y β1 · · · wη1 · · · z γ1
= (wx1 )α1 (wy1 )β1 · · · wη1 · · · (wz1 )γ1
= w1r (xα1 1 y1β1 · · · w10 · · · z1γ1 ) = w1r p1 = wr p1
and
q : = xα2 y β2 · · · wη2 · · · z γ2
= (wx1 )α2 (wy1 )β2 · · · wη2 · · · (wz1 )γ2
= w1r (xα1 2 y1β2 · · · w1s−r · · · z1γ2 ) = w1r q1 = wr q1 .
Thus I1 := I R1 = (p1 , q1 )R1 is the transform of I in R1 . We have
(3)
ordR1 (I1 ) ≤ ordR1 (p1 )
and
ordR1 (p1 ) = r − η1 ≤ ordR (p).
Hence if η1 > 0, then ordR1 (I1 ) < ordR (I).
It is observed in [HKT, Lemma 3.6] that ordR1 (I R1 ) ≤ ordR0 (I) even without
the assumption that I is a monomial ideal.
Remark 4.5. Let notation be as in Setting 4.2, and let V be a valuation domain
that dominates Rn . The following are equivalent:
(1) The transform from R0 to R1 is in the x-direction.
(2) v(x) < v(w) for each w ∈ ∆(m) with w 6= x.
(3) m V = xV .
Assume that the transform from R0 to R1 is in the x-direction. By relabeling the
variables, we may assume that
v(y) = min v(w) | w ∈ ∆(m)
with w 6= x .
With this assumption, the following are equivalent:
(a) The set Dy is nonempty.
(b) The set Dy ∪ · · · ∪ Dz is nonempty.
(c) There exists a positive integer s in {1, . . . , n − 1} such that the transform
from R0 to Rs is in the x-direction and the transform from Rs to Rs+1 is in
the y-direction.
(d) We have sv(x) < v(y) < (s + 1)v(x) for some s in {1, . . . , n − 1}.
Remark 4.6. Let notation be as in Remark 4.5. Assume that the sets Dx , Dy , . . . , Dz
are all nonempty. Label the variables x, y, . . . , z in ∆(m) as w1 , w2 , . . . wd , where
14
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
i < j if the sequence {Ri }ni=0 contains a transform in the wi -direction before it con-
tains a transform in the wj -direction. Thus with the assumptions of Remark 4.5,
we have w1 = x, w2 = y, and v(wi ) < v(wj ) if and only if i < j.
Definition 4.7. Let V be a valuation domain, and let b1 , . . . , bs be nonzero elements
√
√
of V . We say that b1 , . . . , bs are comparable in V if b1 V = · · · = bs V , that is,
the ideals bi V and bj V have the same radical for all i and j between 1 and s.
Proposition 4.8. Let notation be as in Remark 4.6. Thus the sets Dx , Dy , . . . , Dz
are all nonempty, and our fixed regular system of parameters (x, y, . . . , z) is also denoted (w1 , w2 , . . . , wd ). Let ai := v(wi ) for 1 ≤ i ≤ d. Thus the tuple (a1 , a2 , . . . , ad )
is the family of v-values of (w1 , w2 , . . . , wd ). Then we have:
(1) 0 < a1 < a2 < · · · < ad .
(2) There exists a positive integer s such that sa1 < a2 < (s + 1)a1 .
(3) We have
(j − 2)aj < a1 + a2 + · · · + aj−1
for j with 3 ≤ j ≤ d.
(4) The elements w1 , . . . , wd are comparable in V .
Proof. Item 1 follows from Remark 4.6, and item 2 follows from Remark 4.5.
In order to give a detailed proof of item 3, we introduce more notation. For each
i ∈ {0, 1, . . . , n} and each positive integer j with 1 ≤ j ≤ d, we define wij as follows:
m = m0 = (x, y, . . . , z)R0 = (w01 , w02 , . . . , w0d )R0
mi = (xi , yi , . . . , zi )Ri = (wi1 , wi2 , . . . , wid )Ri .
To prove item 3, it suffices to show that if (ℓ − 2)aℓ ≥ a1 + a2 + · · · + aℓ−1 for some
ℓ with 3 ≤ ℓ ≤ d, then the set Dwℓ is empty. Let
a01 , a02 , . . . , a0d = v(w01 ), v(w02 , . . . , v(w0d )),
ak1 , ak2 , . . . , akd = v(wk1 ), v(wk2 ), . . . , v(wkd )
for each k ≥ 1.
With this notation, we show (ℓ − 2)akℓ ≥ ak1 + · · · + akℓ−1 for all k ≥ 0 by induction
on k. By hypothesis (ℓ − 2)a0ℓ ≥ a01 + · · · + a0ℓ−1 . Assume the assertion holds for
k and let ek := min{ak1 , ak2 , . . . , akℓ−1 }. Then we have
(ℓ − 2)ak+1ℓ
= (ℓ − 2)(akℓ − ek ) = (ℓ − 2)akℓ − (ℓ − 2)ek
≥ ak1 + · · · + akℓ−1 − (ℓ − 2)ek
= ak+11 + · · · + ak+1ℓ−1 .
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
15
Therefore the assumption that Dwℓ is nonempty implies that
(ℓ − 2)aℓ < a1 + a2 + · · · + aℓ−1 .
This proves item 3. Item 4 now follows from items 1, 2 and 3.
5. Infinite Directed Unions of Monomial Local Quadratic Transforms
Remark 5.1. Let (R, m) be a d-dimensional regular local ring with d ≥ 2, and let
V be a valuation domain birationally dominating R and zero-dimensional over R.
Let v be a valuation associated to V . Then there exists a uniquely defined infinite
sequence
(4)
{(Rn , mn )}n≥0 ≡ R =: R0 ⊂ R1 ⊂ R2 ⊂ · · · ⊂ Rn ⊂ · · · ⊂ V
of local quadratic transforms of R along V .
(1) If the v-values v(x), v(y), . . . , v(z) of a regular system of parameters for m are
rationally independent real numbers, then the sequence given in Equation 4
is monomial with respect to this regular system of parameters. However,
as Shannon shows in [S, Example 4.17], in this situation it is often the case
S
that S := n≥0 Rn is properly contained in V .
(2) If v is rank one discrete, then the sequence given in Equation 4 switches
S
strongly infinitely often, that is V = n≥0 Rn . To see this, let pq ∈ V , with
p, q ∈ R. Assume that from R to R1 , we divide by x, where v(x) > 0. Let
p1 = p/x and q1 = q/x, so that
p
q
=
p1
q1 ,
where v(q1 ) < v(q). A simple
induction argument shows that we can write
p
q
=
pn
qn ,
where pn , qn ∈ Rn and
v(qn ) = 0, for some n. See for example [HRW, Proposition 23.3]. Thus the
sequence {(Rn , mn )}n≥0 switches strongly infinitely often.
(3) If V is rank one and the rational rank of V is less than d, it is possible that
the field extension R/ m ⊂ V / mv is an infinite algebraic field extension. This
is illustrated in Example 4 of page 104 in [ZS2].
(4) Assume that V has rank one and rational rank r and that the local quadratic sequence {(Rn , mn )}n≥0 along V switches strongly infinitely often.
Then for each sufficiently large integer n, there exists a sequence of elements
w1 , . . . , wr that are part of a regular system of parameters for mn and are
such that v(w1 ), . . . , v(wr ) are rationally independent.
Proof. Since V has rational rank r, there exist elements a1 , . . . , ar in R such
S
that v(a1 ), . . . , v(ar ) are rationally independent. Since V = n≥0 Rn , it
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
16
follows from [S, Prop. 4.18] that for sufficiently large integers n, the product a1 · · · ar is a monomial with respect to a regular system of parameters
w1 , . . . , wd of Rn . It follows that v(a1 ), . . . , v(ar ) are in the Q-linear subspace L of R spanned by v(w1 ), . . . , v(wd ). Since v(a1 ), . . . , v(ar ) are linearly
independent over Q, the vector space L has rank at least r. By re-labeling
the wi , it follows that v(w1 ), . . . , v(wr ) are rationally independent.
We observe in Proposition 5.2 that an infinite sequence {(Rn , mn )}n≥0 of local
quadratic transforms that switches strongly infinitely often is eventually monomial
S
if V = n≥0 Rn has maximal rational rank .
Proposition 5.2. Let (R, m), be a d-dimensional regular local ring and let V be a
rank 1, rational rank d valuation domain birationally dominating R. Consider the
infinite sequence
R =: R0 ⊂ R1 ⊂ · · · ⊂ Rn ⊂ · · · ⊂ V
S
of local quadratic transforms of R along V . If V = ∞
n=0 Rn , then there exists an
integer n ≥ 0 and a regular system of parameters x, y, . . . , z of Rn such that the
sequence {Ri }i≥n is monomial in these regular parameters.
Proof. By Remark 5.1.4, there exists an integer n ≥ 0 and a regular system of
parameters x, y, . . . , z of Rn such that v(x), v(y), . . . , v(z) are rationally independent.
It follows that the sequence {Ri }i≥n is monomial with respect to x, y, . . . , z.
In Setting 5.3, we extend the notation of Setting 4.2 to an infinite sequence.
Setting 5.3. Let (R, m) and x, y, . . . , z be as in Setting 4.2. We consider an infinite
sequence of regular local rings
(5)
{(Ri , mi )}i≥0 ≡ R =: R0 ⊂ R1 ⊂ R2 ⊂ · · · ⊂ Ri ⊂ · · · ,
where Ri+1 is a monomial local quadratic transformation of Ri for each i ≥ 0. As in
Setting 4.2, let Dx (respectively, Dy , . . . , Dz ) denote the set of nonnegative integers
i for which the transform from Ri to Ri+1 is in the x-direction (respectively, the
y-direction, . . . , the z-direction).
S
Set S := i≥0 Ri . It is straightforward to prove that S is a normal local domain
S
with field of fractions Q(R) and maximal ideal mS := i≥0 mi . Our assumption that
the local quadratic transforms are monomial implies that there is no residue field
S
extension, that is, i≥0 (Ri / mi ) = R/ m. Moreover, if S is not a valuation domain,
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
17
it is known that there exist infinitely many valuation domains V that birationally
dominate S and have positive dimension over S, that is, the residue field of V
has positive transcendence degree over the field S/ mS ; see [A, Lemma 7] or [SH,
Exercise 6.24].
Theorem 5.4. Let notation be as in Setting 5.3, and let V be a valuation domain
that birationally dominates S. The following are equivalent:
(1) The sets Dx , Dy , . . . , Dz are all infinite.
(2) {(Rn , mn )}n≥0 switches strongly infinitely often.
(3) S = V and V has rank 1.
Proof. (2) ⇒ (1) : Assume that item 1 fails. Then at least one of the sets Dx , Dy , . . . , Dz
is finite. By relabeling, we may assume that Dx is finite. Then there exists a pos-
itive integer j such that for n ≥ j, the transform from Rn to Rn+1 is not in the
x-direction. Let mj := (xj , yj , . . . , zj )Rj . Then pj := (xj )Rj has the property that
S
n≥0 Rn ⊆ Rj p . Thus {(Rn , mn )}n≥0 does not switch strongly infinitely often.
j
(1) ⇒ (2) : Assume that item 1 holds. To prove that {(Rn , mn )}n≥0 switches
strongly infinitely often it suffices to show for each integer j ≥ 0 and nonzero element
f ∈ Rj that there exists an integer n ≥ j such that the transform (f Rj )Rn = Rn ,
cf. Remark 3.2. Let r := ordRj (f ). To prove that the transform (f Rj )Rn = Rn for
some n, it suffices to prove that the order of the transform (f Rj )Rs is less than r
for some integer s > j, and this is immediate from Theorem 4.3. (2) ⇒ (3) : This
is clear by Remark 3.3. (3) ⇒ (2) : This is shown by Granja [Gr, Theorem 13].
Lemma 5.5. Let notation be as in Setting 5.3, and let p and q be distinct monomials
in R. If the sets Dx , Dy , . . . , Dz are all infinite, then there exists an integer t such
that either p/q ∈ mt or q/p ∈ mt .
Proof. Let I := (p, q)R. We prove Lemma 5.5 by induction on r := ordR (I). The
case where r = 0 is clear. Assume the assertion of Lemma 5.5 holds for all nonnegative integers less than r.
We may assume ordR (p) ≤ ordR (q). Then r = ordR (p). Assume notation as in
Remark 4.4. Thus the transform from R to R1 is in the w-direction. If η1 > 0, then
as shown in Equation 3, we have ordR1 (p1 ) < r. Hence by our inductive hypothesis
applied to the ideal I1 in R1 , we have, for some positive integer t, either
or
q
p
=
q1
p1
∈ mt .
p
q
=
p1
q1
∈ mt
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
18
If η1 = 0, then ordR1 (p1 ) = ordR (p) = r. If ordR1 (q1 ) < r, then again by
induction applied to the ideal I1 in R1 , we have either
p
q
=
p1
q1
for some positive integer t.
∈ mt or
q
p
=
q1
p1
∈ mt
If η1 = 0 and ordR1 (q1 ) ≥ r, then the ideal I1 = (p1 , q1 )R1 satisfies the same
hypothesis as the ideal I = (p, q)R. Let I2 = (p2 , q2 )R2 denote the transform of
p2
q2
mt or pq
I1 in R2 , where
p
q
=
p2
q2
∈
= pq . If ordR2 (I2 ) < r, then for some positive integer t, either
=
q2
p2
∈ mt . If ordR2 (I2 ) = r, we define I3 = (p3 , q3 )R3 and
continue. Let Ii = (pi , qi )Ri denote the transform of I in Ri , where
pi
qi
= pq , for each
i ≥ 1.
If ordRi (Ii ) = r for all i, then the transform from Ri−1 to Ri is never in a
direction where the associated exponent of p is positive. This is impossible since
the sets Dx , Dy , . . . , Dz are all infinite and the monomial p contains at least one
positive exponent. We conclude by induction that there exists an integer t such that
either p/q ∈ mt or q/p ∈ mt .
Corollary 5.6. Assume notation as in Theorem 5.4. The valuation domain V has
rational rank d = dim R, and V is a zero-dimension over R. If v is a valuation
associated to V , then v(x), v(y), . . . , v(z) are rationally independent. In particular,
v is a monomial valuation
Proof. Assume that
a1 v(x) + b1 v(y) + · · · + c1 v(z) = a2 v(x) + b2 v(y) + · · · + c2 v(z),
with (a1 , b1 , . . . , c1 ) and (a2 , b2 , . . . , c2 ) being d-tuples of nonnegative integers. Then
we have
v(xa1 y b1 · · · z c1 ) = v(xa2 y b2 · · · z c2 ),
and hence xa1 y b1 · · · z c1 = xa2 y b2 · · · z c2 by Lemma 5.5. Thus we have
a1 = a2 , b1 = b2 , . . . , c1 = c2 .
By [SH, Theorem 6.6.7], we have
rat. rank v + tr. degR/ m k(v) ≤ dim(R).
Since d ≤ rat. rank v, we have rat. rank v = d. It follows that tr. degR/ m k(v) =
0.
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
19
6. Infinite Directed Unions of Local Quadratic Transforms
Related to Item 2 of Remark 5.1, Example 6.1 demonstrates the existence of a
nondiscrete rational rank 1 valuation domain V that birationally dominates and
is zero-dimensional over a 3-dimensional regular local ring R and is such that the
sequence {(Rn , mn )}n≥0 of local quadratic transforms of R along V as in Equation 4
S
fails to switch strongly infinitely infinitely often, that is, we may have n≥0 Rn ( V .
Example 6.1. Let x, y be indeterminates over a field k, and let A0 = k[[x, y]] be the
formal power series ring in x and y over k. We describe a sequence {(An , nn )}n≥0
of local quadratic transforms of A0 , such that nn = (xn , yn )An is the maximal ideal
of An , where for each even integer n ≥ 0, we set An+1 = An [ xynn ](xn , yn −1) , and for
xn
each odd integer n ≥ 1, we set An+1 = An [ xynn ](yn , xn ) . Thus for each integer k ≥ 0,
yn
we have
x2k+1 = x2k ,
y2k+1 =
y2k − x2k
,
x2k
x2k+2 =
x2k+1
,
y2k+1
and
y2k+2 = y2k+1 .
Notice that An+1 is the localization of An [xn+1 , yn+1 ] at the maximal ideal generated
by xn+1 , yn+1 .
S
Let W = ∞
n=0 An . By Remark 3.1, W is a valuation domain that birationally
dominates A0 and has residue field k. Let w denote the valuation associated to W
such that w(x) = 1. The fact that W dominates every An implies that W is rational
k
rank 1 nondiscrete and that w(x2k ) = w(x2k+1 ) = w(y2k ) = w(y2k−1 ) = 21 , for
k
k+1
, it follows that
all k ≥ 0. Since for k ≥ 0, w(n2k ) = 12 and w(n2k+1 ) = 21
1
X
X 1 k X 1 k+1
1
2
w(nk ) =
+
= 3.
+
=
2
2
1 − 12
1 − 21
k≥0
k≥0
k≥0
Let p(x) ∈ k[[x]] be such that x and p(x) are algebraically independent over k
and let z = x4 p(x). Let R = k[x, y, z](x,y,z) , V = W ∩ k(x, y, z), and v = w|R(x,y,z) .
It follows that V is a rational rank 1 nondiscrete valuation domain that birationally
dominates R and has residue field k. Let {(Rn , mn )}n≥0 be the sequence of local
quadratic transforms of R along V . Since v(z) ≥ 4, we have mn = (xn , yn , zn ),
where xn , yn are the generators of nn as defined above and zn is defined inductively
by zn+1 =
z-direction,
zn
xn
S
for n even and zn+1 =
n≥0 Rn
zn
yn
⊂ R(zR) , and thus
S
for n odd. Since we never divide in the
n≥0 Rn
(V.
By stringing together infinitely many finite sequences of monomial local quadratic
transforms as in Theorem 4.3 and interspersing between each two a local quadratic
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
20
transform that is not monomial, we obtain examples of infinite sequences {Ri }i≥0
S
that switch strongly infinitely often and have the property that V = i≥0 Ri has
rational rank less than d = dim R. We give explicit examples with d = 3 in Exam-
ple 6.3. We use the following remark.
Remark 6.2. Let (R, m) be a Noetherian local domain, and let V be a valuation domain dominating R.
that dimR/ m
m / m2
Let v be a valuation associated to V .
Assume
= d, and that there exist elements x1 , . . . , xd in m such that
v(x1 ) < v(x2 ) < · · · < v(xd ) < v(m2 ).
v(y) ∈ {v(x1 ), . . . , v(xd )}.
Then for each y ∈ m \ m2 , we have
To see this, let R = I0 ) I1 ) I2 ) . . . be the sequence of valuation ideals of V
in R, where In+1 = {a ∈ R | v(a) > v(In )}. For 1 ≤ i ≤ d, define Ji = {a ∈ R |
v(a) ≥ v(xi )}. Define J0 = R and Jd+1 = {a ∈ R | v(a) ≥ v(m2 )}. The assumption
that v(xd ) < v(m2 ) implies that Jd+1 ( Jd . Thus we have the chain of inclusions of
ideals,
R = J0 ) J1 ) . . . ) Jd ) Jd+1 ⊃ m2 .
Since λR (R/ m2 ) = d + 1, it follows that for 0 ≤ i ≤ d, Ji /Ji+1 is a simple Rmodule, and it follows that Jd+1 = m2 . Thus for 0 ≤ i ≤ d + 1, Ii = Ji , and for each
y ∈ m \ m2 , v(y) ∈ {v(I1 ), . . . , v(Id )} = {v(x1 ), . . . , v(xd )}.
Example 6.3. Let (R, m) be a 3-dimensional regular local ring, and let m =
(x, y, z)R. Define:
y z
m
m1 = (x1 , y1 , z1 )R1 = (x, , )R1
R1 = R[ ](x, y , z ) and
x
x
x
x x
m1
x2 y z
R2 = R1 [ ]( x1 ,y1 , z1 ) and
m2 = (x2 , y2 , z2 )R2 = ( , , )R2
y1 y1 y1
y x y
m2
x2 y 2 z
R3 = R2 [ ]( x2 , y2 ,z2 ) and
m3 = (x3 , y3 , z3 )R3 = ( , , )R3
z2 z2 z2
z xz y
x2 y 2
z2
m3
R4 = R3 [ ](x3 , y3 −1, z3 −1) and m4 = (x4 , y4 , z4 )R4 = ( , 3 − 1, 2 − 1)R4
x3
x3
x3
z x
x y
For each valuation domain V birationally dominating R4 , we have v(y) = 23 v(x) and
v(z) = 74 v(x).
Starting from (R4 , m4 ) and m4 = (x4 , y4 , z4 )R4 , we make a sequence R4 ⊂ R5 ⊂
R6 ⊂ R7 of monomial local quadratic transforms with respect to the fixed basis
x4 , y4 , z4 of m4 such that R4 to R5 is in the x-direction, R5 to R6 is in the ydirection, and R6 to R7 is in the z-direction. With (x7 , y7 , z7 )R7 the monomial
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
21
regular system of parameters for m7 , we define the local quadratic transform R7 to
R8 in a manner similar to that from R3 to R4 . Thus x8 := x7 , y8 :=
z8 :=
z7
x7
y7
x8
− 1, and
− 1. If V is a valuation birationally dominating R8 , then v(y4 ) = 32 v(x4 )
and v(z4 ) = 47 v(x4 ).
In a similar manner, for each positive integer n, we inductively define the sequence
of local quadratic transforms R4n ⊂ R4n+1 ⊂ R4n+2 ⊂ R4n+3 ⊂ R4n+4 . For each
valuation domain V that birationally dominating R4n+4 , we have v(y4n ) = 23 v(x4n )
and v(z4n ) = 74 v(x4n ).
Theorem 4.3 implies that the sequence {Ri }i≥0 switches strongly infinitely ofS
ten. Hence V = i≥0 Ri is a rank one valuation domain by Remark 3.3 and
V is a zero-dimension over R. The rational rank of V is 1, for if the rational
rank of V were greater than 1, then by Remark 5.1.4, there exists an integer
n > 0 and elements w1 , w2 ∈ R4n such that ordR4n (w1 ) = ordR4n (w2 ) = 1 and
v(w1 ), v(w2 ) are rationally independent. Since v(x4n ) < v(y4n ) < v(z4n ) < v(m24n )
and v(x4n ), v(y4n ), v(z4n ) are rationally dependent, Remark 6.2 implies that v(w1 )
and v(w2 ) are in {v(x4n ), v(y4n ), v(z4n )} and hence are rationally dependent.
Theorem 3.5 implies the following result for an infinite sequence of monomial local
quadratic transforms.
Construction 6.4. Let notation be as in Theorem 5.4 and assume that the sets
Dx , Dy , . . . , Dz are all infinite. Then:
(1) There exists an infinite decreasing chain {In }n≥0 of v-ideals in R such that
each In is a monomial ideal and λR (In /In+1 ) = 1 for each n ≥ 0.
T
(2) We have n≥1 In = (0).
(3) If p is a monomial in R, then pV ∩ R = In for some n ≥ 0.
S
(4) V = n≥0 Rn is a rank one valuation domain.
Proof. Since the sequence {Rn }n≥0 is monomial, we have R/ m = Rn / mn for all
n ≥ 0. Since the sets Dx , Dy , . . . , Dz are all infinite, the sequence {Rn }n≥0
S
switches infinitely often by Theorem 5.4. Remark 3.3 implies that V = n≥0 Rn is
a rank one valuation domain. Theorem 3.5 implies items 1, 2 and 3.
7. Properties of infinite sequences of local quadratic transforms
Setting 7.1. Let (R, m) be a d-dimensional regular local ring with d ≥ 2 and
let V be a zero-dimensional real valuation domain birationally dominating R with
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
22
a corresponding valuation v. Consider the infinite sequence {(Ri , mi )}i≥0 of local
quadratic transforms of R along V ,
R = R0 ⊂ R1 ⊂ · · · ⊂ Rn ⊂ · · · ⊂ ∪n≥0 Rn ⊂ V.
Let s denote the infinite sum,
s=
∞
X
v(mi ).
i=0
Thus s is either ∞ or a positive real number.
In Theorem 7.2, we describe the value of s in the monomial case.
Theorem 7.2. Assume the notation of Setting 7.1 and that m = (x, y, . . . , z)R,
where v(x), v(y), . . . , v(z) are rationally independent. Then,
P∞
v(x)+v(y)+···+v(z)
(1) s =
. In particular, s is finite.
i=0 v(mi ) ≤
(d−1)
(2) The following are equivalent:
(a) Equality holds in Item 1.
(b) The sequence {(Rn , mn )}n≥0 switches strongly infinitely often.
(c) lim v(wn ) = 0, for each variable w ∈ ∆(m0 ).
n→∞
Proof. For item 1, we prove by induction that the following equation holds for every
integer n ≥ 0.
(6)
n−1
X
!
v(mi )
i=0
+
v(x) + v(y) + · · · + v(z)
v(xn ) + v(yn ) + · · · + v(zn )
=
.
d−1
d−1
Equality is clear in the case where n = 0. Assume that the claim is true for n.
By re-arranging variables, we may assume that v(xn ) = v(mn ), so that the local
quadratic transform from Rn to Rn+1 is in the x-direction. Hence for every variable
w that is not x, wn+1 := wn /xn , so v(wn+1 ) = v(wn ) − v(xn ). Thus,
!
n
X
v(xn+1 ) + v(yn+1 ) + · · · + v(zn+1 )
v(mi ) +
d−1
i=0
!
n−1
X
(v(xn ) + v(yn ) + · · · + v(zn )) − (d − 1)v(xn )
v(mi ) + v(xn ) +
=
d−1
i=0
!
n−1
X
v(xn ) + v(yn ) + · · · + v(zn )
v(mi ) +
=
d−1
i=0
=
v(x) + v(y) + · · · + v(z)
.
d−1
By taking limits, we conclude that s =
P∞
i=0 v(mi )
≤
v(x)+v(y)+···+v(z)
.
(d−1)
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
For item 2, we let En :=
we have
Pn−1
i=0
23
v(mi ). We first show (a) ⇒ (c): By assumption,
v(xn ) + v(yn ) + · · · + v(zn )
v(x) + v(y) + · · · + v(z)
thus lim
= 0.
n→∞
n→∞
d−1
d−1
For each w ∈ ∆(m0 ), the sequence {v(wn )}n≥0 is a nonincreasing sequence of positive
lim En =
real numbers. Hence each lim v(wn ) exists and is a nonnegative real number. Since
n→∞
d ≥ 2, we have lim {v(xn )+ v(yn )+ · · ·+ v(zn )} = 0, and thus we have lim v(wn ) =
n→∞
n→∞
0, for each variable w ∈ ∆(m0 ).
(c) ⇒ (a): . Assume that lim v(wn ) = 0, for each variable w ∈ ∆(m0 ). That is,
n→∞
we have
lim v(xn ) = 0, lim v(yn ) = 0, · · · , lim v(zn ) = 0.
n→∞
n→∞
n→∞
By Equation 6 we have
v(x) + v(y) + · · · + v(z)
v(xn ) + v(yn ) + · · · + v(zn )
En =
−
.
d−1
d−1
Hence
n v(x) + v(y) + · · · + v(z) o
n v(x ) + v(y ) + · · · + v(z ) o
n
n
n
lim En = lim
− lim
n→∞
n→∞
n→∞
d−1
d−1
v(x) + v(y) + · · · + v(z)
=
.
d−1
The last equality follows by assumption.
(c) ⇒ (b): Assume that lim v(wn ) = 0, for each variable w ∈ ∆(m0 ). Suppose
n→∞
by way of contradiction that the sequence {(Rn , mn )}n≥0 does not switch strongly
infinitely often. By Theorem 5.4, there exists a variable w ∈ ∆(m0 ) such that Dw
is finite. That is, there exists positive integer n0 such that the sequence Rn0 ⊂
Rn0 +1 ⊂ Rn0 +2 ⊂ · · · never goes in the w-direction. By replacing n0 by zero, we
may assume that the sequence {(Rn , mn )}n≥0 never goes in the w-direction. That
is,
v(wn ) 6= min{v(xn ), v(yn ), . . . , v(wn ), . . . , v(zn )} for every integer n ≥ 0.
From Equation 6, we have
v(w) = v(wn ) + En
and
[ + · · · + v(z) = v(xn ) + v(yn ) + · · · + v(w
\
v(x) + v(y) + · · · + v(w)
n ) + · · · + v(zn ) + (d − 2)En .
By assumption, we have
v(w) = lim En
n→∞
[ + · · · + v(z) = (d − 2) lim En .
and v(x) + v(y) + · · · + v(w)
n→∞
[ + · · · + v(z) = (d − 2)v(w),
Using both equalties, we have v(x) + v(y) + · · · + v(w)
which is a contradiction, since v(x), v(y), . . . , v(z) are rationally independent.
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
24
(b) ⇒ (c): By item (1), we have s :=
P∞
i=0 v(mi )
is finite. Hence lim v(mn ) =
n→∞
lim En+1 − lim En = 0. Fix w ∈ ∆(m0 ), we have v(wn ) = v(mn ) for every n such
n→∞
n→∞
that the transform from Rn to Rn+1 is in the w-direction. By assumption, Dw = ∞,
and hence lim v(wn ) = lim v(mn ) = 0
n→∞
n→∞
Proposition 7.3. Assume the notation of Setting 7.1. Then we have the following:
(1) If V is a DVR, then s = ∞, and ∪n≥0 Rn = V .
(2) If s = ∞, then the sequence {Rn }n≥0 switches strongly infinitely often.
(3) If V has rational rank r ≥ 2, then s < ∞. Thus if s = ∞, then V has
rational rank r = 1.
(4) If V has rational rank 1 and V is not a DVR, then both s = ∞ and s < ∞
are possible.
Proof. If V is discrete, we may assume the value group of V is Z, and hence v(mn ) ≥
1 for all n ≥ 0. Thus item 1 is clear.
Item 2 is proved by Granja, Martinez and Rodriguez in [GMR, Proposition 23].
To see Item 3, assume that V has rational rank r ≥ 2 and suppose by way of
contradiction that s = ∞. By Item 2, the sequence switches strongly infinitely
often. By Remark 5.1.4, there is some n > 0 and elements x, y that are part of some
regular system of parameters for Rn such that v(x), v(y) are rationally independent.
By replacing R by Rn , we may assume that there are elements x, y in some regular
system of parameters for R such that v(x), v(y) are rationally independent.
We show that s ≤ v(x) + v(y) by inductively proving that for all n ≥ 0, there are
elements xn , yn of some regular system of parameters for Rn such that v(xn ), v(yn )
P
n−1
are rationally independent and
i=0 v(mi ) + v(xn ) + v(yn ) ≤ v(x) + v(y).
Taking x0 = x and y0 = y, the base case n = 0 is clear.
Assume the claim is true for n. Thus we have elements xn , yn ∈ mn such that
v(xn ), v(yn ) are rationally independent and
!
n−1
X
v(mi ) + v(xn ) + v(yn ) ≤ v(x) + v(y).
(7)
i=0
Let z ∈ mn denote an element of minimal v-value. Then v(z) and at least one of
v(xn ), v(yn ) are rationally independent, so we may assume without loss of generality
that v(z), v(yn ) are rationally independent. Thus z, yn are part of a regular system of
parameters for Rn . Set xn+1 = z and yn+1 =
yn
z ,
so v(xn+1 ), v(yn+1 ) are rationally
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
25
independent, and v(yn+1 ) = v(yn ) − v(z). We have v(mn ) = v(z) ≤ v(xn ). Thus,
!
!
n−1
n
X
X
v(mi ) + v(z) + v(z) + v(yn ) − v(z)
v(mi ) + v(xn+1 ) + v(yn+1 ) =
i=0
≤
i=0
n−1
X
i=0
!
v(mi )
+ v(xn ) + v(yn )
≤ v(x) + v(y)
where the last inequality follows from Equation 7.
We conclude that s < ∞, in contradiction to the assumption that s = ∞. This
proves Item 3.
Example 6.3 shows that s < ∞ is possible in the case where the rational rank of V
1
1
+ 16
)+. . .,
is 1 and V is not a DVR. The pattern gives s = (1+ 12 + 41 + 41 )+( 14 + 18 + 16
P∞ 1
P∞ 1
so s = n=0 2n + 2 n=1 4n = 38 . To complete the proof of item 4, we show in
Examples 7.11 and 7.12 that s = ∞ is possible in the case where the rational rank
S
of T = n≥0 An is 1 and T is not a DVR.
Setting 7.4. Let (R, m) be a d-dimensional regular local ring with d ≥ 2, and with
R := R0 , let {(Rn , mn )}n≥0 be an infinite sequence of local quadratic transforms
with dim Rn = dim R for all n. Let yn ∈ ∆(mn ) be such that the transform from
S
Rn to Rn+1 is in the yn -direction. Let S := n≥0 Rn . It is well known that S is a
S
normal local domain with unique maximal ideal mS := n≥0 mn . Assume that for
some nonnegative integer j there exists a regular prime element x in Rj such that
S ⊂ (Rj )xRj . In considering properties of the directed union, we may assume the
sequence starts at Rj . Thus with a change of notation, we assume that j = 0. Let
P := xRxR ∩ S and let T := S/P and mT := mS /P . For each n ≥ 0, let xn be the
S
transform of x in Rn . Then P = n≥0 xn Rn . For each n ≥ 0, let
mn
Rn
and nn :=
.
x n Rn
x n Rn
Each (An , nn ) is a (d − 1)-dimensional regular local ring. Moreover {(An , nn )}n≥0 is
S
an infinite sequence of local quadratic transforms. Hence T = n≥0 An is a normal
S
local domain with maximal ideal mT = n≥0 nn . Let νx be the x-adic valuation
An :=
associated with the valuation domain RxR . For an element g ∈ S, let g denote the
image of g in T .
Theorem 7.5. Let notation be as in Setting 7.4. The following are equivalent.
(1) The sequence {(Rn , mn )}n≥0 is height one directed.
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
26
(2) S is a valuation domain.
(3) T is a rank one valuation domain with associated real valuation ν such that
P
n≥0 ν(nn ) = ∞.
(4) The sequence {(An , nn )}n≥0 switches strongly infinitely often, and for the
P
associated real valuation ν we have n≥0 ν(nn ) = ∞.
Proof. (1) ⇒ (2): This is proved by Granja in [Gr, Theorem 13].
(2) ⇒ (1): By [Gr, Proposition 7], the valuation domain S has rank either 1 or
2. By assumption, the sequence {(Rn , mn )}n≥0 does not switch strongly infinitely
often. Hence by [Gr, Theorem 13], the sequence {(Rn , mn )}n≥0 is height one directed
and S has rank 2.
(2) ⇒ (3): Assume that S is a valuation domain. Then T = S/P is a valuation
domain of the field RxR /xRxR . Since P is a nonzero nonmaximal prime ideal in S
P
and dim S = 2, we have dim T = 1. It remains to show that s := n≥0 ν(nn ) = ∞.
Suppose by way of contradiction that s < ∞. Let z ∈ ∆(m0 ) be such that νx (z) = 0,
(that is zR 6= xR). By the Archimedean property, there exists an integer N > 0
such that N ν(z) > s. Consider the following element
f := x − z N .
For simplicity of notation, we set g := z N . Then ν(g) = ν(z N ) > s. For each n ≥ 0,
let fn be the transform f in Rn . Let νf be the f -adic valuation associated with the
valuation domain Rf R . To see S ⊂ Rf R , we prove by induction on n the following
claim;
Claim 7.6. For every integer n ≥ 0, there exists gn ∈ Rn such that
fn = xn − gn ,
where νf (fn ) = 1,
νf (gn ) = 0 and ν(gn ) >
∞
X
ν(ni ).
i=n
Proof. The case where n = 0 is clear by construction. Assume that the claim holds
for n. Since the transform from Rn to Rn+1 is in the yn -direction, we have
fn
xn gn
gn
fn+1 =
=
−
= xn+1 − .
yn
yn
yn
yn
gn
We set gn+1 := yn . Then we have
νf (gn+1 ) = νf (gn ) − νf (yn ) = 0 and
∞
∞
g
X
X
n
ν(gn+1 ) = ν
= ν(gn ) − ν(yn ) >
ν(ni ).
ν(ni ) − ν(nn ) =
yn
i=n
i=n+1
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
27
By Claim 7.6, we have Rn ⊂ Rf R for each positive integer n. Hence S ⊂ Rf R .
Since RxR 6= Rf R , this contradicts the fact that the sequence {(Rn , mn )}n≥0 is
height one directed.
(3) ⇒ (2): Assume that T is a rank one valuation domain with associated real
P
valuation ν such that n≥0 ν(nn ) = ∞. Let W be the composite valuation domain
defined by the valuations ν and νx . Thus
W = {α ∈ Q(R) | νx (α) > 0
or
νx (α) = 0 and
ν(e
α) ≥ 0},
where α
e denotes the image of α in RxR /xRxR . We have S ⊂ W ⊂ RxR . We prove
the following claim :
Claim 7.7. W ⊂ S.
Proof. Let α ∈ W . Then α = xt hg ,where g, h ∈ R \ xR are relatively prime in R
and t = νx (α) ≥ 0. We consider two cases
(Case i): Assume that t = νx (α) = 0. Since α ∈ W , we have ν(e
α) ≥ 0. Let
r0 := min{ ordR0 (g), ordR0 (h) }.
If r0 = 0, then at least one of g or h is a unit in R. If h is a unit in R then α ∈ R ⊂ S.
If g is a unit in R, then ν(e
g ) = 0. Since ν(e
α) = ν(e
g ) − ν(e
h) ≥ 0, we must also have
ν(e
h) = 0, and α ∈ R. Assume that r0 > 0 and set g0 := g and h0 := h. Since the
transform from R = R0 to R1 is in the y = y0 -direction there exist elements g1 and
h1 in R1 such that g = y r0 g1 and h = y r0 h1 . Then α =
g1
h1 .
Notice that the ideal
(g1 , h1 )R1 is the transform in R1 of the ideal (g, h)R.
Let r1 := min{ ordR1 (g1 ), ordR1 (h1 ) }. If r1 = 0, then α ∈ R1 . If r1 > 0, then
since the transform from R1 to R2 is in the y1 -direction there exist elements g2 and
h2 in R2 such that g1 = y1r1 g2 and h1 = y1r1 h2 . Then α =
g1
h1
=
g2
h2 .
The ideal
(g2 , h2 )R2 is the transform in R2 of the ideal (g1 , h1 )R1 and also the transform of
the ideal (g, h)R.
Suppose the transform of (g, h)R in Rn is a proper ideal in Rn for every integer
n ≥ 0. Then rn := min{ ordRn (gn ), ordRn (hn ) } is positive for all n ≥ 0. This gives
infinite sequences {gn+1 }n≥0 , {hn+1 }n≥0 such that gn+1 , hn+1 ∈ Rn+1 \ xn+1 Rn+1 ,
where gn = ynrn gn+1 and hn = ynrn hn+1 . Then for every integer n ≥ 0 we have
ν(gg
yn ) = ν(gen ) − rn ν(nn ) ≤ ν(gen ) − ν(nn ),
n+1 ) = ν(gen ) − rn ν(f
f
fn ) − rn ν(nn ) ≤ ν(h
fn ) − ν(nn ).
ν(h]
yn ) = ν(h
n+1 ) = ν(hn ) − rn ν(f
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
28
Hence for every integer n ≥ 1 we have
ν(gen ) ≤ ν(ge0 ) −
By taking limits, we have
lim ν(gen ) ≤ ν(ge0 ) −
n→∞
Since
P∞
i=0 ν(ni )
n−1
X
i=0
fn ) ≤ ν(f
ν(ni ) and ν(h
h0 ) −
∞
X
ν(ni )
and
i=0
n−1
X
ν(ni ).
i=0
fn ) ≤ ν(f
lim ν(h
h0 ) −
n→∞
∞
X
ν(ni )
i=0
= ∞, this is a contradiction. Hence rn = min{ ordRn (gn ), ordRn (hn ) } =
0 for some integer n ≥ 0. Then either gn is a unit in Rn or hn is a unit in Rn . If hn
is a unit in Rn , then α ∈ Rn , and hence α ∈ S. If gn is a unit in Rn , then ν(gen ) = 0.
fn ) ≥ 0, we must also have ν(h
fn ) = 0, and α ∈ R.
Since ν(e
α) = ν(gen ) − ν(h
(Case ii): Assume that t = νx (α) > 0. Let β0 := hg . If ν(βe0 ) ≥ 0, then by Case i,
β0 ∈ S and hence xt β0 = α ∈ S. If ν(βe0 ) < 0, let β1 := y0t β0 . Then
α = xt β0 = xt1 y t β0 = xt1 β1 .
If ν(βe1 ) ≥ 0, then by (Case i), β1 ∈ S and hence xt1 β1 = α ∈ S. If ν(βe1 ) < 0, let
β2 := y1t β1 . Then
α = xt1 β1 = xt2 y1t β1 = xt2 β2
β2 = y1t y0t β0
and
For each positive integer n, we define βn+1 = ynt βn . It follows that
α = xtn βn = xtn+1 ynt βn = xtn+1 βn+1
and
βn+1 = (yn yn−1 · · · y0 )t β0 .
Thus we have
ν(β]
n+1 ) = t
Since
P∞
i=0 ν(ni )
n
X
i=0
n
X
ν(yei ) + ν(βf
)
=
t
ν(n
)
+ ν(βe0 ).
0
i
i=0
fn ) ≥ 0 for all sufficiently large integers n. Then
= ∞, we have ν(β
by Case i, we have βn ∈ S and hence xtn βn = α ∈ S. This completes the proof of
Claim 7.7.
Hence by Claim 7.7 we have S = W , and we conclude that S is a valuation
domain.
(3) ⇔ (4): This equivalence follows from Shannon([S, Proposition 4.18]) and
Granja([Gr, Theorem 13]).
Remark 7.8. Let (R0 , m0 ) be a d-dimensional regular local ring with d ≥ 2,
and let {(Rn , mn )}n≥0 be an infinite sequence of local quadratic transforms with
S
dim Rn = dim R0 for all n. Assume that S := n≥0 Rn is a valuation domain of
rank greater than one. If R0 is excellent and equicharacteristic zero, Granja proves
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
29
in [Gr, Theorem 17] that there exists for some nonnegative integer j a regular prime
x in Rj such that S ⊂ (Rj )xRj .
Remark 7.9. Let notation be as in Setting 7.4. Assume that S is a valuation
domain. Let G be the value group of a valuation v associated with S. and let H be
the value group of a valuation ν associated with T . Then we have the following :
(1) S has rational rank 2 and G ∼
= Z ⊕ H.
(2) T has rational rank 1.
(3) If T is DVR, then G ∼
= Z2 .
Proof. To see item 1, by Granja [Gr, Proposition 14], the valuation domain S has
rational rank 2. By [ZS2, Theorem 15, page 40] the set GP := { ±v(α) | α ∈ S \ P }
is an isolated subgroup of G, (that is, GP is a segment and a proper subgroup of
G). By [ZS2, Theorem 17, page 43] the group H and the group GP are isomorphic
as ordered groups. Since SP = RxR is a DVR, the value group of SP is Z. Hence
by [ZS2, Theorem 17, page 43] the groups G/GP and Z are order isomorphic. It
follows that G ∼
= Z ⊕ H. Items 2 and 3 follow from item 1.
= Z ⊕ GP ∼
Remark 7.10. Let notation be as in Setting 7.4.
S
(1) If d = 3, then T = n≥0 An is a valuation domain by [A, Lemma 12].
(2) Let T be as in Setting 7.4. If T is a rank one valution domain and s =
P∞
i≥0 ν(ni ) < ∞, then there exist infinitely many choices for the positive
integer N and hence infinitely many nonassociate regular prime elements
f = x − z N in R such that S ⊂ Rf R . Let Q = S ∩ f Rf R . Then S/Q is an
infinite directed union of d − 1-dimensional regular local domains. If d = 3,
then by item 1, each of the infinitely many S/Q is a valuation domain.
In Example 7.11 we construct an infinite directed sequence {(Rn , mn )}n≥0 of
local quadratic transforms of a 3-dimensional regular local ring R = R0 such that
S
S = n≥0 Rn is a rank 2 valuation domain with value group Z ⊕ H, where H is
rational rank 1 but not discrete.
Example 7.11. Let (R0 , m0 ) be a 3-dimensional regular local ring, and let m0 =
(x, y, z)R0 . We define a sequence {(Rn , mn )}n≥0 of local quadratic transforms of R0
as follows. The sequence from R0 to R3 is
y
z
y′
R := R0 ⊂ R1 ⊂ R2 ⊂ R3
30
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
defined as
m
R1 = R0 [ ]( xy ,y, yz ) and
m1 = (x1 , y1 , z1 )R1
y
x1 y 1
m1
m2 = (x2 , y2 , z2 )R2 = ( , , z1 )R2
R2 = R1 [ ]( x1 , y1 ,z1 ) and
z1 z1 z1
z1 z1
m2
x2
z2
R3 = R2 [ ]( x2 ,y2 , z2 −1) and
m3 = (x3 , y3 , z3 )R3 = ( , y2 ,
− 1)R3
y2 y2 y2
y2
y2
Starting from (R3 , m3 ) and m3 = (x3 , y3 , z3 )R3 , we define a sequence
y
y
z
y′
R3 ⊂ R4 ⊂ R5 ⊂ R6 ⊂ R7
of local quadratic transform with respect to the regular system of parameters x3 , y3 , z3
of m3 such that the 21 transforms from R3 to R5 are both monomial in the ydirection, the transform from R5 to R6 is monomial in the z-direction, and the
transform from R6 to R7 is defined in a manner similar to that from R2 to R3 . Thus
we have
m7 = (x7 , y7 , z7 )R7 =
x
6
y6
, y6 ,
x y3 z2
z6
3
3
3
− 1 R7 =
,
,
−
1
R7 .
y6
y33 z3 y35
Let t0 := 0 and t1 := 3. For each integer n ≥ 2, let tn = 2tn−1 − tn−2 + 2n−2 .
We inductively define a sequence of local quadratic transforms with respect to the
regular system of parameters xtn , ytn , ztn of mtn as follows:
2n times
}|
{
z
y
y
y
y′
z
Rtn ⊂ Rtn +1 ⊂ · · · ⊂ Rtn +2n ⊂ Rtn +2n +1 ⊂ Rtn +2n +2 ,
The 2n transforms from Rtn to Rtn +2n are all monomial in the y-direction, the
transform from Rtn +2n to Rtn +2n +1 is monomial in the z-direction, and the transform
from Rtn +2n +1 to Rtn +2n +2 is defined in a manner similar to that from R6 to R7 .
Thus we have
2n times
z }| {
x
ztn
y
y
tn
mtn = (xtn , ytn , ztn )Rtn → · · · → mtn +2n =
,
y
,
n
n Rtn +2n
t
n
yt2n
yt2n
n
x
yt2n +1 ztn
z
tn
n
→ mtn +2 +1 =
,
, 2n Rtn +2n +1
ztn
ztn
ytn
n
x
yt2n +1
zt2n
y′
tn
,
→ mtn +2n +2 =
−
1
Rtn +2n +2
,
n
2(2n )+1
ztn
yt2n +1
ytn
S
S
Let S := n≥0 Rn and mS := n≥0 mn . Since we never divide in the x-direction, we
have S ⊂ RxR . Let P := xRxR ∩ S and let T := S/P and mT := mS /P . For each
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
n ≥ 0, let xn be the transform of x in Rn . Then P =
let
An :=
Rn
x n Rn
and
nn :=
S
n≥0 xn Rn .
31
For each n ≥ 0,
mn
.
x n Rn
Each (An , nn ) is a 2-dimensional regular local ring with maximal ideal nn generated
by the images of yn and zn in An . By identifying yn and zn with their images in An ,
we have nn = (yn , zn )An . Moreover {(An , nn )}n≥0 is an infinite sequence of local
quadratic transforms of (A0 , n0 ), where n0 := (y, z)A0 . By [A, Lemma 12], the ring
S
S
T = n≥0 An is a valuation domain with maximal ideal mT = n≥0 nn . Let ν be a
valuation associated with the valuation domain T . The sequence {(An , nn )}n≥0 is
determined by the sequence {(Rn , mn )}n≥0 . With the integers tn as defined in the
construction of the sequence {(Rn , mn )}n≥0 , we have
2n times
z
}|
{
y
y
y
y′
z
Atn ⊂ Atn +1 ⊂ · · · ⊂ Atn +2n ⊂ Atn +2n +1 ⊂ Atn +2n +2 ,
The 2n transforms from Atn to Atn +2n are all monomial in the y-direction, the
transform from Atn +2n to Atn +2n +1 is monomial in the z-direction, and we have
2n times
ntn
Since
z }| {
zt
y
y
= (ytn , ztn )Atn → · · · → ntn +2n = ytn , 2nn Atn +2n
ytn
n
y 2 +1 z
z
t
→ ntn +2n +1 = tn , 2nn Atn +2n +1
ztn
ytn
n +1
2
y
z2
y′
→ ntn +2n +2 = tn , 2(2tnn)+1 − 1 Atn +2n +2
ztn
ytn
ztn +2n +1
ytn +2n +1
=
zt2n
2(2n )+1
yt n
∈ Atn +2n +2 \ ntn +2n +2 , we have ν(ztn ) =
2(2n )+1
ν(ytn ).
2
Assume that ν(y) = 1. Then we have:
s :=
∞
X
i=0
ν(ni ) = 1 +
1 1
1 1
1 1
1 1 1 1
+ +{ + }+ + +{ + + + }
2 2
2 2
4 4
4 4 4 4
z
23 times
}|
{
z
24 times
}|
{
1
1
1
1
1
1 1
1
1 1
+{ +
+ · · · + } + · · · = ∞.
+ + + { + + ··· + }+ +
8 8
8 8
8
16 16
16 16
16
By Theorem 7.5, the ring S is the valuation domain, the sequence {(Rn , mn )}n≥0 is
height one directed, and S has rank 2. Let G be the value group of S and let H
be the value group of T . By Remark 7.9, S has rational rank 2 and G ∼
= Z ⊕H.
Clearly, T has rational rank 1, and T is not a DVR.
32
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
In Example 7.12 we construct an example similar to that of Example 7.11, but
with dim R = 4 and dim A = 3.
Example 7.12. Let (R0 , m0 ) be a 4-dimensional regular local ring, and let m0 =
(x, y, z, w)R0 . We define a sequence {(Rn , mn )}n≥0 of local quadratic transforms of
R0 as follows: The sequence from R0 to R5 is
y
y
z
y′
w
R := R0 ⊂ R1 ⊂ R2 ⊂ R3 ⊂ R4 ⊂ R5
defined by
m
m1 = (x1 , y1 , z1 , w1 )R1
R1 = R0 [ ]( xy ,y, yz , wy ) and
y
m1
R2 = R1 [ ]( x1 ,y1 , z1 , w1 ) and
m2 = (x2 , y2 , z2 , w2 )R2
y1 y1 y1 y1
m2
R3 = R2 [ ]( x2 , y2 ,z2 , w2 ) and
m3 = (x3 , y3 , z3 , w3 )R3
z2
z2 z2 z2
m3
R4 = R3 [ ]( x3 , y3 , z3 ,w3 ) and
m4 = (x4 , y4 , z4 , w4 )R4
w3 w3 w3 w3
m4
m5 = (x5 , y5 , z5 , w5 )R5
R5 = R4 [ ]( x4 ,y4 , z4 −1, w4 −1) and
y4
y4 y4 y4
x
x y3 z2
z4
w4
w2
4
=
, y4 ,
− 1,
− 1 R5 =
,
,
−
1,
−
1
R5
y4
y4
y4
y3 w y5
y3z
Starting from (R5 , m5 ) and m5 = (x5 , y5 , z5 , w5 )R5 , we define a sequence
y
y
y
y
z
w
y′
R5 ⊂ R6 ⊂ R7 ⊂ R8 ⊂ R9 ⊂ R10 ⊂ R11 ⊂ R12
of local quadratic transform with respect to the regular system of parameters x5 , y5 , z5 , w5
of m5 such that the 22 transforms from R5 to R9 are all monomial in the y-direction,
the transform from R9 to R10 is monomial in the z-direction, the transform from
R10 to R11 is monomial in the w-direction, and the transform from R11 to R12 is
defined in a manner similar to that from R4 to R5 Thus we have
x y5 z2
x
z11
w11
w52
5
11
5
5
, y11 ,
−1,
−1 R12 =
,
,
−1,
−1
R12 .
m12 = (x12 , y12 , z12 , w12 )R12 =
y11
y11
y11
y55 w5 y59
y55 z5
Let t0 := 0, t1 := 5, and t2 := 12 = 2t1 − t0 + 2. For each n ≥ 3, we let tn :=
2tn−1 − tn−2 + 3 · 22(n−2) . We inductively define a sequence of local quadratic
transforms with respect to the fixed regular system of parameters xtn , ytn , ztn , wtn
of mtn as follows:
22n times
}|
{
z
y
y
y
y′
z
w
Rtn ⊂ Rtn +1 ⊂ · · · ⊂ Rtn +22n ⊂ Rtn +22n +1 ⊂ Rtn +22n +2 ⊂ Rtn +22n +3
The 22n transforms rom Rtn to Rtn +22n are all monomial in the y-direction, the
transform from Rtn +22n to Rtn +22n +1 is monomial in the z-direction, the transform
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
33
from Rtn +22n +1 to Rtn +22n +2 is monomial in the w-direction, and the transform from
Rtn +22n +2 to Rtn +22n +3 is defined in a manner similar to that from R4 to R5 . Thus
with mtn = (xtn , ytn , ztn , wtn )Rtn , we have
22n times
z }| {
x
ztn wtn
y
y
tn
,
y
,
mtn → · · · → mtn +22n =
t
2n
2n ,
2n Rtn +22n
n
yt2n
yt2n yt2n
2n
x
yt2n +1 ztn wtn
z
tn
→ mtn +22n +1 =
,
, 22n ,
Rtn +22n +1
ztn
ztn
ztn
ytn
x
y 2 +1
z2
wtn
→ mtn +22n +2 =
, tn
, 22ntn ,
Rtn +22n +2
wtn
wtn
ytn wtn ztn
2n
x
yt2n +1
wt2n
zt2n
y′
tn
−
1,
−
1
Rtn +22n +3
,
→ mtn +22n +3 =
,
2n
2n
2(22n )+1
wtn
yt2n +1
yt2n +1 ztn
ytn
S
S
Let S := n≥0 Rn and mS := n≥0 mn . Since we never divide in the x-direction, we
w
2n
tn
have S ⊂ RxR . Let P := xRxR ∩ S and let T := S/P and mT := mS /P . For each
S
n ≥ 0, let xn be the transform of x in Rn . Then P = n≥0 xn Rn . For each n ≥ 0,
let An :=
Rn
xn Rn
mn
xn Rn .
and nn :=
Each (An , nn ) is a 3-dimensional regular local ring
with maximal ideal nn generated by images of yn , zn and wn in An . By identifying
yn , zn and wn with their images in An , we have nn = (yn , zn , wn )An . Moreover
{(An , nn )}n≥0 is an infinite sequence of local quadratic transforms of (A0 , n0 ), where
S
n0 := (y, z, w)A0 . Then T = n≥0 An is a normal local domain with maximal
S
ideal mT = n≥0 nn . The sequence {(An , nn )}n≥0 is determined by the sequence
{(Rn , mn )}n≥0 . Thus the local quadratic transforms from A0 to A5 are:
y
y
z
w
y′
A0 ⊂ A1 ⊂ A2 ⊂ A3 ⊂ A4 ⊂ A5
where
z
y3 z2
w4
w2
4
n5 = (y5 , z5 , w5 )A5 = y4 ,
− 1,
− 1 A5 =
, 5 − 1, 3 − 1 A5
y4
y4
w y
y z
Let W be a valuation domain birationally dominating T and let ω be a valuation
associated with W . Since
z4
y4
=
z2
y5
and
w4
y4
=
w2
y3 z
are in A5 \n5 , we have ω(z) = 25 ω(y)
and ω(w) = 32 ω(y) + 12 ω(z). The transforms from A5 to A12 are:
y
y
y
y
z
w
y′
A5 ⊂ A6 ⊂ A7 ⊂ A8 ⊂ A9 ⊂ A10 ⊂ A11 ⊂ A12
where
y5 z2
w2
z11
w11
5
n12 = (y12 , z12 , w12 )A12 = y11 ,
− 1,
− 1 A12 =
, 59 − 1, 5 5 − 1 A12 .
y11
y11
w5 y5
y5 z5
WILLIAM HEINZER, MEE-KYOUNG KIM1 , AND MATTHEW TOENISKOETTER
34
Since
z12
y12
ω(w5 ) =
z52
y59
=
and
w11
y11
1
5
2 ω(y5 ) + 2 ω(z5 ).
=
y52
y55 z5
are in A12 \ n12 , we have ω(z5 ) =
9
2 ω(y5 )
and
With the integers tn as defined in the construction of the
sequence {(Rn , mn )}n≥0 , we have
22n times
z
}|
{
y
y
y
y′
z
w
Atn ⊂ Atn +1 ⊂ · · · ⊂ Atn +22n ⊂ Atn +22n +1 ⊂ Atn +22n +2 ⊂ Atn +22n +3 .
The 22n transforms from Atn to Atn +22n are all monomial in the y-direction, the
transform from Atn +22n to Atn +22n +1 is monomial in the z-direction, the transform from Atn +22n +1 to Atn +22n +2 is monomial in the w-direction, and with ntn =
(ytn , ztn , wtn )Rtn , we have
22n times
z }| {
ztn wtn
y
y
ntn → · · · → ntn +22n = ytn , 22n
, 22n Atn +22n
ytn
ytn
2n
y 2 +1 z
wtn
z
tn
, 22n
,
Atn +22n +1
→ ntn +22n +1 = tn
ztn
ztn
ytn
y 22n +1
z2
wtn
w
→ ntn +22n +2 = tn
, 22ntn ,
Atn +22n +2
wtn
ytn wtn ztn
y 22n +1
wt2n
zt2n
y′
→ ntn +22n +3 = tn
−
1,
−
1
Atn +22n +3
, 2(22n
2n
)+1
wtn
yt2n +1 ztn
ytn
Since
zt2n
2(22n )+1
yt n
and ω(wtn ) =
and
wt2n
22n +1
zt n
yt n
22n +1
2 ω(ytn )
in Atn +22n +3 \ntn +22n +3 , we have ω(ztn ) =
2(22n )+1
ω(ytn )
2
+ 12 ω(ztn ). Assume that ω(y) = 1. Then we have
22·1 times
}|
{
z
∞
n1
X
1 n1
1o n1
1
1
1o 1
1o
ω(ni ) = 1 + 1 + + 2 + 2 + 2 + 2 + 2 + 2 + 3 + 4 + 4
s:=
2
2
2
2
2
2
2
2
2
2
i=0
22·2 times
22·3 times
z
}|
{
}|
{
z
n1
n
1
1o 1
1o n1
1
1o 1
1
+ 4 + 4 + · · · + 4 + 5 + 6 + 6 + 6 + 6 + · · · + 6 + 7 + · · · = ∞.
2
2
2
2
2
2
2
2
2
2
By Proposition 7.3, the sequence {(An , nn )}n≥0 switches strongly infinitely often.
S
Hence by Remark 3.3, the ring T = n≥0 An is a valuation domain and T has rank
1. Notice that T = W . By Theorem 7.5, the ring S is a valuation domain, the
sequence {(Rn , mn )}n≥0 is height one directed, and S has rank 2. Let G be the
value group of S and H be the value group of T . By Remark 7.9, the ring S has
rational rank 2 and G ∼
= Z ⊕H. Clearly, T has rational rank 1, and T is not a DVR.
DIRECTED UNIONS OF LOCAL QUADRATIC TRANSFORMS
35
References
[A]
S.S. Abhyankar, On the valuations centered in a local domain, Amer. J. Math. 78 (1956)
321-348.
[GR]
A. Granja and C. Rodrı́guez, Proximity relations for real rank one valuations dominating
a local regular ring, Rev. Mat. Iberoamericana 19 (2003), no. 2, 393-412.
[Gr]
A. Granja, Valuations determined by quadratic transforms for a regular ring, J. Algebra,
280 (2004), 699-718.
[GMR] A. Granja, M.C. Martı́nez, and C. Rodrı́guez, Valuation dominating regular local rings
and proximity relations, J. of Pure and Algebra, 209 (2007), 371-382.
[HKT] W. Heinzer, M.-K. Kim, and M. Toeniskoetter, Finitely supported *-simple complete ideals
in a regular local ring , J. Algebra, 401 (2014), 76-106
[HRW] W. Heinzer, C. Rotthaus and S. Wiegand, Examples Using Power Series Over Noetherian
Integral Domains, to appear.
[HubS] R. Hübl and I. Swanson, Adjoints of ideals, Michigan Math. J., 57 (2008), 447-462.
[KS]
K. Kiyek and J. Stückrad, Integral closure of monomial ideals on regular sequences , Rev.
Mat. Iberoamericana, 19 (2003), 483-508.
[L]
J. Lipman, On complete ideals in regular local rings, Algebraic Geometry and Commutative
Algebra in Honor of Masayoshi Nagata, (1986), 203-231.
[M]
H. Matsumura, Commutative Ring Theory, Cambridge Univ. Press, Cambridge, 1986.
[N]
M. Nagata, Local Rings, Interscience, New York, 1962.
[S]
D. Shannon, Monoidal transforms of regular local rings, Amer. J. Math. 95 (1973), 294-320.
[SH]
I. Swanson and C. Huneke, Integral Closure of Ideals, Rings, and Modules, London Math.
Soc. Lecture Note Series 336, Cambridge Univ. Press, Cambridge, 2006.
[ZS2]
O. Zariski and P. Samuel,Commutative Algebra, Vol. 2, D. Van Nostrand, New York, 1960.
Department of Mathematics, Purdue University, West Lafayette, Indiana 47907
U.S.A.
E-mail address: [email protected]
Department of Mathematics, Sungkyunkwan University, Jangangu Suwon 440-746,
Korea
E-mail address: [email protected]
Department of Mathematics, Purdue University, West Lafayette, Indiana 47907
U.S.A.
E-mail address: [email protected]
| 0math.AC
|
1
The Importance of System-Level Information in
Multiagent Systems Design: Cardinality and
Covering Problems
arXiv:1710.07460v1 [cs.GT] 20 Oct 2017
Dario Paccagnan1 and Jason R. Marden2
Abstract
A fundamental challenge in multiagent systems is to design local control algorithms to ensure a
desirable collective behaviour. The information available to the agents, gathered either through communication or sensing, naturally restricts the achievable performance. Hence, it is fundamental to identify
what piece of information is valuable and can be exploited to design control laws with enhanced
performance guarantees. This paper studies the case when such information is uncertain or inaccessible
for a class of submodular resource allocation problems termed covering problems. In the first part of this
work we pinpoint a fundamental risk-reward tradeoff faced by the system operator when conditioning the
control design on a valuable but uncertain piece of information, which we refer to as the cardinality, that
represents the maximum number of agents that can simultaneously select any given resource. Building
on this analysis, we propose a distributed algorithm that allows agents to learn the cardinality while
adjusting their behaviour over time. This algorithm is proved to perform on par or better to the optimal
design obtained when the exact cardinality is known a priori.
S
I. I NTRODUCTION
EVERAL social and engineering systems can be thought of as a collection of multiple
subsystems or agents, each taking local decisions in response to available information. A
central goal in this field is to design control algorithms for the individual subsystems to ensure
that the collective behaviour is desirable with respect to a global objective. Achieving this goal is
particularly challenging because of the restriction on the information available to each agent and
to the large scale of typical systems. Examples include, but are not limited to, power grid networks
[2], charging of electric vehicles [3], transportation network [4], task assignment problems [5],
sensor allocation [6], robotic networks [7]. A considerable bulk of the research has focused on
This research was supported by SNSF Grant #P1EZP2-172122 and by AFOSR Grant #FA9550-12-1-0359, ONR Grant
#N00014-15-1-2762, NSF Grant #ECCS-1351866.
1
D. Paccagnan is with the Department of Electrical and Computer Engineering, University of California, Santa Barbara,
USA and with the Automatic Control Laboratory, Swiss Federal Institute of Technology, Zurich, Switzerland. Email:
[email protected].
2
J. R. Marden is with the Department of Electrical and Computer Engineering, University of California, Santa Barbara, USA.
Email: [email protected].
A much abbreviated conference version of this work appeared in [1].
October 23, 2017
DRAFT
2
the design of local control algorithms in a framework where the information at agents’ disposal
is itself a fixed datum of the problem. A non exhaustive list includes [8], [9] and references
therein. Understanding the impact of information availability on the achievable performances is
a seemingly important but less tracked problem [10], [11], [12].
Of particular interest is to recognise what supplementary piece of information could coordinate
the agents to improve the system performance, and further how to incorporate this additional
knowledge into a control algorithm. It is important to highlight that providing each agent with
all the information available to the system is in principle beneficial, but not necessarily desirable.
Indeed, the communication costs associated with propagating additional information through the
system might overcome the performance gains that the knowledge of additional information
gives. Therefore, the previous question has to be understood within this context. Ideally, one is
interested in a piece of information that gives a significant performance enhancement, and is
simple to obtain. Loosely speaking, we measure the value of an additional piece of information
with the performance gain that the best controller can offer, using that supplementary piece of
information.
Relative to the class of resource allocation problems termed covering problems, [11], [13] show
that the maximum number of agents that can simultaneously select a resource (which we term
cardinality, see Equation (1) for a formal definition) constitutes a valuable piece of information.
More precisely, when the system operator is aware of the cardinality of the problem, he can devise
distributed algorithms with improved performance guarantees. Nevertheless, the knowledge of
the exact cardinality is in many applications uncertain, not available or may require excessive
communication to be determined. Following this observation, a system operator would like to
understand how to operate when the knowledge of the exact cardinality is not available. What
is the risk associated with using the wrong cardinality in the control design? What is the reward
for using the correct one? Further and more fundamental: when the cardinality is not available
at all, can the agents learn it while simultaneously adjusting their behaviour?
The paper proceeds by considering covering problems [14], [15], a class of resource allocation
problems where agents are assigned to resources in order to maximise the total value of covered
items. Examples include vehicle-target assignment problems [16], sensor allocation [6], task
assignment [17], among others. Due to the inherent limitations in sensing and communication, in
all these applications the control algorithms are required to rely only on local information. Thus,
we model distributed covering problems as strategic-form games, where the system operator
has the ability to assign local objective functions to each agent. Indeed, as shown in [18], [10],
Game Theory lends itself to analyse distributed systems where individual agents adjust their
behaviour in response to partial information. Such game theoretic approach offers the possibility
to build upon existing tools to quantify the system performance, as well as the opportunity to
exploit readily available algorithms to compute equilibria in a distributed fashion [5], [13]. The
October 23, 2017
DRAFT
3
overarching goal of the system operator is to design local utilities in order to render the equilibria
of the game as efficient as possible. Agents can then be guided towards an equilibrium of such
game by means of existing distributed algorithms [5], [19]. It is important to highlight that we
are not modelling agents as competing units, but we are rather designing their utilities to achieve
the global objective.
Building on the previous results of [11], [13], we contribute as follows.
i) We study the problem of optimally designing the utility functions in the case when the
true cardinality is not known, but only an upper bound is available.1 We further perform
a risk-reward analysis in the case when the information on the cardinality of the game is
uncertain. When the goal is to guard the system against the worst case performances, the
right choice is to design the utilities as if the true cardinality was the given upper bound.
Different designs will offer potential benefits, but come with a certain degree of risk. These
results are presented in Theorem 1.
ii) Motivated by the potential advantages and inherent shortcomings presented in the riskreward analysis, we propose a distributed and asynchronous algorithm that dynamically
updates the utility functions while agents adjust their behaviour over time. Such algorithm
requires no initial information, and is certified to perform on par or better (in a worst case
sense) to the optimal design possible, had we known the cardinality in the first place. These
results are summarised in Theorem 2.
iii) We compare, instance by instance, the performance of the proposed learning algorithm with
the performance of the optimal design obtained with full knowledge of the cardinality. We
show that it is not possible to deem one approach superior to the other on all instances
of covering problems, in that there are instances where one outperforms the other, and the
reverse too. These results are presented in Theorem 3.
The remaining of the paper is organised as follows. The next section introduces the covering
problem, its formulation as a strategic game and the metric used to measure the system-level
performance. Section III studies the utility design problem when a sole upper bound on the
cardinality is available and presents the risk-reward tradeoff associated with the use of uncertain
information. Section IV shows the possibility of dynamically adjusting the utility functions to
improve the performance. Numerical simulations and conclusions follow.
Notation
For any two positive integers p ≤ q, denote [p] = {1, ..., p} and [p, q] = {p, ..., q}; given
(a , . . . , an ), denote a−i = (a1 , . . . , ai−1 , ai+1 , . . . , an ). We use N and R≥0 to denote the set of
1
natural numbers (excluding zero) and the set of non-negative real numbers, respectively.
1
A simple bound is given by the number of agents.
October 23, 2017
DRAFT
4
II. D ISTRIBUTED COVERING VIA GAME THEORY
In this section we present the covering problem and the associated covering game. We further
define the performance metric used throughout the paper and recap previous results.
A. Model
Let us consider the problem of assigning a collection of agents N = {1, . . . , n} to a finite set
of resources R = {r1 , . . . , rm } with the goal of maximising the value of covered resources. The
feasible allocations for each agent i ∈ N are the elements of the action set ai ∈ Ai ⊆ 2R , while
every resource r ∈ R is associated with a non-negative value vr ≥ 0. Observe that ai ⊆ R. The
welfare of an allocation a = (a1 , . . . , an ) ∈ A1 × · · · × An is measured by the total value of
covered resources
X
W (a) :=
vr ,
r : |a|r ≥1
where |a|r denotes the number of agents that choose resource r in allocation a. The covering
problem C = {N, R, {Ai }i∈N , {vr }r∈R } consists in finding an optimal allocation, that is an
assignment ao ∈ arg maxa∈A W (a).2 Given a covering problem C, we define its cardinality as
the maximum number of players that can concurrently select the same resource, that is
max |a|r .
r∈R, a∈A
(1)
Instead of directly specifying a distributed algorithm, we shift the focus to the design of local
utility functions for each agent, as proposed first for distributed welfare games by [22], [5]
and successively by [11]. Within this framework, each agent i ∈ N is associated with a utility
function of the form
ui (ai , a−i ) :=
X
vr ·f (|a|r ) .
(2)
r∈Ai
The function f : [n] → R≥0 constitutes our design choice and is called distribution rule as it
represents the fractional benefit an agent receives from each resource he selects. The advantages
of using utilities of the form (2) are twofold. First, ui (ai , a−i ) is local as it depends only on the
resources agent i selects, their value and the number of agents that selects the same resources.
Second, (2) allows to construct a distribution rule irrespective of {Ai }i∈N and {vr }r∈R so that
the final design is scalable and applies to different choices of the action sets and of the resource
valuations .
2
Approximation algorithms for finding a near optimal solution to submodular optimization problems have been extensively
studied in the literature [20], [21]. The focus of this literature is predominantly on centralized algorithms for finding near optimal
allocations. In contrast, our focus is on distributed solutions where each decision-making entity has incomplete information about
the system as a whole.
October 23, 2017
DRAFT
5
Given a covering problem C and a distribution rule f : [n] → R≥0 , we consider the associated
covering game G := {C, f } = {N, R, {Ai }i∈N , {vr }r∈R , f }, where Ai is the set of feasible
allocations and the utility of agent i ∈ N is as in equation (2).
We do not aim at designing f using information on the specific instance of covering problem
at hand, as such information is often not available to the system designer. Our goal is rather to
construct a distribution rule that behaves well for a large class of problems. Hence, we consider
the set of covering problems for which the cardinality is smaller or equal to k ∈ N, k ≤ n.
Given a distribution rule f : [k] → R≥0 , we define the set of associated games as
Gfk := {G = {C, f } : max |a|r ≤ k} .
r∈R, a∈A
Our objective is to design f : [k] → R≥0 so that the efficiency of all the equilibria of games in
Gfk is as high as possible. Note that for fixed f , any game G is potential [22]. Hence existence of
equilibria is guaranteed and distributed algorithms, such as the best response scheme, converge
to them [23]. Throughout the paper, we focus on pure Nash equilibria [24], which we will refer
to in the following just as equilibria.
Definition 1 (Pure Nash equilibrium). Given a game G, an allocation ae ∈ A is a pure Nash
i i −i
i
i
equilibrium iff ui (aie , a−i
e ) ≥ u (a , ae ) for all deviations a ∈ A and for all players i ∈ N . In
the following we use NE(G) to denote the set of Nash equilibria of G.
For a given distribution rule, we evaluate the efficiency of the Nash equilibria of games in
Gfk , adapting the concept of Price of Anarchy from [25] as
mina∈NE(G) W (a)
PoA(f, k) := inf
≤ 1.
maxa∈A W (a)
G∈Gfk
(3)
In essence, the quantity PoA(f, k) bounds the inefficiency of the worst equilibrium (and thus of
all equilibria) over games in Gfk , that is over games with distribution rule set to f and cardinality
bounded by k.3 The higher the price of anarchy, the better the performance guarantees we can
provide.
B. Related Work and Performance Guarantees
The problem of designing a distribution rule so as to maximise PoA(f, k) has been studied in
[11] and [13]. Both works impose a natural constraint on the admissible f , requiring f (1) = 1
and f : [k] → R≥0 to be non-increasing. The optimal distribution rule is explicitly derived in
3
The quantity W (a) appearing in Equation (3) does depend on which game instance G we are considering, since the resource
valuations do. Hence, a more formal notation would entail using W (a; G). In the interest of readability, we avoid the latter and
simply use W (a) when no ambiguity arise.
October 23, 2017
DRAFT
6
the former work, while the latter shows how PoA(f, k) is fully characterised by a single scalar
quantity χ(f, k) defined in (4), measuring how fast the distribution rule f decreases. We intend
to build upon these results, which are summarised in the following proposition. Given k and a
distribution rule f , we define
χ(f, k) := max {j · f (j) − f (j + 1), (k − 1) · f (k)} .
j≤k−1
(4)
Proposition 1 ([11], [13]). Consider a non-increasing distribution rule f : [k] → R≥0 , with
f (1) = 1.
i) The price of anarchy over the class Gfk is
PoA(f, k) =
1
.
χ
1 + (f, k)
ii) The price of anarchy over the class Gfk is maximised for
Pk−1 1
1
+ i=j
(k−1)(k−1)!
i!
?
fk (j) = (j − 1)!
Pk−1 1 ,
1
+ i=1 i!
(k−1)(k−1)!
j ∈ [k]
(5)
with corresponding
(6)
χ(fk? , k) = (k − 1)·fk? (k) .
iii) The optimal price of anarchy is a decreasing function of the cardinality k
PoA(fk? , k) = 1 −
1
1
(k−1)(k−1)!
+
Pk−1
1
i=1 i!
.
(7)
III. T HE CASE OF UNKNOWN CARDINALITY: A R ISK -R EWARD TRADEOFF
When the cardinality k defining the class of games Gfk is known, Proposition 1 gives a
conclusive answer on which distribution rule agents should choose to achieve the best worst case
performance. In spite of that, the knowledge of the exact cardinality is in many applications not
available or may require excessive communications between the agents to be determined.
Motivated by this observation, we study in the following the problem of designing a distribution
rule when the cardinality k defining the class of games Gfk is not known, but an upper bound k ≤
ku is available. Observe that a universal upper bound for such quantity can be easily computed as
the number n of agents. Potentially tighter bounds can be derived for specific applications. Our
objective is to design a distribution rule f : [ku ] → R≥0 with the best performance guarantees
possible with the sole knowledge of ku . Once such a distribution rule has been designed, one
can use existing distributed algorithms to find an equilibrium as discussed in the introduction.
Two natural questions arise in this context:
1) How should we select the distribution rule?
2) What performance can we guarantee?
October 23, 2017
DRAFT
7
We will show how selecting fk?u guards us against the worst case performance but will not
guarantee the same efficiency of fk? , when k < ku . We will then present the potential benefits
and risks associated with a more aggressive choice. These results motivate Section IV, where
we will present a dynamic scheme that overcomes the difficulties encountered here, offering the
same performances of fk? at no risk.
A. Two alternative distributions
A natural choice when an upper bound on the cardinality is available consists in designing
the distribution rule exactly at the upper bound, that is using fk?u . A different choice might
entail constructing a distribution rule where the entries [kd ] are designed as if the cardinality
was kd < ku , while the remaining entries [kd + 1, ku ] are optimally filled. The latter suggestion
is inspired by the observation that the optimal system level performance (measured by the price
of anarchy) is a decreasing function of k as per (7). This distribution is denoted with fk0 d and is
constructed fixing fk0 d (j) = fk?d (j) for j ∈ [kd ]. The tail entries corresponding to j ∈ [kd + 1, ku ]
are chosen to mitigate the risk taken. Formally, for any 1 < kd < ku we define the distribution
rule fk0 d : [ku ] → R≥0 as a solution of the following optimisation problem
fk0 d ∈ arg max PoA(f, ku )
f ∈F
s.t.
f (j) =
fk?d (j)
(8)
∀j ∈ [kd ] ,
where F = {f : [ku ] → R≥0 with f (1) = 1, f (j + 1) ≤ f (j), ∀j ∈ [ku − 1]} is the set of
admissible distributions.
Note that we do not define fk0 d for kd = 1 or kd = ku as it would reduce in both cases to
fk?u . Further observe that the constraint fk0 d (j) = fk?d (j), ∀j ∈ [kd ] is equivalent to requiring
fk0 d ∈ arg maxf ∈F PoA(f, kd ).
The next proposition characterises explicitly fk0 d .
Proposition 2. For any 1 < kd < ku , the distribution fk0 d defined in (8) is given by
fk?d (j)
j ∈ [kd ]
fk0 d (j) = (j−1)! f ? (kd ) − χ(f 0 , ku ) Pj−1−kd (j−1)! + 1
k
k
h=1
(k
−1)!
(j−h−1)!
d
d
d
j ∈ [kd + 1, ku ]
(9)
where the expression for χ(fk0 d , ku ) can be found in Equation (13) in the Appendix.
The proof can be found in the appendix.
Remark. In [11] a distribution rule f was required to satisfy the constraint j ·f (j) ≤ 1 for all
j. Loosely speaking the above requirement guarantees that a distribution rule does not overpay
October 23, 2017
DRAFT
8
1
fk?u
fk0 d
0.8
0.6
f (j)
0.4
0.2
0
1
2
3
4
5
6
7
8
9
10
j
Fig. 1: Example of distribution rules fk?u and fk0 d as defined respectively in (5) and (9); ku = 10,
kd = 2.
P
the players, in the sense that i∈N ui (ai , a−i ) ≤ W (a) for all allocations. Observe that such
constraint might be important for economic applications, but it is irrelevant in the design of
engineering systems. While [11] shows that the distribution rule fk?u satisfies this property, the
next lemma proves that also fk0 d verifies this condition even if this was not requested a priori.
Lemma 1. For any 1 < kd < ku the distribution fk0 d : [ku ] → R≥0 satisfies j ·fk0 d (j) ≤ 1 for all
j ∈ [ku ].
The proof is provided in the appendix.
B. Performance comparison
Based on the metric introduced in (3), we compare in this section the performance of fk?u with
the performance of fk0 d . Theorem 1 constitutes the main result of this section.
Theorem 1. Consider the set of games Gfk , where k ≤ ku .
i) The distribution fk?u has performance
PoA(fk?u , k) = PoA(fk?u , ku ) .
Such performance is strictly worse than the one achieved by the distribution fk? if k < ku
and equal if k = ku .
ii) For 1 < kd < k the distribution fk0 d has performance
PoA(fk0 d , k) = PoA(fk0 d , ku ) ,
October 23, 2017
DRAFT
9
which is strictly worse than the one achieved by fk?u .
For k ≤ kd < ku the distribution fk0 d has performance
PoA(fk0 d , k) = PoA(fk?d , kd ) ,
which is strictly better than the one achieved by fk?u .
The proof can be found in the Appendix.
Remark. Claim i) in Theorem 1 shows that the performance of the distribution fk?u on the class
of games with cardinality bounded by k is independent on the actual value of k, as for any
k ≤ ku , such performance is governed by PoA(fk?u , ku ). Claim ii) in Theorem 1 ensures that
the distribution fk0 d outperforms fk?u for ku > kd ≥ k and the opposite when kd < k. In each
of these cases the performance is independent on the actual value of k, but only depends on
whether kd is above or below k. Loosely speaking, if we underestimate k by designing kd < k,
the performance guarantees offered by fk0 d are worse than what fk?u can achieve. The reverse
holds in the case when we overestimate the cardinality as in ku > kd ≥ k.
In Figure 2 we compare the performance of fk0 d with the performance of fk?u . It is important
to note that the performance degradation (incurred whenever kd < k) always dominates the
potential gains (achieved when kd ≥ k). This motivates the next section where we will introduce
a dynamic algorithm capable of offering the benefits of fk0 d for kd ≥ k at no risk.
IV. B EYOND THE R ISK -R EWARD TRADEOFF
The previous section has focused on the design of a distribution rule when an upper bound
on the true cardinality is known. We have demonstrated how fk?u guards against worst case
performance while fk0 d could give potential benefits, but comes with a certain degree of risk. In
both cases the performance is equal or inferior to what we could achieve if we knew the true
cardinality.
In this section we show how to overcome such difficulties, when we are given a game G ∈ Gfk
with unknown cardinality k. We propose a distributed and asynchronous implementation of the
best response algorithm that dynamically updates which distribution rule to use (see footnote 5
for further details). The upshot is that we guarantee an equal or superior performance to what
we could achieve if we knew k.
In the following, we allow distribution rules to depend on an additional variable xr ∈ [n]
defined for r ∈ R, which we will dynamically update to coordinate the agents. In particular, we
generalise the utilities of (2) to
ui (ai , a−i ; x) :=
X
vr f (xr , |a|r ) ,
(10)
r∈ai
October 23, 2017
DRAFT
10
Performance of fk0 d relative to fk?u , k = 3
0.25
0
−0.25
−0.5
−0.75
−1
2
3
4
5
6
7
8
9
kd
Performance of fk0 d relative to fk?u , k = 4
0.25
0
−0.25
−0.5
−0.75
−1
2
3
4
5
6
7
8
9
kd
Fig. 2: The bars represent the difference PoA(fk0 d , k) − PoA(fk?u , k), normalized by its largest
value. As such, it describes the normalized difference in performance between fk0 d and fk?u for
various values of 1 < kd < ku = 10. In the first figure k = 3, while k = 4 in the second figure.
where x = {xr }r∈R and f : [n] × [n] → R≥0 might be different across the resources, depending
on the value of xr .
One could question wether the improved performance we will obtain comes from the additional
degree of freedom introduced allowing resource specific distribution rules. Nevertheless [11]
shows that it is not the case, in that the best resource specific and non resource specific distribution
perform equally (in the worst case sense). The only rationale to introduce resource dependent
October 23, 2017
DRAFT
11
rules is the distributability of the algorithm. Indeed, similar results could have been achieved
dynamically updating a single distribution rule shared by all resources, but such algorithm would
have not been distributed.
A. Algorithm description and distributedness
In the following t ∈ N describes the time step of the algorithm and at ∈ A the corresponding
allocation. With slight abuse of notation, for every resource r ∈ R we introduce the quantity
xr (t) that associates r ∈ R to the maximum number of agents that chose such resource until
time t ∈ N. Further, we define f`alg : [n] → R≥0 for every ` ∈ N as a distribution rule matching
the optimal in equation (5) for j ∈ [`] and constant in between [`, n]4
f ? (j) j ∈ [`] ,
`
alg
:=
f` (j)
f ? (`) j ∈ [` + 1, n] .
`
(11)
Algorithm 1 Asynchronous cardinality learning
1:
Initialise a1 ; t ← 1;
2:
while not converged do
xr (t) ← |a1 |r
∀r ∈ R
. Best response
3:
i ← t mod n
4:
ait+1 ← arg maxai ∈Ai
5:
ait+1 ← (ait+1 , a−i
t )
P
r∈ai
vr fxalg
(|at |r )
r (t)
. Update kt and thus f on every resource
6:
xr (t + 1) ← max{xr (t), |at+1 |r }
7:
t←t+1
8:
∀r ∈ R
end while
Through the additional variable xr (t), the algorithm keeps track of the maximum number of
players that visited every resource until the current time t, and selects consequently a resource
specific distribution rule. In particular on every r ∈ R, the algorithm uses f`alg with ` set to the
maximum number of players that visited that resource until time t (lines 4 and 6). Following a
4
The rule f`alg is a valid distribution rule, being non increasing and such that f`alg (1) = 1. It will in general not satisfy
j · f`alg (j) ≤ 1, but this was neither requested, nor has relevance in the design of engineering systems.
October 23, 2017
DRAFT
12
round-robin rotation, players i is selected to best respond and update the allocation (lines 3 to
5). The procedure repeats until convergence.5
The algorithm is distributed in the sense that every agent needs to keep track of xr (t) only
for those resources he has access to i.e. for r ∈ Ai . Further, it is asynchronous as players need
not to update their allocation in a specified order, but can spontaneously revise their strategies.
It is important to highlight that the communnication requirements of Algorithm 1 are the same
of those needed by the best response algorithm applied for instance to distribution rules fk?u or
fk0 d . That is, Algorithm 1 better exploits the information that is already available.
For ease of exposition we have presented the case where the distribution rules depend on the
history xr (t), but the same across the players. It is simple to extend these results to the case
of agent specific distribution rules. Every player would use a resource specific distribution rules
that depend on the maximum number of players that visited every resource up until his last visit.
Similar convergence guarantees and performance certificates will follow.
B. Convergence and quality of equilibria
The following theorem is the main result of this section. Claim i) shows convergence of
Algorithm 1 to a Nash equilibrium. Claim ii) proves that the quality of such equilibrium is
higher or equal to what the optimal distribution fk? could achieve.
Theorem 2. Consider a covering game G with cardinality k and initial assignment a1 ∈ A.
i) Algorithm 1 converges in a finite number of steps to ae := limt→∞ at ∈ A. The allocation
ae is a Nash equilibrium of the game with resource specific distribution rules fixed to fxalg
∞
r
∞ :=
for r ∈ R, where xr
limt→∞ xr (t).
?
ii) Let kM := maxr∈R x∞
r . The quality of the equilibrium ae is higher than PoA(fkM , kM ) and
thus of PoA(fk? , k)
W (ae )
≥ PoA(fk?M , kM ) ≥ PoA(fk? , k) .
W (ao )
(12)
These statements hold for any initial condition, even if the allocation to which the Algorithm
1 converges may be different. The proof is detailed in the Appendix.
5
Note that the best response strategy is not guaranteed to be unique. To overcome this issue, in the following we assume the
existence of a tie-breaking rule selecting a single optimal allocation, should these be multiple. Nevertheless, we observe that
neither this, nor requiring players to best respond in a round-robin is fundamental. It is still possible to show that Algorithm 1
converges almost surely if the players best responding are uniformly randomly selected form [n] and a single optimal allocation
is uniformly randomly extracted from the set of best responses. This will produce a totally asynchronous algorithm.
October 23, 2017
DRAFT
13
C. Instance by instance analysis
The previous theorem shows that Algorithm 1 achieves a higher or equal worst case performance than the optimal distribution fk? . While worst case analysis has been and still is a fruitful
tool to measure and improve on algorithms’ performance, the computer science community has
recently showed interest in moving beyond it [26]. Inspired by this, the question arises as to
whether Algorithm 1 performs better than fk? , instance by instance. More formally, we would
like to understand if Algorithm 1 yields higher welfare than the optimally designed rule on all
the remaining instances (the non worst case ones). We show that neither this, nor the opposite
holds.
Theorem 3. Let C be an instance of covering problem defined in Section II. Further denote with
NEalg (C) the set equilibria obtained using Algorithm 1 on C, and G? = {C, fk? } the associated
game where the optimal distribution fk? has been selected.
i) There exists an instance C of the covering problem such that
min
a∈NEalg (C)
W (a) >
min
a∈NE(G? )
W (a) .
ii) There exists an instance C of the covering problem such that
min
a∈NEalg (C)
W (a) <
min
a∈NE(G? )
W (a) .
The proof is constructive and is presented in the Appendix. Note that both statements in
Theorem 3 compare the performances of a given covering problem C and associated game
G? . Observe that this metric is significantly different from (3), where we additionally take the
infimum over problems with cardinality bounded by k.
V. S IMULATIONS
In this section we provide simulations to compare the performance of different distribution
rules. More specifically, given a coverage problem of cardinality k, we compare the performance
of Algorithm 1 with the performance of the best response dynamics (BR) using the distribution
rule fk?u or the optimal distribution rule fk? .
For this numerical study, we randomly generated 3000 instances of the covering problem with
n = 10 agents, m = 20 resources. Each agent is equipped with an action set whose elements
are singletons, that is ai ∈ R for all i ∈ N . We believe this is not restrictive in assessing the
performance, as the worst cases characterized in [11] and [13] are of this form. The value of the
resources is randomly selected according to a uniform distribution with minimum and maximum
value of 1 and 100, respectively. For every instance generated, we randomly selected 50 different
October 23, 2017
DRAFT
14
initial allocations used as starting points for the algorithms.
The following table presents the results.
TABLE I: Performance Comparison
Algorithm
Average Poa
Average min Poa
BR with fk?u
0.9701
0.9517
fk?
0.9723
0.9533
Algorithm 1
0.9734
0.9538
BR with
W (ae )
In the first column we show the mean price of anarchy, i.e. the quantity W
averaged over
(ao )
all the instances and all the initial conditions. The second column presents the average-minimum
price of anarchy, that is for every covering problem we compute the minimum of
W (ae )
W (ao )
over all
the initial conditions, and we then take the average over the instances.
We note that Algorithm 1 as well as the distribution rules fk?u , fk? perform similarly, when
W (ae )
looking at a randomly generated instance. The efficiency values W
obtained are much higher
(ao )
compared to the analytical worst case, hinting at the fact that such instances are very few. Given
that the average performance is similar, but the distribution fk?u is proven to have inferior worst
case performance (Theorem 1), one might want to use either the optimal distribution fk? or
Algorithm 1. Recall indeed that the worst case performance of Algorithm 1 is on par or better
to fk? (Theorem 2). Nevertheless, the use of fk? require knowledge of the cardinality k, while
the algorithm proposed does not.
To conclude: Algorithm 1 achieves similar average performances compared to the other
distributions tested, but has a better worst case performance than fk?u and a better-equal worst
case performance than fk? even if it does not require the knowledge of k.
VI. C ONCLUSION
In this work we studied how additional information impacts the optimal design of local utility
functions, when the goal is to improve the overall efficiency of a class of multiagent systems.
Relative to covering problems, in the first part of the manuscript we highlighted an inherent
tradeoff between potential risks and rewards when such additional information is uncertain. In
the second part, we showed how it is possible to fully eliminate the risks by using a distributed
algorithm that dynamically updated the local utility functions. The methodology used suggests
that similar results could be obtained for a broader class of resource allocation problems than
the one studied here.
October 23, 2017
DRAFT
15
A PPENDIX A
P ROOF OF P ROPOSITION 2
Before proving the proposition, we report the expression for χ(fk0 d , ku ), appearing in Equation
(9)
χ(fk0 , ku ) =
d
fk?d (kd )
(ku − 1)(ku − 1)!
.
Pku −1−kd (ku −1)!
(kd − 1)!
ku + (ku − 1) h=1
(ku −h−1)!
(13)
Thanks to result i) in Proposition 1 maximising PoA(f, ku ) is equivalent to minimising χ(f, ku )
u
and fk0 d can be computed by the following linear program (LP) in the unknowns x, {f (j)}kj=1
min
x≥0, f ∈F
x
s.t. jf (j) − f (j + 1) ≤ x
j ∈ [ku − 1] ,
(14)
(ku − 1)f (ku ) ≤ x ,
f (j) = fk?d (j)
j ∈ [kd ] .
We remove the constraints x ≥ 0, f ∈ F as well as jf (j) − f (j + 1) ≤ x for j ∈ [kd − 1] and
introduce the following relaxed linear program
min x
x, f
s.t. jf (j) − f (j + 1) ≤ x
j ∈ [kd , ku − 1] ,
(15)
(ku − 1)f (ku ) ≤ x ,
f (j) = fk?d (j)
j ∈ [kd ] .
The proof is divided in two subproofs:
i) We show that a solution to the relaxed program (15) is given by (9) and (13).
ii) We show that the solution to the relaxed program obtained in i) is feasible for the original
problem too.
Proof. i) The proof proceeds by showing that a solution of (15) can be obtained transforming
all the inequality constraint into equalities. This will produce the expressions (9) and (13).
Let us define vj = f (j) for j ∈ [p+1, ku ] and introduce the cost function J(x, vkd +1 , . . . , vku ) =
x. We further introduce the constraint functions g1 (x, vkd +1 ) = −x−vkd +1 and gi (x, vp+i−1 , vp+i ) =
−x + j vp+i−1 − vp+i for i ∈ [2, ku − p] and gku −kd +1 (x, vku ) = −x + (ku − 1)vku . With these
October 23, 2017
DRAFT
16
definitions the LP (15) is equivalent to the following where we have removed the decision
variables that are already determined
min
x,vkd +1 ,...,vku
s.t.
J(x, vkd +1 , . . . , vku ) ,
g1 (x, vkd +1 ) ≤ −kd fk?d (kd ) ,
gi (x, vp+i−1 , vp+i ) ≤ 0
i ∈ [2, ku − p] ,
gku −kd +1 (x, vku ) ≤ 0 .
Thanks to the convexity of the cost function and to the polytopic constraints, the Karush-KuhnTucker conditions are necessary and sufficient for optimality [27]. Consequently, a feasible point
?
, . . . , vn? ) is an optimiser iff there exists µi ∈ R so that
(x? , vk+1
?
∇J +
ku −k
d +1
X
µi ∇gi? = 0
i=1
gi? ≤ 0,
µi ≥ 0,
µi gi? = 0 i ∈ [ku − kd + 1]
where we used ∇J ? to indicate ∇J(x? , vk?d +1 , . . . , vk?u ), and similarly for gi? , ∇gi? . Observe that
the distribution rule in (9) and the corresponding χ(fk0 d , ku ) in (13) are the unique solution of
the linear system gi? = 0 for all i ∈ [ku − kd + 1], that is
jfk0 d (j) − fk0 d (j + 1) − χ(fk0 d , ku ) = 0 j ∈ [kd , ku − 1] ,
(ku − 1)fk0 d (ku ) − χ(fk0 d , ku ) = 0 ,
f 0 (j) = f ? (j)
j ∈ [kd ] .
kd
kd
(16)
Primal feasibility and complementarity slackness are hence naturally satisfied. We are only left
P u −kd +1
µi ∇gi? = 0. We proceed by writing
to prove that there exists µi ≥ 0 such that ∇J ? + ki=1
the stationarity conditions explicitly and show that this is indeed the case. Note that both the
cost function and the constraints are linear so that their derivatives are constant functions
∇J ? = (1, 0, . . . , 0)
∇g1? = (−1, −1, 0, . . . , 0)
∇g2? = (−1, kd + 1, −1, 0, . . . , 0)
..
.
∇gk?u −kd −1 = (−1, 0, . . . , 0, ku − 2, −1, 0)
∇gk?u −kd = (−1, 0, . . . , 0, ku − 1, −1)
∇gk?u −kd +1 = (−1, 0, . . . , 0, ku − 1)
October 23, 2017
DRAFT
17
Solving the stationarity condition in a recursive fashion starting from last component gives
u −1)!
µ i
= µku −kd +1 (ku(k−1)(k
i ∈ [ku − kd ]
d +i−1)!
Pku −kd +1 µi = 1 .
i=1
Substituting the first equation into the second one and solving yields
Pku −kd −1 (ku −1)(ku −1)! −1
(ku −1)(ku −1)!
µ i =
i ∈ [ku − kd ],
i=1
(kd +i−1)!
(kd +i−1)!
Pku −kd −1 (ku −1)(ku −1)! −1
µku −kd +1 =
.
i=1
(kd +i−1)!
Since µi ≥ 0 for all i ∈ [ku − kd + 1], we conclude that (9) and (13) solve the relaxed program
(15).
Proof. ii) The proof proceeds by showing that (9) and (13) satisfy the constraints removed when
transforming the original program (14) into (15).
Using (13) and (9), it is trivial to verify that χ(fk0 d , ku ) ≥ 0 and fk0 d (1) = 1, fk0 d (j) ≥ 0.
We proceed to prove that fk0 d is non increasing. Note that for j ≤ kd , fk0 d coincides with
fk?d , which was proven to be non increasing in [11]. Further, from Lemma 2 we know that
jfk0 d (j) − (j + 1)fk0 d (j + 1) ≥ 0 for j ∈ [kd , ku − 1]. Thus
jfk0 d (j) − jfk0 d (j + 1) ≥ jfk0 d (j) − (j + 1)fk0 d (j + 1) ≥ 0
for j ∈ [kd , ku − 1], which guarantees that fk0 d is non increasing for j ∈ [kd , ku ] too.
We are left to show that jfk0 d (j) − fk0 d (j + 1) ≤ χ(fk0 d , ku ) for j ∈ [kd − 1]. Since j ∈ [kd − 1],
it holds that
jfk0 d (j) − fk0 d (j + 1) = jfk?d (j) − fk?d (j + 1).
Note that jfk?d (j)−fk?d (j +1) ≤ χ(fk?d , kd ) by definition of χ(fk?d , kd ) in (4). Further, χ(fk?d , kd ) ≤
χ(fk? , ku ) for any kd ≤ ku since the price of anarchy is a monotonically decreasing function
u
(Proposition 1). Finally, Lemma 3 shows that for any kd ≤ ku , it holds χ(fk?u , ku ) ≤ χ(fk0 d , ku ).
Hence jfk0 d (j) − fk0 d (j + 1) ≤ χ(fk0 d , ku ) for j ∈ [kd − 1]. It follows that fk0 d is feasible for the
original problem (14).
Thanks to this, and to the fact that fk0 d is optimal for (15), we conclude that fk0 d is a solution
of the original problem.
P ROOF OF L EMMA 1
Proof. The result of Lemma 2 implies that for all kd ≤ ku
kd fk0 d (kd ) − jfk0 d (j) ≥ 0
October 23, 2017
∀j ∈ [kd , ku − 1] .
DRAFT
18
Further we know from [11] that kd fk?d (kd ) ≤ 1. Since fk?d (kd ) = fk0 d (kd ), we conclude that for
j ∈ [kd , ku − 1]
1 ≥ kd fk?d (kd ) = kd fk0 d (kd ) ≥ jfk0 d (j) .
For j ∈ [kd ], it holds fk0 d (j) = fk?d (j) and we already know that the optimal distribution fk?d
does not overpay the players [11]. This concludes the proof.
P ROOF OF T HEOREM 1
Proof. i) Thanks to Proposition 1, the performance of fk?u on the class of games with cardinality
k can be computed as PoA(fk?u , k) =
4 to χ(fk? , k) and conclude that
1
.
1+χ(fk?u ,k)
Since k ≤ ku , we can apply part i) of Lemma
u
1
.
χ
1 + (fk?u , ku )
on the set of games with cardinality k is the same
PoA(fk?u , k) =
This means that the performance of fk?u
performance of the distribution fk?u on the set of games with cardinality ku ≥ k, and
PoA(fk?u , k) = PoA(fk?u , ku ) ≤ PoA(fk? , k) ,
where the last inequality holds since PoA(fk? , k) is a decreasing function of k as seen in part
iii) of Proposition 1. The inequality is tight if and only if k = ku .
ii) Consider kd ∈ [ku − 1]. The performance of fk0 d on the class of games with cardinality k
can be computed as PoA(fk0 d , k) =
4 to conclude that
1
.
1+χ(fk0 ,k)
Since kd ∈ [k − 1], we apply part ii) of Lemma
d
PoA(fk0 d , k) =
1
.
χ
1 + (fk0 d , ku )
Hence, for kd ∈ [ku − 1], the performance of fk0 d in the class of games with cardinality k
is the same of the performance in the class of games with cardinality ku i.e. PoA(fk0 d , k) =
PoA(fk0 d , ku ). Finally, by Lemma 3 we conclude that such performance is worse than what fk?u
can offer
PoA(fk0 d , k) <
1
= PoA(fk?u , k) .
?
χ
1 + (fku , ku )
Consider kd ∈ [k, ku ]. Since k ≤ kd , only the first k entries of fk0 d will determine the
performance and these are identical to fk?d by definition of fk0 d . Hence PoA(fk0 d , k) =
Further kd ∈ [k, ku ] and part i) of Lemma 4 applies
1
= PoA(fk?d , kd ) ,
PoA(fk0 d , k) =
1 + χ(fk?d , kd )
1
.
1+χ(fk? ,k)
d
so that fk0 d has the same performance of fk?d . Using the fact that the optimal price of anarchy is
a decreasing function, for any p ∈ [k, ku ] we get
PoA(fk0 d , k) = PoA(fk?d , kd ) ≥ PoA(fk?u , ku ) = PoA(fk?u , k) .
The inequality is tight if and only if p = ku .
October 23, 2017
DRAFT
19
P ROOF OF T HEOREM 2
Proof. i) Consider xr (t) for fixed r ∈ R. The integer sequence {xr (t)}∞
t=1 is upper bounded by
the true cardinality k (by definition of cardinality) and is non decreasing in t thanks to its update
rule (line 6 in Algorithm 1). Hence, after a finite number of steps, xr (t) has converged to x∞
r .
Repeating the same reasoning for all the resources r ∈ R shows that the map xr (t) converges
in a finite number t̂ of steps. Hence, for t ≥ t̂ the distribution rule used in the algorithm is
fixed. Consequently the game is potential as it can be formulated as a standard congestion game
[11], [23]. Since for t ≥ t̂ agents are playing round-robin best response on a potential game,
their strategy will converge in a finite number of steps to a Nash equilibrium of the game with
resource specific distribution rules fixed to fxalg
∞ for r ∈ R.
r
ii) Let us define ke = maxr∈R |ae |r (note that in general ke 6= kM ). To ease the notation, in
i i −i
) to
the following we will simply use f (xr , |a|r ) to indicate fxalg
∞ (|a|r ), and similarly u (a , a
r
i
i
refer to ui (ai , a−i ; {x∞
r }r∈R ). Further, we define Ae = ∪i ae and Ao = ∪i ao
i i
−i
By definition of equilibrium we have for all i ∈ [n], ui (aie , a−i
e ) ≥ u (ao , ae ) and hence
X
X
0≤
ui (aie , a−i
)
−
ui (aio , a−i
(17)
e
e ).
i∈[n]
i∈[n]
Using the definition of payoff, the first term can be rewritten as
X
XX
ui (aie , a−i
)
=
f (xr , |ae |r )vr
e
i∈[n] r∈aie
i∈[n]
=
X
|ae |r f (xr , |ae |r )vr =
ke
X
X
(18)
jf (xr , j)vr .
j=1 r∈Ae
|ae |r =j
r∈Ae
With a similar manipulation the second term becomes
X
XX
ui (aio , a−i
f (xr , |(aio , a−i
)
=
e
e )|r )vr
i∈[n] r∈aio
i∈[n]
XX
≥
f (xr , min{k, |ae |r + 1})vr ,
i∈[n] r∈aio
i
−i
this holds because for all resources |(aio , a−i
e )|r ≤ |ae |r + 1, |(ao , ae )|r ≤ k and f is non
increasing in its second argument. For resources r ∈ ao it holds |ao |r ≥ 1, and so
XX
f (xr , min{k, |ae |r + 1})vr
i∈[n] r∈aio
=
X
|ao |r f (xr , min{k, |ae |r + 1})vr
r∈Ao
≥
X
f (xr , min{k, |ae |r + 1})vr .
r∈Ao
October 23, 2017
DRAFT
20
The second term in (17) can thus be lower bounded by
X
X
ui (aio , a−i
)
≥
f (xr , min{k, |ae |r + 1})vr
e
r∈Ao
i∈[n]
=
(19)
ke
X
X
f (xr , min{k, j + 1})vr .
j=0 r∈Ao
|ae |r =j
Substituting (18) and (19) in (17) gives
X
X
0≤
ui (aie , a−i
)
−
ui (aio , a−i
e
e )
i∈[n]
≤
=
ke
X
i∈[n]
X
jf (xr , j)vr −
ke
X
X
f (xr , min{k, j + 1})vr
j=1 r∈Ae
|ae |r =j
j=0 r∈Ao
|ae |r =j
ke
X
X
ke
X
X
jf (xr , j)vr −
j=1 r∈Ao
|ae |r =j
j=1 r∈Ae
|ae |r =j
−
X
f (xr , min{k, j + 1})vr
vr f (xr , 1)
r∈Ao \Ae
=
ke
X
X
jf (xr , j)vr +
j=1 r∈Ae \Ao
|ae |r =j
−
ke
X
X
j=1 r∈Ae ∩Ao
|ae |r =j
ke
X
X
f (xr , min{k, j + 1})vr −
j=1 r∈Ao
|ae |r =j
=
ke
X
X
jf (xr , j)vr −
ke
X
X
X
vr f (r, 1)
r∈Ao \Ae
j=1 r∈Ae \Ao
|ae |r =j
+
jf (xr , j)vr
X
vr
r∈Ao \Ae
jf (xr , j) − f (xr , min{k, j + 1}) vr ,
(20)
j=1 r∈Ao
|ae |r =j
where we have used the fact that f (xr , 1) = 1 for all resources. We intend to bound the first and
the third term in the last expression. In the summands of (20) j = |ae |r ≤ x∞
r due to the update
of xr (t) in Algorithm 1 and recall that f (xr , |a|r ) = fxalg
∞ (|a|r ). Hence we can apply Lemma 5
r
to the first term in (20)
ke
X
X
j=1 r∈Ae \Ao
|ae |r =j
October 23, 2017
jf (xr , j)vr ≤
ke
X
X
j=1 r∈Ae \Ao
|ae |r =j
(χ(fk?M , kM ) + 1)vr .
(21)
DRAFT
21
Similarly for the third term in (20)
ke
X
X
jf (xr , j) − f (xr , min{k, j + 1}) vr
j=1 r∈Ao
|ae |r =j
≤
ke
X
X
(22)
χ(fk? , kM )vr = χ(fk? , kM )
M
M
X
vr .
r∈ae ∩ao
j=1 r∈Ao
|ae |r =j
Hence combining (20) with the bounds from (21) and (22)
0 ≤(χ(fk?M , kM ) + 1)
ke
X
X
(23)
vr
j=1 r∈Ae \Ao
|ae |r =j
+ χ(fk?M , kM )
X
r∈Ae ∩Ao
=(χ(fk?M , kM ) + 1)
X
vr −
vr
r∈Ao \Ae
X
(24)
vr
r∈Ae \Ao
+ χ(fk?M , kM )
X
r∈Ae ∩Ao
=(χ(fk?M , kM ) + 1)
X
vr −
vr
r∈Ao \Ae
X
(25)
vr
r∈Ae \Ao
X
+ (χ(fk?M , kM ) + 1)
X
r∈Ae
vr −
X
vr
r∈Ao
r∈Ae ∩Ao
=(χ(fk?M , kM ) + 1)
X
vr −
vr
r∈Ao
=(χ(fk?M , kM ) + 1)W (ae ) − W (ao ).
Hence (χ(fk?M , kM ) + 1)W (ae ) − W (ao ) ≥ 0 and rearranging
W (ae )
1
≥
= PoA(fk?M , kM ) ≥ PoA(fk? , k) ,
?
χ
W (ao )
1 + (fkM , kM )
where the last inequality follows from the fact that the price of anarchy is a decreasing function,
and kM ≤ k by definition of cardinality.
P ROOF OF T HEOREM 3
Proof. i) Consider the covering problem depicted in the following figure (a), composed of players
p1 , p2 , p3 represented by a solid dot; resources r1 , r2 , r3 , r4 represented by a circle with values
v1 , v2 , v3 , v4 such that
v1 > v3 > v4 > v2
October 23, 2017
and v1 f3? (2) < v2 < v1 f2? (2) < v4
DRAFT
k
2
k
e
n
r1
✓++
. . . xn xn+1
. . . x2n
1
x2n
v3
r3
p3
v1
22
v3
✓
✓p1
✓p+2
✓++
p2
p1
{ron
1 , r2 , r3 }, {r2 , r3 , r4 }, {r1 , r2 , r3 , r4 }, respectively i.e. each player can only choose one arrow
v2
x1 x2 . . .
pointing
outwards from himself.
v2
. . . xn xn+1
r2
off
xn+1 xn+2 . . .
p1
v1
v2
v3
1
p1
r2
. . . x2n
p2
r1
✏
r1
As an example take v = (11, 5, 7, 6). Each player p1 , p2 , p3 can choose only one resource from
r3
2
r3
p3
v1
n
r2
r3
1
24 ✏
1
x2n
v
p2
p3
2
✏
p3
v1
p1
p2
v2
v3
v4
p3
(a) Original game
r4
(b) Equilibrium ae
The cardinality is k = 3 since all players could choose simultaneously r2 or r3 , hence the
optimal distribution rule is f ? . Amongst the equilibria obtained with f3? there is ae = (r2 , r3 , r1 ),
a project, HeatReserves and .. This work was not supported by Swiss Nano-Tera
3
project, HeatReserves and ..
ygeros are with the
Switzerland. Email:
.ee.ethz.ch.
MK: ask your funding from John
depicted
in the previous figure (b). This configuration is an equilibrium since v2 > v1 f3? (2) >
D. Paccagnan, M. Kamgarpour and J. Lygeros are with the
Control Laboratory, ETH Zürich,
and v1 Email:
> v2 f3? (2), v1 > v3 f3? (2), v1 > v4 . Such equilibrium
v3 > v4 , v3 > v2 f3? (2)Switzerland.
v3Automatic
f3? (2) and
{dariop,mkamgar,lygeros}@control.ee.ethz.ch.
gives a welfare of v1 + v2 + v3 that is less than the optimal v1 + v3 + v4 , since v2 < v4 . We intend
to show that for any initial condition and for any execution, Algorithm 1 will converge to an
optimal allocation. This suffices to prove that the worst equilibrium obtained with Algorithm 1
performs better than the worst equilibrium obtained with f3? , which is not optimal as shown above.
Observe that the conditions v1 > v3 > v4 > v2 and v4 > v1 f2? (2) ensure that an allocation with
two or more agents covering the same resource is never an equilibrium.
This holds regardless
4
of the distribution used. Hence, the welfare can potentially take = 4 different values,
3
since the binomial represents the number of subsets with 3 elements (agents allocations) that
can be extracted from a set of 4 elements (set of resources). These different welfare values are
obtained for (r1 , r2 , r4 ), (r2 , r3 , r4 ), (r1 , r3 , r4 ), (r1 , r2 , r3 ), or feasible permutations of each. The
allocation (r1 , r2 , r4 ) is never an equilibrium since player p3 can improve moving to r3 because
v3 > v4 . Similarly for any feasible permutation of (r1 , r2 , r4 ), the player selecting resource r4
can always improve moving to r3 . The allocation (r2 , r3 , r4 ) is never an equilibrium since player
p3 can improve moving to r1 since v1 is the highest. Similarly for any feasible permutation
of (r2 , r3 , r4 ), there exists a player that can improve moving to r1 . This holds regardless of
what distribution rule is used. The allocation (r1 , r3 , r4 ) (or any feasible permutation) is optimal.
We are thus left to show that Algorithm 1 never converges to (r1 , r2 , r3 ), or any other feasible
permutation. We show this by enumeration.
October 23, 2017
DRAFT
23
The allocation (r1 , r2 , r3 ) can not be an equilibrium since player p2 can improve moving to
r4 because v4 > v2 . The allocation (r1 , r3 , r2 ) can not be an equilibrium since player p3 can
improve moving to r4 . The allocation (r3 , r2 , r1 ) can not be an equilibrium since player p2 can
improve moving to r4 . We are left to check ae = (r2 , r3 , r1 ), depicted in the previous figure (b).
On characterizing
Controllability
of 1,abecause
Population
of for
Thermostatically
bility
of a This
Population
of Thermostatically
l = 1, 2 and so player
can not be an
equilibrium
of Algorithm
v2 < v1 f`alg (2)
p2 could improve moving toControlled
r1 . The fact thatLoads
the algorithm uses l ≤ 2 on resource r1 holds
ntrolled Loads
because the maximum number of players on r1 is two, and so kt (1) ≤ 2 at any time step t ∈ [n].
We conclude
thatLygeros
all
the equilibria
towards
which the and
algorithm
converges give optimal welfare,
Dario
Paccagnan,
Maryam
Kamgarpour,
John Lygeros
aryam Kamgarpour,
and John
?
while f3 also produces the suboptimal equilibrium ae ; the claim follows. Observe that this is
not I.
a worst
case instance because the price of anarchy with the example values v = (11, 5, 7, 6)
I NTRODUCTION
is
10k(x)
=
µ10 (x,
2 t)
on
on n = ke
k
2
2
n
k
n = ke
k
n
i) Consider the covering problem depicted in the following figure (a), composed of players
✓ 2
ke
k ✓+
ke
✓+
n
2
ke
p1 , p2 , p3 represented
by a solid dot; resources r1 , r2 , r3 represented by an empty circle with
01 (x)
n
n
k
ke
2
k
ke
2 t)
µ01 (x,
values v1 , v2 , v3 such that
off
off
✓
Fig.
Soft boundaries transition rates
(x) at
01 1.
As an 2example
the bottom.
++
W (ae )
11 + 5 + 7
23
7
?
k=
=
=
> PoA(f
= ke .
n
2
3 , 3)
W (ao )
11 + 6 + 7
24
11
ke
v3 f3? (2) < v1 < v2 < v3 /2 <kv3 .
e
n
2
k
k
take v = [9, 9.5, 20]. Each player p1 , p2 , p3 can choose only one resource from
10 (x)
at the top and
01 (x)
n
at
{r1 , r2 },
i.e. each player
can only
ke arrow pointing
e
2
k choose one
2 {r2 , r3 }, k{r1 , r2 , r3 },krespectively
outwards from himself.
r1
p3
v1
✓
✓
p1✓+
n+1 x1
r3
r1
v3
v1
p2✓++
on
v2
. . . xn xn+1
r2
x2 . . .
off
xn+1 xn+2 . . .
1 x2n
n
n
. . .game
x2n
(a) Original
1
x2n
r3
p3
v3
p2
p1
v2
r2
(b) Equilibrium a2
The cardinality is k = 3 since all players could choose simultaneously r1 , hence the optimal
?
distributionr1rule is fr23? . All the
r3 equilibria obtained with f3 are completely spread i.e. they feature
one and only one player on each resource. Any allocation where there are two or more players
1
1
2
✏
in one resource is not an equilibrium for f3? , as detailed in the following.
If all three players selected resource r2 , p2 could improve moving to r3 since v3 > v2 . If p1 and
p
p
p
2
3
p3 selected r1 ,1 depending
on the choice of p2 , either p1 or p3 could improve moving respectively
to r2 or r3 since v2 > v1 and v3 > v1 . If p2 and p3 selected r3 , depending on the choice
2
October 23, 2017
✏
r4
DRAFT
24
of p1 , either p2 or p3 could improve moving respectively to r2 or r1 since v3 f3? (2) < v2 and
v3 f3? (2) < v1 . If p1 , p3 selected both r2 , regardless of the choice of p2 , p3 could improve moving
to r3 since v3 > v2 . If p2 , p3 selected both r2 , regardless of the choice of p1 , p3 could improve
moving to r3 since v3 > v2 . Finally, if p1 , p2 selected r2 , regardless of the choice of p3 , p2 could
always improve moving to r3 since v3 > v2 . Thus all equilibria obtained with f3? (including the
worst) give a welfare of v1 + v2 + v3 .
Let us consider Algorithm 1 and initialise it at a1 = [r2 , r3 , r1 ], giving k1 (r) = 1 for all r.
Player p3 updates and since v3 · 1 > v1 · 1, he selects r3 , giving a2 = [r2 , r3 , r3 ] and k2 (r) = 1
for r1 , r2 and k2 (r3 ) = 2. This allocation is depicted in the previous figure (b) and is an
equilibrium configuration. Indeed p1 can not improve since v2 > v1 ; p2 can not improve since
v3 f2? (2) =
v3
2
> v2 · 1; p3 can not improve since v3 f2? (2) =
v3
2
> v2 · 1 and v3 f2? (2) =
v3
2
> v1 · 1.
Such equilibrium has a welfare of v2 + v3 .
In conclusion, all equilibria obtained with f3? give a better welfare than a2 and thus of the
worst equilibrium obtained with Algorithm 1.
L EMMATA
Lemma 2. Let kd ∈ [ku ]. The distribution fk0 d satisfies
jfk0 d (j) − (j + 1)fk0 d (j + 1) ≥ 0
j ∈ [kd , ku − 1] .
Proof. Recall that fk0 d is obtained from equation (16). Using χ(fk0 d , ku ) from (13), one can
reconstruct the tail entries of fk0 d (j) with the following backward recursion
jfk0 d (j) − fk0 d (j + 1) = χ(fk0 d , ku ) j ∈ [kd , ku − 1] ,
(ku − 1)fk0 d (ku ) = χ(fk0 d , ku ) .
χ(f 0 ,ku )
k
, the first equation gives for j ≥ kd
kX
u −1
j!
j!
0
0
+
.
jfkd (j) = χ(fkd , ku ) 1 +
i!
(k
−
1)(k
−
1)!
u
u
i=j+1
Starting from fk0 d (ku ) =
d
ku −1
Hence
1
0
0
χ(fk0 , ku ) (jfkd (j) − (j + 1)fkd (j + 1)) =
d
kX
u −1
kX
u −1
(j + 1)!
j! − (j + 1)!
j!
=
−
+
=
i! i=j+2
i!
(ku − 1)(ku − 1)!
i=j+1
=
kX
u −2
i=j+1
October 23, 2017
j! (j + 1)!
−
i!
(i + 1)!
j!
1
j+1
+
1+
−
.
(ku − 1)!
ku − 1 ku − 1
DRAFT
25
Note that for i > j, one has
j!
1
=
i!
i(i − 1) . . . (j + 1)
and so
j! (j + 1)!
−
> 0.
i!
(i + 1)!
Further
1
j+1
ku − j − 1
+
=
≥0
ku − 1 ku − 1
ku − 1
since j ≤ ku − 1 by assumption. Hence we conclude that
1+
jfk0 d (j) − (j + 1)fk0 d (j + 1) ≥ 0
j ∈ [kd , ku − 1] .
Lemma 3. For any 1 < kd < ku it holds χ(fk?u , ku ) < χ(fk0 d , ku ).
Proof. The expression of χ(fk?u , ku ) in (6) and of χ(fk0 d , ku ) in equation (13) can be rewritten as
χ(fk? , ku ) =
u
χ(fk0 , ku ) =
d
(ku − 1)(ku − 1)!
P u −1
1 + (ku − 1)(ku − 1)! ki=1
ku +
,
(ku − 1)(ku − 1)!
Pku −1−kd (ku −1)(ku −1)! β(kd ) ,
h=1
where
β(kd ) :=
1
i!
(ku −h−1)!
1
1+
Pkd −1
(kd −1)(kd −1)!
h=1
h!
.
Instead of showing χ(fk?u , ku ) < χ(fk0 d , ku ), in the following we equivalently prove that
1
χ(f 0 ,ku ) i.e., that
1
χ(f ? ,ku )
>
ku
kd
kX
u −1
k −2
u
X
1
1
= ku + (ku − 1)(ku − 1)!
>
1 + (ku − 1)(ku − 1)!
i!
i!
i=1
i=1
ku −1−k
kX
d −1
X d (ku − 1)(ku − 1)!
(kd − 1)(kd − 1)!
ku +
1+
.
(ku − h − 1)!
h!
h=1
h=1
The previous inequality can be rewritten as
kX
ku −1−k
u −2
Xd
1
1
(ku − 1)(ku − 1)!
−
>
i!
(k
−
h
−
1)!
u
i=1
h=1
ku −1−k
d
d −1
X (ku − 1)(ku − 1)! kX
(kd − 1)(kd − 1)!
ku +
.
(ku − h − 1)!
h!
h=1
h=1
P d −1 1
P d −1
Since the left hand side is equal to (ku −1)(ku −1)! kh=1
, we can simplify the term kh=1
h!
1
h!
to get
(ku − 1)(ku − 1)!
>
(kd − 1)(kd − 1)!
October 23, 2017
ku −1−k
X d (ku − 1)(ku − 1)!
ku +
,
(ku − h − 1)!
h=1
DRAFT
26
which is finally equivalent to
ku −1−k
Xd
1
1
ku
>
+
.
(kd − 1)(kd − 1)!
(ku − 1)(ku − 1)
(ku − h − 1)!
h=1
(26)
We use induction to show that inequality (26) holds for 1 < kd < ku , as required. We start from
kd = ku − 1 and apply induction backwards until we reach kd = 2.
i) For kd = ku − 1 and kd > 1 inequality (26) reads as
kd + 1
1
>
⇐⇒ kd 2 > kd 2 − 1 ,
(kd − 1)(kd − 1)!
kd · kd !
which is always satisfied.
ii) Let us assume the inequality holds for a generic kd ≤ ku − 1, we show that it holds also
for kd − 1 (with kd > 1). That is, we assume
ku −1−k
Xd
1
ku
1
>
+
,
(kd − 1)(kd − 1)!
(ku − 1)(ku − 1)
(k
u − h − 1)!
h=1
(27)
and want to show
ku −k
Xd
ku
1
1
>
+
.
(kd − 2)(kd − 2)!
(ku − 1)(ku − 1) h=1 (ku − h − 1)!
(28)
We can rewrite the right hand side of (28) and use (27) to upper bound it
kX
u −kd
ku
1
+
=
(ku − 1)(ku − 1)
(ku − h − 1)!
h=1
ku −k
d −1
X
ku
1
1
+
+
<
(ku − 1)(ku − 1)
(k
(k
u − h − 1)!
d − 1)!
h=1
(29)
1
1
kd
+
=
<
(kd − 1)(kd − 1)! (kd − 1)!
(kd − 1)(kd − 1)!
1
.
(kd − 2)(kd − 2)!
The last inequality holds since it is equivalent to
kd
1
<
⇐⇒ kd 2 − 2kd < (kd − 1)2 ,
2
(kd − 1)
kd − 2
which is always satisfied. Comparing the first and last term in (29) gives (28).
This completes the induction and thus the proof.
Lemma 4.
i) For any 1 ≤ l ≤ m, m ∈ N it holds
?
?
χ(fm
, m) .
, l) = χ(fm
October 23, 2017
DRAFT
27
ii) For any k ∈ [ku ] and 1 < kd < k it holds
χ(fk0 , k) = χ(fk0 , ku ) .
d
d
Proof. i) If l = m, the result holds trivially. Hence in the following we consider l ∈ [m − 1].
?
By definition of χ(fm
, l) in (4), one has
?
χ(fm
, l) = min x
x≥0
?
?
s.t. jfm
(j) − fm
(j + 1) ≤ x
j ∈ [l − 1] ,
?
(l − 1)fm
(l) ≤ x .
?
is derived in [11, Theorem 2] solving the following recursion
Note that fm
?
?
?
jfm
(j) − fm
(j + 1) = χ(fm
, m)
j ∈ [m − 1]
(30)
?
?
(l − 1)fm
(l) = χ(fm
, m) .
?
, m).
Since m > l, it follows that any feasible x from the LP above has to satisfy x ≥ χ(fm
?
?
(l) ≤ x is also
, m), the constraint (l − 1)fm
In the following we show that setting x = χ(fm
?
?
satisfied. This will be enough to conclude that χ(fm , l) = χ(fm , m).
?
Since fm
is non increasing, one has
?
?
?
?
?
(l − 1)fm
(l) − χ(fm
, m) = lfm
(l) − fm
(l) − χ(fm
, m) ≤
?
?
?
≤ lfm
(l) − fm
(l + 1) − χ(fm
, m) = 0 ,
where the equality holds applying (30) for j = l ∈ [m − 1].
ii) We intend to compute
χ(fk0 , k) = min x
d
x≥0
s.t.
jfk0 d (j) − fk0 d (j + 1) ≤ x
j ∈ [k − 1]
(k − 1)fk0 d (k) ≤ x .
For any feasible x, it must be x ≥ χ(fk0 d , ku ) due to how fk0 d (j) is recursively defined for
j > kd in Equation (16). Similarly to what shown before, one can prove that x = χ(fk0 , ku )
d
will also satisfy the constraint (k − 1)fk0 d (k) ≤ x. Hence χ(fk0 d , k) = χ(fk0 d , ku ) and the proof is
concluded.
Lemma 5. For all resources r ∈ R, the distribution rules fxalg
∞ are such that
r
October 23, 2017
?
j fxalg
∞ (j) ≤ χ(fk , kM ) + 1
M
r
j ∈ [x∞
r ]
(31)
alg
?
j fxalg
∞ (j) − fx∞ (min{k, j + 1}) ≤ χ(fk , kM )
M
r
r
j ∈ [x∞
r ]
(32)
DRAFT
28
where kM = maxr∈R x∞
r .
∞
Proof. We start from (31) and examine fxalg
∞ for a fixed r ∈ R. Consider j ∈ [xr − 1], by
r
?
definition of fxalg
∞ and the fact that fx∞ is non increasing
r
r
?
?
j fxalg
∞ (j) − 1 = j fx∞ (j) − fx∞ (1)
r
r
r
≤ j fx?∞
(j) − fx?∞
(j + 1) ≤ χ(fx?∞
, x∞
r )
r
r
r
?
∞
j fxalg
∞ (j) ≤ χ(fx∞ , xr ) + 1 ,
r
r
=⇒
where the last inequality holds thanks to the definition (4). Since x∞
r ≤ kM and the price
?
∞
?
of anarchy is a decreasing function, one has χ(fx∞
, xr ) ≤ χ(fkM , kM ) and so j fxalg
∞ (j) ≤
r
r
χ(fk? , kM ) + 1 for j ∈ [x∞
r − 1].
M
In a similar fashion when j = x∞
r
alg
∞ ?
?
x∞
(x∞
(x∞
(1)
r f x∞
r ) − 1 = x r f x∞
r ) − f x∞
r
r
r
?
?
≤ x∞
(x∞
(x∞
r f x∞
r ) − f x∞
r )
r
r
=
(x∞
r
−
1)fx?∞
(x∞
r )
r
(33)
χ ?
= χ(fx?∞
, x∞
r ) ≤ (fkM , kM ) ,
r
where the only difference is in the last equality that comes from equation (6). Repeating the
same reasoning for all r ∈ R, one has proven (31).
In the remaining, we show that (32) holds. Consider fxalg
∞ for a fixed resource r ∈ R and
r
∞
∞
recall that xr ≤ k. Thus for j ∈ [xr − 1] one has min{k, j + 1} = j + 1 and the claim reads
alg
?
as j fxalg
∞ (j) − fx∞ (j + 1) ≤ χ(fk , kM ). This holds since
r
r
M
alg
?
?
j fxalg
∞ (j) − fx∞ (j + 1) = j fx∞ (j) − fx∞ (j + 1)
r
r
r
r
χ ?
≤ χ(fx?∞
, x∞
r ) ≤ (fkM , kM )
r
where the first inequality holds thanks to definition (4) and the last since the price of anarchy
is non increasing (x∞
r ≤ kM ).
∞
In the remaining we focus on j = x∞
r and divide the proof in two subparts. When k = xr ,
min{k, j + 1} = x∞
r and the claim follows from
alg
alg
?
∞ ?
(x∞
(x∞
x∞
(x∞
(x∞
r )
r ) = x r f x∞
r ) − f x∞
r f x∞
r ) − f x∞
r
r
r
r
?
χ ? , x∞
χ ?
= (x∞
(x∞
r ) ≤ (fkM , kM ) ,
r − 1)fx∞
r ) = (fx∞
r
r
∞
similarly to (33). When k > x∞
r , then min{k, j + 1} = xr + 1 and the claim holds if we show
alg
alg
χ ? , x∞
x∞
(x∞
(x∞
r f x∞
r ) − f x∞
r + 1) ≤ (fx∞
r ).
r
r
r
October 23, 2017
DRAFT
29
For this to hold, one has to require
∞
?
∞
∞
∞ alg
fxalg
∞ (xr + 1) ≥ xr fx∞ (xr ) − χ(fx∞ , xr )
r
r
r
?
χ ? , x∞
(x∞
= x∞
r )
r ) − (fx∞
r f x∞
r
r
alg
(x∞
= fx?∞
(x∞
r ),
r ) = f x∞
r
r
∞
?
where the second equality sign follows form χ(fx?∞
, x∞
(x∞
r ) = (xr − 1)fx∞
r ), that is form
r
r
equation (4). Hence we need to impose
alg
∞
∞
fxalg
∞ (xr + 1) ≥ fx∞ (xr ) ,
r
r
∞
but at the same time we are limited to non increasing distribution rules. Hence we set fxalg
∞ (xr +
r
alg
∞
from Equation (11). The proof is completed by observing
1) = fxalg
∞ (xr ) as by definition of f`
r
that the same reasoning can be repeated for any resource r ∈ R.
R EFERENCES
[1] D. Paccagnan and J. R. Marden, “The risks and rewards of conditioning noncooperative designs to additional information,”
in Proceedings of the 55th Allerton Conference on Communication, Control, and Computing, 2017.
[2] D. D. Siljak, Decentralized control of complex systems.
Courier Corporation, 2011.
[3] D. Paccagnan, M. Kamgarpour, and J. Lygeros, “On aggregative and mean field games with applications to electricity
markets,” in 2016 European Control Conference (ECC), June 2016, pp. 196–201.
[4] P. N. Brown and J. R. Marden, “Studies on robust social influence mechanisms: Incentives for efficient network routing
in uncertain settings,” IEEE Control Systems, vol. 37, no. 1, pp. 98–115, Feb 2017.
[5] J. R. Marden and A. Wierman, “Distributed welfare games,” Operations Research, vol. 61, no. 1, pp. 155–168, 2013.
[6] S. Martinez, J. Cortes, and F. Bullo, “Motion coordination with distributed information,” IEEE Control Systems, vol. 27,
no. 4, pp. 75–88, Aug 2007.
[7] M. Pavone, A. Arsie, E. Frazzoli, and F. Bullo, “Distributed algorithms for environment partitioning in mobile robotic
networks,” IEEE Transactions on Automatic Control, vol. 56, no. 8, pp. 1834–1848, Aug 2011.
[8] C. Langbort, R. S. Chandra, and R. D’Andrea, “Distributed control design for systems interconnected over an arbitrary
graph,” IEEE Transactions on Automatic Control, vol. 49, no. 9, pp. 1502–1519, Sept 2004.
[9] B. Bamieh, F. Paganini, and M. A. Dahleh, “Distributed control of spatially invariant systems,” IEEE Transactions on
Automatic Control, vol. 47, no. 7, pp. 1091–1107, Jul 2002.
[10] J. R. Marden, “The role of information in multiagent coordination,” in 53rd IEEE Conference on Decision and Control,
Dec 2014, pp. 445–450.
[11] M. Gairing, Covering Games: Approximation through Non-cooperation. Berlin, Heidelberg: Springer Berlin Heidelberg,
2009, pp. 184–195.
[12] C. Langbort and V. Gupta, “Minimal interconnection topology in distributed control design,” SIAM Journal on Control
and Optimization, vol. 48, no. 1, pp. 397–413, 2009.
[13] V. Ramaswamy, D. Paccagnan, and J. R. Marden, “The impact of local information on the performance of multiagent
systems,” ArXiv:1710.01409, 2017.
[14] V. V. Vazirani, Approximation Algorithms.
October 23, 2017
New York, NY, USA: Springer-Verlag New York, Inc., 2001.
DRAFT
30
[15] U. Feige, “A threshold of ln n for approximating set cover,” J. ACM, vol. 45, no. 4, pp. 634–652, July 1998.
[16] R. A. Murphey, Target-Based Weapon Target Assignment Problems.
Boston, MA: Springer US, 2000, pp. 39–53.
[17] U. Feige and J. Vondrak, “Approximation algorithms for allocation problems: Improving the factor of 1 - 1/e,” in 2006
47th Annual IEEE Symposium on Foundations of Computer Science (FOCS’06), Oct 2006, pp. 667–676.
[18] H.-L. Chen, T. Roughgarden, and G. Valiant, “Designing network protocols for good equilibria,” SIAM Journal on
Computing, vol. 39, no. 5, pp. 1799–1832, 2010.
[19] B. Gentile, F. Parise, D. Paccagnan, M. Kamgarpour, and J. Lygeros, “Nash and wardrop equilibria in aggregative games
with coupling constraints,” arXiv preprint arXiv:1702.08789, 2017.
[20] G. Nemhauser, L. Wolsey, and M. Fisher, “An analysis of approximations for maximizing submodular set functions – I,”
Mathematical Programming, vol. 14, no. 1, pp. 265–294, 1978.
[21] A. Krause and C. Guestrin, “Near-optimal observation selection using submodular functions,” in in Proceedings of the
22nd Conference on Artifical Intelligence, 2007.
[22] J. R. Marden, G. Arslan, and J. S. Shamma, “Cooperative control and potential games,” IEEE Transactions on Systems,
Man, and Cybernetics, Part B (Cybernetics), vol. 39, no. 6, pp. 1393–1407, Dec 2009.
[23] D. Monderer and L. S. Shapley, “Potential games,” Games and Economic Behavior, vol. 14, no. 1, pp. 124 – 143, 1996.
[24] J. F. Nash, “Equilibrium points in n-person games,” Proceedings of the National Academy of Sciences, vol. 36, no. 1, pp.
48–49, 1950.
[25] E. Koutsoupias and C. Papadimitriou, “Worst-case equilibria,” in Proceedings of the 16th Annual Conference on Theoretical
Aspects of Computer Science, ser. STACS’99.
Berlin, Heidelberg: Springer-Verlag, 1999, pp. 404–413.
[26] M.-F. Balcan, B. Manthey, H. Röglin, and T. Roughgarden, “Analysis of Algorithms Beyond the Worst Case (Dagstuhl
Seminar 14372),” Dagstuhl Reports, vol. 4, no. 9, pp. 30–49, 2015.
[27] S. Boyd and L. Vandenberghe, Convex optimization.
October 23, 2017
Cambridge university press, 2004.
DRAFT
| 3cs.SY
|
Closure properties of classes of multiple testing procedures
arXiv:1607.04327v4 [math.ST] 30 Apr 2017
Georg Hahn
Department of Mathematics, Imperial College London
Abstract
Statistical discoveries are often obtained through multiple hypothesis testing. A variety
of procedures exists to evaluate multiple hypotheses, for instance the ones of BenjaminiHochberg, Bonferroni, Holm or Sidak. We are particularly interested in multiple testing
procedures with two desired properties: (solely) monotonic and well-behaved procedures.
This article investigates to which extent the classes of (monotonic or well-behaved) multiple
testing procedures, in particular the subclasses of so-called step-up and step-down procedures, are closed under basic set operations, specifically the union, intersection, difference
and the complement of sets of rejected or non-rejected hypotheses. The present article
proves two main results: First, taking the union or intersection of arbitrary (monotonic
or well-behaved) multiple testing procedures results in new procedures which are monotonic
but not well-behaved, whereas the complement or difference generally preserves neither property. Second, the two classes of (solely monotonic or well-behaved) step-up and step-down
procedures are closed under taking the union or intersection, but not the complement or
difference.
Keywords: Multiple Hypothesis Testing, Statistical Significance, Step Up Procedure, Set Operations, Monotonicity
1
Introduction
Multiple testing is a widespread tool to evaluate scientific studies (Westfall and Young, 1993;
Hsu, 1996; Hochberg and Tamhane, 2008). We are interested in testing m ∈ N hypotheses
H01 , . . . , H0m with corresponding p-values p1 , . . . , pm for statistical significance while controlling
an error criterion such as the familywise error (FWER) or the false discovery rate (FDR).
Following Gandy and Hahn (2016), we define a multiple testing procedure as a mapping
h : [0, 1]m × [0, 1] → P({1, . . . , m})
whose input is a vector of m p-values p ∈ [0, 1]m and a significance level α ∈ [0, 1] and whose
output is the set of indices of rejected hypotheses, where P denotes the power set.
Many procedures of the above form are available in the literature in order to correct for
multiple tests, for instance the procedures of Bonferroni (1936), Sidak (1967), Holm (1979),
Hochberg (1988) or Benjamini and Hochberg (1995). Many common procedures, including the
ones aforementioned, belong to a certain class of procedures, called step-up and step-down
procedures (Romano and Shaikh, 2006). It is assumed throughout the article that only the
m p-values which serve as input to h are used as a basis for making decisions, dependencies
between elementary hypotheses are not considered explicitly. Apart from defining properties on
p imposed by those multiple testing procedures to which the results of this article are applied,
no additional conditions on p are required.
1
This article focuses on two types of multiple testing procedures: monotonic procedures defined in Roth (1999) and Tamhane and Liu (2008) as well as well-behaved procedures (Gandy and
Hahn, 2016). We investigate to which extent the class of solely monotonic and the class of wellbehaved multiple testing procedures is closed under the computation of the union, intersection,
difference or the complement of sets of rejected or non-rejected hypotheses.
A multiple testing procedure is said to be monotonic if smaller p-values (Tamhane and Liu,
2008) or a higher significance level (Roth, 1999) lead to more rejections. Gandy and Hahn
(2016) call a monotonic multiple testing procedure well-behaved if p-values corresponding to
rejected hypotheses can be lowered and p-values corresponding to non-rejected hypotheses can
be increased while leaving all rejections and non-rejections invariant.
For a set of given hypotheses, the closed testing procedure (CTP) of Marcus et al. (1976)
(also referred to as the closure principle) and the partitioning principle (PP) of Finner and
Strassburger (2002) provide means to efficiently construct a simultaneous hypothesis test controlling the FWER. The CTP is based on enforcing coherence (Gabriel, 1969): An intersection
hypothesis HI , that is a hypothesis of the form HI = ∩i∈I Hi for I ⊆ {1, . . . , m}, is rejected if
and only if all intersection hypotheses implying HI are rejected by their local tests (Hommel
et al., 2007). Many common procedures such as the one of Holm (1979) can be constructed
using the CTP. The PP divides the parameter space underlying the hypotheses of interest into
disjoint subsets which are then tested independently at level α. Since the partitioned hypotheses
are disjoint, no multiplicity correction is necessary and at most one of the mutually exclusive
hypotheses is true. Whereas CTP and PP can only be used to construct procedures with FWER
control, the present article offers a means to combine procedures controlling several criteria such
as the FDR into one procedure (see the example in Section 4.5). In case of the CTP, the exponential number of tests to be carried out might also pose a problem: The present article
considers the direct construction of step-up and step-down procedures which allow for efficient
testing of multiple hypotheses.
The motivation for the present article is as follows:
1. Investigating closure properties (in a set theoretical sense) of a class, in the case of the
present article certain classes of multiple testing procedures, is of interest in its own right:
The closure of step-up and step-down procedures allows us to construct new multiple testing
procedures of the same (step-up/step-down) form from existing ones; moreover, the resulting
procedure will be given explicitly.
2. Being able to perform set operations with multiple testing procedures is useful in practice:
Many multiple testing procedures exist to test hypotheses according to various criteria, each
of which might prove beneficial in certain applications. Whereas hypotheses can also be tested
sequentially using several procedures, it is nontrivial a priori that procedures can be combined to
test multiple hypotheses in a single run while drawing benefits of several criteria simultaneously.
This feature is similar to using (stepwise) “shortcut procedures” (Romano and Wolf, 2005;
Hommel et al., 2007) which aim to reduce the (potentially) exponential number of tests required
by the CTP for FWER control to a polynomial number of tests.
3. Monotonic and well-behaved procedures have already been of interest in the literature. For
instance, Gordon (2007) uses the idea of monotonicity to show that there is no monotonic stepup procedure which improves upon the Bonferroni (1936) procedure in the sense that it always
returns the same rejections or possibly more. Gordon and Salzman (2008) show that the classical
Holm (1979) procedure dominates all monotonic step-down multiple testing procedures in the
2
above sense. Proving that certain classes of procedures (for instance, monotonic procedures) are
closed renders the applicability of known results more apparent.
4. The results discussed in this paper extend the methodology developed in Gandy and
Hahn (2014) and Gandy and Hahn (2016) which relies on well-behaved procedures. Briefly, the
authors consider a scenario in which the p-value underlying each hypothesis is unknown, but can
be estimated through Monte Carlo samples drawn under the null, for instance using bootstrap
or permutation tests. Instead of using estimated p-values to obtain ad-hoc decisions on all
hypotheses, the authors prove that it is possible to improve existing algorithms designed for
Monte Carlo based multiple testing (Besag and Clifford, 1991; Lin, 2005; van Wieringen et al.,
2008; Guo and Peddada, 2008; Sandve et al., 2011): the proposed modifications guarantee that
the test results of published algorithms are identical (up to an error probability pre-specified by
the user) to the ones obtained with the unknown p-values. This ensures the repeatability and
objectivity of multiple testing results even in the absence of p-values.
The article is structured as follows. Section 2 provides formal definitions of the two properties of a multiple testing procedure under investigation. Section 3 considers arbitrary (solely
monotonic or well-behaved) multiple testing procedures and demonstrates that solely the monotonicity is preserved when taking unions and intersections. The difference and complement are
neither monotonic nor well-behaved. Section 4 focuses on step-up and step-down procedures and
shows that both classes of (solely monotonic or well-behaved) step-up and step-down procedures
are closed under the union or intersection operation, but not the complement or difference. The
article concludes with a short discussion in Section 5. All proofs are given in Appendix A. In
the entire article, | · | and k · k denote the absolute value and the Euclidean norm, respectively,
and M := {1, . . . , m}.
2
Basic definitions
Consider a step-up (hu ) and step-down (hd ) procedure
hu (p, α) = i ∈ {1, . . . , m} : pi ≤ max{p(j) : p(j) ≤ τα (j)} ,
hd (p, α) = i ∈ {1, . . . , m} : pi < min{p(j) : p(j) > τα (j)} ,
returning the set of indices of rejected hypotheses (Gandy and Hahn, 2016), where p(1) ≤ p(2) ≤
· · · ≤ p(m) refers to the ordered p-values. Any procedure of the above form is fully characterised
by a threshold function τα : {1, . . . , m} → [0, 1] returning the critical value τα (i) each p(i) is
compared to. A step-up procedure first determines the largest j ∈ M such that the p-value p(j)
lies below τα (j) and then rejects all hypotheses having p-values up to p(j) . Likewise, a step-down
procedure non-rejects all those hypotheses with p-values larger or equal to the smallest p-value
above the threshold function.
We now consider two useful properties of arbitrary multiple testing procedures. The first
one, monotonicity, states that smaller p-values (Tamhane and Liu, 2008) or a higher significance
level (Roth, 1999) lead to more rejections:
Definition 1. A multiple testing procedure h is monotonic if h(p, α) ⊆ h(q, α0 ) for p ≥ q and
α ≤ α0 .
The monotonicity in α introduced by Roth (1999), also called α-consistency (Hommel and
Bretz, 2008), is a natural property desired for any testing procedure since testing at a more
3
stringent significance level should never result in more rejections (Dmitrienko and Tamhane,
2013).
Gandy and Hahn (2016) introduce another useful property, the class of well-behaved multiple
testing procedures. Such procedures, in connection with a generic algorithm presented in Gandy
and Hahn (2016), allow to use p-value estimates obtained with independent samples under
the null to compute test results which are proven to be identical (up to a pre-specified error
probability) to the ones obtained with the unknown p-values. A monotonic multiple testing
procedure h is well-behaved if it additionally satisfies the following condition.
Condition 1.
1. Let p, q ∈ [0, 1]m and α ∈ R. If qi ≤ pi ∀i ∈ h(p, α) and qi ≥ pi ∀i ∈
/ h(p, α),
then h(p, α) = h(q, α).
2. Fix p∗ ∈ [0, 1]m and α∗ ∈ [0, 1]. Then there exists δ > 0 such that p ∈ [0, 1]m , α ∈ [0, 1]
and max(kp − p∗ k, |α − α∗ |) < δ imply h(p, α) = h(p∗ , α∗ ).
Well-behaved procedures stay invariant if rejected (non-rejected) p-values are replaced by
smaller (larger) values. Moreover, well-behaved procedures are constant on a δ-neighbourhood
around fixed inputs p∗ and α∗ .
The level α is a parameter in Condition 1 to account for settings in which α is unknown
a-priori: This can occur, for instance, when the significance level depends on an estimate of
the proportion of true null hypotheses which is often a functional of p (Gandy and Hahn, 2016,
Section 2.2). Condition 1 is a generalisation of (Gandy and Hahn, 2014, Condition 1) which
states the same invariance property for the case that α is a given constant: In this case, h is
solely a function of p and the condition |α − α∗ | < δ in the second part of Condition 1 can be
ignored.
3
Arbitrary multiple testing procedures
We define the union, intersection, difference and the complement of two procedures to be the
equivalent operations on the sets of rejected hypotheses returned by the two procedures. Formally, for two multiple testing procedures h1 and h2 we define
h1 ∪ h2 : [0, 1]m × [0, 1] → P({1, . . . , m}),
h1 ∪ h2 (p, α) := h1 (p, α) ∪ h2 (p, α),
and similarly h1 ∩ h2 , h1 \ h2 and the complement hi (p, α)c := {1, . . . , m} \ hi (p, α), where
i ∈ {1, 2}.
In what follows, we sometimes drop the dependence of h(p, α) on p, on α, or on both
parameters. The following lemma summarises the results.
Lemma 1. Let h1 and h2 be two well-behaved multiple testing procedures.
1. h1 ∪ h2 and h1 ∩ h2 are monotonic and satisfy part 2. of Condition 1.
2. hi (p, α)c and h1 \ h2 are not monotonic, i ∈ {1, 2}.
As well-behaved procedures are also monotonic, the complement or difference of two procedures is also not well-behaved.
Although by Lemma 1, both the union and the intersection are monotonic, they do not necessarily allow to lower the p-values of rejected hypotheses or to increase the p-values of non-rejected
hypotheses (first part of Condition 1) as demonstrated in the following two counterexamples.
4
Example 1. Let p∗ = (0.034, 0.06, 1) and α∗ = 0.1. Let h1 be the Benjamini and Hochberg
(1995) step-up procedure, h2 be the Sidak (1967) step-down procedure and h(p, α) = h1 (p, α) ∩
h2 (p, α). Then h1 (p∗ , α∗ ) = {1, 2}, h2 (p∗ , α∗ ) = {1} and thus 2, 3 ∈
/ h(p∗ , α∗ ). However,
∗
∗
∗
increasing p to q = (0.034, 1, 1) results in h1 (q, α ) = ∅ and thus h(q, α ) = ∅ =
6 h(p∗ , α∗ ).
Example 2. Let p∗ and α∗ be as in Example 1. Let h1 be a step-up procedure which uses
the same threshold function as the (step-down) Sidak (1967) correction, and likewise h2 be a
step-down procedure using the same threshold function as the (step-up) Benjamini and Hochberg
(1995) procedure – using (Gandy and Hahn, 2016, Lemma 3), it is straightforward to show that
both procedures are well-behaved. Let h(p, α) = h1 (p, α) ∪ h2 (p, α). Then h1 (p∗ , α∗ ) = {1},
h2 (p∗ , α∗ ) = ∅ and thus h(p∗ , α∗ ) = {1}. However, decreasing p∗ to q = (0, 0.06, 1) results in
h2 (q, α∗ ) = {1, 2} and thus h(q, α∗ ) = {1, 2} =
6 h(p∗ , α∗ ).
Examples 1 and 2 also demonstrate that both the union and the intersection of a well-behaved
step-up and a well-behaved step-down procedure are not necessarily well-behaved any more.
Although neither the class of well-behaved multiple testing procedures of general form nor
the combination of a well-behaved step-up and a well-behaved step-down procedure is closed
under the four set operations aforementioned, the next section proves that this holds true for
the special classes of well-behaved step-up and step-down procedures individually (when taking
unions and intersections).
4
Step-up and step-down procedures
Gandy and Hahn (2016) show that any step-up or step-down procedure (characterised by its
threshold function τα ) which satisfies the following condition is well-behaved:
Condition 2.
1. τα (i) is non-decreasing in i for each fixed α.
2. τα (i) is continuous in α and non-decreasing in α for each fixed i.
Furthermore, Gandy and Hahn (2016) verify that a large variety of commonly used procedures satisfies Condition 2 and is hence well-behaved, among them the procedures of Bonferroni
(1936), Sidak (1967), Holm (1979), Hochberg (1988) or Benjamini and Hochberg (1995).
Even though (Gandy and Hahn, 2016, Lemma 3) only prove that Condition 2 is sufficient
for a procedure to be well-behaved, the condition is actually also necessary:
Lemma 2. Any well-behaved step-up or step-down procedure satisfies Condition 2.
Consider two step-up procedures hu and h̃u with threshold functions ταu and τ̃αu as well as
two step-down procedures hd and h̃d with threshold functions ταd and τ̃αd .
In the following subsections we separately investigate whether the classes of step-up (stepdown) procedures are closed under each of the four set operations (union, intersection, difference
and complement). Moreover, we investigate whether the subclasses of well-behaved step-up
(step-down) procedures are closed. To this end, by Lemma 2, it suffices to show that the classes
of step-up (step-down) procedures satisfying Condition 2 are closed.
5
0.15
0.10
0.00
0.05
p−values and thresholds
0.15
0.10
0.05
p−values and thresholds
0.00
1
2
3
4
5
6
1
rank of hypothesis
2
3
4
5
6
rank of hypothesis
Figure 1: Combined threshold function (bold) for the computation of the union (left) and the
intersection (right) of the Bonferroni (1936) correction (vertical solid line) and the Hochberg
(1988) (dashed line) procedure. The Bonferroni (1936) correction was applied with significance
level 0.2, the Hochberg (1988) procedure with level 0.1. P-values of rejected (crosses) and
non-rejected (triangles) hypotheses.
4.1
Union
The class of step-up procedures is closed under the union operation: To be precise, if hu and
h̃u are two step-up procedures, their union is computed by another step-up procedure h with
threshold function τα (i) = max(ταu (i), τ̃αu (i)) as visualised in Fig. 1 (left).
This is seen as follows: As ταu (i), τ̃αu (i) ≤ τα (i) for all i ∈ M , all hypotheses rejected by either
u
h or h̃u are also rejected by h, that is hu ∪ h̃u ⊆ h. Likewise, as τα (i) takes precisely one of the
values ταu (i) or τ̃αu (i) for each i ∈ M , any p-value belonging to the non-rejection area of both
procedures hu and h̃u also stays non-rejected in h, hence (hu )c ∩ (h̃u )c ⊆ hc .
Moreover, the subclass of well-behaved step-up procedures is also closed under the union
operation as proven in the following lemma.
Lemma 3. If hu and h̃u are two step-up procedures which satisfy Condition 2 then so does the
union hu ∪ h̃u .
Similarly, the union of two step-down procedures hd and h̃d (having threshold functions
ταd and τ̃αd ) is obtained through another step-down procedure characterised by the threshold
function τα (i) = max(ταd (i), τ̃αd (i)). Since the proof of Lemma 3 does not use any properties of
ταu and τ̃αu other than that both satisfy Condition 2, the maximum of two step-down threshold
functions likewise leads to a threshold function satisfying Condition 2.
4.2
Intersection
Similarly to Section 4.1, the intersection of two step-up procedures hu and h̃u is again a step-up
procedure h, characterised by the new threshold function τα (i) = min(ταu (i), τ̃αu (i)) as visualised
in Fig. 1 (right).
This is seen as follows: As ταu (i), τ̃αu (i) ≥ τα (i) for all i ∈ M , any hypothesis non-rejected by
either procedure hu or h̃u is also non-rejected by h, that is (hu )c ∪ (h̃u )c ⊆ hc . Likewise, as τα (i)
takes precisely one of the values ταu (i) or τ̃αu (i) for each i ∈ M , any p-value in the rejection area
of both procedures remains rejected when tested with h, thus hu ∩ h̃u ⊆ h.
6
Similarly to Lemma 3, the subclass of well-behaved step-up procedures is again closed under
the intersection operation.
Lemma 4. If hu and h̃u are two step-up procedures which satisfy Condition 2 then so does the
intersection hu ∩ h̃u .
The intersection of two step-down procedures hd and h̃d is again obtained with another
step-down procedure using the threshold function τα (i) = min(ταd (i), τ̃αd (i)). Analogously to
Section 4.1, the proof of Lemma 4 does not use any properties of ταu and τ̃αu other than that
both satisfy Condition 2, thus the minimum of two step-down threshold functions again leads
to a threshold function satisfying Condition 2.
4.3
Complement
Whereas the complement is generally neither well-behaved nor monotonic, it can be computed
for step-up and step-down procedures using the following construction.
Let α be a known constant. We re-consider the step-up procedure hu with threshold function
u
τα . Then the step-down procedure hd (1 − p) with threshold function ταd (i) = 1 − ταu (m + 1 − i)
applied to 1 − p (instead of p) computes the complement of hu (p), where 1 − p for p ∈ [0, 1]m is
understood coordinate-wise.
The reasoning behind this is as follows: For any hypothesis with p-value p(i) below ταu (i),
1 − p(i) (having rank m + 1 − i in the sorted sequence of values 1 − p) is above ταd (m + 1 − i) by
construction of ταd . Therefore, all former rejections of hu turn into non-rejections of hd and vice
versa.
Likewise, the complement of a step-down procedure hd with threshold function ταd and constant α is computed by a step-up procedure hu with threshold function ταu (i) = 1 − ταd (m + 1 − i).
Condition 2 is again satisfied:
Lemma 5. Let α be a known constant. If the step-up procedure hu with threshold function ταu
satisfies Condition 2, then so does its step-down complement hd (defined with threshold function
ταd (i) = 1 − ταu (m + 1 − i)).
The requirement that α be a known constant is crucial since ταd is not non-decreasing in α
for a fixed i as required in the second part of Condition 2. However, Lemma 5 is made possible
by the fact that for a given constant α (that is, if h and the threshold function seize to be a
function of α), all the parts in Condition 1 (and likewise, Condition 2) which involve α can be
ignored (see remark at the end of Section 2).
4.4
Difference
Following the notation of Section 3, the difference h1 \ h2 of two procedures h1 and h2 can
equivalently be written as h1 ∩ hc2 using the complement of h2 . If h2 is a step-up procedure, hc2
turns into a step-down procedure (see Section 4.3).
Therefore, in case both h1 and h2 are step-up (step-down) procedures satisfying Condition 2,
Lemma 1 yields that h1 \ h2 is still monotonic but not well-behaved any more. However, if h1 is
a step-down and h2 is a step-up procedure (or vice versa), the results from Section 4.2 apply and
yield that h1 \ h2 a well-behaved step-up/step-down procedure with explicit threshold function.
7
4.5
Example
Suppose we are interested in testing H01 , . . . , H0m for statistical significance while ensuring FDR
control at a pre-specified level 0.05, for instance using the Benjamini and Hochberg (1995) procedure. Additionally, we are interested in only selecting those k ∈ N hypotheses having the lowest
p-values (assuming there are no ties), for instance due to the fact that budget constraints only
allow follow-up studies for k hypotheses. We thus look to construct an intersection procedure
which returns the indices of hypotheses satisfying both requirements simultaneously.
To this end, let h1 be the Benjamini and Hochberg (1995) step-up procedure controlling the
FDR at level 0.05, defined through the threshold function τ 1 (i) = 0.05 · i/m for i ∈ {1, . . . , m}.
Moreover, let h2 be the (step-up) Bonferroni (1936) correction with constant but p-dependant
threshold function τp2 (i) = p(k) for i ∈ {1, . . . , m}, where p(k) denotes the k’th smallest entry
of vector p = (p1 , . . . , pm ). By construction, all rejected hypotheses by h2 are precisely the
ones with the k lowest p-values. Threshold functions τα for which α = α(p) is a function of p
are widely used in practice, for instance when using an estimate of the proportion of true null
hypotheses to correct the level α (see, for instance, Example 1 in Gandy and Hahn (2016)).
Both the Benjamini and Hochberg (1995) procedure h1 and the Bonferroni (1936) correction h2
satisfy Condition 2 and are thus well-behaved.
Following Section 4.2, the step-up procedure h defined through the threshold function τp (i) =
min(τ 1 (i), τp2 (i)) = min(0.05 · i/m, p(k) ) computes h1 ∩ h2 . Moreover, h is well-behaved by
Lemma 4.
Consider the numerical example of 15 ordered p-values (here denoted as p̃) given in Section 3.2 of Benjamini and Hochberg (1995). In agreement with Benjamini and Hochberg (1995),
who test p̃ while controlling the FDR at level 0.05 and observe four rejections (of the first four
hypotheses), h1 applied to p̃ yields h1 (p̃) = {1, 2, 3, 4}. Applying the intersection procedure h
constructed above with k = 3 to p̃ yields h(p̃) = {1, 2, 3}, that is h indeed yields those k = 3
hypotheses having the lowest p-values which are also significant under FDR control at level 0.05.
5
Discussion
This article investigates closure properties of general multiple testing procedures, step-up and
step-down procedures as well as subclasses of (solely) monotonic and well-behaved procedures
under four set operations (union, intersection, complement and difference).
The article shows that for general multiple testing procedures, solely the class of monotonic
procedures is closed under taking the union and intersection. However, the subclass of wellbehaved step-up (step-down) procedures is closed under taking the union and intersection.
The implications of the closure properties proven in this article are threefold: They provide
a tool to construct new procedures of known form and with known properties, they render
theoretical results (Gordon, 2007; Gordon and Salzman, 2008) instantly applicable to a large
class of multiple testing procedures and they allow to combine the benefits of various multiple
testing procedures in practice.
A
Proofs
The appendix contains all proofs sorted by section.
8
A.1
Proofs of Section 3
Proof of Lemma 1. We prove both assertions.
1. Monotonicity. If p ≤ q and α ≤ α0 then h1 (q, α) ⊆ h1 (p, α0 ), h2 (q, α) ⊆ h2 (p, α0 ) and thus
h1 (q, α) ∪ h2 (q, α) ⊆ h1 (p, α0 ) ∪ h2 (p, α0 ) as well as h1 (q, α) ∩ h2 (q, α) ⊆ h1 (p, α0 ) ∩ h2 (p, α0 ).
The second statement of Condition 1. As h1 satisfies Condition 1, there exists δ1 such that
max(kp − p∗ k, |α − α∗ |) < δ1 implies h1 (p, α) = h1 (p∗ , α∗ ). Likewise for h2 with a suitable
δ2 . For δ = min(δ1 , δ2 ) and max(kp − p∗ k, |α − α∗ |) < δ, we have h1 (p, α) = h1 (p∗ , α∗ ) and
h2 (p, α) = h2 (p∗ , α∗ ) and thus h1 ∪ h2 (p, α) = h1 ∪ h2 (p∗ , α∗ ). Likewise for the intersection.
2. Fix α. If q ≤ p then hi (p, α) ⊆ hi (q, α), but hi (p, α)c ⊇ hi (q, α)c for i ∈ {1, 2}. The
complement is thus not monotonic. The operation h1 (p, α) \ h2 (p, α) is equivalent to h1 (p, α) ∩
(h2 (p, α))c and thus also not monotonic.
A.2
Proofs of Section 4
Proof of Lemma 2. Let h be a step-up (step-down) procedure characterised through its threshold
function τα . We now verify Condition 2.
1. We show that τα (i) must be non-decreasing in i for a fixed α. Indeed, suppose τα is
decreasing for some i. Then h cannot be monotonic for all inputs: Assume that m = 2,
p = (0.5, 0.5) and h is of step-up type with τα (1) = 1 and τα (2) = 0. Then h(p) = {1} but
increasing p to q = (1, 0.5) results in h(q) = {2} 6⊆ h(p), thus contradicting monotonicity.
2. We show that τα (i) must also be non-decreasing in α for any fixed i. Indeed, for a fixed
i, suppose τα (i) > τα0 (i) for α < α0 . Then h can again not be monotonic for all inputs: Assume
we test m = 1 hypothesis H01 with p-value p = τα (1) > τα0 (1). Then H01 is rejected at τα (1)
but non-rejected at τα0 (1) even though α < α0 , thus contradicting monotonicity.
3. We show that τα (i) is continuous in α for a fixed i. Let > 0 be given. Fix i and α∗ . We
show continuity of the threshold function at α∗ as α → α∗ .
Case 1: α∗ > α. Then τα∗ (i) ≥ τα (i) by monotonicity. Define p∗ = (0, . . . , 0, p∗i , 1, . . . , 1) for
any p∗i ∈ [0, τα∗ (i)) (i.e., p∗ contains p∗i as ith entry, zeros before and ones after). Since h is wellbehaved it satisfies the second part of Condition 1, hence for the fixed p∗ and α∗ there exists δ > 0
such that for all α and p satisfying |α−α∗ | < δ, kp−p∗ k < δ we have h(p, α) = h(p∗ , α∗ ). Assume
|α − α∗ | < δ. Define p = (0, . . . , 0, p∗i − γ, 1, . . . , 1) for any 0 < γ < min(δ, ). Since |α − α∗ | < δ
and kp − p∗ k = γ < δ, h(p, α) = h(p∗ , α∗ ) by Condition 1: As the ith hypothesis is rejected in
h(p∗ , α∗ ) and hence also in h(p, α), it follows that τα∗ (i) ≥ τα (i) ≥ pi = p∗i − γ. This holds true
for all p∗i ∈ [0, τα∗ (i)), thus τα∗ (i) ≥ τα (i) ≥ τα∗ (i) − γ and hence |τα∗ (i) − τα (i)| ≤ γ < .
Case 2: α∗ ≤ α. Then τα∗ (i) ≤ τα (i). Using p∗ = (0, . . . , 0, p∗i , 1, . . . , 1) with p∗i ∈ (τα∗ (i), 1]
and p = (0, . . . , 0, p∗i + γ, 1, . . . , 1) with 0 < γ < min(δ, ), the same argument as in Case 1 yields
τα∗ (i) ≤ τα (i) < τα∗ (i) + γ.
Proof of Lemma 3. Let h = hu ∪ h̃u be defined through τα (i) = max(ταu (i), τ̃αu (i)). First, h is
monotonic by Lemma 1. We now verify Condition 2.
9
1. The function τα (i) is non-decreasing in i: Suppose w.l.o.g. τα (i) = ταu (i). If ταu (i + 1) ≥
+ 1) then τα (i) = ταu (i) ≤ ταu (i + 1) = τα (i + 1) by definition of τα as the maximum of ταu
and τ̃αu . If ταu (i + 1) < τ̃αu (i + 1) then τα (i) = ταu (i) ≤ ταu (i + 1) < τ̃αu (i + 1) = τα (i + 1).
τ̃αu (i
2. τα is continuous in α as the maximum of two continuous functions (in this case in α) is
continuous. The function τα is also non-decreasing in α: Indeed, fix i, let α ≤ α0 and suppose
w.l.o.g. τα (i) = ταu (i). If ταu0 (i) ≤ τ̃αu0 (i) then τα (i) = ταu (i) ≤ ταu0 (i) ≤ τ̃αu0 (i) = τα0 (i) by definition
of τα as the maximum of ταu and τ̃αu . Otherwise, τα (i) = ταu (i) ≤ ταu0 (i) = τα0 (i).
Proof of Lemma 4. Let h = hu ∩ h̃u be defined through τα (i) = min(ταu (i), τ̃αu (i)). Again, h is
monotonic by Lemma 1. We now verify Condition 2.
1. The function τα (i) is non-decreasing in i: Suppose w.l.o.g. τα (i) = ταu (i). If ταu (i + 1) ≥
τ̃αu (i + 1) then τα (i) = ταu (i) ≤ τ̃αu (i) ≤ τ̃αu (i + 1) = τα (i + 1) by definition of τα as the minimum
of ταu and τ̃αu . If ταu (i + 1) < τ̃αu (i + 1) then τα (i) = ταu (i) ≤ ταu (i + 1) = τα (i + 1).
2. τα is continuous in α as the minimum of two continuous functions (in this case in α) is
continuous. The function τα is also non-decreasing in α: Indeed, fix i, let α ≤ α0 and suppose
w.l.o.g. τα (i) = ταu (i). If ταu0 (i) ≤ τ̃αu0 (i) then τα (i) = ταu (i) ≤ ταu0 (i) = τα0 (i). Otherwise,
τα (i) = ταu (i) ≤ τ̃αu (i) ≤ τ̃αu0 (i) = τα0 (i) (by definition of τα as the minimum).
Proof of Lemma 5. Since ταu (i) is non-decreasing in i, it is immediate to verify that ταd (i) is also
non-decreasing in i. For a given constant α, the second part of Condition 2 can be ignored
as shown in (Gandy and Hahn, 2014, Condition 1) and is hence automatically satisfied (see
Section 2).
References
Benjamini, Y. and Hochberg, Y. (1995). Controlling the false discovery rate: A practical and
powerful approach to multiple testing. J Roy Stat Soc B Met, 57(1):289–300.
Besag, J. and Clifford, P. (1991). Sequential Monte Carlo p-values. Biometrika, 78(2):301–304.
Bonferroni, C. (1936). Teoria statistica delle classi e calcolo delle probabilità. Pubblicazioni del
R Istituto Superiore di Scienze Economiche e Commerciali di Firenze, 8:3–62.
Dmitrienko, A. and Tamhane, A. (2013). General theory of mixture procedures for gatekeeping.
Biom J, 55(3):402–419.
Finner, H. and Strassburger, K. (2002). The partitioning principle: a powerful tool in multiple
decision theory. Ann Stat, 30(4):1194–1213.
Gabriel, K. (1969). Simultaneous Test Procedures – Some Theory of Multiple Comparisons.
Ann Math Statist, 40(1):224–250.
Gandy, A. and Hahn, G. (2014). MMCTest – A Safe Algorithm for Implementing Multiple
Monte Carlo Tests. Scand J Stat, 41(4):1083–1101.
10
Gandy, A. and Hahn, G. (2016). A framework for Monte Carlo based Multiple Testing. Scand
J Stat, 43(4):1046–1063.
Gordon, A. (2007). Unimprovability of the Bonferroni procedure in the class of general step-up
multiple testing procedures. Stat Probab Lett, 77(2):117–122.
Gordon, A. and Salzman, P. (2008). Optimality of the Holm procedure among general step-down
multiple testing procedures. Stat Probab Lett, 78(13):1878–1884.
Guo, W. and Peddada, S. (2008). Adaptive choice of the number of bootstrap samples in large
scale multiple testing. Stat Appl Genet Mol Biol, 7(1):1–16.
Hochberg, Y. (1988). A sharper Bonferroni procedure for multiple tests of significance.
Biometrika, 75(4):800–802.
Hochberg, Y. and Tamhane, A. (2008). Multiple Comparison Procedures. Wiley.
Holm, S. (1979). A simple sequentially rejective multiple test procedure. Scand J Stat, 6(2):65–
70.
Hommel, G. and Bretz, F. (2008). Aesthetics and power considerations in multiple testing – a
contradiction? Biom J, 50(5):657–666.
Hommel, G., Bretz, F., and Maurer, W. (2007). Powerful short-cuts for multiple testing procedures with special reference to gatekeeping strategies. Stat Med, 26(22):4063–4073.
Hsu, J. (1996). Multiple Comparisons: Theory and Methods. Chapman and Hall/CRC.
Lin, D. (2005). An efficient Monte Carlo approach to assessing statistical significance in genomic
studies. Bioinformatics, 21(6):781–787.
Marcus, R., Peritz, E., and Gabriel, K. (1976). On closed testing procedures with special
reference to ordered analysis of variance. Biometrika, 63(3):655–660.
Romano, J. and Shaikh, A. (2006). Stepup procedures for control of generalizations of the
familywise error rate. Ann Stat, 34(4):1850–1873.
Romano, J. and Wolf, M. (2005). Exact and Approximate Stepdown Methods for Multiple
Hypothesis Testing. J Am Stat Assoc, 100(469):94–108.
Roth, A. (1999). Multiple comparison procedures for discrete test statistics. J Stat Plan Infer,
82(1-2):101–117.
Sandve, G., Ferkingstad, E., and Nygard, S. (2011). Sequential Monte Carlo multiple testing.
Bioinformatics, 27(23):3235–3241.
Sidak, Z. (1967). Rectangular confidence regions for the means of multivariate normal distributions. J Am Stat Assoc, 62(318):626–633.
Tamhane, A. and Liu, L. (2008). On weighted Hochberg procedures. Biometrika, 95(2):279–294.
van Wieringen, W., van de Wiel, M., and van der Vaart, A. (2008). A test for partial differential
expression. J Am Stat Assoc, 103(483):1039–1049.
Westfall, P. and Young, S. (1993). Resampling-based multiple testing: Examples and methods
for p-value adjustment. Wiley.
11
| 10math.ST
|
Integrating Research Data Management into
Geographical Information Systems
Christian T. Jacobs, Alexandros Avdis,
Simon L. Mouradian, and Matthew D. Piggott
arXiv:1509.04729v1 [cs.DL] 15 Sep 2015
Department of Earth Science and Engineering, South Kensington Campus,
Imperial College London, London SW7 2AZ, United Kingdom
{c.jacobs10,a.avdis,simon.mouradian06,m.d.piggott}@imperial.ac.uk
http://www.imperial.ac.uk/engineering/departments/earth-science
Abstract. Ocean modelling requires the production of high-fidelity computational meshes upon which to solve the equations of motion. The
production of such meshes by hand is often infeasible, considering the
complexity of the bathymetry and coastlines. The use of Geographical
Information Systems (GIS) is therefore a key component to discretising
the region of interest and producing a mesh appropriate to resolve the
dynamics. However, all data associated with the production of a mesh
must be provided in order to contribute to the overall recomputability
of the subsequent simulation. This work presents the integration of research data management in QMesh, a tool for generating meshes using
GIS. The tool uses the PyRDM library to provide a quick and easy way
for scientists to publish meshes, and all data required to regenerate them,
to persistent online repositories. These repositories are assigned unique
identifiers to enable proper citation of the meshes in journal articles.
Keywords: Geographical Information Systems, Research Data Management, Digital Curation, Reproducibility, Digital Object Identifier,
Online Repositories
1
Introduction
Computer simulations of ocean dynamics are becoming ever more important to
predict the effects of global-scale hazards such as tsunamis [13], the influence of
marine renewable energy turbines on sediment transport [20], and the dispersal
range of nuclear contaminants [6], to name just a few applications. The underlying numerical model behind such simulations often requires a mesh upon which
the equations describing the flow dynamics are solved, thereby transitioning from
a continuous description of the region of interest (also known as the domain) to a
discrete one. An example focussing on the area around the Orkney and Shetland
Isles is shown in Figure 1. A mesh for ocean simulations must be of high enough
quality to resolve the intricate coastlines and bathymetry [12]. However, creating
such a mesh manually is infeasible for large-scale, high-resolution simulations.
Geographical Information Systems (GIS) offer an effective way of processing
bathymetry and coastline data to create a geometry with which to work [19].
2
Data Management in Geographical Information Systems
Fig. 1. An example of an unstructured computational mesh which discretises the marine area around the North-East coast of Scotland. The resolution is highest around
the Scottish coastline and around the Orkney and Shetland Isles.
A method of producing a computational mesh from this geometry is then required to perform a simulation on it. QMesh [2] is a software package currently
being developed at Imperial College London for this purpose. QMesh reads in a
geometry defined in the QGIS Geographical Information System software [21],
and then converts the geometry into a readable format for the Gmsh mesh generation software [11], which in turn generates the mesh to provide a discrete
representation of the domain. Ocean simulations may then be performed with a
computational fluid dynamics package.
Publications that are dependant on numerical simulation often provide details of the simulation setups to improve reproducibility and indeed recomputability. However, while a description of the domain may also be given, the
mesh that discretises this domain is rarely provided as a supplementary material.
This lack of data availability has also been highlighted in many other areas of
science [27], [1], [26]. Furthermore, citations to the software used to produce the
mesh typically only refer to a generic user manual and contain no information
about which version was used. For the purpose of recomputability and reproducibility, it is crucial that researchers provide all the data files, as well as the
precise version of the software’s source code used to produce the output in the
first place [8], [5]. In the case of this work, the input data is the geographical
information defining the domain, the output data is the computational mesh,
and the software is QMesh (and its dependencies).
Data Management in Geographical Information Systems
3
Despite the need for a more open research environment where software and
datasets are shared freely, the level of motivation amongst researchers to do
this is generally quite low. This is in part due to the extra effort and time
required to gather and publish the data [18], whilst typically gaining little from
the process. To encourage the sharing of data and improve its reproducibility
and recomputability, it is therefore important to make the publication process
more straight forward and swift. This can be effected by the development of
research data management tools that readily capture the datasets involved and
information about the software being used [25], [18].
This paper describes the integration of a research data management tool,
which uses the PyRDM library [14], into the QMesh software. The tool automates the publication of the QMesh source code, as well as the input and
output data for a specified QGIS project, to online, citable and persistent repositories such as those provided by Figshare (figshare.com), Zenodo (zenodo.org)
and DSpace-based (dspace.org) hosting services. The tool has both a command
line and a graphical user interface, and allows users to publish the software and
data at the ‘push of a button’, thereby facilitating sharing and a more open
research environment. In contrast to other software tools that also facilitate
the publication of code and datasets, such as Fidgit [24], rfigshare [4], and dvn
[17], the QMesh publishing tool incorporates application-specific knowledge to
provide a greater amount of automation. For example, the tool is able to parse
QGIS project files to automatically determine the relevant input data to publish,
rather than the user having to specify the data files manually. Furthermore, this
work represents a novel application of research data management and curation
software within a GIS environment.
Section 2 describes in greater detail the extensions made to the QMesh software to automate the publication process for the software itself, the input files
(for a given QGIS project) and any output files (i.e. the computational mesh).
Section 3 presents a realistic example of a scientific workflow involving production of a mesh of a UK coastal region. The data files are read in to QGIS and
a mesh is produced. Both the QGIS data and mesh are subsequently published
to an online repository provided by Figshare, and a DOI is assigned which can
be used to properly cite the data in journal articles. Finally, some concluding
remarks are made in Section 4.
2
Integration with QMesh
QMesh features a command line interface (CLI), as well as a graphical user interface (GUI) via a QGIS plugin through which users can select relevant geometry
objects and produce a mesh. The integration of research data management techniques into QMesh was achieved by adding a PyRDM-based publishing tool to
both of these interfaces.
The tool provides the option of publishing the QMesh software source code
and data required to reproduce the mesh to separate online repositories. Users
are presented with a simple interface and only have to provide a minimal amount
4
Data Management in Geographical Information Systems
of information; this is illustrated in Figure 2. The publication process itself is
handled by the PyRDM library [14] which communicates with an online repository hosting service via its Application Programming Interface (API). The publication process results in a Digital Object Identifier (DOI) [7] being assigned to
the repository, with which users can properly cite their research outputs.
Fig. 2. The QMesh publisher tool, which is part of the QMesh QGIS plugin. Users
choose the online repository service that they wish to use; by default this is set to
Figshare. In addition to the input data files associated with the QGIS project, users
may also publish the output data file (i.e. the resulting computational mesh) produced
by QMesh, if they so desire. By default, the publication is made public unless the
user decides otherwise; in the case of private publication, a DOI is still assigned to the
repository, but will not be made active/‘live’ until the repository is made public.
The publication of data is handled separately to the publication of the QMesh
software. In the former case, when a suitable mesh has been produced and is
ready to be published, users simply have to provide the QMesh publishing tool
with the location of the QGIS project file on the computer’s file system when
using the CLI. When using the GUI, this location is provided automatically when
the project is opened in QGIS. The tool then searches for the <datasource> tags
in the XML-based project file to determine the location of all the files that the
project comprises; these may include shape files that define various layers in the
Data Management in Geographical Information Systems
5
geometry, data files in NetCDF format [23] which define the bathymetry of the
ocean, and a multitude of other data formats. Optionally, the location of the
Gmsh mesh file may also be provided, thereby publishing the resultant output
data along with the files required to produce it. The locations of all these data
files, including the QGIS project file itself, are then provided to PyRDM which
automatically creates a repository on the hosting service and uploads the files
via the service’s API. The service then returns a publication ID and a DOI,
which is presented to the user for citation purposes. This process is illustrated
in Figure 3.
The publication of software involves a similar process, but can currently only
be accomplished via the CLI. The user only has to provide the QMesh publishing
tool with the location of the software’s source code on the computer’s file system.
The PyRDM library then handles the rest; it determines the exact version of
QMesh currently in use using the Git version control system (git-scm.com) [22],
and then checks to see whether that version has been published already1 . If it
has, PyRDM retrieves the existing DOI for re-use. If it has not, then PyRDM
publishes the source code in a similar fashion to the case of publishing data, as
shown in Figure 3. Note that publications in journals would need to reference
both the software repository’s DOI and the data repository’s DOI. There is
currently no explicit link that is made between the software and data repositories,
unless specified manually.
As demonstrated by Figure 3, the QMesh publishing tool requires minimal
user interaction and is largely automated by the PyRDM library. This is important for encouraging the sharing of software and data files, in order to achieve a
more open research environment.
3
Workflow Example
To demonstrate an example of a scientific workflow involving mesh generation
using GIS, the Orkney and Shetland Isles considered in [2] and [3] are used.
The researcher first has to describe the geography of the domain in QGIS and
then decide on the area they wish to create a mesh for. The QGIS project for the
Orkney and Shetland Isles comprises a number of geometrical layers which define
the coastlines (and potentially coastal engineering structures such as marine
power turbines), in addition to a NetCDF file which defines the bathymetry of
the ocean floor, and another NetCDF file which defines the desired resolution
throughout the mesh. These files are shown in Figure 4 beside the area that will
be meshed.
The mesh that QMesh produces for this domain (shown in Figure 1) is then
used by the researcher in their marine simulations. Once the researcher is ready
to publish their results, they upload the data files associated with the production
1
Repository searching is only available when using the Figshare repository service,
due to API limitations explained later in Section 4. PyRDM will publish the software
regardless of whether it has been published before when Zenodo or a DSpace-based
service is chosen.
6
Data Management in Geographical Information Systems
Fig. 3. The processes behind publishing the QGIS data files (left) and QMesh software
source code (right) to Figshare.
Fig. 4. Screenshot of the UK region visualised in QGIS. The solid dark purple line
defines the area that will be meshed (in this case it contains the Orkney and Shetland
Isles). The different files that make up the layers of the geometry are specified in the
column on the left-hand side.
Data Management in Geographical Information Systems
7
of the simulation’s mesh to an online repository using the QMesh publishing tool
shown in Figure 2 (the CLI may also be used instead of the graphical interface).
In this example, it uploads all the files previously mentioned to Figshare. Once
uploaded, the files can be downloaded from the Figshare website (see Figure 5)
and a DOI is presented to the researcher to share with colleagues and for use in
journal publications (see Figure 6).
Fig. 5. A screenshot of the resulting repository on the Figshare website, with the files
readily available to download. The QMesh publishing tool automatically assigns a title
and tags to the repository based on the QGIS project’s name.
The version of the QMesh software’s source code that is used should also
be published, in a separate repository to the data. However, it should be noted
that publishing the QMesh source code may not be enough to reproduce the
exact same mesh without also knowing the versions of its dependencies. For
example, different versions of Gmsh may produce slightly different meshes as a
result of algorithmic improvements within the software. It is therefore important
that such information be recorded in some way to further improve the degree of
reproducibility. For example, ideally Gmsh would also have a similar system for
publishing the current version of its source code in use.
8
Data Management in Geographical Information Systems
Fig. 6. A Figshare publication ID and a DOI are assigned to each repository, and
presented to the researcher once the publication process is complete.
4
Discussion and Conclusions
Throughout the production of the PyRDM-based publishing tool for QMesh,
several issues were encountered which largely stemmed from a lack of standardisation and support in the repository hosting services’ APIs. For example, in
order for PyRDM to attribute authors to the software repository on Figshare,
all authors of QMesh must provide their Figshare author IDs in the AUTHORS
file that is part of the QMesh source code. Unfortunately, another different set
of author IDs would need to be provided when using a different repository service such as Zenodo, which is inconvenient and requires all authors of QMesh
to have accounts across all the supported services. A more standardised way of
identifying and attributing authors to research software and data would be to
use ORCID (orcid.org) researcher IDs. Figshare has recently added support for
authenticating with ORCID IDs via its web interface [9], and it is hoped that
ORCID authentication via the Figshare API will also be added for the benefit
of PyRDM. Another example, this time involving lack of API support, is the
current inability to search for an existing repository with the Zenodo API. Further developments are necessary in this area to enrich the publication process
and improve automation.
The production of meshes can involve proprietary and/or private data which
cannot be published openly, but at the same time sharing all research output
is becoming a common requirement imposed by research funders. The QMesh
publishing tool comes with the option of publishing the data to private repositories. However, with some services the private storage space is rather limited,
and typically not large enough to store high quality mesh files for realistic ocean
simulations. For example, the free private storage space offered by Figshare is 1
GB at the time of writing this paper, with a 250 MB individual file size limit2 .
Furthermore, only a maximum of 5 collaborators can be given access to a private
repository. In contrast, the integration of Figshare for Institutions [10] offers a
more suitable platform for larger-scale research data management. This project
enables researchers at an institution to publish to private repositories hosted
in the cloud. This is considerably more sustainable for GIS projects and mesh
2
http://figshare.com/pricing
Data Management in Geographical Information Systems
9
generation that can involve very large file sizes, both public and private data,
and collaboration amongst many researchers and research groups.
In conclusion, the integration of a publishing tool in a Geographical Information System has helped to mitigate one of the reasons why researchers tend not
to publish their software and data; that is, it is time-consuming to do so with
little reward. The new QMesh publishing tool makes publishing a computational
mesh and associated data files easy and largely effortless through the addition
of a significant amount of automation. Furthermore, the use of online repository
services enable more formal citation of all research outputs through the use of
DOIs. However, it is the responsibility of the scientific community to encourage
and provide incentives for the openness and public availability of this software
and data, in order to overcome the barrier of lack of motivation to publish.
Acknowledgments CTJ was funded by an internal grant entitled “Research
data management: Where software meets data” from the Research Office at
Imperial College London. Part of the work presented in this paper is based
on work first presented in poster form at the International Digital Curation
Conference (IDCC) in February 2015 [16], and in a PyRDM project report [15].
The authors would like to thank the two anonymous reviewers of this paper for
their feedback.
References
1. Alsheikh-Ali, A.A., Qureshi, W., Al-Mallah, M.H., Ioannidis, J.P.A.: Public Availability of Published Research Data in High-Impact Journals. PLoS ONE 6(9),
e24357 (2011)
2. Avdis, A., Hill, J., Jacobs, C.T., Kramer, S.C., Candy, A.S., Gorman, G.J., Piggott, M.D.: Efficient unstructured mesh generation for renewable tidal energy using
Geographical Information Systems (In Preparation)
3. Avdis, A., Jacobs, C.T., Hill, J., Piggott, M.D., Gorman, G.J.: Shoreline and
Bathymetry Approximation in Mesh Generation for Tidal Renewable Simulations.
In: Proceedings of the 11th European Wave and Tidal Energy Conference (Accepted)
4. Boettiger, C., Chamberlain, S., Ram, K., Hart, E.: rfigshare: an R interface to
figshare.com. (2014), http://CRAN.R-project.org/package=rfigshare, r package
version 0.3-1
5. Buckheit, J.B., Donoho, D.L.: WaveLab and Reproducible Research. In: Antoniadis, A., Oppenheim, G. (eds.) Wavelets and Statistics, Lecture Notes in Statistics, vol. 103, pp. 55–81. Springer, New York (1995)
6. Choi, Y., Kida, S., Takahashi, K.: The impact of oceanic circulation and phase
transfer on the dispersion of radionuclides released from the Fukushima Dai-ichi
Nuclear Power Plant. Biogeosciences 10, 4911–4925 (2013)
7. Davidson, L.A., Douglas, K.: Digital Object Identifiers: Promise and Problems for
Scholarly Publishing. Journal of Electronic Publishing 4(2) (1998)
8. de Leeuw, J.: Reproducible Research: the Bottom Line. Department of Statistics Papers, University of California (2001), http://escholarship.org/uc/item/
9050x4r4
10
Data Management in Geographical Information Systems
9. Figshare: figshare ORCID integration. Figshare blog, http://figshare.com/blog
(2013)
10. Figshare: Loughborough University, figshare, Arkivum and Symplectic announce pioneering research data management solution. Figshare blog,
http://figshare.com/blog (2014)
11. Geuzaine, C., Remacle, J.F.: Gmsh: A 3-D finite element mesh generator with builtin pre- and post-processing facilities. International Journal for Numerical Methods
in Engineering 79(11), 1309–1331 (2009)
12. Gorman, G.J., Piggott, M.D., Wells, M.R., Pain, C.C., Allison, P.A.: A systematic
approach to unstructured mesh generation for ocean modelling using GMT and
Terreno. Computers & Geosciences 34(12), 1721–1731 (2008)
13. Hill, J., Collins, G.S., Avdis, A., Kramer, S.C., Piggott, M.D.: How does multiscale
modelling and inclusion of realistic palaeobathymetry affect numerical simulation
of the Storegga Slide tsunami. Ocean Modelling 83, 11–25 (2014)
14. Jacobs, C.T., Avdis, A., Gorman, G.J., Piggott, M.D.: PyRDM: A Python-based
library for automating the management and online publication of scientific software
and data. Journal of Open Research Software 2(1), e28 (2014)
15. Jacobs, C.T., Avdis, A., Gorman, G.J., Piggott, M.D.: RDM Green Shoots Project
Report: Research data management: Where software meets data (2014), http:
//dx.doi.org/10.6084/m9.figshare.1269127
16. Jacobs, C.T., Avdis, A., Gorman, G.J., Piggott, M.D.: PyRDM: A library to facilitate the automated publication of software and data in computational science.
Poster presentation at the 10th International Digital Curation Conference (2015),
http://dx.doi.org/10.6084/m9.figshare.1318710
17. Leeper, T.J.: Archiving Reproducible Research with R and Dataverse. The R Journal 6(1) (2014), http://journal.r-project.org/archive/2014-1/leeper.pdf
18. LeVeque, R.J., Mitchell, I.M., Stodden, V.: Reproducible Research for Scientific
Computing: Tools and Strategies for Changing the Culture. Computing in Science
& Engineering 14(4), 13–17 (2012)
19. Li, R.: Data Models for Marine and Coastal Geographic Information Systems,
chap. 3. CRC Press (2000)
20. Martin-Short, R., Hill, J., Kramer, S.C., Avdis, A., Allison, P.A., Piggott, M.D.:
Tidal resource extraction in the Pentland Firth, UK: potential impacts on flow
regime and sediment transport in the Inner Sound of Stroma. Renewable Energy
76, 596–607 (2015)
21. QGIS Development Team: QGIS Geographic Information System. Open Source
Geospatial Foundation (2009), http://qgis.osgeo.org
22. Ram, K.: Git can facilitate greater reproducibility and increased transparency in
science. Source Code for Biology and Medicine 8(7) (2013)
23. Rew, R.K., Davis, G.P.: NetCDF: an interface for scientific data access. IEEE
Computer Graphics and Applications 10(4), 76–82 (1990)
24. Smith, A.: Fidgit - DOIs for code. figshare (2013), http://dx.doi.org/10.6084/
m9.figshare.828487
25. Stodden, V., Bailey, D., Borwein, J., LeVeque, R.J., Rider, W., Stein, W.: Setting the Default to Reproducible: Reproducibility in Computational and Experimental Mathematics. Tech. rep., Institute for Computational and Experimental Research in Mathematics (ICERM) (2013), http://www.davidhbailey.com/
dhbpapers/icerm-report.pdf
26. Vines, T.H., Andrew, R.L., Bock, D.G., Franklin, M.T., Gilbert, K.J., Kane, N.C.,
Moore, J.S., Moyers, B.T., Renaut, S., Rennison, D.J., Veen, T., Yeaman, S.: Man-
Data Management in Geographical Information Systems
11
dated data archiving greatly improves access to research data. The FASEB Journal
27(4), 1304–1308 (2013)
27. Whitlock, M.C., McPeek, M.A., Rausher, M.D., Rieseberg, L., Moore, A.J.: Data
Archiving. The American Naturalist 175(2), 145–146 (2010)
| 5cs.CE
|
Approximating Directed Steiner Problems via Tree Embedding
Bundit Laekhanukit∗.†
arXiv:1511.06559v3 [cs.DS] 29 Feb 2016
March 1, 2016
Abstract
Directed Steiner problems are fundamental problems in Combinatorial Optimization and
Theoretical Computer Science. An important problem in this genre is the k-edge connected
directed Steiner tree (k-DST) problem. In this problem, we are given a directed graph G on n
vertices with edge-costs, a root vertex r, a set of h terminals T and an integer k. The goal is
to find a min-cost subgraph H ⊆ G that connects r to each terminal t ∈ T by k edge-disjoint
r, t-paths. This problem includes as special cases the well-known directed Steiner tree (DST)
problem (the case k = 1) and the group Steiner tree (GST) problem. Despite having been
studied and mentioned many times in literature, e.g., by Feldman et al. [SODA’09, JCSS’12],
by Cheriyan et al. [SODA’12, TALG’14] and by Laekhanukit [SODA’14], there was no known
non-trivial approximation algorithm for k-DST for k ≥ 2 even in the special case that an input
graph is directed acyclic and has a constant number of layers. If an input graph is not acyclic,
the complexity status of k-DST is not known even for a very strict special case that k = 2 and
|T | = 2.
In this paper, we make a progress toward developing a non-trivial approximation algorithm
for k-DST. We present an O(D · k D−1 · log n)-approximation algorithm for k-DST on directed
acyclic graphs (DAGs) with D layers, which can be extended to a special case of k-DST on “general graphs” when an instance has a D-shallow optimal solution, i.e., there exist k edge-disjoint
r, t-paths, each of length at most D, for every terminal t ∈ T . For the case k = 1 (DST), our algorithm yields an approximation ratio of O(D log h), thus implying an O(log3 h)-approximation
algorithm for DST that runs in quasi-polynomial-time (due to the height-reduction of Zelikovsky
[Algorithmica’97]). Our algorithm is based on an LP-formulation that allows us to embed a solution to a tree-instance of GST, which does not preserve connectivity. We show, however, that
one can randomly extract a solution of k-DST from the tree-instance of GST.
Our algorithm is almost tight when k, D are constants since the case that k = 1 and D = 3
is NP-hard to approximate to within a factor of O(log h), and our algorithm archives the same
approximation ratio for this special case. We also remark that the k 1/4−ǫ -hardness instance
of k-DST is a DAG with 6 layers, and our algorithm gives O(k 5 log n)-approximation for this
special case. Consequently, as our algorithm works for general graphs, we obtain an O(D ·k D−1 ·
log n)-approximation algorithm for a D-shallow instance of the k edge-connected directed Steiner
subgraph problem, where we wish to connect every pair of terminals by k edge-disjoint paths.
∗
The Weizmann Institute of Science, Israel, email:[email protected].
The work was partly done while the author was at McGill University, Simons Institute for the Theory of Computing and the Swiss AI Lab IDSIA. Partially supported by the ERC Starting Grant NEWNET 279352 and by Swiss
National Science Foundation project 200020 144491/1.
†
1
1
Introduction
Network design is an important class of problems in Combinatorial Optimization and Theoretical
Computer Science as it formulates scenarios that appear in practical settings. In particular, we
might wish to design an overlay network that connects a server to clients, and this can be formulated
as the Steiner tree problem. In a more general setting, we might have an additional constraint that
the network must be able to function after link or node failures, leading to the formulation of the
survivable network design problem. These problems are well-studied in symmetric case where a
network can be represented by an undirected graph. However, in many practical settings, links in
networks are not symmetric. For example, we might have different upload and download bandwidths
in each connection, and sometimes, transmissions are only allowed in one direction. This motivates
the study of network design problems in directed graphs, in particular, directed Steiner problems.
One of the most well-known directed network design problem is the directed Steiner tree problem (DST), which asks to find a minimum-cost subgraph that connects a given root vertex to each
terminal. DST is a notorious problem as there is no known polynomial-time algorithm that gives
an approximation ratio better than polynomial. A polylogarithmic approximation can be obtained
only when an algorithm is allowed to run in quasi-polynomial-time [CCC+ 99, Rot11, FKK+ 14]. A
natural generalization of DST, namely, the k edge-connected directed Steiner tree (k-DST) problem,
where we wish to connect a root vertex to each terminal by k edge-disjoint paths, is even more mysterious as there is no known non-trivial approximation algorithm, despite having been studied and
mentioned many times in literature, e.g., by Feldman et al. [FKN12], by Cheriyan et al. [CLNV14]
and by Laekhanukit [Lae14].
The focus of this paper is in studying the approximability of k-DST. Let us formally describe
k-DST. In k-DST, we are given a directed graph G with edge-costs {ce }e∈E(G) , a root vertex r
and a set of terminals T ⊆ V (G). The goal is to find a min-cost subgraph H ⊆ G such that H
has a k edge-disjoint directed r, t-paths from the root r to each terminal t ∈ T . Thus, removing
any k − 1 edges from H leaves at least one path from the root r to each terminal t ∈ T , and
DST is the case when k = 1 (i.e., we need only one path). The complexity status of k-DST tends
to be negative. It was shown by Cheriyan et al. [CLNV14] that the problem is at least as hard
1−ǫ
as the label cover problem. Specifically, k-DST admits no 2log n -approximation, for any ǫ > 0,
unless NP ⊆ DTIME(2polylog(n) ). Laekhanukit [Lae14], subsequently, showed that k-DST admits
no k1/4−ǫ -approximation unless NP = ZPP. The integrality gap of a natural LP-relaxation for kDST is Ω(k/ log k) which holds even for a special case of connectivity-augmentation where we wish
to increase a connectivity of a graph by one. All the lower bound results are based on the same
construction which are directed acyclic graphs (DAGs) with diameter 5, i.e., any path in an input
graph has length (number of edges) at most 5 (we may also say that it has 6 layers). Even for a very
simple variant of k-DST, namely (1, 2)-DST, where we have two terminals, one terminal requires
one path from the root and another terminal requires 2 edge-disjoint paths, it was not known
whether the problem is NP-hard or polynomial-time solvable. To date, the only known positive
result for k-DST is an O(nkh )-time (exact) algorithm for k-DST on DAGs [CLNV14], which thus
runs in polynomial-time when kh is constant, and a folk-lore |T |-approximation algorithm, which
can be obtained by computing min-cost k-flow for |T | times, one from the root r to each terminal
t and then returning the union as a solution. We emphasize that there was no known non-trivial
approximation algorithm even when an input graph is “directed acyclic” and has “constant number
of layers”. Also, in contrast to DST, in which an O(2|T | poly(n))-time (exact) algorithm exists for
2
general graphs, it is not known whether k-DST for k = 2 and |T | = 2 is polynomial-time solvable
if an input graph is not acyclic. This leaves challenging questions whether ones can design a nontrivial approximation algorithm for k-DST on DAGs with at most D layers, and whether ones can
design a non-trivial approximation algorithm when an input graph is not acyclic.
In this paper, we make a progress toward developing a non-trivial approximation algorithm for
k-DST. We present the first “non-trivial” approximation algorithm for k-DST on DAGs with D
layers that achieves an approximation ratio of O(D · kD−1 · log n). Our algorithm can be extended
to a special case of k-DST on “general graphs” where an instance has a D-shallow optimal solution,
i.e., there exist k edge-disjoint r, t-paths, each of length at most D, for every terminal t ∈ T . Consequently, as our algorithm works for a general graph, we obtain an O(D · kD−1 · log n)-approximation
algorithm for a D-shallow instance of the k edge-connected directed Steiner subgraph problem, where
we wish to connect every pair of terminals by k edge-disjoint paths, i.e., the set of terminal T is
required to be k-edge connected in the solution subgraph (there is no root vertex in this problem).
Our algorithm is almost tight when k and D are constants because the case that k = 1 and
D = 3 is essentially the set cover problem, which is NP-hard to approximate to within a factor
of O(log h) [LY94, Fei98], and our algorithm achieves the same approximation ratio. We also
remark that the k1/4−ǫ -hardness instance of k-DST is a DAG with 6 layers, and our algorithm gives
O(k5 log n)-approximation for this special case. For k = 1, we obtain a slightly better bound of
O(D log h), thus giving an LP-based O(log3 h)-approximation algorithm for DST as a by product.
The key idea of our algorithm is to formulate an LP-relaxation with a special property that a
fractional solution can be embedded into a tree instance of the group Steiner tree problem (GST).
Thus, we can apply the GKR Rounding algorithm in [GKR00] for GST on trees to round the
fractional solution. However, embedding of an LP-solution to a tree instance of GST does not
preserve connectivity. Also, it does not lead to a reduction from k-DST to the k edge-connected
variant of GST, namely, k-GST. Hence, our algorithm is, although simple, not straightforward.
1.1
Our Results
Our main result is an O(D · kD−1 · log n)-approximation algorithm for k-DST on a D-shallow
instance, which includes a special case that an input graph is directed acyclic and has at most D
layers.
Theorem 1. Consider the k edge-connected directed Steiner tree problem. Suppose an input instance has an optimal solution H ∗ in which, for every terminal t ∈ T , H ∗ has k edge-disjoint
r, t-paths such that each path has length at most D. Then there exists an O(D · kD−1 · log n)approximation algorithm. In particular, there is an O(D · kD−1 · log n)-approximation algorithm for
k-DST on a directed acyclic graph with D layers.
For the case k = 1, our algorithm yields a slightly better guarantee of O(D log h). Thus, we
have as by product an LP-based approximation algorithm for DST. Applying Zelikovsky’s heightreduction theorem [Zel97, HRZ01], this implies an LP-based quasi-polynomial-time O(log3 h)approximation algorithm for DST. (The algorithm runs in time O(poly(nD ) and has approximation
ratio O(h1/D · D 2 log h).)
Theorem 2 also implies an algorithm of the same (asymptotic) approximation ratio for a Dshallow instance of the k edge-connected directed Steiner subgraph problem, where we wish to find
a subgraph H such that the set of terminal T is k-edge-connected in H. To see this, we invoke
the algorithm in Theorem 2 as follows. Take any terminal t∗ ∈ T as a root vertex of a k-DST
3
instance. Then we apply the algorithm for k-DST to find a subgraph H out such that every terminal
is k edge-connected from t∗ . We apply the algorithm again to find a subgraph H in such that every
terminal is k edge-connected to t∗ . Then the set of terminal T is k-edge connected in the graph
H out ∪ H in by transitivity of edge-connectivity. The cost incurred by this algorithm is at most
twice that of the algorithm in Theorem 2. Thus, we have the following theorem as a corollary of
Theorem 2
Theorem 2. Consider the k edge-connected directed Steiner subgraph problem. Suppose an input
instance has an optimal solution H ∗ in which, for every pair of terminals s, t ∈ T , H ∗ has k edgedisjoint s, t-paths such that each path has length at most D. Then there exists an O(D ·kD−1 ·log n)approximation algorithm.
Overview of our algorithm The key idea of our algorithm is to embed an LP solution for
k-DST to a standard LP of GST on a tree. (We emphasize that we embed the LP solution of
k-DST to that of GST not k-GST.) At first glance, a reduction from k-DST to GST on trees is
unlikely to exist because any such reduction would destroy all the connectivity information. We
show, however, that such tree-embedding exists, but we have to sacrifice running-time and cost to
obtain such embedding.
The reduction is indeed the same as a folk-lore reduction from DST to GST on trees. That is, we
list all rooted-paths (paths that start from the root vertex) of length at most D in an input graph
and form a suffix tree. In the case of DST, if there is an optimal solution which is a tree of height
D, then it gives an approximation preserving reduction from GST to DST which blows up the size
(and thus the running time) of the instance to O(nD ). Unfortunately, for the case of k-DST with
k > 1, this reduction does not give an equivalent reduction from k-DST to k-GST on trees. The
reduction is valid in one direction, i.e., any feasible solution to k-DST has a corresponding feasible
solution to the tree-instance of k-GST. However, the converse is not true as a feasible solution to
the tree-instance of k-GST might not give a feasible solution to k-DST. Thus, our reduction is
indeed an “invalid” reduction from k-DST to a tree instance of “GST” (the case k = 1).
To circumvent this problem, we formulate an LP that provides a connection between an LP
solution on an input k-DST instance and an LP solution of a tree-instance of GST. Thus, we can
embed an LP solution to an LP-solution of GST on a (very large) tree. We then round the LP
solution using the GKR Rounding algorithm for GST on trees [GKR00]. This algorithm, again,
does not give a feasible solution to k-DST as each integral solution we obtain only has “connectivity
one” and thus is only feasible to DST. We cope with this issue by using a technique developed by
Chalermsook et al. in [CGL15]. Specifically, we sample a sufficiently large number of DST solutions
and show that the union of all these solutions is feasible to k-DST using cut-arguments.
Each step of our algorithm and the proofs are mostly standard, but ones need to be careful in
combining each step. Otherwise, the resulting graph would not be feasible to k-DST.
Organization. We provide definitions and notations in Section 2. We start our discussion by
presenting a reduction from DST to GST in Section 3. Then we discuss properties of minimal
solutions for k-DST in Section 4. We present standard LPs for k-DST and GST in Section 5
and formulate a stronger LP-relaxation for k-DST in Section 6. Then we proceed to present our
algorithm in Section 7. Finally, we provide some discussions in Section 8.
4
2
Preliminaries
We use standard graph terminologies. We refer to a directed edge (u, v), shortly, by uv (i.e., u
and v are head and tail of uv, respectively), and we refer to an undirected edge by {u, v}. For a
(directed or undirected) graph G, we denote by V (G) and E(G) the sets of vertices and edges of G,
respectively. If a graph G is associated
with edge-costs {ce }e∈E(G) , then we denote the cost of any
P
subgraph H ⊆ G by cost(H) = e∈E(H) ce . For any path P , we use length to mean the number of
edges in a path P and use cost to mean the total costs of edges in P .
In the directed Steiner tree problem (DST), we are given a directed graph G with edge-costs
{ce }e∈E(G) , a root vertex r and a set of terminals T ⊆ V (G). The goal is to find a min-cost subgraph
H ⊆ G such that H has a directed path from the root r to each terminal t ∈ T . A generalization
of DST is the k edge-connected directed Steiner tree problem (k-DST). In k-DST, we are given the
same input as in DST plus an integer k. The goal is to find a min-cost subgraph H that has k
edge-disjoint paths from the root r to each terminal t ∈ T . The k edge-connected directed Steiner
subgraph problem is a variant of k-DST, where there is no root vertex, and the goal is to find a
min-cost subgraph H such that the set of terminals T is k edge-connected in H.
The problems on undirected graphs that are closely related to of DST and k-DST are the group
Steiner tree problem (GST) and the k edge-connected group Steiner tree problem (k-GST). In GST,
we are given an undirected graph G with edge-costs {ce }e∈E(G) , a root vertex r and a collection of
subset of vertices {Ti }hi=1 called groups. The goal is to find a a min-cost subgraph H that connects
r to each group Ti . In k-GST, the input consists of an additional integer k, and the goal is to find
a min-cost subgraph H with k edge-disjoint r, Ti -paths for every group Ti .
Consider an instance of DST (resp., k-DST). We denote by Q the set of all paths in G that start
from the root r. The set of paths in Q that end with a particular pattern, say σ = (v1 , . . . , vq ),
is denoted by Q(σ). This pattern σ can be a vertex v, an edge e or a path σ = (v1 , . . . , vq ) in G.
For example, Q(u, v, w) consists of paths P of the form P = (r, . . . , u, v, w). We say that a path P
ends at a vertex v (resp., an edge e) if v (resp., e) is the last vertex (resp., edge) of P .
We may consider only paths with particular length, say D. We denote by QD the set of paths
that start at r and has length at most D. The notation for QD is analogous to Q, e.g., QD (uv) ⊆ QD
is the set of paths in QD that end at an edge uv. A concatenation of a path p with an edge e
or a vertex v are denoted by p + e and p + v, respectively. For example, (u1 , . . . , uℓ ) + vw =
(u1 , . . . , uℓ , v, w).
Given a subset of vertices S, the set of edges entering S is denoted by
δ− (S) = {uv ∈ E : u ∈ S, v 6∈ S}
The indegree of S is denoted by indeg(S) = |δ− (S)|. Analogously, we use δ+ (S) and outdeg(S) for
the set of edges leaving S. For undirected graphs, we simply use the notations δ(S) and deg(S).
We say that a feasible solution H to k-DST is D-shallow if, for every terminal t ∈ T , there exists
a set of k edge-disjoint r, t-paths in H such that every path has length at most D. An instance
of k-DST that has an optimal D-shallow solution is called a D-shallow instance. We also use the
term D-shallow analogously for k-GST and the k edge-connected Steiner subgraph problem.
To distinguish between the input of k-DST (which is a directed graph) and k-GST (which is
an undirected graph), we use script fonts, e.g., G, to denote the input of k-GST. Also, we use Q
to denote the set of all paths from the root r to any vertexP
v in the graph G. The cost of a set
of edges F (or a graph) is defined by a function cost(F ) = e∈F ce . At each point, we consider
5
only one instance of k-DST (respectively, k-GST). So, we denote the cost of the optimal solution
to k-DST by optkDST (respectively, optkGST ).
3
Reduction from Directed Steiner Tree to Group Steiner Tree
In this section, we describe a reduction R from DST to GST. We recall that Q denotes all the
r, v-paths in a DST instance G. The reduction is by simply listing paths in the directed graph G
as vertices in a tree G = R(G) and joining each path p to p + e if p + e is a path in G. In fact,
R(G) is a suffix tree of paths in Q. To be precise,
V (G) = {p : p is an r, v-path in G}
E(G) = {{p, p + e} : both p and p + e are paths in G starting at r}
We set the cost of edges of G to be c{p,p+e} = ce . Since the root r has no incoming edges in G,
r maps to a unique vertex (r) ∈ G, and we define (r) as the root vertex of the GST instance. We
will abuse r to mean both the root of DST and its corresponding vertex of GST. For each terminal
ti ∈ T , define a group of the GST instance as
Ti := Q(ti ) = {p ⊆ G : p is an r, ti -path in G}
It is easy to see that the reduction R produces a tree, and there is a one-to-one mapping between
a path in the tree G = R(G) and a path in the original graph G. Thus, any tree in G corresponds
to a subtree of R(G) (but not vice versa), which implies that the reduction R is approximationpreserving (i.e., optDST = optGST ). Note, however, that the size of the instance blows up from
O(n + m) to O(nD ), where D is the length of the longest path in G. The reduction holds for general
graphs, but it is approximation-preserving only if the DST instance is D-shallow, i.e., it has an
optimal solution H ∗ such that any r, ti -path in H ∗ has length at most D, for all terminals ti ∈ T .
However, Zelikovsky [Zel97, HRZ01] showed that the metric completion of G always contains a
D-shallow solution with cost at most D|V (G)|1/D of an optimal solution to DST. (This is now
known as Zelikovsky’s height reduction theorem.) Thus, we may list only paths of length at most
D from the metric completion. We denote the reduction that lists only paths of length at most D
by RD .
4
Properties of Minimal Solutions to k-DST
In this section, we provide structural lemmas which are building blocks in formulating a strong
LP-relaxation for k-DST. These lemmas characterize properties of a minimal solution to k-DST.
Lemma 3. Let H be any minimal solution to k-DST. Then H has at most k edge-disjoint r, v-paths,
for any vertex v ∈ V (H).
Proof. Suppose to a contrary that H has k + 1 edge-disjoint r, v-paths, for some vertex v ∈ V (H).
Then v must have indegree at least k + 1 in H. We take one of the k − 1 edges entering v, namely,
uv. By minimality of H, removing uv results in a graph H ′ = H − uv that has less than k edgedisjoint r, ti -paths for some terminal ti ∈ T . Thus, by Menger’s theorem, there must be a subset
of vertices S ⊆ V such that ti ∈ S, r ∈ V − S and indegH ′ (S) ≤ k − 1. Observe that we must have
6
−
(S) because H is a feasible solution to k-DST, which means that v ∈ S. Since we remove
uv in δH
only one edge uv from H, the graph H ′ must have k edge-disjoint r, v-paths. But, this implies that
indegH ′ (S) ≥ k, a contradiction.
Lemma 4. Let H be any minimal solution to k-DST. Any vertex v ∈ V (H) has indegree exactly
λ(v), where λ(v) is the maximum number of edge-disjoint r, v-paths in H.
Proof. The proof follows a standard uncrossing argument. Assume a contradiction that v has
indegree at least λ(v) + 1 in H. By Menger’s theorem, there is a subset of vertices U ⊆ V such
that indegH (U ) = λ(v), v ∈ U and r 6∈ U that separates v from r. We assume that U is a minimum
−
(U ), i.e.,
such set. Since indegH (v) > λ(v), there is an edge uv ∈ E(H) that is not contained in δH
u, v ∈ U .
By minimality of H, removing uv results in the graph H ′ = H − uv such that H ′ has less than
k edge-disjoint r, ti -path for some terminal ti ∈ T . Thus, by Menger’s theorem, there is a subset of
−
(W ) and indegH (W ) = k. (The latter is because H
vertices W such that ti ∈ W , r 6∈ W , uv ∈ δH
is a feasible solution to k-DST.)
Now we apply an uncrossing argument to U and W . By submodularity of indegH , we have
indegH (U ) + indegH (W ) ≥ degH (U ∩ W ) + degH (U ∪ W )
Observe that v ∈ U ∩ W , t ∈ U ∪ W and r 6∈ S ∪ S ′ . So, by the edge-connectivity of v and t,
indegH (U ∩ W ) ≥ λ(v)
and
indegH (U ∪ W ) ≥ k
(1)
The sum of the left-hand side of Eq (1) is
indegH (U ) + indegH (W ) = k + λ(v).
So, we conclude that
indegH (U ∩ W ) = λ(v)
and
indegH (U ∪ W ) = k
Consequently, we have the set U ′ = U ∩ W such that indegH (U ′ ) = λ(v), v ∈ U ′ and r 6∈ U ′ that
separates v from r. Since u 6∈ W , we know that U ′ is strictly smaller than U . This contradicts to
the minimality of U .
The following is a corollary of Lemma 3 and Lemma 4
Corollary 5. Let H be a minimal solution to k-DST. Then any vertex v ∈ V (H) has indegree at
most k.
The next lemma follows from Corollary 5.
Lemma 6. Consider any minimal solution H to k-DST (which is a simple graph). For any edge
e ∈ E(H) and ℓ ≥ 2, there are at most kℓ−2 paths in H with length at most ℓ that start at the root
ℓ−2 for all e ∈ E(H), where QH (e) is the set of r, v-paths of
r and ends at e. That is, |QH
ℓ
ℓ (e)| ≤ k
length ℓ in H.
7
Proof. We prove by induction. The base case ℓ = 2 is trivial because any rooted path of length at
most 2 cannot have a common edge.
ℓ−3 for some ℓ ≥ 3. Consider any edge vw ∈ E(H).
Assume, inductively, that |QH
ℓ−1 (e)| ≤ k
By Corollary 5, v has indegree at most k. Thus, there are at most k edges entering v, namely,
u1 v, . . . , ud v, where d = indeg(v). By the induction hypothesis, each edge is the last edge of at
ℓ−3 ≤ k ℓ−2 paths that end at uv. That is,
most kℓ−3 paths in QH
ℓ−1 . Thus, we have at most d · k
|QH
ℓ (vw)| ≤
d
X
|QH
ℓ−1 (ud v)| ≤
kℓ−3 = d · kℓ−3 ≤ kℓ−2 .
i=1
i=1
5
d
X
Standard LPs for k-DST and GST
In this section, we describe standard LPs for k-DST and GST. Each LP consists of two sets of
variables, a variable xe on each edge e and a variable fpi on each path p and a terminal ti . The
variable xe indicates whether we choose an edge e in a solution. The variable fpi is a flow-variable
on each path and thus can be written in a compact form using a standard flow formulation.
P
min
e∈E(G) ce xe
P
s.t.
f i ≤ xe ∀e ∈ E(G), ∀ti ∈ T
p∈Q(t
i ):e∈E(p) p
P
i
∀ti ∈ T
p∈Q(ti ) fp ≥ k
LP-k-DST
xe ≤ 1 ∀e ∈ E(G)
x
∀e ∈ E(G)
e ≥0
fpi ≥ 0 ∀p ∈ Q(ti ), ∀ti ∈ T.
The standard LP for GST is similar to LP-k-DST.
P
min
e∈E(G) ce xe
P
P
s.t.
fpi ≤ xe
v∈Ti P p∈Q(v):e∈E(p)
P
i
v∈Ti
p∈Q(v) fp ≥ 1
LP-GST
xe ≤ 1
xe ≥ 0
fpi ≥ 0
6
∀e ∈ E(G), ∀i = 1, 2, . . . , h
∀i = 1, 2, . . . , h
∀e ∈ E(G)
∀e ∈ E(G)
∀p ∈ Q, ∀i = 1, 2, . . . , h
A Strong LP-relaxation for for k-DST
In this section, we formulate a strong LP-relaxation for k-DST that allows us to embed a fractional
solution into an LP solution of LP-GST on a tree.
8
P
min P
e∈E ce xe
s.t.
fi
p∈Q(t
i ):e∈E(p) p
P
i
p∈Q(ti ) fp
P
i
f
p∈Q(t
P i ):q⊆p p
LP-k-DST*
p∈Qℓ (e) yp
xe
xe
fpi
yp
≤ xe
≥k
≤ yq
≤ max{1, kℓ−2 } · xe
≤1
≥0
≥0
≥0
∀e ∈ E(G), ∀ti ∈ T
∀ti ∈ T
∀q ∈ Q, ∀ti ∈ T
∀e ∈ E, ∀ℓ ≥ 1
∀e ∈ E(G)
∀e ∈ E(G)
∀p ∈ Q(ti ), ∀ti ∈ T
∀p ∈ Q
(Subflow Capacity)
(Aggregating k-Flow)
For D-shallow instances of k-DST, we replace Q by QD to restrict length of paths to be at most
D. The next lemma shows that LP-k-DST* is an LP-relaxation for k-DST.
Lemma 7. LP-k-DST* is an LP-relaxation for k-DST. Moreover, replacing Q by QD gives an
LP-relaxation for k-DST on D-shallow instances.
Proof. LP-k-DST* is, in fact, obtained from LP-k-DST (which is a standard LP) by adding a new
variable yp and two constraints.
X
(1) Subflow-Capacity:
fpi ≤ yq , ∀q ∈ Q, ∀ti ∈ T .
p∈Q(ti ):q⊆p
(2) Aggregating k-Flow:
X
yp ≤ max{1, kℓ−2 } · xe , ∀e ∈ E, ∀ℓ ≥ 1.
p∈Qℓ (e)
To show that these two constraints are valid for k-DST, we take a minimal feasible (D-shallow)
solution H of k-DST. We define a solution (x, f, y) to LP-k-DST as below.
1 if e ∈ E(H)
1 if p ⊆ H ∧ p ∈ Q
xe =
yp =
0 otherwise
0 otherwise
1
if
p
⊆
H
∧
p
∈
Q(t
)
i
fpi =
0 otherwise
By construction, fpi = 1 implies that yp = 1. Thus, (x, f, y) satisfies the Subflow-Capacity constraint. By minimality of H, Corollary 6 implies that even if we list all the paths of length ℓ ≥ 2
in H, at most kℓ−2 of them end at the same edge, and we know that rooted paths of length one
share no edge (given that H is a simple graph). Thus, (x, f, y) satisfies the Aggregating k-Flow
constraint. Consequently, these two constraints are valid for k-DST.
On the other hand, any integral solution that is not feasible to k-DST could not satisfy the
constraints of LP-k-DST* simply because LP-k-DST* contains the constraints of LP-k-DST, which
is a standard LP for k-DST. Thus, LP-k-DST* is an LP-relaxation for k-DST.
The proof for the case of D-shallow instances is the same as above except that we take H as a
minimal D-shallow solution and replace Q by QD .
9
7
An Approximation Algorithm for k-DST
In this section, we present an approximation algorithm for k-DST on a D-shallow instance. Our
algorithm is simple. We solve LP-k-DST* on an input graph G and then embed an optimal
fractional solution (x, f, y) to an LP-solution (x̂, fˆ) of LP-GST on the tree R(G). We lose a factor
of O(kD−2 ) in this process. As we now have a tree-embedding of an LP-solution, we can invoke the
GKR Rounding algorithm [GKR00] to round an LP-solution on the tree R(G). Our embedding
guarantees that any edge-set of size k − 1 in the original graph G never maps to an edge-set in
the tree G = R(G) that separates r and Ti = Q(ti ) in G. So, the rounding algorithm still outputs
a feasible solution to GST with constant probability even if we remove edges in the tree G that
correspond to a subset of k − 1 edges in G. Consequently, we only need to run the algorithm for
O(D · k · log n) times to boost the probability of success so that, for any subset of k − 1 edges and
any terminal ti ∈ T , we have at least one solution that contains an r, ti -path using none of these
k − 1 edges. In other words, the union of all the solutions satisfies the connectivity requirement.
Our algorithm is described in Algorithm 1.
Algorithm 1 Algorithm for k-DST
Solve LP-k-DST* and obtain an optimal solution (x, f, y).
Map (x, f, y) to a solution (x̂, fˆ) to LP-GST on G = R(G).
for i = 1 to 2Dk log2 n do
Run GKR Rounding on (x̂, fˆ) to get a solution Zi .
Map Zi back to a subgraph Zi of G.
end for
S
return H = i Zi as a solution to k-DST.
We map a solution (x, f, y) of LP-k-DST* on G to a solution (x̂, fˆ) of LP-GST on the tree
G = R(G) as below. Note that there is a one-to-one mapping between a path in G and a path in
the tree G.
x̂{p,p+e} := yp+e for all p + e ∈ Q
fˆpi := fpi
for all p ∈ Q and for all ti ∈ T
7.1
Cost Analysis
We show that cost(x̂, fˆ) ≤ kD−2 · cost(x, f, y).
Lemma 8. Consider a solution (x̂, fˆ) to LP-GST, which is mapped from a solution (x, f ) of LPk-DST* when an input k-DST instance is D-shallow, for D ≥ 2. The cost of (x̂, fˆ) is at most
cost(x̂, fˆ) ≤ kD−2 · cost(x, f ).
P
Proof. By the constraint p∈Qℓ (e) fp ≤ max{1, kℓ−2 } · xe , we have that
P
P
P
= e∈E(G)
c x̂
cost(x̂, fˆ) = e′ ∈E(G) ce′ x̂e′
P
P
P
P{p,p+e}∈E(G) e {p,p+e}
= e∈E(G)
ce fp+e
= e∈E(G)
p∈Q(e) ce fp
p+e∈Q
P
P
P
= e∈E(G) ce · p∈Q(e) fp
≤ e∈E(G) ce · kD−2 · xe
= kD−2 · cost(x, f ).
10
It can be seen from Algorithm 1 and Lemma 8 that the algorithm outputs a solution H with
cost at most O(Dk D−1 log n) · cost(x, f ). Thus, H is an O(Dk D−1 log n)-approximate solution. It
remains to show that H is feasible to k-DST.
7.2
Feasibility Analysis
Now we show that H is feasible to k-DST with at least constant probability. To be formal, consider
any subset F ⊆ E(G) of k − 1 edges. We map F to their corresponding edges F in the tree G.
Thus, F := {{P, P + e} : P + e ∈ Q ∧ e ∈ F }.
Observe that no vertices in G − F correspond to paths that contain an edge in F . Thus, we can
define an LP solution (y F , z F ) for LP-GST on the graph G − F as follows.
i
x̂e if e 6∈ F
fˆp if E(p) ∩ F = ∅
F,i
F
ye =
zp
=
0 otherwise
0 otherwise
We show that (y F , z F ) is feasible to LP-GST on G − F.
Lemma 9. For any subset of edges F ⊆ E(G), define (y F , z F ) from (x̂, fˆ) as above. Then (y F , z F )
is feasible to LP-GST on G − F.
Proof. First, observe that zpF,i > 0 only if a path p contains no edges in F. So, by construction,
(y F , z F ) satisfies zpF,i = fˆpi ≤ x̂e = yeF for all e ∈ E(p). Hence, (y F , z F ) satisfies the capacity
constraint.
Next we show that (y F , z F ) satisfies the connectivity constraint. Consider the solution (x, f, y)
to LP-k-DST*. By the feasibility of (x, f, y) and the Max-Flow-Min-Cut theorem, the graph G − F
with P
capacities {xe }e∈G−F can support a flow of value one from r to any terminal ti . This implies
that p∈Q(ti ):E(p)∩F =∅ fpi ≥ 1. Consequently, we have
P
i
p∈Q(ti ):E(p)∩F =∅ fp
P
P
ˆi
= p∈Ti :E(p)∩F =∅
p′ ∈Q(v) fp′
P
P
fˆi′
= v∈Ti
′
′
Pp ∈Q(v):E(p )∩F =∅ pF,i
P
= v∈Ti
′
′ )∩F =∅ zp′
Pp ∈Q(v):E(p
P
F,i
= v∈Ti
p′ ∈Q(v) zp′
≥ 1.
All the other constraints are satisfied because (y F , z F ) is constructed from (x̂, fˆ). Thus, (y F , z F )
is feasible to LP-GST on G − F.
Lemma 9 implies that we can run the GKR Rounding algorithm on (y F , z F ). The following is
the property of GKR Rounding.
Lemma 10 ([GKR00]). There exists a randomized algorithm such that, given a solution (x̂, fˆ) to
LP-GST on a tree G with height D, the algorithm outputs a subgraph H ⊆ G so that the probability
that any subset of vertices U ⊆ V (G) is connected to the root is at least
P
P
ˆi
v∈U
p∈Q(v) fp
Pr[H has an r, U -path.] ≥
O(D)
Moreover, the probability that each edge is chosen is at most x̂e . That is, E[cost(H)] = cost(x̂, fˆ).
The running time of the algorithm is O(|E(G)| + |V (G)|).
11
Since (y F , z F ) ≤ (x̂, fˆ) (coordinate-wise), we can show that running GKR Rounding on (x̂, fˆ)
simulates the runs on (y F , z F ) for all F ⊆ E(G) with |F | ≤ k − 1, simultaneously.
Lemma 11. Let H be a subgraph of G obtained by running GKR Rounding on (x̂, fˆ), and let H
be a subgraph of G corresponding to H. Then, for any subset of edges F ⊆ E(G) with |F | ≤ k − 1
and for any terminal ti ∈ T ,
Pr[H − F has an r, ti -path] ≥
1
.
O(D)
Proof. Let us briefly describe the work of GKR Rounding. The algorithm marks each edge e in the
tree with probability xe /x̺(e) , where ̺(e) is the parent of an edge e in G, which is unique. Then
the algorithm picks an edge e if all of its ancestors are marked. Clearly, removing any set of edges
F from G only affects paths that contain an edge in F.
Let (y F , z F ) be defined from (x̂, fˆ) as above. This LP solution is defined on a graph G − F.
Thus, the probability of success is not affected by removing F from the graph. By Lemma 9, we
can run GKR Rounding on (y F , z F ) and obtain a subgraph HF of G − F. Since zpF ≤ fˆp for all
6 ∅, we have from Lemma 10 and Lemma 9 that
paths p ∈ Q and zpF = 0 for all p ∈ Q : E(p) ∩ F =
Pr[H − F has an r, ti -path] = Pr[H − F has an r, Ti -path]
F has an r, T -path]
≥ Pr[H
i
P
P
≥
≥
v∈Ti
1
O(D) .
F
p∈Q(v) zp
O(D)
Finally, we recall that Algorithm 1 employs GKR Rounding on (x̂, fˆ) for 2Dk log2 n times. So,
for any subset of k − 1 edges F ⊆ E(G) and for any terminal ti ∈ T , there exists one subgraph
that has an r, ti -path that contains no edge in F with large probability. In particular, the union is
a feasible solution to k-DST with at least constant probability.
S
Lemma 12. Consider the run of Algorithm 1. The solution subgraph H = i Zi is a feasible
solution to k-DST with probability at least 1/n.
Proof. For i = 1, 2, . . . , 2Dk log2 n, let Zi be a subgraph of G obtained by running GKR Rounding
on (x̂, fˆ) and mapping the solution back to a subgraph of G as in Algorithm 1. By Lemma 11,
Zi − F has an r, ti -path with probability Ω(1/D). Since each Zi is sampled independently, we have
Pr[∀i Zi − F has no r, ti -path] ≤
1
1−
O(D)
2Dk log2 n
2k log2 n
1
≤
≤ n−2k .
e
We have at most |E(G)|k−1 ≤ n2(k−1) such sets F and at most |T | ≤ n terminals. So, there are at
most n2k−1 bad events where there exists an edge-set ofSsize k − 1 that separates the root r and
some terminal ti ∈ T . Therefore, by union bound, H = i Zi is a feasible solution to k-DST with
probability at least 1/n.
This completes the proof of Theorem 2. Note that, for the case of DST (k = 1), we only need
to run GKR Rounding for O(D log h) times, thus implying an approximation ratio of O(D log h).
12
8
Conclusion and Discussion
We presented the first non-trivial approximation algorithm for k-DST in a special case of a Dshallow instance, which exploits the reduction from DST to GST. We hope that our techniques will
shed some light in designing an approximation algorithm for k-DST in general case and perhaps
lead to a bi-criteria approximation algorithm in the same manner as in [CGL15].
One obstruction in designing an approximation algorithm in directed graphs is that there is no
“true” (perhaps, probabilistic) tree-embedding for directed graphs. Both devising a tree-embedding
for directed graphs and designing an approximation algorithm for k-DST with k ≥ 2 are big open
problems in the area. Another open problem, which is considered as the most challenging one
by many experts, is whether there exists a polynomial-time algorithm for DST that yields a subpolynomial approximation ratio.
Acknowledgements. Our work was inspired by the works of Rothvoß [Rot11] and Friggstad et al. [FKK+ 14] and by discussions with Joseph Cheriyan and Lap Chi Lau. We also thank
Zachary Friggstad for useful discussions.
References
[CCC+ 99] Moses Charikar, Chandra Chekuri, To-Yat Cheung, Zuo Dai, Ashish Goel, Sudipto
Guha, and Ming Li. Approximation algorithms for directed steiner problems. J. Algorithms, 33(1):73–91, 1999. 2
[CGL15]
Parinya Chalermsook, Fabrizio Grandoni, and Bundit Laekhanukit. On survivable set
connectivity. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015, pages 25–
36, 2015. 4, 12
[CLNV14] Joseph Cheriyan, Bundit Laekhanukit, Guyslain Naves, and Adrian Vetta. Approximating rooted steiner networks. ACM Transactions on Algorithms, 11(2):8:1–8:22, 2014. 2
[Fei98]
Uriel Feige. A threshold of ln n for approximating set cover. J. ACM, 45(4):634–652,
1998. 3
[FKK+ 14] Zachary Friggstad, Jochen Könemann, Young Kun-Ko, Anand Louis, Mohammad
Shadravan, and Madhur Tulsiani. Linear programming hierarchies suffice for directed
steiner tree. In Integer Programming and Combinatorial Optimization - 17th International Conference, IPCO 2014, Bonn, Germany, June 23-25, 2014. Proceedings, pages
285–296, 2014. 2, 13
[FKN12]
Moran Feldman, Guy Kortsarz, and Zeev Nutov. Improved approximation algorithms
for directed steiner forest. J. Comput. Syst. Sci., 78(1):279–292, 2012. 2
[GKR00]
Naveen Garg, Goran Konjevod, and R. Ravi. A polylogarithmic approximation algorithm for the group steiner tree problem. J. Algorithms, 37(1):66–84, 2000. 3, 4, 9,
11
13
[HRZ01]
Christopher S. Helvig, Gabriel Robins, and Alexander Zelikovsky. An improved approximation scheme for the group steiner problem. Networks, 37(1):8–20, 2001. 3, 6
[Lae14]
Bundit Laekhanukit. Parameters of two-prover-one-round game and the hardness of
connectivity problems. In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2014, Portland, Oregon, USA, January 5-7, 2014,
pages 1626–1643, 2014. 2
[LY94]
Carsten Lund and Mihalis Yannakakis. On the hardness of approximating minimization
problems. J. ACM, 41(5):960–981, 1994. 3
[Rot11]
Thomas Rothvoß. Directed steiner tree and the lasserre hierarchy. CoRR, abs/1111.5473,
2011. 2, 13
[Zel97]
Alexander Zelikovsky. A series of approximation algorithms for the acyclic directed
steiner tree problem. Algorithmica, 18(1):99–110, 1997. 3, 6
14
| 8cs.DS
|
1
Incentive Mechanism Design for Wireless Energy
Harvesting-Based Internet of Things
arXiv:1712.06250v1 [cs.IT] 18 Dec 2017
Zhanwei Hou, He Chen, Yonghui Li, and Branka Vucetic
Abstract—Radio frequency energy harvesting (RFEH) is a
promising technology to charge unattended Internet of Things
(IoT) low-power devices remotely. To enable this, in future IoT
system, besides the traditional data access points (DAPs) for
collecting data, energy access points (EAPs) should be deployed
to charge IoT devices to maintain their sustainable operations.
Practically, the DAPs and EAPs may be operated by different
operators, and the DAPs thus need to provide effective incentives
to motivate the surrounding EAPs to charge their associated IoT
devices. Different from existing incentive schemes, we consider
a practical scenario with asymmetric information, where the
DAP is not aware of the channel conditions and energy costs
of the EAPs. We first extend the existing Stackelberg gamebased approach with complete information to the asymmetric
information scenario, where the expected utility of the DAP is
defined and maximized. To deal with asymmetric information
more efficiently, we then develop a contract theory-based framework, where the optimal contract is derived to maximize the
DAP’s expected utility as well as the social welfare. Simulations
show that information asymmetry leads to severe performance
degradation for the Stackelberg game-based framework, while
the proposed contract theory-based approach using asymmetric
information outperforms the Stackelberg game-based method
with complete information. This reveals that the performance of
the considered system depends largely on the market structure
(i.e., whether the EAPs are allowed to optimize their received
power at the IoT devices with full freedom or not) than on
the information availability (i.e., the complete or asymmetric
information).
Index Terms—Internet of Things, Wireless energy harvesting,
Stackelberg game, contract theory, incentive mechanism
I. I NTRODUCTION
A. Background and Motivations
By connecting objects, physical devices, vehicles, animals
and other items to the Internet, Internet of Things (IoT) has
emerged as a new paradigm to enable ubiquitous and pervasive
communications [1]–[3]. Wireless sensing service is one of the
fundamental applications of IoT, which enables systems and
users to continuously monitor ambient environment [4].
One of the major hurdles for implementing the wireless
sensing application is the limited lifetime of traditional batterypowered sensors, which are costly and hard to maintain [5],
[6]. For example, frequent recharging or battery replacement is
inconvenient in deserts or remote areas, and is even impossible
for some scenarios, such as toxic environment or implanted
medical applications [7]. To tackle this problem, radio frequency energy harvesting (RFEH) has recently been proposed
Z.W. Hou, H. Chen, Y.H. Li and B. Vucetic are with the Centre of Excellence in Telecommunications, School of Electrical and Information Engineering, University of Sydney, Sydney, NSW, Australia (e-mail: {zhanwei.hou,
he.chen, yonghui.li, branka.vucetic}@sydney.edu.au).
as an attractive technology to prolong the operational lifetime
of sensors, enhance the deployment flexibility, and reduce the
maintenance costs [7], [8].
In this paper, we consider a radio frequency energy harvesting based IoT system consisting of a data access point
(DAP) and several energy access points (EAPs). The DAP
collects information from its associated sensors. EAPs can
provide wireless charging services to sensors via the RF
energy transfer technique. The sensors are assumed to have
no embedded energy supply, but they can harvest energy from
radio frequency (RF) signals radiated by the surrounding EAPs
to transmit the data to the DAP [9].
There are some research considering the deployment of
dedicated EAPs in the existing cellular network, such that
the upgraded network can provide both wireless access and
wireless charging services [10]–[17]. However, it was assumed
that the EAPs are deployed by the same operator of the
existing network. In practice, the DAP and EAPs may be
operated by different operators1 . To effectively motivate these
third-party and self-interested EAPs to help charge the sensors,
effective incentive mechanisms are required to improve the
payoff of the DAP as well as those of EAPs. While there are
several initial work designing the incentive mechanism [18]–
[20] for the EAPs belonging to different operators, complete
information was considered in these schemes. Specifically, it
was assumed that the EAPs will truthfully report their private
information to the DAP, e.g., their energy costs and channel
gains between EAPs and sensors. This happens when there
exists a supervising entity in the network, which is capable of
monitoring and sharing all behaviours and network conditions
of the DAP and EAPs to ensure that they always report
the trustful information. However, without such a supervising
entity, EAPs’ private information might be not aware to the
DAP, which is normally called information asymmetry in
the literature [21]. A rational EAP may provide misleading
information maliciously and pretend to be an EAP with better
channel condition and/or higher energy cost to cheat for more
rewards. A malicious EAP can succeed in cheating to get more
benefits because of information asymmetry in the RF energy
trading process.
To address above issues, in this paper we will design
effective incentive mechanisms to maximize the utilities of the
DAP and EAPs under scenarios with asymmetric information.
1 This could happen when a resource-limited operator cannot provide radio
frequency energy charging service in some certain area due to limited budgets,
or lack of site locations, or lack of licensed spectrum for energy harvesting.
Therefore, it has to resort to third-party operators.
2
To this end, the following important questions should be
addressed under asymmetric information:
Which EAPs the DAP should hire, how much energy should
be requested from the hired EAPs, and how many rewards
should be given to the hired EAPs?
The above questions are non-trivial to answer because the
hierarchical interactions between multiple parities should be
modeled and analyzed: the cooperations between the DAP,
the DAP’s sensors and the EAPs, and the competition among
EAPs with heterogeneous private information. Moreover, the
information asymmetry make the problem even more challenging, because it is difficult for the DAP to hire the effective
EAPs without knowing EAPs’ private information, such as
energy costs and channel condition towards its sensors.
B. Solution and Contribution
To answer the above questions, we apply the wellestablished economic theories to model the conflicted interests
among the multiple parities in the considered RFEH-based IoT
system. Specifically, we first extend the existing Stackelberg
game-based approach with complete information to the considered case with asymmetric information, such that we can
evaluate the performance degradation caused by information
asymmetry to this approach. More specifically, due to lack of
the complete information, the expected utility function of the
DAP is defined and optimized in the Stackelberg game with
asymmetric information. Considering that contract theory is a
powerful tool originated from economics to deal with information asymmetry in a monopoly market, we apply contract
theory to develop an optimal contract to effectively motivate
the EAPs under asymmetric information. In our contract, the
RF energy trading market is analogous as a monopoly labor
market in economics. The DAP is modeled as the employer
who offers a contract to each EAP. The contract is composed
of a serious of contract items, which are combinations of
energy-reward pairs. Each contract item is an agreement
about how many rewards an EAP will get by contributing a
certain amount of RF energy. Various heterogeneous EAPs are
classified into different types according to their energy costs
and instantaneous channel conditions. The EAPs are regarded
as labors in the market, which will choose a contract item best
meeting their interests. By properly designing the contract, an
EAP’s type will be revealed through its selection. Thus the
DAP can capture each EAP’s private information to a certain
extent and thus relieve the issue of information asymmetry.
To the best knowledge, this is the first paper that systematically studies the RFEH-based IoT system under asymmetric
information. The main contributions of this paper are summarized as follows.
• We develop the framework of RF energy trading in
the RFEH-based IoT system and systematically design
the incentive mechanisms for a practical scenario with
asymmetric information.
• To explore the performance degradtion due to lack of
full information, we first extend the existing Stackelberg
game-based approach to the considered case without
instantaneous channel condition and energy costs of the
EAPs by optimizing the expected utility function of the
DAP. As contract theory is a powerful economic theory
for designing incentive mechanism with asymmetric information. We then reformulate the problem by using
contract theory. In our contract design, we characterize
the necessary and sufficient conditions for the contract
feasibility, i.e, individual rationality (IR) conditions and
incentive capability (IC) conditions [21]. Subject to the
IR and IC constraints, the optimal contract under information asymmetry is achieved by maximizing the DAP’s
expected utility as well as the social welfare.
• To compare the performance of the proposed contract
theory-based approach using asymmetric information
with that of the existing Stackelberg game-based method
with complete information, we generalize the existing
Stackelberg game formation with unified pricing to the
case with discriminative pricing and derive the new Stackelberg equilibrium in closed-form. Here discriminative
pricing means that we set different energy prices for different EAPs to fully exploit their potentials for the enery
charging service. Numerical simulation results show that
information asymmetry can lead to severe performance
degradation for the Stackelberg game-based framework,
while the proposed contract theory-based scheme using asymmetric information outperforms the Stackleberg
game-based approach with complete information. This
implies that the performance of the considered system
depends largely on the market structure (i.e., whether the
EAPs are allowed to optimize their received power at
the IoT devices with full freedom or not) than on the
information availability (i.e., the complete or asymmetric
information).
Note that part of the work was presented in our previous
conference paper [22]. In this journal version, we extend our
previous work by considering both scenarios of complete information and asymmetric information and explore the impacts
of information availability and market structure.
The rest of this paper is organized as follows: In Section
II, we review the related literature. The system model is presented in Section III. The incentive mechanisms in asymmetric
information in complete information are proposed in Section
IV. The benchmark schemes are elaborated in Section V.
Numerical results are presented in Section VI, and conclusions
are drawn in Section VII.
II. R ELATED W ORK
A. EAP assisted Wireless Energy Harvesting
The idea of deploying a dedicated wireless energy network,
that can provide wireless charging service to the terminals by
using RFEH technology, was originally proposed by Huang
et al. [10], [11]. The dedicated power transmitters are called
power beacons or EAPs. Using stochastic geometry, the tradeoff between the densities of the base stations and EAPs was
analyzed in [10]. There are many works exploiting EAPs to
enable services of both wireless information and energy access
in existing cellular networks [12]–[17]. Stochastic geometry
was used to analyze the network performance with the EAPs
3
in [12]–[14]. In [15], beamforming was introduced in the EAPs
assisted cellular network to reduce the interference resulting
from the EAPs. Leveraging finite-length information theory,
the system performance in the finite blocklength regime was
analyzed in [16]. The security issue with EAP in the presence
of a passive eavesdropper was investigated in [17]. In all
above works, the DAP and EAPs are assumed to belong to
the same operator. In such a network, the devices belonging
to the same operator with extra energy were assumed to
voluntarily assist other devices. However, in practice, the
DAP and EAPs may be operated by different operators. To
successfully motivate self-interested EAPs to provide help,
effective incentive mechanisms are required. There are several
prior research works in designing the incentive mechanism
[18]–[20] for the EAPs, where [18], [19] adopted Stackelberg
game and [20] used auction to design the incentive mechanism.
However, the existing incentive mechanisms only considered
complete information scenario.
B. Contract Theory
The contract theory has been employed to address incentive
design problems in wireless communication areas, such as
mobile edge computing [23], device-to-device (D2D) communications [24] and cooperative spectrum sharing [25]. To
the best knowledge of the authors, we are the first to apply
contract theory in the RF energy trading process in RFEHbased IoT systems. To design the incentive mechanism in such
a scenario is challenging because the DAP needs to choose and
reward the most efficient EAPs without knowing their channel
conditions and energy costs.
C. Stackelberg Game
Stackelberg game has been widely used in wireless communications to model the interactions of steatitic parties, such as
physical layer security [26], resource management for LTEunlicensed [27], cognitive radio [28] and wireless energy
harvesting [18], [19], [29]–[31]. In [29], the authors considered
cooperative spectrum sharing with one primary user (PU) and
one secondary user (SU), which harvests energy from ambient
radio signal. The Stackelberg game was used to design the
the SU’s optimal cooperation strategy. In [30], simultaneous
wireless information and power transfer (SWIPT) in relay
interference channels was considered, where multiple sourcedestination pairs communicate through their dedicated energy
harvesting relays. The optimal power splitting ratios for all
relays were derived by the formulated Stackelberg game. In
[31], the authors formulated a stochastic Stackelberg game to
study the delay optimal power allocation scheme. There is
a recent paper addressing the EAP assisted wireless energy
harvesting by using Stackelberg game [19], but their system
settings are different from those in our work. Only one EAP
with multiple antennas was considered in this paper, and
the EAP acts as the seller and the base station (BS) as the
buyer in behalf of its sensors. A more relevant work is [18],
where an incentive mechanism was designed for the system
with the similar setup where monetary reward with unified
pricing were provided by the DAP to motivate third-party
EAP 1
EAP 2
DAP
Backhaul
EAP N
Server
Information
Transmission
Radio Frequency
Energy Transfer
Active Sensor
Inactive Sensor
Fig. 1: System model for the radio frequency energy harvesting
assisted Internet of Things network
EAPs to assist the charging process. Here the unified pricing
means prices per unit energy for different EAPs are the same.
However, in this paper, we consider discriminative pricing
scheme of Stackelberg game for the heterogenous EAPs in our
work, including unified pricing scheme as a special case. Here
discriminative pricing means that we set different energy prices
for different EAPs to fully exploit their potentials for the enery
charging service. Moreover, we extend the Stackelberg game
to asymmetric information scenario by optimizing expected
utility function of the DAP, instead of optimizing instantaneous
utility function of the DAP in the classical Stackelberg game.
III. S YSTEM M ODEL
Consider a wireless energy harvesting-based IoT system
consisting of one DAP and N EAPs belonging to different
operators, which are connected to constant power supplies and
connected to the server by backhauls, as shown in Fig. 1. The
DAP is responsible for collecting various data from several
wireless-powered sensors within its serving region. Without
embedded energy supplies, the wireless-powered sensors fully
rely on the energy harvested from the RF signals emitted by
the EAPs to transmit its information to the DAP. For simplicity,
we consider that the RF energy transfer and information
transmission are performed over orthogonal bandwidth. For
analytical tractability, time division-based transmission among
sensors is adopted, i.e., there is only one active sensor during
each transmission block. Hereafter, we refer to this active
sensor as the information source. Besides, all the nodes in
the system are assumed to be equipped with single antenna
and operate in the half-duplex mode.
We consider that the energy-carrying signals sent by the
EAPs are independent and identically distributed (i.i.d.) random variables with zero mean and unit variance. Note that no
coordination between the EAPs is needed since independent
signals are transmitted. All channels are assumed to experience
independent slow and flat fading, where the channel gains
remain constant during each transmission block and change
4
independently from one block to another2. The information
source rectifies the RF signals received from the EAPs and
uses the harvested energy to transmit its information. The time
duration of every transmission block is normalized to one. So
we use “energy” and “power” interchangeably hereafter. The
amount of energy harvested by the information source during
one transmission block can be expressed as
Es = η
N
X
pm Gm,s ,
(1)
m=1
where 0 < η < 1 is the energy harvesting efficiency, pm is
the charging power of the mth EAP, and Gm,s is the channel
power gain between the mth EAP and the information source.
Note that the noise is ignored in (1) since it is practically
negligible at the energy receiver.
The harvest-use protocol is considered in this paper [32].
More specifically, the information source will use the harvested energy to perform instantaneous information transmission to the DAP. We consider a battery-free design which
indicates that the sensor only has a storage device like supercapacitor to hold the harvested energy for a short period of
time, e.g., among its scheduled transmission block. Hence the
sensor exhausts all the harvested energy in each transmission
block, so the sensor’s energy storage device is emptied at
the beginning of the transmission block. This battery-free
design can reduce the complexity and costs of the sensors,
which is particularly suitable for the considered IoT sensing
applications and has been adopted by other applications [33],
[34]. The transmit power of the information source is thus
given by
Ps = Es .
(2)
Then, the received signal-to-noise ratio (SNR) at the DAP is
given by
ps Ga,s
,
(3)
β=
N0
where N0 is the noise power at the DAP, and Ga,s is the
channel power gain from the information source to the DAP.
Note that the time duration for each transmission block is normalized as one, such that the channel capacity and throughput
can be used interchangeably. Hence the achievable throughput
(bps) from the information source to the DAP can be expressed
by
RSD = W log2 (1 + β)
!
N
(4)
ηGa,s X
pm Gm,s ,
= W log2 1 +
N0 m=1
where W is the bandwidth. We define the received signal
power at the active sensor contributed by the mth EAP3 as
2 Pilots are broadcasted by the active sensor to allow the DAP and EAPs
to estimate the channels. So the DAP is aware of the channel gain from the
DAP to the sensor and each EAP is aware of the channel gain from this EAP
to the sensor. But the DAP generally is not aware of the channel gains from
the EAPs to the active sensor. The energy consumption of channel estimation
is ignored.
3 Note that the received power contributed by each EAP is assumed to
be distinguishable by considering that the EAPs work in disjoint narrow
bandwidth.
qm = pm Gm,s , and set γ = ηGa,s /N0 for notation simplicity.
We can thus simplify (4) as
!
N
X
(5)
qm .
RSD = W log2 1 + γ
m=1
As we mentioned before, the EAPs considered in the system
belong to different operators and act strategically, so they
would not help the DAP voluntarily. To address this issue, the
DAP needs to provide rewards to motive the EAPs to charge its
sensors. In this paper, we mainly focus on monetary rewards as
the incentive between operators. Other forms of rewards, such
as physical resources (e.g., spectrum), or free offloading data
between operator can also be used. To efficiently exploit the
EAPs to achieve a good throughput, the following questions
need to be answered in asymmetric information: Which EAPs
the DAP should hire, how much energy should be requested
from the hired EAPs, and how many rewards should be given
to the hired EAPs?
IV. I NCENTIVE M ECHANISMS WITH A SYMMETRIC
I NFORMATION
To answer the above questions in the practical scenario
with asymmetric information, we first model the strategic
interactions between the DAP and EAPs as a Stackelberg
game. We will first re-design and re-analyze the existing
Stackelberg game into the considered scenario by defining
and optimizing expected utility of the DAP. In economic
theories, contract theory is a powerful tool to design incentive
mechanism in information asymmetry. As such, we will then
reformulate the incentive mechanism problem into an optimal
contract design problem.
A. Stackelberg Game with Asymmetric Information
In this part, we will first explore how to design a Stackelberg game to model the interactions between the DAP and
the EAPs, and then derive the optimal energy prices under
asymmetric information. In the proposed Stackelberg game
with asymmetric information, the DAP provides rewards to the
EAPs for charging its sensors. The DAP is the leader of the
formulated Stackelberg game, which imposes energy prices for
the EAPs. The DAP optimizes the energy prices to maximize
its expected utility function defined as the difference between
the benefits obtained from the achievable throughput and its
total payment to the EAPs. The EAPs are the followers which
optimize their utility functions defined as the payment received
from the DAP minus its energy cost.
1) Stackelberg Game Formulation: The channel conditions
and energy costs of EAPs are different, so the efficiencies
of EAPs to charge the sensor are distinct. To fully exploit
the potential of the EAPs, a discriminative pricing strategy
is considered, i.e., the DAP can impose different prices of
per unit energy harvested from different EAPs. Let q =
[q1 , q2 , . . . , qN ]T as the vector of the active sensor’s received
power from EAPs, with qm denoting the received power from
the mth EAP, and let λ = [λ1 , λ2 , . . . , λN ]T as the vector of
prices per unit energy harvested from EAPs, with λm ≥ 0
5
denoting the price per unit energy harvested from the mth
EAP. The total payment of the DAP to the EAPs is
λ, q ) =
Λ(λ
N
X
λm qm ,
(6)
m=1
where qm is the received energy from mth EAP. Since the aim
of the DAP is to achieve higher throughput at the cost of less
rewards to the EAPs, the utility function of the DAP can be
defined as
S
λ , q ) = RSD − cΛ(λ
λ , q ).
(λ
UDAP
(7)
where RSD is the achievable throughput defined in (4) and (5),
c is the unit cost of the DAP, which is normalized as c = 1
without loss of generality hereafter.
Each EAP is modeled as a follower which would like to
maximize its individual profit, the utility of which is defined
as
(8)
UkS (λk , qk ) = λk qk − Ck (pk ),
where pk = qk /Gk,s is the transmit power of the kth EAP,
and Ck (·) is used to model the energy cost of the kth EAP,
given by
Ck (x) = ak x2 ,
(9)
where ak > 0 is the energy cost coefficient. Note that the
above quadratic function has been widely adopted in the
energy trading market to model the energy cost [35]. The
utility function of the kth EAP becomes
ak
(10)
UkS (λk , qk ) = λk qk − 2 qk2 ,
Gk,s
Since the DAP is not aware of each EAP’s exact energy
cost coefficient and channel gain, it can sort EAPs into some
discrete types and use the statistical distributions of the types
of EAPs from historical data to optimize the expected utility
of the DAP. Specifically, we define the type of the kth EAP
as
G2k,s
,
(11)
θk :=
ak
which suggests that the larger the channel gain Gk,s between
the EAP and the information source, and/or the lower the unit
energy cost coefficient ak , the higher the type of the EAP.
Without loss of generality, we assume that there are totally K
types of EAPs with θ1 < θ2 < · · · < θK . In this definition,
the higher type EAP has better channel quality and/or lower
energy cost coefficient. Note that since ak > 0 and Gk,s > 0,
θ > 0 holds. Using (11), the EAP’s utility can be rewritten as
qk2
.
(12)
θk
Assume
PKthere are Nk EAPs belonging to the kth type, we thus
have k=1 Nk = N . We then can rewrite the DAP’s utility
according to the types of EAPs as
!
K
K
X
X
S
Nk λk qk .
Nk qk −
UDAP (λk , qk ) = W log2 1 + γ
UkS (λk , qk ) = λk qk −
k=1
k=1
(13)
In this section, we consider a scenario with strong information asymmetry. In such a scenario, the DAP is only aware
of the total number of EAPs (i.e., N ) and the distribution
of each type. But it does not know each EAP’s private
type and thus it does not know the exact number of EAPs
belonging to each type k (i.e., Nk ). As such, the DAP needs
to optimize its expected utility over the possibilities of all
possible combinations of Nk . The expected utility of the DAP
with N EAPs is given by
S
λ , q )} =
E{UDAP
(λ
−n1
N NX
X
N−
···
n1 =0 n2 =0
(
"
Φn1 ,...,nK W log2
1+γ
K
X
k=1
PK−2
i=0
X
ni
nK−1 =0
nk qk
!
−
K
X
k=1
nk λk qk
#)
,
(14)
PK−1
where nK = N −
n
is
known
after
giving
i
i=0
n1 , n2 , . . . , nK−1 since the DAP knows the total number
N of EAPs, and Φn1 ,...,nK is the probability of a certain
combination of the number of EAPs belonging to each type
(i.e., Nk , {k = 1, 2, . . . , K}). We assume that all types are
uniformly distributed. The probability of one EAP belonging
to each type is the same, which is 1/K. In this case, Φn1 ,...,nK
can be calculated as
Φn1 ,...,nK = Pr (N1 = n1 , N2 = n2 , . . . , Nk = nk )
(15)
N!
=
n1 !n2 ! . . . nK !K N
Since the DAP is not aware of the EAPs’ private information, it can only optimize the expectation of DAP’s utility
function by using the statistical knowledge of the EAP’s
private information. So the optimization problem for the DAP
or the leader-level game can be formulated as
(P4.1) :
S
λ , q )}
max E{UDAP
(λ
λ
s.t. λ ≥ 0
(16)
Accordingly, the optimization problem for the EAP with kth
type or the follower-level game can be formulated as
(P4.2) :
max UkS (λk , qk )
qk
(17)
s.t. qk ≥ 0
Note that although the DAP does not know the EAP’s exact
type, it knows the type set of EAPs.
The Stackelberg game for the considered system has been
formulated by combining problems (P4.1) and (P4.2). In this
game, the DAP is the leader who aims to solve problem
(P4.1), while the EAPs are the followers who aim to solve
their individual problem (P4.2). Once a game is formulated,
the subsequent task is to find its equilibrium point(s). For the
solution of the formulated game, the most well-known concept
is the Stackelberg equilibrium (SE), which can be formally
defined as follows:
Definition 1 (Stackelberg equilibrium (SE)). We use λ ∗ =
∗ T
] to denote the
[λ∗1 , λ∗2 , . . . , λ∗N ]T and q ∗ = [q1∗ , q2∗ , . . . , qN
solutions of problems (P4.1) and (P4.2), respectively. Then,
λ ∗ , q ∗ ) is a SE of the formulated game if the following
(λ
conditions are satisfied
S
S
λ ∗ , q ∗ ) ≥ UDAP
λ, q ∗ ) ,
UDAP
(λ
(λ
(18)
6
S
∗
S
Um
(qm
, λ∗m ) ≥ Um
(qm , λ∗m ) ,
(19)
for all λ ≥ 0 and q ≥ 0 .
2) Analysis of the Proposed Game: In this part, we will
analyze the SE of the proposed Stackelberg game with asymmetric information.
It can be observed from (10) that for given values of λk , the
utility function of the the EAP with the kth type is a quadratic
function of its contributed power qk to the active sensor and
the constraint is affine, which indicates that the problem (P4.2)
is a convex optimization problem. Thus, it is straightforward
to obtain its optimal solution given in the following lemma:
Lemma 1. For given values of λk , the optimal qk∗ of the EAP
with kth type for problem (P4.2) is given by
θk λk
.
(20)
2
Proof. The proof of this lemma follows by noting that the objective function of problem (P4.2) given in (17) is a quadratic
function in terms of qk .
qk∗ =
It can be observed from Lemma 1 that for the same energy
price, an EAP with better channel gain and/or less energy cost
would like to contribute more power to the sensors.
Then we replace qk with qk∗ in problem (P4.1), the optimization problem at the DAP side can be expressed as
S
λ, q ∗ )}
max E{UDAP
(λ
(P4.3) :
λ
(21)
s.t. λ ≥ 0
S
λ , q ∗ )} is given by
where E{UDAP
(λ
S
λ , q ∗ )}
E{UDAP
(λ
=
−n1
N NX
X
N−
···
n1 =0 n2 =0
(
"
Φn1 ,...,nK W log2
−
K
1X
2
k=1
nk θk λ2k
i=0
X
ni
nK−1 =0
K
γX
1+
nk θk λk
2
k=1
#)
PK−2
!
(22)
,
where Φn1 ,...,nK is given in (15).
We can observe that problem (P4.3) is a concave function in
terms of vector λ . This is because each term in the summation
is composed by a logarithm function (concave) and quadratic
functions (concave), and the summation of concave functions
are still a concave function. Moreover, the constraint is affine.
Problem (P4.3) is then a convex optimization problem. So
we can numerically solve the system of equations given by
the KKT conditions to get the solution of problem (P4.3).
According the KKT conditions, we can also get some insight
about the structure of the solution and thus have the following
proposition.
Proposition 1. The optimal solution to problem (P4.3) have
the following structure:
λ∗1 = λ∗2 = · · · = λ∗N .
Proof. See Appendix A.
(23)
We surprisingly find that the optimal energy prices for
different EAPs are the same, even if we impose discriminative prices for different EAPs in the original design of the
Stackelberg game. This is because the energy price of unit
received power is used in our pricing scheme. The DAP has no
motivation to treat the received power from EAPs differently,
so a unified pricing per unit received power is achieved.
Lack of complete information, the performance of the
Stackelberg game with asymmetric information is worse than
that with complete information. Note that in the considered
scenario, there are N EAPs in the market. In each channel
realization, each EAP in the market selects one EAP type
from a EAP type set randomly. In each channel realization,
the Stackelberg game under complete information can adapt
to the instantaneous combination of EAP types and calculate
an optimal price for each instantaneous combination of EAP
types by optimizing the instantaneous utility function. While
the Stackelberg game under asymmetric information cannot
adapt to instantaneous combination of EAP types, since it can
only calculate a single price for all possible combinations of
EAP types. Therefore, the reason that Stackelberg game under
asymmetric information is worse than Stackelberg game under
complete information is that it fails to adapt to the change of
the instantaneous combinations of EAP types, i.e., the change
of wireless channel conditions. This deduction will be verified
later in the simulation part.
B. Optimal Contract with Asymmetric Information
As we mentioned above, the performance of the Stackelberg
game is degraded under asymmetric information. To improve
the performance under asymmetric information, the DAP
could design and offer a contract to effectively motivate the
EAPs to charge its sensors. Note that in Stackelberg game, the
EAP has the freedom to optimize its own utility by choosing
any amount of received signal power at the active sensor when
the DAP imposes some given energy price. Different from
Stackelberg game, limited options are allowed for EAPs to
select in contract theory. Specifically, a group of energy-reward
pairs (referred to as contract items) are designed. A contract
consisting of a group of contract items is provided to the
EAPs. The EAPs will choose a contract item at its discretion
to maximize its benefit. By properly designing the contract
item, the DAP can induce the EAP to expose its type by its
selection of the contract item and thus relieve the information
asymmetry.
In the following, we will formulate the optimal contract,
characterize its feasibility conditions and provide optimal
solution for the formulated contract.
1) Contract Formulation: In this part, we will formulate
a contract for the RF energy trading between the DAP and
EAPs, characterize its feasibility conditions, and derive the
optimal contract subject to the feasibility conditions.
A contract including a series of energy-reward pairs (qk , πk )
is designed to maximize the expectation of the DAP’s utility.
For the kth type EAP, qk is the received power contributed
by kth EAP and πk is the reward paid to the kth EAP as the
incentive for the corresponding contribution.
7
We first rewrite the utility functions of the DAP and EAPs
according to contract items. The DAP’s utility function is thus
given by
! K
K
X
X
C
π , q ) = W log2 1 + γ
Nk πk , (24)
Nk qk −
UDAP (π
k=1
k=1
where π = [π1 , π2 , . . . , πK ]T is the reward paid by the DAP
to the EAP with the kth type for its corresponding contribution
q = [q1 , q2 , . . . , qK ]T . Similar to (14), the expectation of
C
π , q ) can be represented as
UDAP
(π
C
π , q )} =
(π
E{UDAP
−n1
N NX
X
N−
···
n1 =0 n2 =0
(
"
Φn1 ,...,nK W log2
1+γ
PK−2
i=0
X
ni
nK−1 =0
K
X
nk qk
k=1
!
−
K
X
nk πk
k=1
#)
,
(25)
PK−1
where nK = N −
n
is
known
after
giving
i
i=0
n1 , n2 , . . . , nK−1 since the DAP knows the total number
N of EAPs, and Φn1 ,...,nK is the probability of a certain
combination of the number of EAPs belonging to each type
(i.e., Nk , {k = 1, 2, . . . , K}), which is given by (15). And then
the utility function of the EAP with the kth type is rewritten
as
q2
UkC (πk , qk ) = πk − k .
(26)
θk
The social welfare is defined as the summation of the utilities
of the DAP and all N EAPs, given by
C
π , q ) = UDAP
π, q ) +
Γ(π
(π
= W log2
K
X
Nk UkC (πk , qk )
k=1
K
X
1+γ
k=1
Nk qk
!
−
K
X
Nk q 2
k
k=1
θk
(27)
.
It can be seen that the internal transfers, i.e., rewards, are
cancelled in the social welfare, which is consistent with
the aim to maximize the efficiency of the whole system,
i.e., achieving more throughput at the cost of less energy
consumptions.
Next, we will figure out the feasibility conditions. In our
design, to encourage the EAPs to participate in the charging
process and ensure that each EAP only chooses the contract
item designed for its type, the following individual rationality
(IR) and incentive compatibility (IC) constraints should be
satisfied [21].
Definition 2 (Individual Rationality (IR)). The contract item
that an EAP chooses should ensure a nonnegative utility, i.e.,
UkC (πk , qk ) = πk −
qk2
≥ 0, ∀k ∈ {1, . . . , K}.
θk
(28)
Definition 3 (Incentive Compatibility (IC)). An EAP of any
type k prefers to choose the contract item (qk , πk ) designed
for its type, instead of any other contract item (qj , πj ), ∀j ∈
{1, . . . , K} and j 6= k, given by
qj2
πk −
≥ πj − , ∀k, j ∈ {1, . . . , K}.
θk
θk
qk2
(29)
The IR condition requires that the received reward of each
EAP should compensate the cost of its consumed energy when
it participates in the energy trading. If Uk ≤ 0, the EAP will
choose not to charge the information source for the DAP.
We define this case as (qk = 0, πk = 0). The IC condition
ensures that each EAP automatically selects the contract item
designed for its corresponding type. The type of each EAP
is thus revealed to the DAP, which is called “self-reveal”. If
a contract satisfies the IR and IC constraints, we refer to the
contract as a feasible contract.
Following the idea of contract theory [21], the DAP aims at
maximizing its expected utility subjecting to the constraints of
IR and IC given in (28) and (29). Thus, the optimal contract
is the solution to the following optimization problem
(P4.4) :
C
π , q )}
max E{UDAP
(π
π ,q
q)
(π
qk2
≥ 0, ∀k ∈ {1, . . . , K},
θk
qj2
q2
πk − k ≥ πj − , ∀k, j ∈ {1, . . . , K},
θk
θk
qk ≥ 0, πk ≥ 0, θk ≥ 0, ∀k ∈ {1, . . . , K}.
(30)
The first two constraints correspond to IR and IC, respectively.
Note that the EAP will reveal its private type truthfully
with the IR and IC constraints. Specifically, the IR condition
ensures the EAP’s participation and the IC condition ensures
that each EAP selects the contract item designed for its
corresponding type to gain highest payoff.
2) Constraint Reduction: There are K IR constraints and
K(K − 1) IC constraints in (30), which are non-convex and
couple different EAPs together. It is hard to solve (30) directly
due to the complicated constraints. Motivated by this, in the
subsection we first reduce the constraints of (30) and transform
it.
We first realize that the following necessary conditions can
be derived from the IR and IC constraints.
s.t. πk −
Lemma 2. For any feasible contract, πi > πj if and only if
qi > qj , ∀i, j ∈ {1, . . . , K}.
Proof. See Appendix B.
Lemma 2 shows that the EAP contributing more received
power at the information source will receive more reward.
Lemma 3. For any feasible contract, πi = πj if and only if
qi = qj , ∀i, j ∈ {1, . . . , K}.
Lemma 3 can be proved by using similar procedures as
Lemma 2, which is omitted for brevity. Lemma 3 indicates
that the EAPs providing the same received power will get the
same amount of reward.
Lemma 4. For any feasible contract, if θi > θj , then πi > πj ,
∀i, j ∈ {1, . . . , K}.
Proof. See Appendix C.
Lemma 3 shows that a higher type EAP should be given
more reward. Together with Lemma 1 and Lemma 2, it can
be duduced that a higher type EAP also contributes more
8
energy to the information source. We define this feature as
monotonicity.
Definition 4 (Monotonicity). If θi ≥ θj , ∀i, j ∈ {1, . . . , K}
and then πi ≥ πj .
Based on the above analysis, we can now use the IC
condition to reduce the IR constraints and have the following
lemma.
Lemma 5. With the IC condition, the IR constraints can be
reduced as
q2
π1 − 1 ≥ 0.
(31)
θ1
Proof. See Appendix D.
We can also reduce the IC constraints and attain the following lemma.
3) Solution to Optimal Contract: We now solve the optimization problem (35) to attain the optimal contract in the
subsequent way: a standard method is first applied to resolve
the relaxed problem without monotonicity and the solution is
then verified to satisfy the condition of the monotonicity. By
iterating the first and second constraints in (35), we have
k
2
q12 X qn2 − qn−1
+
θ1 n=2
θn
k
1 2 X
1
1
2
= qk +
qn−1
,
−
θk
θ
θ
n−1
n
n=2
πk =
C
π , q )},
where ∀k ∈ {2, . . . , K}. Substitute (36) into E{UDAP
(π
and all πk , ∀k ∈ {1, . . . , K} are removed from the optimization problem (35), which becomes
Lemma 6. With monotonicity, the IC condition can be reduced
as the local downward incentive compatibility (LDIC), given
by
q2
q2
πi − i ≥ πi−1 − i−1 , ∀i ∈ {2, . . . , K},
(32)
θi
θi
and the local upward incentive compatibility (LUIC), given by
q2
qi2
≥ πi+1 − i+1 , ∀i ∈ {1, . . . , K − 1},
θi
θi
Proof. See Appendix E.
πi −
π ,q
q)
(π
q12
≥ 0,
2θ1
q2
q2
πi − i ≥ πi−1 − i−1 , ∀i ∈ {2, . . . , K},
θi
θi
2
qi+1
qi2
≥ πi+1 −
, ∀i ∈ {1, . . . , K − 1},
πi −
θi
θi
πK ≥ πK−1 ≥ · · · ≥ π1 ,
s.t. π1 −
qk ≥ 0, πk ≥ 0, θk ≥ 0, ∀k ∈ {1, . . . , K}.
(34)
The LDIC and the LUIC in (34) can be combined as shown
in Lemma 8.
Lemma 7. Since the optimization objective function is an increasing function of qk and a decreasing function of πk , ∀k ∈
{1, . . . , K}, the above optimal problem can be further simplified as
(P4.6) :
C
π , q )}
max E{UDAP
(π
π ,q
q)
(π
q12
= 0,
θ1
q2
q2
πk − k = πk−1 − k−1 , ∀k ∈ {2, . . . , K},
θk
θk
πK ≥ πK−1 ≥ · · · ≥ π1 ,
s.t. π1 −
qk ≥ 0, πk ≥ 0, θk ≥ 0, ∀k ∈ {1, . . . , K}.
(35)
Proof. See Appendix F.
q
−n1
N NX
X
N−
× W log2 1 + γ
−
K−1
X
k=1
PK−2
···
n1 =0 n2 =0
"
s.t.
C
π , q )}
max E{UDAP
(π
(P4.5) :
max
(33)
By using the reduced IR and IC constraints, the optimization
problem (30) can be transformed as
(36)
i=0
X
ni
Φn1 ,...,nK
nK−1 =0
K
X
k=1
nk qk
!
K
K
1 X
1 X
ni −
ni
θk
θk+1
i=k
i=k+1
!
qk2
#
nK 2
−
q
,
θK K
qk ≥ 0, ∀k ∈ {1, . . . , K}.
(37)
Note that (37) is composed of logarithmic functions and
quadratic functions, both of which are concave functions. And
the positive summation of all these concave function is still a
concave function. Besides, the constraint set is a convex set.
So we can leverage standard convex optimization tools in [36]
to solve it to get qk , and then πk can be calculated by (36).
Moreover, monotonicity is met automatically when the type is
uniformly distributed [21]. So far, we have derived the optimal
contract (qk , πk ), ∀k ∈ {1, . . . , K}, which can maximize the
utility of the DAP and satisfy the constraints of IR and IC.
C. Practical Implementation
To implement the proposed approach in a practical radio
frequency energy harvesting-based IoT system. The following
steps should be followed.
First, the DAP needs to collect the information it requires
by the computation of the optimal contract. The active sensor
will broadcast pilots to allow the DAP and EAPs to estimate
the channels such that the DAP is aware of the channel gain
from the DAP to the sensor. From historical data, the DAP
can obtain empirical values of the energy harvesting efficiency
factor and noise power, and thus it can attain the value of
parameter γ. With the known values of other public system
parameters including the channel bandwidth, the user number,
and the set of EAP types, the DAP can calculate the optimal
contract.
Next, the DAP will broadcast the optimal contract to the
candidate EAPs via the corresponding backhauls. By evaluating the contract, the EAPs will decide whether to participate
in the cooperation. If it decides to participate in the current
energy trading, it will send a feedback to the DAP. After the
9
DAP received the feedback, it will sign a contract with the
EAP.
Finally, after the contracts are signed, the EAPs will perform
the contracts by establishing an energy transfer link towards
the active sensor and charge it according to the agreed transmit
power. When the DAP detects that the EAPs have fulfilled its
contractual obligation, the DAP will pay the EAP with agreed
amount of rewards via the backhaul connecting the operators.
To investigate the impacts from information scenarios and
compare the proposed schemes with the existing schemes under complete information, we first extend existing Stackelberg
game from unified pricing strategy into discriminative pricing
strategy. And then we present the centralized optimization
scheme under complete information as the reference for the
proposed incentive mechanisms.
A. Stackelberg Game Formulation
To fully exploit the potentials of EAPs with distinct channel
conditions and energy costs, a discriminative pricing strategy
is considered, i.e., the DAP can impose different prices of per
unit energy harvested from different EAPs. The utility function
of the DAP can be rewritten as
= RSD −
N
X
λm qm .
(38)
m=1
where q = [q1 , q2 , . . . , qN ]T is the vector of the active sensor’s
received power from EAPs, with qm denoting the received
power from the mth EAP, λ = [λ1 , λ2 , . . . , λN ]T is the vector
of prices per unit energy harvested from EAPs, with λm ≥ 0
denoting the price per unit energy harvested from m EAP, and
RSD is the achievable throughput defined in (4) and (5). The
optimization problem for the DAP or the leader-level game
can be formulated as
(P5.1) :
S
λ, q )
max UDAP
(λ
λ
s.t. λ ≥ 0
(39)
Note that the optimization problem (P5.1) is different from
(P4.1) under asymmetric information, the instantaneous utility
of the DAP is optimized here, instead of expected utility of
the DAP in asymmetric information.
Each EAP is modeled as a follower which would like to
maximize its individual profit, the utility of which is rewritten
as
am 2
S
,
(40)
Um
(λm , qm ) = λm qm − 2 qm
Gm,s
where am > 0 and Gm,s are the energy cost coefficient
and channel gain of the mth EAP, respectively. Thus, the
optimization problem for the EAP m or the follower-level
game is given by
(P5.2) :
S
max Um
(λm , qm )
qm
s.t. qm ≥ 0
In this subsection, we will derive the SE of the formulated
game by analyzing the optimal strategies for the DAP and
EAPs to maximize their own utility functions. A closedform solution is derived by using Karush-Kuhn-Tucker (KKT)
conditions.
∗
First, the optimal qm
of the mth EAP is similar to that in
(20), which is given by the following Lemma:
Lemma 8. For given λm , the optimal solution for problem
(P5.2) is given by
V. B ENCHMARK S CHEMES WITH C OMPLETE
I NFORMATION
S
λ, q )
UDAP
(λ
B. Analysis of the Formulated Stackelberg Game
(41)
∗
qm
=
G2m,s λm
.
2am
(42)
Proof. The proof of this lemma follows by noting that the
objective function of problem (P5.2) given in (41) is a concave
function in terms of qm .
It can be observed from Lemma 8 that for the same energy
price, an EAP with better channel gain and/or less energy cost
would like to contribute more power to the active sensor.
Subsequently, we need to solve problem (P5.1) by replacing
∗
qm with qm
given in (42). The optimization problem at the
DAP side can be expressed as
(P5.3) :
S
λ, q ∗ )
max UDAP
(λ
λ
(43)
s.t. λ ≥ 0
S
λ, q ∗ ) is given by
where UDAP
(λ
S
λ, q ∗ )
UDAP
(λ
= W log2
N
X
G2m,s
λm
1+γ
2am
m=1
N
X
G2m,s 2
−
λ .
2am m
m=1
!
(44)
We can observe that problem (P5.3) is a concave function
in terms of vector λ since the former part in (44) is a
logarithm function (concave) and the latter parts in (44) are the
summation of quadratic functions (concave), and the constraint
is affine. So problem (P5.3) is a convex optimization problem.
By using KKT conditions to solve problem (P5.3), the closedform solution for λ is derived in the following proposition.
Proposition 2. The optimal solution to problem (P5.3) is given
by
p
log2 (e)γ 2 W Θ + 1 − 1
∗
∗
∗
λ1 = λ2 = · · · = λN =
, (45)
γΘ
where e is the base of the natural logarithm, Θ is given by
Θ=
Proof. See Appendix G.
N
X
G2m,s
.
am
m=1
(46)
Proposition 2 shows that the optimal prices for the Stackelberg game with complete information are the same. This result
is consistent with that of the Stackelberg game with asymmetric information. As we explained before, this is because
the received power price is used in the Stackelberg games
10
with complete and asymmetric information. The DAP has no
motivation to treat the received power from EAPs differently.
Note that the Stackelberg game under complete information
can calculate an optimal price for each instantaneous channel
realization and equivalently the combination of EAPs’ types.
As such, it can adapt to the change of channel conditions.
As a comparison, the Stackelberg game under asymmetric
information can only calculate a single price no matter of
the change of the channel conditions, i.e., the change of the
combinations of the EAPs’ types.
TABLE I: System Settings
Parameters
1MHz
Energy cost coefficient am
[0.1,1]
dm,s
[5m,10m]
da,s
[15m,25m]
2
Power attenuation at reference distance of 1m
30dB
Noise power N0
10−8 mW
0.1
C
π , q )}
max E{UDAP
(π
π ,q
q)
(π
q2
s.t. πk − k ≥ 0, ∀k ∈ {1, . . . , K}
θk
(47)
-0.1
-0.2
peak values
-0.3
type 3
type 6
type 9
-0.4
C
π , q )} is given in (25).
where E{UDAP
(π
Since the DAP knows exactly the types of the EAPs, the
optimal prices are given by
qk2
, ∀k ∈ {1, . . . , K}
θk
0
Utilities of EAPs
In this part, the performance of centralized optimization
scheme, i.e., the optimal contract with complete information,
where the DAP knows exactly the types of the EAPs, is
presented. The centralized optimization problem is given as
follows.
πk∗ =
0.5
Bandwidth W
Path-loss coefficient α
C. Centralized Optimization
(P5.4) :
Values
Energy harvesting efficiency η
(48)
-0.5
1
2
3
4
5
6
7
8
9
10
Contract item for a certain type EAP
Fig. 2: Utilities of EAPs with type 3, type 6 and type 9 as
functions of contract items designed for all kinds of EAPs
from type 1 to type 10. We set N = 5 and K = 10.
We substitute πk in (47) with πk∗ and get
VI. S IMULATIONS
C
π ∗ , q )}
max E{UDAP
(π
(P5.5) :
q
(49)
s.t. q ≥ 0
C
π ∗ , q )} is given by
where E{UDAP
(π
C
π ∗ , q )}
E{UDAP
(π
=
−n1
N NX
X
N−
···
n1 =0 n2 =0
(
"
Φn1 ,...,nK W log2
1+γ
PK−2
K
X
k=1
i=0
X
ni
nK−1 =0
nk qk
!
−
K
X
k=1
q2
nk k
θk
#)
.
(50)
Note that (50) is exactly the expectation of the social
welfare, which is defined in (27). Although we originally
optimize the utility function of the DAP in problem (P5.4),
it is consistent with the optimization of the social welfare,
which is similar case in the design of contract theory as we
mentioned before.
It can also be observed that problem (P5.5) is a convex optimization problem. This is because each term in the summation
is composed a logarithm function (concave) and quadratic
functions (concave), the summation of concave functions are
still a concave function, and the constraint is affine. We can
get the solution of problem (P5.5) by solving the system of
equations given by KKT conditions, which is omitted here as
it is similar to that in Appendix A.
AND
D ISCUSSIONS
In this section, we first evaluate the feasibility of the
proposed contract, and then compare the performance of the
proposed incentive mechanisms. The performance of centralized optimization scheme is also simulated as the upper bound.
The main system parameters are shown in Table I. Since
θ = G2m,s /am and γ = ηGa,s /N0 , the practical ranges of θ
and γ can be determined by the parameters shown in Table I. In
the simulations, K types of EAPs are first generated randomly
and used as the set of EAP types. Then each of N EAPs in
the market will choose one type from the set of EAP types
uniformly, and thus the DAP’s type θ is uniformly distributed.
The unit of achievable throughput is set as Mbps.
To verify the feasibility (i.e., IR and IC) of the proposed
scheme under information asymmetry, the utilities of EAPs
with type 3, type 6 and type 9 are plotted in Fig. 2 as
functions of all contract items (qk , πk ), k ∈ 1, 2, . . . , K. We
can see from Fig. 2 that each of the utility achieves its peak
value only when it chooses the contract item designed for its
corresponding type, which indicates that the IC constraint is
satisfied. For example, for the type 6 EAP, its utility achieves
the peak value only when it selects the contract item (q6 , π6 ),
which is exactly designed for its type. If the type 6 EAP selects
any other contract item (qk , πk ), k ∈ 1, 2, . . . , K and k 6= 6,
its utility will reduce. Moreover, when each of above type
EAPs (i.e., type 3, type 6 and type 9) chooses the contract
11
2.2
3.5
2
1.8
Centralizd optimization, Complete information
Contract theory,Asymmetric infomation
Stackelberg game, Asymmetric infomation
Stackelberg game, Complete information
Centralizd optimization, Complete information
Contract theory, Asymmetric infomation
Stackelberg game, Asymmetric infomation
Stackelberg game, Complete information
3
1.6
Social Welfare
Social Welfare
2.5
1.4
1.2
2
1
1.5
0.8
0.6
1
0.4
0.2
0.8
0.5
1
1.2
1.4
1.6
1.8
2
2
2.2
Fig. 3: Social welfare as a function of γ. We set N = 2 and
K = 5.
3
4
5
6
7
8
9
10
Fig. 5: Social welfare as a function of N . We set K = 2,
γ = 2.2 and K = 2, 3, . . . , 10.
0.9
1
0.85
0.8
0.75
0.7
Normalized Social Welfare
Normalized Social Welfare
0.9
Contract theory, Asymmetric infomation
Stackelberg game, Asymmetric infomation
Stackelberg game, Complete information
0.8
0.65
0.6
0.55
0.5
0.7
0.6
0.5
0.4
0.45
0.3
0.4
0.8
1
1.2
1.4
1.6
1.8
2
2.2
Contract theory, Asymmetric infomation
Stackelberg game, Asymmetric infomation
Stackelberg game, Complete information
0.2
2
Fig. 4: Normalized social welfare as a function of γ. We set
N = 2 and K = 5.
item designed for its corresponding type, the utilities are
nonnegative. Note that similar phenomenon can be observed
for all other types of EAPs when they select the contract item
designed for their corresponding types, which are not shown
in Fig. 2 for brevity. In this sense, the IR condition is satisfied.
It can be concluded that utilizing the proposed scheme, EAPs
will automatically reveal its type to the DAP after selecting
the contract item. This means that using the proposed scheme,
the DAP can capture the EAPs’ private information (i.e., its
type), and thus effectively address the problem of information
asymmetry.
To evaluate the performance of the proposed schemes, we
compare the social welfare of the contract, Stackelberg games
and the upper bound. Fig. 3 plots the social welfare of these
schemes as a function of γ. It can be observed from Fig.
3 that the utilities achieved by allPschemes increase with
N
γ. This is because with the same m=1 qm , the larger the
value of γ, the larger the achievable throughput Rsa (refer
3
4
5
6
7
8
9
10
Fig. 6: Normalized social welfare as a function of N . We set
K = 2, γ = 2.2 and N = 2, 3, . . . , 10.
to (5)), and thus larger social welfare (refer to (27)). The
performance of the optimal scheme with complete information
providing the best performance serving as the upper bound.
The performance of contract scheme is generally better than
that of two Stackelberg games. This is because in contract
theory, the EAPs have limited contract items to choose from
and thus by using the contract theory, the DAP extracts more
benefits from the EAPs and leave less surplus for the EAPs.
However, in Stackelberg games, the EAPs have the freedom
to optimize its individual utility function and thus can reserve
more surplus. So the performance of the Stackelberg games are
inferior than that of the contract scheme. We can also observe
that the Stackelberg game with asymmetric information is
inferior than that with complete information. This is because
without complete information, the Stackelberg game fails to
adapt to the change of the channel, and thus the performance
becomes worse.
12
Fig. 4 shows the normalized social welfare as a function of
γ, where social welfare of the contract and Stackelberg games
are normalized by the upper bound. It can be seen in Fig. 4
that when γ is small, the social welfare of contract can initially
achieve more than 85% of that of the centralized optimization
scheme with complete information, and gradually approach to
it with the increasing of γ. This demonstrates that the proposed
incentive mechanism can effectively mitigate the effects of
information asymmetry by leveraging contract theory. While
the performance of the Stackelberg game with complete information is generally less than 75% of that of the optimal
scheme with complete information. Moreover, the performance
of the Stackelberg game with asymmetric information is even
worse, which is generally less than 50% of that of the optimal
scheme with complete information. The above results show
that by using the monopoly position in contract theory to
provide limited contract items, the contract can achieve good
performance close to the optimal centralized optimization with
complete information. However, in Stackelberg games, the
DAP grants some freedom for the EAPs to do optimizations,
which are selfish and do not care about social welfare. As
such, its performance in terms of social welfare is degraded.
To explore the impact of total EAP number N in the market,
we plot the curves of the social welfare and the normalized
social welfare of the contract, Stackelberg games and the upper
bound in Fig. 5 and Fig. 6. In Fig. 5, the social welfare of
these schemes is plotted as a function of N . We can observe
from Fig. 5 that the utility functions achieved by all three
schemes of upper bound, contract, Stackelberg with complete
information increase with N . This is because the overall social
welfare increases with the number of EAPs in the market.
The more EAPs in the market, the larger the summation
of utility functions of all the DAP and EAPs. However,
the Stackelberg game under asymmetric information decrease
slightly as N increases. This is because the Stackelberg game
under asymmetric information fails to adapt to the change
of the combinations of the EAPs’ types. As we mentioned
before, it can only calculate one single price for all the
combinations of EAPs’ types. The more EAPs in the market,
the more diverse combinations of the EAPs’ types. As such the
performance of the Stackelberg under asymmetric information
become worse. As a comparison, the Stackelberg game with
complete information can calculate a price targeting a certain
combination of EAPs’ types in the market. So it provides
better performance than that of its asymmetric counterpart.
While the contract leverages its monopoly status in the market
structure to provide a limited group of contract items for
the EAPs to choose from. Therefore, contract theory-based
scheme provides the better performance than that of both
Stackelberg games and close to the performance of the upper
bound.
In Fig. 6, the normalized social welfare of these schemes is
plotted as a function of N , where social welfare of the contract
and Stackelberg games are normalized by the upper bound. It
can be seen in Fig. 6 that when N = 2, the social welfare
of contract can initially gain more than 95% of that of the
centralized optimization scheme with complete information,
and gradually approach to it with the increasing of N . This
proves that the effects of information asymmetry can be
mitigated successfully by leveraging contract theory. While the
Stackelberg game with complete information can only provide
the normalized social welfare of less than 75%. Besides,
the performance of the Stackelberg game with asymmetric
information is even worse, which is generally less than 50%
of that of the optimal scheme with complete information
and decrease significantly with the increasing of N in the
market. This is because the more EAPs in the market, the
more diverse the combinations of EAPs’ types will be. The
Stackelberg game under asymmetric information cannot adapt
to the change of the combinations of EAPs’ types as it can
only calculate a single price for all possible combinations of
EAPs’ types.
VII. C ONCLUSIONS
AND
F UTURE W ORK
In this paper, we developed incentive mechanisms under
complete and asymmetric information to unveil the impact
of information asymmetry and market structure. Specifically,
we developed a contract based incentive mechanisms for the
wireless energy trading in radio frequency energy harvesting
(RFEH) based Internet of Things (IoT) systems under asymmetric information. In the asymmetric information scenario,
a Stackelberg game based scheme is also formulated as a
comparison. In complete information, the existing Stackelberg
game is extended from unified pricing into discriminative
pricing as a comparison. In the simulations, it was shown that
the Stackelberg game degrades significantly without complete
information, and the performance of the contract scheme under
asymmetric information is better than that of the Stackelberg
scheme with complete information. It can be concluded that
the performance of the considered system depends largely on
the market structure (i.e., whether the EAPs are allowed to
optimize their received power at the IoT devices with full
freedom or not) than on the information scenarios (i.e., the
complete or asymmetric information).
In our future work, we could consider both information
asymmetry as well as hidden action. In this scenario, the
DAP is not aware of the private information of EAPs and
it cannot distinguish the actions taken by different EAPs, i.e.,
the received power contributed by different EAPs. Because the
actions of EAPs are hidden from the DAP, some EAPs may get
the reward of the group without paying any efforts, which leads
to the free-rider problem. In this case, another mathematical
tool from the economics, known as the moral hazard in teams,
has a good potential to design effective incentive mechanisms
for this new scenario.
A PPENDIX
A. Proof of Proposition 1
In this part, we will prove the proposition 1. Because
the problem (P4.3) is a convex optimization problem, KKT
conditions are the sufficient and necessary conditions for the
optimal solution. The KKT conditions of problem (P4.3) are
given as follows.
13
The first-order necessary condition are given by
−W log (e)γn θ
X
1 1
2
Φn1 ,...,nK
+
n
θ
λ
1
1
1
K
P
n1 ,...,nK
n
θ
λ
2
+
γ
k
k
k
k=1
−
µ
=
0
1
−W log (e)γn θ
X
2 2
2
+
n
θ
λ
Φn1 ,...,nK
2
2
2
K
P
n ,...,n
1
K
nk θk λk
2+γ
k=1
−
µ
=
0
2
..
.
−W log (e)γn θ
X
N N
2
+
n
θ
λ
Φ
N
N
N
n
,...,n
1
K
K
P
n1 ,...,nK
n
θ
λ
2
+
γ
k k k
k=1
− µN = 0
(51)
where θk ≥ 0 are the types of EAPs, µk ≥ 0 are KKT
multipliers, λm ≥ 0 are the prices, and k = 1, 2, . . . , K. The
complementary slackness condition is given by
µ1 λ1 + µ2 λ2 + · · · + µK λK = 0.
(52)
Since µk ≥ 0 and λk ≥ 0, k = 1, 2, . . . , K hold, (52) becomes
µ1 λ1 = 0
µ2 λ2 = 0
(53)
..
.
µK λK = 0
To get the optimal solution of the KKT conditions, we need
to solve the equation system consists of (51) and (53) in terms
of µk ≥ 0 and λk ≥ 0, k = 1, 2, . . . , N , which is a system of
quadratic equations. Now we will discuss the combinations of
active or inactive constraints in KKT condition. Let first test
if µ1 = µ2 = · · · = µK = 0 leads to a valid solution. By
substitute µ1 , µ2 , . . . , µK with µ1 = µ2 = · · · = µK = 0 in
(51), we have
λ)
λ1 ∆ = W log2 (e)γΩ(λ
λ2 ∆ = W log2 (e)γΩ(λ
λ)
(54)
.
..
λ)
λK ∆ = W log2 (e)γΩ(λ
where ∆ is given by
∆=
X
Φn1 ,...,nK = 1,
(55)
The above system of equations in (54) can be solved
numerically. Note that the right term of each equation in (54)
are the same, so we can conclude that
λ ).
λ1 = λ2 = · · · = λK = W log2 (e)γΩ(λ
(57)
Because problem (P4.3) is a convex optimization problem,
we can conclude that this solution given by KKT conditions
is the solution of original optimization problem.
B. Proof of Lemma 2
The proof is conducted in two parts. First, we prove if qi >
qj , then πi > πj . Due to the IC constraints in (30), we have
πi −
qj2
qi2
≥ πj − ,
θi
θi
(58)
and equivalently,
θi (πi − πj ) ≥ qi2 − qj2 = (qi + qj )(qi − qj ).
(59)
Since qi > qj , we have
θi (πi − πj ) ≥ qi2 − qj2 = (qi + qj )(qi − qj ) > 0,
(60)
and thus πi > πj .
Next we prove if πi > πj , then qi > qj . Due to the IC
constraints in (30), we have
πj −
qj2
q2
≥ πi − i ,
θj
θj
(61)
and equivalently,
(qi + qj )(qi − qj ) = qi2 − qj2 ≥ θi (πi − πj ).
(62)
Since πi > πj , then we have
(qi + qj )(qi − qj ) = qi2 − qj2 ≥ θi (πi − πj ) > 0,
(63)
and thus qi > qj . This completes the proof.
C. Proof of Lemma 4
We prove this by contradiction. Suppose that there exists
πi < πj when θi > θj . We have
(πi − πj )(θi − θj ) < 0.
(64)
Due to the IC constraints, we also have
πi −
qj2
qi2
≥ πj − ,
θi
θi
(65)
πj −
qj2
q2
≥ πi − i .
θj
θj
(66)
and
n1 ,...,nK
λ ), λ = [λ1 , λ2 , . . . , λK ]T is given by
and Ω(λ
X
Φn1 ,...,nK
λ) =
.
Ω(λ
K
P
n1 ,...,nK 2 + γ
nk θk λk
k=1
Combine (65) and (66), we have
(πi − πj )(θi − θj ) ≥ 0,
(56)
(67)
which is in contradiction with (64). So if θi > θj , then πi >
πj .
14
D. Proof of Lemma 5
Due to the IC condition, ∀k ∈ {2, . . . , K}, we have
qk2
q2
≥ π1 − 1 .
(68)
θk
θk
Since we have defined that θ1 < θ2 < · · · < θK , we also have
πk −
π1 −
q12
q12
≥ π1 − .
θk
θ1
Combine (68) and (69), we have
(69)
q2
qk2
≥ π1 − 1 ≥ 0.
(70)
θk
θ1
Note that (70) shows that with the IC condition, if the IR
condition of the EAP with type θ1 holds, the IR condition of
the other K − 1 types will also hold. So the other K − 1 IR
conditions can be bind into the IR condition of the EAP with
type θ1 .
So far, we have proved that type θi+1 will prefer contract item
(qi+1 , πi+1 ) rather than contract item (qi−1 , πi−1 ). By using
(79), it can be extended downward until type θ1 , and thus all
DIC holds.
πi+1 −
2
q2
qi+1
q2
≥ πi−1 − i−1 ≥ . . . ≥ π1 − 1 , ∀i.
θi+1
θi+1
θ1
(80)
So we conclude that with the monotonicity and the LDIC, the
DIC holds. Similarly, we can prove that with the monotonicity
and the LUIC, the UIC holds.
πk −
E. Proof of Lemma 6
There are K(K − 1) IC constraints in (30), which can be
divided into K(K − 1)/2 downward incentive compatibility
(DIC)4 , given by
πi −
qj2
qi2
≥ πj − , ∀i, j ∈ {2, . . . , K}, i > j,
θi
θi
(71)
and K(K − 1)/2 upward incentive compatibility (UIC), given
by
qj2
qi2
≥ πj − , ∀i, j ∈ {2, . . . , K}, i < j,
(72)
θi
θi
Let’s first prove the DIC can be reduced as the LDIC. By
using the LDIC for three continuous types, θi−1 < θi <
θi+1 , ∀i ∈ {2, . . . , K − 1}, we have
πi −
πi+1 −
2
qi+1
q2
≥ πi − i ,
θi+1
θi+1
(73)
q2
qi2
≥ πi−1 − i−1 , ∀i.
(74)
θi
θi
By applying the monotonicity, i.e., if θi > θj , then πi > πj ,
∀i, j ∈ {1, . . . , K}, we have
πi −
(θi+1 − θi )(πi − πi−1 ) ≥ 0,
(75)
θi+1 (πi − πi−1 ) ≥ θi (πi − πi−1 ),
(76)
F. Proof of Lemma 7
We will first prove that the LDIC can be simplified as πk −
2
qk2 /θk = πk−1 − qk−1
/θk , which together with monotonicity
can ensure the LUIC hold.
For the reduced IR constraint π1 − q12 /θ1 ≥ 0 in (34), the
DAP will lower the π1 as possible as it can to improve the
optimization objective function E{UDAP }, until π1 − q12 /θ1 =
2
0. As for the LDIC, which is πi −qi2 /θi ≥ πi−1 −qi−1
/θi , ∀i ∈
{2, . . . , K}. Notice that the LDIC will still hold if both πi
and πi−1 are lowered by the same amount. To maximize the
optimization objective function, the DAP will lower all π − j
2
as much as possible until πi − qi2 /θi = πi−1 − qi−1
/θi . Note
that this process will not affect on other type’s LDIC. So the
2
LDIC can be reduced to πi − qi2 /θi = πi−1 − qi−1
/θi , ∀k ∈
{2, . . . , K}.
2
Next, we show that if πi − qi2 /θi = πi−1 − qi−1
/θi , ∀k ∈
{2, . . . , K} and the monotonicity hold, the LUIC holds. Since
2
/θi , ∀k ∈ {2, . . . , K}, and
we have πi − qi2 /θi = πi−1 − qi−1
equally it becomes
2
θi (πi − πi−1 ) = qi2 − qi−1
.
Because of monotonicity, i.e., if θi ≥ θi−1 , then πi ≥ πi−1 ,
we further have
θi (πi − πi−1 ) ≥ θi−1 (πi − πi−1 ).
(77)
Equally, (77) becomes
q2
q2
πi − i ≥ πi−1 − i−1 .
θi+1
θi+1
(78)
Combine (78) and (73), we have
q2
q2
πi+1 − i+1 ≥ πi−1 − i−1 .
θi+1
θi+1
(82)
Combine (81) and (82), we have
2
θi (πi − πi−1 ) = qi2 − qi−1
≥ θi−1 (πi − πi−1 ),
(83)
and equally we have
2
θi−1 πi − qi2 ≤ θi−1 πi−1 − qi−1
,
Combine (74) and (76), we have
2
θi+1 (πi − πi−1 ) ≥ θi (πi − πi−1 ) ≥ qi2 − qi−1
.
(81)
πi −
q2
qi2
≤ πi−1 − i−1 ,
θi−1
θi−1
(84)
(85)
which is exactly the LUIC condition. So the LUIC can be
removed from the constraints in (34).
G. Proof of Proposition 2
(79)
4 Note that K(K − 1)/2 is still an integer. Because K(K − 1) is the
multiplication of two continuous integers, which must be an even number. So
it is divisible by two.
In this part, we will prove the proposition 2. Since the problem (P5.3) is a convex optimization problem, KKT conditions
will be the sufficient and necessary conditions for the optimal
solution. To solve problem (P5.3), the KKT conditions are
given as follows.
15
The first-order necessary condition are given by
−W log2 (e)γθ1
+ θ1 λ1 − µ1 = 0
N
P
2+γ
θm λm
m=1
−W log2 (e)γθ2
+ θ2 λ2 − µ2 = 0
N
P
2+γ
θm λm
m=1
..
.
−W log2 (e)γθN
+ θN λN − µN = 0
N
P
2+γ
θm λm
ACKNOWLEDGMENT
The authors would like to thank Prof. Zhu Han for his
helpful discussion. The authors also thank the editor and
anonymous reviewers for their valuable comments and suggestions, which improved the quality of the paper.
(86)
m=1
G2m,s
am ,
where θm =
µm ≥ 0 are KKT multipliers, and λm ≥
0, m = 1, 2, . . . , N .
The complementary slackness condition is given by
µ1 λ1 + µ2 λ2 + · · · + µN λN = 0.
(87)
Since µm ≥ 0 and λm ≥ 0, m = 1, 2, . . . , N hold, (87)
becomes
µ1 λ1 = 0
µ2 λ2 = 0
(88)
..
.
µN λN = 0
To get the optimal solution of the KKT conditions, we need
to solve the equation system consists of (86) and (88) in terms
of µm ≥ 0 and λm ≥ 0, m = 1, 2, . . . , N , which is a system
of quadratic equations.
Now we will discuss the combinations of active or inactive
constraints in KKT condition. Let first test if µ1 = µ2 = · · · =
µN = 0 leads to a valid solution. By substitute µ1 , µ2 , . . . , µN
with µ1 = µ2 = · · · = µN = 0 in (86), we have
!
N
X
θm λm = W log2 (e)γ
λ1 2 + γ
m=1
!
N
X
λ2 2 + γ
θm λm = W log2 (e)γ
(89)
m=1
.
..
!
N
X
θm λm = W log2 (e)γ
λN 2 + γ
m=1
By solving the system of equations of (89), we can get a
solution as given by
p
log2 (e)γ 2 W Θ + 1 − 1
λ1 = λ2 = · · · = λN =
, (90)
γΘ
where e is the base of the natural logarithm, Θ is given by
Θ=
N
X
θm .
(91)
m=1
Because problem (P5.3) is a convex optimization problem,
we can conclude that this solution given by KKT conditions
is the solution of original optimization problem.
R EFERENCES
[1] J. Gubbi, R. Buyya, S. Marusic, and M. Palaniswami, “Internet of things
(IoT): A vision, architectural elements, and future directions,” Future
Generation Computer Systems, vol. 29, no. 7, pp. 1645–1660, Sep. 2013.
[2] J. Lin, W. Yu, N. Zhang, X. Yang, H. Zhang, and W. Zhao, “A survey
on internet of things: Architecture, enabling technologies, security and
privacy, and applications,” IEEE Internet of Things Journal, vol. 4, no. 5,
pp. 1125–1142, Oct 2017.
[3] W. Hou, W. Li, L. Guo, Y. Sun, and X. Cai, “Recycling edge devices
in sustainable Internet of Things networks,” IEEE Internet of Things
Journal, vol. 4, no. 5, pp. 1696–1706, Oct 2017.
[4] Y. Peng, F. Al-Hazemi, R. Boutaba, F. Tong, I. S. Hwang, and C. H.
Youn, “Enhancing energy efficiency via cooperative MIMO in wireless
sensor networks: State of the art and future research directions,” IEEE
Communications Magazine, vol. 55, no. 11, pp. 47–53, Nov. 2017.
[5] H. Kawabata, K. Ishibashi, S. Vuppala, and G. T. F. de Abreu, “Robust
relay selection for large-scale energy-harvesting IoT networks,” IEEE
Internet of Things Journal, vol. 4, no. 2, pp. 384–392, April 2017.
[6] Z. Yang, W. Xu, Y. Pan, C. Pan, and M. Chen, “Energy efficient resource
allocation in machine-to-machine communications with multiple access
and energy harvesting for IoT,” IEEE Internet of Things Journal, vol. PP,
no. 99, pp. 1–1, 2017.
[7] P. Kamalinejad, C. Mahapatra, Z. Sheng, S. Mirabbasi, V. C. Leung,
and Y. L. Guan, “Wireless energy harvesting for the Internet of Things,”
IEEE Commun. Magazine, vol. 53, no. 6, pp. 102–108, Jun. 2015.
[8] D. Niyato, D. I. Kim, P. Wang, and L. Song, “A novel caching
mechanism for Internet of Things (IoT) sensing service with energy
harvesting,” in Proc. ICC, May. 2016, pp. 1–6.
[9] H. Chen, C. Zhai, Y. Li, and B. Vucetic, “Cooperative strategies
for wireless-powered communications: An overview,” IEEE Wireless
Communications, avaliable: https://arxiv.org/abs/1610.03527, 2017.
[10] K. Huang and V. K. Lau, “Enabling wireless power transfer in cellular
networks: Architecture, modeling and deployment,” IEEE Trans. Wireless Commun., vol. 13, no. 2, pp. 902–912, 2014.
[11] K. Huang and X. Zhou, “Cutting the last wires for mobile communications by microwave power transfer,” IEEE Commun. Mag., vol. 53,
no. 6, pp. 86–93, 2015.
[12] C. Zhong, X. Chen, Z. Zhang, and G. K. Karagiannidis, “Wirelesspowered communications: Performance analysis and optimization,”
IEEE Trans. Commun., vol. 63, no. 12, pp. 5178–5190, 2015.
[13] Y. Chen, D. B. da Costa, and H. Ding, “Interference analysis in wireless
power transfer,” IEEE Communications Letters, vol. 21, no. 10, pp.
2318–2321, Oct 2017.
[14] L. Chen, W. Wang, and C. Zhang, “Stochastic wireless powered communication networks with truncated cluster point process,” IEEE Trans.
Veh. Technol., vol. PP, no. 99, pp. 1–1, 2017.
[15] J. H. Park, Y. S. Jeon, and S. Han, “Energy beamforming for wireless
power transfer in MISO heterogeneous network with power beacon,”
IEEE Communications Letters, vol. 21, no. 5, pp. 1163–1166, May 2017.
[16] T. A. Khan, R. W. Heath, and P. Popovski, “Wirelessly powered
communication networks with short packets,” IEEE Trans. Wireless
Commun., vol. PP, no. 99, pp. 1–1, 2017.
[17] X. Jiang, C. Zhong, X. Chen, T. Q. Duong, T. A. Tsiftsis, and Z. Zhang,
“Secrecy performance of wirelessly powered wiretap channels,” IEEE
Trans. Commun., vol. 64, no. 9, pp. 3858–3871, 2016.
[18] H. Chen, Y. Li, Z. Han, and B. Vucetic, “A stackelberg game-based
energy trading scheme for power beacon-assisted wireless-powered
communication,” in Proc. ICASSP, Apr. 2015, pp. 3177–3181.
[19] S. Sarma, K. Kandhway, and J. Kuri, “Robust energy harvesting based
on a stackelberg game,” IEEE Wireless Communications Letters, vol. 5,
no. 3, pp. 336–339, 2016.
[20] Y. Ma, H. Chen, Z. Lin, Y. Li, and B. Vucetic, “Distributed and
optimal resource allocation for power beacon-assisted wireless-powered
communications,” IEEE Trans. Commun., vol. 63, no. 10, pp. 3569–
3583, Aug. 2015.
[21] P. Bolton and M. Dewatripont, Contract theory. MIT press, 2005.
16
[22] Z. Hou, H. Chen, Y. Li, Z. Han, and B. Vucetic, “A contract-based
incentive mechanism for energy harvesting-based internet of things,” in
Proc. ICC, May 2017, pp. 1–6.
[23] T. Liu, J. Li, F. Shu, M. Tao, W. Chen, and Z. Han, “Design of contractbased trading mechanism for a small-cell caching system,” IEEE Trans.
Wireless Commun., vol. 16, no. 10, pp. 6602–6617, Oct 2017.
[24] Y. Zhang, L. Song, W. Saad, Z. Dawy, and Z. Han, “Contract-based
incentive mechanisms for device-to-device communications in cellular
networks,” IEEE J. Sel. Areas Commun., vol. 33, no. 10, pp. 2144–2155,
May 2015.
[25] L. Duan, L. Gao, and J. Huang, “Cooperative spectrum sharing: A
contract-based approach,” IEEE Trans. Mobile Commun., vol. 13, no. 1,
pp. 174–187, Nov. 2014.
[26] R. Zhang, L. Song, Z. Han, and B. Jiao, “Physical layer security for
two-way untrusted relaying with friendly jammers,” IEEE Trans. Veh.
Technol., vol. 61, no. 8, pp. 3693–3704, 2012.
[27] H. Zhang, Y. Xiao, L. X. Cai, D. Niyato, L. Song, and Z. Han, “A
multi-leader multi-follower stackelberg game for resource management
in lte unlicensed,” IEEE Trans. Wireless Commun., vol. 16, no. 1, pp.
348–361, 2017.
[28] Y. Xu and S. Mao, “Stackelberg game for cognitive radio networks
with mimo and distributed interference alignment,” IEEE Trans. Veh.
Technol., vol. 63, no. 2, pp. 879–892, 2014.
[29] S. Yin, E. Zhang, Z. Qu, L. Yin, and S. Li, “Optimal cooperation strategy
in cognitive radio systems with energy harvesting,” IEEE Trans. Wireless
Commun., vol. 13, no. 9, pp. 4693–4707, 2014.
[30] H. Chen, Y. Li, Y. Jiang, Y. Ma, and B. Vucetic, “Distributed power
splitting for swipt in relay interference channels using game theory,”
IEEE Trans. Wireless Commun., vol. 14, no. 1, pp. 410–420, 2015.
[31] T. Zhang, W. Chen, and F. Yang, “Balancing delay and energy efficiency
in energy harvesting cognitive radio networks: A stochastic stackelberg
game approach,” IEEE Trans. Cogn. Commun. Netw, vol. 3, no. 2, pp.
201–216, June 2017.
[32] I. Krikidis, G. Zheng, and B. Ottersten, “Harvest-use cooperative networks with half/full-duplex relaying,” in Proc. WCNC, Apr. 2013, pp.
4256–4260.
[33] H. Ju and R. Zhang, “Throughput maximization in wireless powered
communication networks,” IEEE Trans. Wireless Commun., vol. 13,
no. 1, pp. 418–428, Dec. 2014.
[34] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless networks
with RF energy harvesting: A contemporary survey,” IEEE Commun.
Surveys Tuts., vol. 17, no. 2, pp. 757–789, Nov. 2015.
[35] A.-H. Mohsenian-Rad, V. W. Wong, J. Jatskevich, R. Schober, and
A. Leon-Garcia, “Autonomous demand-side management based on
game-theoretic energy consumption scheduling for the future smart
grid,” IEEE Trans. Smart Grid, vol. 1, no. 3, pp. 320–331, Nov. 2010.
[36] S. Boyd and L. Vandenberghe, Convex optimization.
Cambridge
university press, 2004.
| 7cs.IT
|
MAXIMAL COHEN-MACAULAY MODULES OVER CERTAIN SEGRE
PRODUCTS
arXiv:1802.04786v2 [math.AC] 7 Mar 2018
LINQUAN MA
Dedicated to Professor Gennady Lyubeznik on the occasion of his 60th birthday
Abstract. We prove some results on the non-existence of rank one maximal CohenMacaulay modules over certain Segre product rings. As an application we show that over
these Segre product rings there do not exist maximal Cohen-Macaulay modules with multiplicity less than or equal to the parameter degree of the ring. This disproves a conjecture
of Schoutens [Sch14].
1. Introduction
A finitely generated module M over a Noetherian local ring (R, m) is called a maximal
Cohen-Macaulay module, abbreviated as MCM or small MCM,1 if depth M = dim R. A
fundamental and long-standing open question in commutative algebra is that whether every
complete local ring (R, m) admits a finitely generated MCM. This question was known to
be true if dim R ≤ 2, if dim R = 3 and R is N-graded over a field of characteristic p > 0
(Hartshorne–Peskine-Szpiro–Hochster, see [Hoc75]), and in some other special cases (e.g.,
affine toric rings [Hoc72]). The importance of this question lies in the fact that an affirmative
answer will imply Serre’s conjecture on positivity of intersection multiplicities [Hoc73].
Schoutens [Sch14] introduced a stronger version of MCM: in equal characteristic,2 M is
called a very small MCM over (R, m) if M is a finitely generated MCM and
e(M) ≤ min{l(R/I)|I is generated by a system of parameters}
where e(M) denotes the multiplicity of M with respect to the maximal ideal m in R and the
right hand side is called the parameter degree of R.
We note that for local rings of dimension ≤ 2 and for affine toric rings, very small MCM
do exist (it is not hard to see that the normalization will do the job in these cases [Hoc72]
[Hoc73]). Some other existence results were established in [Sch14]. Moreover, it turns out
that under mild conditions, if very small MCM exist in general in characteristic p > 0, then
a reduction to p > 0 technique will guarantee the existence in characteristic 0 [Sch14] (such
reduction process is not known for small MCM). Based on these facts, Schoutens made the
following conjecture:
The author is partially supported by NSF Grant DMS #1600198, and NSF CAREER Grant DMS
#1252860/1501102.
1“Small” here only means M is finitely generated. This notion was introduced in comparison with the
notion of big Cohen-Macaulay modules [Hoc75]: modules that are not necessarily finitely generated but are
Cohen-Macaulay in a natural sense. However, throughout this article we only deal with finitely generated
modules, hence sometimes we simply use MCM instead of small MCM.
2In mixed characteristic the definition of very small MCM is slightly different, we will not work with mixed
characteristic in this paper.
1
Conjecture 1.1 (cf. Conjecture 1.1 in [Sch14]). Any compete local ring admits a very small
MCM.
We will make use of the Segre product of graded rings to obtain a family of counterexamples of Conjecture 1.1. The basic idea is as follows: we first prove in Section 2 that
under some conditions on the I-invariant and multiplicity, very small MCM must have rank
one, and then in Section 3 we do computations over a family of Segre product rings to show
that they satisfy the conditions in Section 2 and that rank one MCM does not exist over
these rings, thus these rings are counter-examples of Conjecture 1.1.
Acknowledgement. I would like to thank Craig Huneke, Anurag Singh, Ehsan Tavanfar,
Bernd Ulrich and Uli Walther for many valuable discussions. I would also like to thank the
referee for his/her comments on Conjecture 3.7 and for pointing out Remark 3.8.
2. Very small MCM and rank one MCM
In this section we make some elementary observations on very small MCM. We will see
that under certain conditions on R, the very small condition will force the module to have
rank one. We begin with the following definition.
Definition 2.1. Let (R, m) be a local ring of dimension n such that Hmi (R) has finite length
for every i < n. Then we define
n−1
X
n−1
l(Hmi (R))
I(R) =
i
i=0
to be the I-invariant of R.
Remark 2.2. It is well-known (for example, see [SV73] [SV78]) that when (R, m) is an excellent local domain that is Cohen-Macaulay on the punctured spectrum, Hmi (R) has finite
length for every i < n, and in this case we always have
0 ≤ l(R/(x)) − e(x, R) ≤ I(R)
for every system of parameters x = x1 , . . . , xn .
As an easy consequence, we observe the following:
Lemma 2.3. Suppose (R, m) is an excellent local domain that is Cohen-Macaulay on the
punctured spectrum. Suppose also that R/m is infinite. If I(R) < e(R), then every very
small MCM has rank one.
Proof. We may pick x = x1 , . . . , xn a minimal reduction of m since R/m is infinite. We know
that e(x, R) = e(R). If M is a very small MCM over R, then by Remark 2.2,
rank(M) · e(R) = e(M) ≤ min{l(R/I)|I is generated by a system of parameters}
≤ l(R/(x)) ≤ e(R) + I(R) < 2 · e(R).
This shows that M must have rank one.
When (R, m) is a normal local domain, every rank one MCM is reflexive, and hence
isomorphic to a pure height one ideal of R, therefore it corresponds to an element in the
class group Cl(R). From this we immediately get:
2
Proposition 2.4. Suppose (R, m) is an excellent local UFD of dimension 3, with R/m
infinite. Suppose R is not Cohen-Macaulay and I(R) < e(R), then R does not admit very
small MCM.
Proof. It is easy to see that R satisfies the hypothesis of Lemma 2.3. Hence every very small
MCM must have rank one. Since R is a UFD, Cl(R) is trivial and thus every rank one
reflexive module is isomorphic to R. Since R is not Cohen-Macaulay, it follows that rank
one MCM does not exists.
In general, constructing UFDs that are not Cohen-Macaulay turns out to be difficult. In
fact, over the complex numbers, complete local UFDs of dimension ≤ 4 are Cohen-Macaulay
[HO74] (this is not true if we do not assume completeness, see [Tav17]). However, examples
of (R, m) that satisfies Proposition 2.4 do exist in positive characteristic, see Example 3 and
Theorem 2.5 in [MS11] (here the I-invariant is 1, so the condition I(R) < e(R) is obvious).
Although this example only works in characteristic 2 and is not complete, it already strongly
suggests that Conjecture 1.1 might be false in general. In the next section we will use Segre
product to construct complete local rings in arbitrary equal characteristic that do not admit
very small MCM.
3. Computations over Segre products
Segre product. Let A and B be N-graded rings over a field A0 = B0 = K. The Segre
product of A and B is the ring
R = A#B = ⊕j≥0 Aj ⊗K Bj
This ring has a natural grading in which Rj = Aj ⊗K Bj . If M and N are graded modules over
A and B respectively, their Segre product is the graded R-module M#N = ⊕j≥0 Mj ⊗K Nj .
If A and B are both normal domains, then R = A#B is also normal since it is a direct
summand of the normal ring A ⊗K B. For reflexive modules M and N over A and B
respectively, we have the Kunneth formula for local cohomology [GW78]:
(3.0.1)
Hmq R (M#N) = M#Hmq B (N) ⊕ Hmq A (M)#N ⊕ (⊕i+j=q+1 Hmi A (M)#Hmj B (N))
where mR , mA , mB denote the homogeneous maximal ideals of the corresponding rings.
Connections with sheaf cohomology. Suppose A and B are normal N-graded K-algebras
generated over K by degree one forms. Let X = Proj A and Y = Proj B, then R is the section
ring of X × Y with respect to the very ample line bundle OX (1) ⊠ OY (1).
Remark 3.1. Now suppose A is a UFD and B is a polynomial ring, both have dimension
≥ 2. Then Cl(X) ∼
= Z and Y is a projective space, so Cl(X × Y ) = Z × Z. It follows from
Excercise II.6.3 in [Har77] that Cl(R) = Z, and is generated by the class of
⊕∞ H 0(X × Y, OX (1 + i) ⊠ OY (i)) ∼
= A[1]#B.
i=0
Main result. Now we can state and prove our result on Segre products. We first recall that the a-invariant of an N-graded Cohen-Macaulay ring R is defined to be aR =
R
sup{j|[Hmdim
(R)]j 6= 0}.
R
Lemma 3.2 (cf. Example 4.4.13 in [GW78]). Suppose dim A ≥ 2 and dim B ≥ 2. Then
A[n]#B is MCM over R = A#B if and only if aA < n < −aB . In particular there exists n
such that A[n]#B is MCM if and only if aA + aB ≤ −2.
3
Proof. This follows directly from (3.0.1): the possible non-vanishing lower local cohomology
B
A
modules of A[n]#B come from A[n]#Hmdim
(B) and Hmdim
(A)[n]#B, and these are both
B
A
zero if and only if aA < n < −aB .
Remark 3.3. Although A[n]#B is not MCM for any n when aA + aB > −2, R = A#B might
still have rank one MCM in many cases. For example, it follows from Proposition 2.3 (and
its proof) in [Ma88] that if X = Proj A and Y = Proj B are smooth projective curves, then
R has a small MCM of rank one.
1 ,...,xn ]
Theorem 3.4. Let A = K[x
be a graded complete intersection with isolated singularity
(f1 ,...,fh )
at mA . Let B = K[s, t] and R = A#B. Set d = deg f1 + deg f2 + · · · + deg fh . If d > n and
bmR do not have any rank one MCM.
dim A = n − h ≥ 4, then R and R
Proof. Since A is a complete intersection that is regular in codimension 3, A is a UFD by
Corollaire XI 3.14 in [Gro68]. By Remark 3.1 we know that Cl(R) = Z and is generated by
the class of
P = A[1]#B ∼
= (x1 s, x2 s, . . . , xn s).
∼
It is easy to check that Q = A[−1]#B = (x1 s, x1 t) represents the inverse of [P ] in Cl(R).
Since Cl(R) = Z, every rank one MCM must be isomorphic to a symbolic power of P or Q,
which is A[n]#B for some n (this is because, under the notation of Remark 3.1, the class of
0
A[n]#B ∼
= ⊕∞
i=0 H (X × Y, OX (n + i) ⊠ OY (i)) is n times the class of P = A[1]#B in the
divisor class group Cl(R), and since it is a rank one reflexive module, it must be isomorphic
to P (n) ). However, since A and B are normal Cohen-Macaulay rings, Lemma 3.2 shows that
there is no such n because aA + aB = d − n − 2 > −2.
bmR , it is enough to observe that Cl(R
bmR ) is still generated by [P R
bmR ].
In the case of R
But this follows from a result of Flenner (see 1.5 of [Fle81]): if R is N-graded, R2 and S3 ,
bmR ) is an isomorphism. Since our ring R is an isolated
then the natural map Cl(R) → Cl(R
singularity with depth R = dim R − 1 = dim A ≥ 4, R satisfies the hypothesis.
As an application of Theorem 3.4, we obtain the following which disproves Conjecture 1.1:
Corollary 3.5. Let R = K[x1f,...,xn ] #K[s, t] with K an arbitrary infinite field. Suppose f is a
degree n + 1 homogeneous polynomial with isolated singularity at the origin with n ≥ 5 (e.g.,
bmR does not admit
one can take f = xn+1
+ xn+1
+ · · · + xn+1
when char(K) ∤ n + 1). Then R
1
2
n
very small MCM.
n]
and B = K[s, t]. Note that dim A = n − 1 and dim R = n. Since
Proof. Set A = K[x1,...,x
f
dimK [R]i = dimK [A]i · dimK [B]i , the leading term in the Hilbert polynomial of R is
e(A) n−2
n + 1 n−1
i
·i=
i .
(n − 2)!
(n − 2)!
Hence we have
bmR ) = e(mR , R) =
e(R
n+1
× (n − 1)! = (n + 1)(n − 1).
(n − 2)!
On the other hand, using (3.0.1) it is easy to see that
Hmq R (R) = 0 for q ≤ n − 2 and Hmn−1
(R) = Hmn−1
(A)#B.
R
A
4
In particular,
b mR ) =
I(R
n−1
X
n−1
i=0
i
bmR )) = lR (Hmn−1 (R))
l(Hmi R (R
R
= dimK ([Hmn−1
(A)]0 ⊗ [B]0 ) + dimK ([Hmn−1
(A)]1 ⊗ [B]1 )
A
A
= n+2
bmR ) < e(R
bmR ). Since R and hence R
bmR are normal isolated singularities,
Therefore I(R
m
b R must have rank one. But since deg f > n
Lemma 2.3 tells us that very small MCM over R
m
b R does not have any rank one MCM.
and n ≥ 5, Theorem 3.4 implies R
Remark 3.6. It is perhaps interesting to point out that in our Theorem 3.4 (or Corollary
bmR do admit small MCM of higher rank. This follows from Proposition 3.2.2
3.5), R and R
of [Han99] and the main result of [HUB91].
Motivated by Theorem 3.4 and Remark 3.6, I hazard the following conjecture:
Conjecture 3.7. For every integer N, there exists a complete local domain R that does not
admit small MCM of rank ≤ N.
Remark 3.8. We point out that, for every N, there exists a complete local domain R that
does not admit non-free MCM of rank ≤ N. Suppose R is a hypersurface with singular locus
1 ,...,xn ]]
of codimension c ≥ 2N + 2 (e.g., one can take R = K[[x
for n ≥ 2N + 3, where K is a
x21 +···+x2n
perfect field of char(K) 6= 2), it then follows from Corollary 2.2 in [Bru81] that any non-free
MCM M has rank at least c−1
> N.
2
References
[Bru81]
[Fle81]
[GW78]
[Gro68]
[Han99]
[Har77]
[HO74]
[HUB91]
[Hoc72]
[Hoc73]
W. Bruns: The Eisenbud-Evans generalized principal ideal theorem and determinantal ideals,
Proc. Amer. Math. Soc. 83 (1981), no. 1, 19–24.
H. Flenner: Divisorenklassengruppen quasihomogener Singularitäten, J. Reine Angew. Math.
328 (1981), 128–160.
S. Goto and K. Watanabe: On graded rings I, J. Math. Soc. Japan 30 (1978), no. 2, 179–213.
A. Grothendieck: Cohomologie Locale des Faisceaux Cohérents et Théorèmes de Lefschetz Locaux et Globaux (SGA2), Augmenté d’un exposé par Michèle Raynaud. Séminaire de Géométrie
Algébrique du Bois-Marie, 1962, Advanced Studies in Pure Mathematics, vol. 2, North-Holland
Publishing Co., Amsterdam; Masson & Cie, Éditeur, Paris, 1968, pp. vii+287.
D. Hanes: Special conditions on maximal Cohen-Macaulay modules, and applications to the theory
of multiplicities, Thesis, University of Michigan (1999).
R. Hartshorne: Algebraic geometry, Springer-Verlag, New York, 1977, Graduate Texts in Mathematics, No. 52.
R. Hartshorne and A. Ogus: On the factoriality of local rings of small embedding codimension,
Comm. Algebra 1 (1974), 415–437.
J. Herzog, B. Ulrich, and J. Backelin: Linear maximal Cohen-Macaualy modules over strict
complete intersections, J. Pure Appl. Algebra 71 (1991), no. 2-3, 187–202.
M. Hochster: Rings of invariants of tori, Cohen-Macaulay rings generated by monomials, and
polytopes, Ann. of Math. (2) 96 (1972), 318–337.
M. Hochster: Cohen-Macaulay modules, Conference on Commutative Algebra (Lawrence,
Kansas, 1972), Lecture Notes in Mathematics, no. 311, Springer-Verlag, Berlin, 1973, pp. 120–
152.
5
[Hoc75] M. Hochster: Big Cohen-Macaulay modules and algebras and embeddability in rings of Witt
vectors, Conference on Commutative Algebra-1975 (Queen’s Univ., Kingston, Ont., 1975), Queen’s
Papers in Pure and Applied Math., no. 42, Queen’s Univ., Kingston, Ont., 1975, pp. 106–195.
[Ma88] F. Ma: Splitting in integral extesions, Cohen-Macaulay modules and algebras, J. Algebra 116
(1988), no. 1, 176–195.
[MS11] A. Marcelo and P. Schenzel: Non-Cohen-Macaulay unique factorization domains in small
dimensions, J. Symbolic Comput. 46 (2011), no. 5, 609–621.
[Sch14] H. Schoutens: Maximal Cohen-Macaulay modules over local toric rings, arXiv: 1408.6220.
[SV78] J. Stückrad and W. Vogel: Toward a theory of Buchsbaum singularities, Amer. J. Math 100
(1978), no. 4, 727–746.
[SV73] J. Stückrad and W. Vogel: Eine Verallgemeinerung der Cohen-Macaulay-Ringe und anwendungen auf ein Problem der Multiplizitätstheorie, J. Math. Kyoto Univ. 13 (1973), 513–528.
[Tav17] E. Tavanfar: Reduction of the small Cohen-Macaulay conjecture to excellent unique factorization
domains, Arch. Math. (Basel) 109 (2017), no. 5, 429–439. 3710763
Department of Mathematics, University of Utah, Salt Lake City, UT 84112
E-mail address: [email protected]
6
| 0math.AC
|
1
Cooperative Localization for Mobile Networks:
A Distributed Belief Propagation – Mean Field
Message Passing Algorithm
arXiv:1512.07782v2 [cs.SY] 3 Apr 2016
Burak Çakmak, Daniel N. Urup, Florian Meyer, Member, IEEE, Troels Pedersen, Member, IEEE,
Bernard H. Fleury, Senior Member, IEEE, and Franz Hlawatsch, Fellow, IEEE
Abstract—We propose a hybrid message passing method
for distributed cooperative localization and tracking of mobile
agents. Belief propagation and mean field message passing are
employed for, respectively, the motion-related and measurementrelated part of the factor graph. Using a Gaussian belief approximation, only three real values per message passing iteration
have to be broadcast to neighboring agents. Despite these very
low communication requirements, the estimation accuracy can
be comparable to that of particle-based belief propagation.
Index Terms—Belief propagation, mean field approximation,
cooperative localization, distributed estimation, information projection, Kullback-Leibler-divergence, mobile agent network.
I. I NTRODUCTION
Cooperative localization is a powerful approach for mobile
networks [1]–[5]. An attractive methodology for cooperative
localization is sequential Bayesian estimation via message
passing algorithms [6]. In particular, distributed belief propagation (BP) message passing algorithms were proposed in
[2], [3], [7]–[11] to localize static or mobile agents. Feasible
implementations involve certain approximations and use, e.g.,
particle methods [2], [3], [8]–[10] or the sigma point technique
[11]. Each message transmitted between neighboring agents
is a set of hundreds or more particles in the former case
[2], [3], [8] and a mean and a covariance matrix, i.e., five
real numbers in 2-D localization, in the latter case. For static
agents, also message passing algorithms based on expectation
propagation [12], [13] or the mean field (MF) approximation
[14] were proposed. Similarly to sigma point BP [11], they
use a Gaussian approximation and the transmitted messages
consist of a mean and a covariance matrix.
In this letter, building on the theoretical framework in
[15], we present a distributed hybrid BP–MF message passing
method for cooperative localization and tracking of mobile
agents. We employ BP and MF [15] for, respectively, the
motion-related and measurement-related part of the underlying
factor graph, and we use a Gaussian belief approximation.
Each BP–MF iteration includes an information projection
Final manuscript, April 5, 2016.
B. Çakmak, T. Pedersen, and B. H. Fleury are with the Department of
Electronic Systems, Aalborg University, Aalborg, Denmark (e-mail: {buc,
troels, bfl}@es.aau.dk). D. N. Urup is with Danish Defence, Denmark (e-mail:
[email protected]). F. Meyer is with CMRE, La Spezia, Italy (e-mail: [email protected]). F. Hlawatsch is with the Institute of Telecommunications, TU Wien, Vienna, Austria (e-mail: [email protected]).
This work was supported by the European Commission in the framework
of the FP7 Network of Excellence NEWCOM# (grant 318306) and by the
Austrian Science Fund (FWF) under grant P27370-N30.
[16] that is efficiently implemented by means of a Newton
conjugate-gradient technique [17]. Our method can achieve
an accuracy comparable to that of BP-based methods with the
same communication cost as the MF method [14], i.e., three
real numbers per transmitted message in 2-D localization.
This letter is organized as follows. The system model is
described in Section II. The hybrid BP–MF scheme is developed in Section III, and the Gaussian belief approximation in
Section IV. Section V presents simulation results.
II. S YSTEM M ODEL
The mobile network at discrete time n ∈ {1, ..., N } is
described by a set of network nodes V n and a set of edges E n
representing the communication/measurement links between
n
the nodes. The set V n is partitioned into a set VM
of mobile
n
agents at unknown positions and a set VA of static anchors at
known positions. An edge (k, l) ∈ E n indicates the fact that
agent or anchor l transmits data to agent k and, concurrently,
agent k acquires a noisy measurement of its distance to agent
n
or anchor l. The edge set E n is partitioned into a set EM
n
of edges between certain agents, i.e., (k, l) ∈ EM implies
n
n
k, l ∈ VM
, and a set EMA
of edges between certain agents
n
n
n
and anchors, i.e., (k, l) ∈ EMA
implies k ∈ VM
and l ∈ VA
.
Information exchange between agents is bidirectional, i.e.,
n
n
(k, l) ∈ EM
implies (l, k) ∈ EM
. We consider a distributed
scenario where each agent knows only its own measurements.
Since the anchors have exact knowledge of their own position,
they do not need to acquire measurements and receive position
information from neighboring nodes. Accordingly, anchors
transmit position information to agents but not vice versa, i.e.,
n
n
.
implies (l, k) ∈
/ EMA
(k, l) ∈ EMA
n
n
Let the vector xk denote the state of
agent
k ∈ VM at1:ntime
n
n
n ∈ {1, ..., N }. Moreover, let x , xk k∈V n and x
,
M
i n
x i=1 . While our approach applies to any linear-Gaussian
motion model, we here consider specifically those two motion
models (MMs) that are most frequently used in practice. In
MM1, xnk = pnk ∈ R2 is the 2-D position of agent k at time
n. If agent k belongs to the network at times n and n−1, i.e.
n−1
n
∩ VM
, then pnk is assumed to evolve according to
k ∈ VM
the Gaussian random walk model [18]
√
pnk = pkn−1 + T v nk .
Here, T is the duration of one time step and v nk ∈ R2 is zeromean Gaussian driving noise with component variance σv2 .
2
Note that v nk can be interpreted as a random velocity. In MM2,
T
xnk = (pnk )T (v nk )T , where v nk ∈ R2 is the 2-D velocity of
n−1
n
agent k at time n. For k ∈ VM
∩ VM
, xnk is assumed to
evolve according to the constant velocity model [18]
xnk = F xkn−1 + Gank .
(1)
Here, ank ∈ R2 is zero-mean Gaussian driving noise (a random
acceleration)
with component
variance
σa2 . Moreover, F =
i
i
h
h
T 2 /2
1 T
0 1
⊗ I2 , where ⊗ denotes the
⊗ I2 and G =
T
Kronecker product and Im is the m×m identity matrix. Note
that in both MM1 and MM2, the state-transition probability
density function (pdf) p(xnk |xkn−1 ) is Gaussian. For agents
that are part of the network at time n but not at time n − 1,
n−1
n
i.e., k ∈ VM
\ VM
, we set p(xnk |xkn−1 ) = p(xnk ), where
the prior pdf p(xnk ) is Gaussian. Under common statistical
independence assumptions on v nk or ank [3], the joint prior
pdf of all agent states up to time n is given by
p(x1:n ) =
n Y
Y
i
i=1 k∈VM
.
p xik |xi−1
k
(2)
n
If (k, l) ∈ E n, agent k ∈ VM
acquires at time n a noisy
measurement of its distance to agent or anchor l,
n
dnk,l = kpnk − pnl k + wk,l
.
(3)
n
is assumed zero-mean Gaussian
The measurement error wk,l
2
with variance σw
. Note that the local likelihood function
n
p(dnk,l |pnk , pnl ) is nonlinear in pnk and pnl . Let d1:n , di i=1
n
with dn , dnk,l (k,l)∈E n . Assuming that all wk,l
are independent, the global likelihood function involving all measurements
and all states up to time n factors according to
p(d1:n |x1:n ) =
n
Y
Y
p dik,l |pik , pil
i
i=1 (k,l)∈EM
Y
p diκ,λ |piκ , p̃iλ ,
i
(κ,λ)∈EMA
(4)
n
.
where p̃nλ denotes the (known) position of anchor λ ∈ VA
III. T HE P ROPOSED M ESSAGE PASSING S CHEME
n
The task of agent k ∈ VM
is to estimate its state xnk from the
1:n
total measurement vector d , for n ∈ {1, . . . , N }. We will
consider the minimum mean-square error (MMSE) estimator
Z
n
x̂nk , xnk p(xnk |d1:n )dxnk , k ∈ VM
.
(5)
Calculating the posterior pdf p(xnk |d1:n ) involved in (5) by
direct marginalization of the joint posterior pdf p(x1:n |d1:n )
is infeasible because of the excessive dimension of integration and because d1:n is not locally available at the agents.
Next, we develop a distributed message passing scheme that
n
approximates p(xnk |d1:n ), k ∈ VM
, n ∈ {1, . . . , N }.
1:n 1:n
By Bayes’ rule, p(x |d ) ∝ p(d1:n |x1:n )p(x1:n ),
where p(x1:n ) and p(d1:n |x1:n ) factor as in (2) and (4),
respectively. This factorization underlies the proposed hybrid
BP–MF message passing scheme, which provides approximate
marginal posterior pdfs (“beliefs”) qk (xnk ) ≈ p(xnk |d1:n ) for
n
. Our scheme is an instance of the general hybrid
all k ∈ VM
BP–MF message passing scheme presented in [15]. We use
BP for the motion-related factors p(xnk |xkn−1 ) and MF for the
measurement-related factors p(dnk,l |pnk , pnl ), and we suppress
all messages sent backward in time (cf. [3]). We thus obtain
the following iterative scheme at time n: In message passing
[t]
iteration t ∈ {1, ..., t∗ }, beliefs qk (xnk ) are calculated as
Y [t]
1
[t]
n
mk→k (xnk )
ml→k (pnk ), k ∈ VM
, (6)
qk (xnk ) =
Z
n
l∈Nk
where Z is a normalization constant and Nkn , {l |(k, l) ∈ E n }
is the set of agents and anchors communicating with agent k
at time n (termed “neighbors”). The factors in (6) are obtained
as
R [t∗ ] n−1
n−1
n n−1
qk (xk )p(xk |xk )dxk ,
n−1
n
k ∈ VM
∩ VM
(7)
mk→k (xnk ) =
n−1
n
n
p(xk ), k ∈ VM \VM
and
Z
[t]
[t−1]
n
n
n
n
n
n
ml→k (pk ) = exp
ql
(xl ) ln p(dk,l |pk , pl )dxl . (8)
(Note that pnl = p̃nl if l is an anchor.) This recursion is
[0]
initialized with qk (xnk ) = mk→k (xnk ).
In a distributed implementation, each agent k broadcasts its
[t−1]
belief qk (xnk ) to its neighbors l ∈ Nkn and receives the
[t−1]
neighbor beliefs ql
(xnl ), l ∈ Nkn . These beliefs are then
[t]
used to calculate the messages ml→k (pnk ), l ∈ Nkn at agent k
as in (8). These messages, in turn, are needed to calculate the
[t]
updated belief qk (xnk ) at agent k according to (6). After t∗
[t∗ ]
iterations, the final belief qk (xnk ) is used for state estimation,
∗
[t ]
i.e. qk (xnk ) is substituted for p(xnk |d1:n ) in (5).
IV. G AUSSIAN B ELIEF A PPROXIMATION
Inspired by [14, Section IV], we introduce an approximation
of the message passing scheme (6)–(8) such that the beliefs
are constrained to a certain class of Gaussian pdfs. This leads
to a significant reduction of both interagent communication
and computational complexity relative to a particle-based
implementation. We first consider MM2. A more detailed
derivation is provided in [19].
A. Gaussian Belief Approximation for MM2
We constrain the beliefs to Gaussian pdfs by using the
information projection approach [16], i.e., substituting for
[t]
qk (·) in (6)
[t]
[t]
q̃k (·) , arg min D g qk .
(9)
g∈G
Here, D gkq , g(x) ln g(x)
q(x) dx is the Kullback-Leibler
divergence and G is the set of 4-D Gaussian pdfs g(x) =
i
h
c c
⊗
N (x; µ, C) with covariance matrix of the form C = p
c cv
I2 . We will denote the mean and covariance matrix of
[t]
n [t]
n
n [t]
q̃k (xnk ) =
in(9) as (µnk )[t] =
N xk ; (µk ) , (Ck )n [t]defined
n [t]
n [t]
(µp,k )
(cp,k ) (ck )
and (C nk )[t] =
⊗ I2 . Because
n [t]
n [t]
n [t]
(µv,k )
R
(ck )
(cv,k )
direct computation of the minimizer (9) is infeasible, we resort
3
to an iterative method. To that end, we first
an analytical
derive
[t]
expression of the objective function D g qk in (9), which
[t]
we abbreviate by Fk (θ) with θ , [µT cp cv c]T. Using the
factorization in (6), this function can be expressed as
X [t]
[t]
Fk (θ) = D[gkmk→k ] −
Gk,l (µp , cp ) + γ ,
(10)
l∈Nkn
where µp is the 2-D vector consisting of the first two entries
of µ, γ is a constant, and
Z
[t]
[t]
Gk,l (µp , cp ) , N (pnk ; µp , cp I2 ) ln ml→k (pnk )dpnk . (11)
To derive an expression of D[gkmk→k ] in (10), we note
[t]
n−1
n
that for k ∈ VM
∩ VM
, due to the Gaussian q̃k (xnk ) and
the∗ linear-Gaussian model (1),∗ the message in (7) (in which
[t ]
[t ]
qk (xkn−1 ) is replaced by q̃k (xkn−1 )) is also Gaussian, i.e.,
n
n
n
mk→k (xk ) = N (xk ; η k , Σnk ). By using (1) and standard
n−1
n
Gaussian integral identities [20], we obtain for k ∈ VM
∩VM
∗
η nk = F(µkn−1 )[t ] ,
∗
Σnk = F (C kn−1 )[t ] F T + σa2 GGT.(12)
n−1
n
For k ∈ VM
\ VM
, η nk and Σnk equal, respectively, the
mean and covariance matrix of the Gaussian prior p(xnk ) =
N (xnk ; η nk , Σnk ). Accordingly, we obtain in either case [20]
D[gkmk→k ] =
1
tr (Σnk )−1 C − ln det(C)
2
+ (µ−ηnk )T (Σnk )−1 (µ−ηnk ) + γ ′, (13)
where γ ′ is a constant. Furthermore, one can express
[t]
Gk,l (µp , cp ) in (11) via an expectation of −(dnk,l −
2
kz nk,l k)2 /σw
, where z nk,l is a 2-D Gaussian random vector
with mean µp − (µnp,l )[t−1] and variance cp + (cnp,l )[t−1] . For
n
l ∈ VA
, in particular, (µnp,l )[t−1] = p̃nl and (cnp,l )[t−1] = 0. By
using expressions of the first-order and second-order moments
of the Rician pdf [21], one obtains [19]
cv =
c=
2
c2
+ n
,
n
cp
Jk,33 + Jk,44
q
n
n )2 c c
1 + 1 + (Jk,13
+ Jk,24
p v
n
n
Jk,13
+ Jk,24
C. Gaussian Belief Approximation for MM1
The results in Sections IV-A and IV-B can be used with
minor changes also for MM1. We here have µ = µp and C =
[t]
cp I2 , and the Gaussian belief approximation reads q̃k (pnk ) =
[t]
N pnk ; (µnp,k )[t] , (cnp,k )[t] I2 . The objective function Fk (θ)
T
(with θ , [µT
p cp ] ) is still given by (10) together with (13)
and (14); however, the expressions (12) are replaced by
d2µ
1
πC
+ γ ′′ , (14)
M − ; 1; −
2
2
2C
where dµ , µp−(µnp,l )[t−1] , C , cp +(cnp,l )[t−1] , M (· ; · ; ·)
denotes the confluent hypergeometric function of the first kind
[22], and γ ′′ is a constant.
B. Iterative Minimization Algorithm for MM2
To derive an iterative
algorithm for computing an approxT
imation of (θnk )[t] = (µnk )[t]T (cnp,k )[t] (cnv,k )[t] (cnk )[t] , i.e.,
[t]
of the minimizer of (10), we set the gradient of Fk (θ) to
zero. This yields the following system of non-linear fixedpoint equations θ = (χnk )[t] (θ), whereof (θ nk )[t] is a solution:
µ=
η nk
+
Σnk
X ∂G[t]
k,l (µp , cp )
l∈Nkn
c2
+
cp =
cv
∂µ
,
−1
n
n
X ∂G[t]
Jk,11
+ Jk,22
k,l (µp , cp )
−
,
2
∂cp
n
l∈Nk
(15)
(16)
∗
Σnk = (C kn−1 )[t ] + T σv2 I2 ,
(19)
n−1 [t ]
n−1 [t ]
) I2 .
where (µkn−1 )[t ] = (µp,k
) and (C kn−1 )[t ] = (cp,k
Finally, fixed point equations in µp and cp are obtained by
[t]
setting to zero the gradient of Fk (θ), and an iterative belief
approximation algorithm is again based on these equations.
∗
r
(18)
;1; x)
=
(16) can be calculated using the relation dM(−1/2
dx
−M (1/2; 2; x)/2 [22], where M (−1/2; 1; x) can be computed efficiently via an approximation [23, Section 4.5].
A Newton conjugate-gradient method [17, Chapter 7.1] is
now applied to (15)–(18) to solve the system θ = (χnk )[t] (θ)
in jmax steps, starting from an initial value θ0 . The method
iteratively computes θ j+1 = (I7 − Ψj )θ j + Ψj (χnk )[t] (θj ),
[t]
where Ψj is the inverse of the Hessian matrix of Fk (θ)
at θ j . The Hessian matrix is approximated via the conjugate
[t]
gradient, which requires only Fk (θ) and its gradient [17].
While the algorithm’s convergence has not been proven so far,
it is suggested by our simulations. The algorithm may produce
[t]
a local minimum of Fk (θ), since this function is not convex
in general. Therefore, the algorithm is run several times with
different values of θ0 , and the result yielding the smallest
[t]
value of Fk (θ) is retained. In our simulations, we used the
generic routine scipy.optimize.fmin_tnc [24].
∗
dnk,l
d2µ + 2cp
+
=−
2
2
2σw
σw
,
n
with Jk,ij
, (Σnk )−1 ij . The partial derivatives in (15) and
η nk = (µkn−1 )[t ] ,
[t]
Gk,l (µp , cp )
(17)
∗
∗
∗
D. Distributed Cooperative Localization Algorithm
The results of the previous subsections lead to a distributed
algorithm for cooperative localization in which only parameters of Gaussian pdfs have to be communicated. At time n,
agent k performs the following operations:
n−1
n
, ηnk and Σnk are
1. Mobility update: For k ∈ VM
∩ VM
n−1 [t∗ ]
n−1 [t∗ ]
calculated from (µk )
and (C k )
as in (12) (for
n−1
n
, η nk and
MM2) or as in (19) (for MM1). For k ∈ VM
\ VM
n
Σk are the mean and covariance matrix of the Gaussian prior
pdf p(xnk ), which are assumed already available at agent k.
2. Iterative message passing: The message passing iterations
are initialized (t = 0) with (µnk )[0] = η nk and (C nk )[0] = Σnk .
At iteration t ∈ {1, . . . , t∗ }, agent k broadcasts (µnp,k )[t−1]
and (cnp,k )[t−1] and receives from the neighbors (µnp,l )[t−1] and
4
100
SBP (t∗ = 30)
SBP (t∗ = 5)
BPMF (t∗ = 5)
NBP (t∗ = 5)
NBP (t∗ = 30)
BPMF (t∗ = 30)
10−1
0
0.5
1
1.5
2
2.5
3
3.5
4
10−1
10−2 0
100
P̂out
P̂out
P̂out
100
NBP (t∗ = 30)
BPMF (t∗ = 5)
BPMF (t∗ = 30)
SBP (t∗ = 30)
SBP (t∗ = 5)
NBP (t∗ = 5)
0.5
1
1.5
2
2.5
3
3.5
4
10−1
10−2 0
NBP (t∗ = 30)
NBP (t∗ = 5)
BPMF (t∗ = 5)
BPMF (t∗ = 30)
SBP (t∗ = 5)
SBP (t∗ = 30)
0.5
1
1.5
2
τ
τ
τ
(a)
(b)
(c)
2.5
3
3.5
4
Fig. 1. Average outage probability versus outage threshold: (a) at n = 1 for both MMs, (b) at n = 30 for MM1, and (c) at n = 30 for MM2.
n
(cnp,l )[t−1] , l ∈ Nkn . Note that the anchors (l ∈ Nkn ∩ VA
)
n
n [t−1]
broadcast their true position, so that (µp,l )
= p̃l and
(cnp,l )[t−1] = 0. Then, new parameters (µnk )[t] and (C nk )[t] are
calculated using the iterative belief approximation algorithm.
After the last iteration (t = t∗ ), an approximation of the MMSE
∗
state estimate x̂nk in (5) is obtained as (µnk )[t ] . This equals
[t∗ ]
the result of (5) with p(xnk |d1:n ) replaced by q̃k (xnk ).
The iterative belief approximation algorithm uses η nk and
Σnk , which are locally available at agent k, and (µnp,l )[t−1] and
(cnp,l )[t−1] , l ∈ Nkn , which were received from the neighbors of
agent k. Therefore, at each message passing iteration t, each
agent k must broadcast to its neighbors l ∈ Nkn only three real
values, namely, two for (µnp,k )[t−1] and one for (cnp,k )[t−1] .
V. S IMULATION R ESULTS
We consider a region of interest (ROI) of size 120m×120m
n
n
with the same |VM
| = 41 agents and |VA
| = 18 anchors at all
N = 30 simulated time steps n. The anchors are regularly
placed within the ROI. To avoid boundary effects, agents
leaving the ROI reenter it at the respective opposite side.
Agents and anchors have a communication radius of 20m;
thereby, each agent communicates with one or two anchors.
The agents measure distances according to (3) with σw = 1m.
For
√ trajectories, we set T = 1s, σv =
√ generating the agent
1.5m/s, and σa = 0.03m/s2 . The initial agent positions
are uniformly drawn on the ROI and, for MM2, the initial
agent velocities are drawn from a Gaussian pdf with mean
[0 0]T and covariance matrix 0.6·I2 . For initializing the various
algorithms, the prior pdf for p0k is chosen Gaussian with mean
µ0p,k and covariance matrix 900·I2. Here, if agent k is adjacent
to one anchor l, then µ0p,k is uniformly drawn from a circle
of radius d0k,l around the true anchor position p̃0l , and if agent
k is adjacent to two anchors l and l′, then µ0p,k is chosen as
(p̃0l + p̃0l′ )/2. For MM2, the pdf for v 0k is chosen Gaussian
with mean [0 0]T and covariance matrix 0.6 · I2 .
We compare the proposed hybrid BP–MF method as stated
in Section IV-D (abbreviated BPMF) with nonparametric BP
(NBP) and sigma point BP (SBP). NBP [8] is an extension
of the particle-based BP method of [2] to mobile agents, and
SBP [11] is a low-complexity sigma-point-based BP scheme
in which, similarly to BPMF, only Gaussian parameters are
communicated. Our simulation of NBP uses 800 particles.
For simulating BPMF, we perform the fixed-point iteration
(with 30 iteration steps) multiple times with different initial
values θ0 . More specifically, 20 initial values
of µ are drawn
[t∗ ]
n
from mk→k (xk ), 20 are drawn from q̃k (xkn−1 ), and, for
each adjacent anchor l, 20 are uniformly drawn from an
annulus of radius dnk,l and radial width 3σw around p̃nl
[2]. Furthermore, the initial values of cp and, for MM2,
of cv ∗and c are always equal to the respective parameters
[t ]
of q̃k (xkn−1 ). Our measure of performance
is the outage
probability Pout , Pr kp̂nk − p̃nk k > τ , where p̃nk is the true
position of agent k at time n, p̂nk is a corresponding estimate,
and τ > 0 is a threshold.
Fig. 1 shows the simulated outage probability P̂out , averaged
over 30 simulation trials, of the three methods versus the
outage threshold τ . It is seen that, at n = 1, BPMF outperforms
NBP and SBP for t∗ = 30; in particular, SBP performs poorly.
Since BPMF and SBP use a Gaussian approximation, one may
conclude that in the case of a noninformative prior (which
is in force at n = 1), the Gaussian approximation degrades
the performance of a pure BP scheme like SBP more than
that of the proposed hybrid BP–MF scheme. At n = 30,
for MM1, BPMF performs as NBP and SBP. However, for
MM2, where the state can be predicted more accurately from
the previous time, SBP outperforms both BPMF and NBP.
Indeed, as previously observed in [11], SBP works very well
when informative prior knowledge is available. We expect that
NBP would be similarly accurate if more particles were used;
however, the complexity of SBP grows quadratically with the
number of particles. It is also seen that for both MMs, contrary
to BPMF, the performance of NBP and SBP at n = 30 does
not improve when t∗ is increased beyond 5. We note that in
less dense networks, where beliefs can be multimodal, NBP
can be expected to outperform SBP and BPMF.
The communication requirements, in terms of number of
real values broadcast per message passing iteration t by each
agent k to adjacent agents l ∈ Nkn , are 3 for BPMF, 5 for SBP,
and 1600 for NBP.
VI. C ONCLUSION
The proposed algorithm for cooperative localization and
tracking combines the advantages of existing BP and MF
methods: its accuracy is similar to that of particle-based BP
although only three real values per message passing iteration
are broadcast by each agent, instead of hundreds of particles.
Our simulations showed that the algorithm performs particularly well relative to pure BP-based methods when the prior
information on the agent positions is imprecise.
5
R EFERENCES
[1] N. Patwari, J. N. Ash, S. Kyperountas, A. O. Hero III, R. L. Moses, and
N. S. Correal, “Locating the nodes: Cooperative localization in wireless
sensor networks,” IEEE Signal Process. Mag., vol. 22, no. 4, pp. 54–69,
Jul. 2005.
[2] A. T. Ihler, J. W. Fisher, R. L. Moses, and A. S. Willsky, “Nonparametric
belief propagation for self-localization of sensor networks,” IEEE J. Sel.
Areas Commun., vol. 23, no. 4, pp. 809–819, Apr. 2005.
[3] H. Wymeersch, J. Lien, and M. Z. Win, “Cooperative localization in
wireless networks,” Proc. IEEE, vol. 97, no. 2, pp. 427–450, Feb. 2009.
[4] S. Mazuelas, A. Bahillo, R. M. Lorenzo, P. Fernandez, F. A. Lago,
E. Garcia, J. Blas, and E. J. Abril, “Robust indoor positioning provided
by real-time RSSI values in unmodified WLAN networks,” IEEE J. Sel.
Topics Signal Process., vol. 3, no. 5, pp. 821–831, Oct. 2009.
[5] D. Dardari, E. Falletti, and M. Luise, Satellite and Terrestrial Radio
Positioning Techniques. Oxford, UK: Academic Press, 2012.
[6] H.-A. Loeliger, “An introduction to factor graphs,” IEEE Signal Process.
Mag., vol. 21, no. 1, pp. 28–41, Jan. 2004.
[7] M. A. Caceres, F. Penna, H. Wymeersch, and R. Garello, “Hybrid
cooperative positioning based on distributed belief propagation,” IEEE
J. Sel. Areas Commun., vol. 29, no. 10, pp. 1948–1958, Dec. 2011.
[8] J. Lien, J. Ferner, W. Srichavengsup, H. Wymeersch, and M. Z. Win,
“A comparison of parametric and sample-based message representation
in cooperative localization,” Int. J. Navig. Observ., 2012.
[9] V. Savic and S. Zazo, “Reducing communication overhead for cooperative localization using nonparametric belief propagation,” IEEE Wireless
Commun. Lett., vol. 1, no. 4, pp. 308–311, 2012.
[10] F. Meyer, O. Hlinka, H. Wymeersch, E. Riegler, and F. Hlawatsch,
“Distributed localization and tracking of mobile networks including
noncooperative objects,” IEEE Trans. Signal Inf. Process. Netw., vol. 2,
no. 1, pp. 57–71, Mar. 2016.
[11] F. Meyer, O. Hlinka, and F. Hlawatsch, “Sigma point belief propagation,”
IEEE Signal Process. Lett., vol. 21, no. 2, pp. 145–149, Feb. 2014.
[12] M. Welling and J. Lim, “A distributed message passing algorithm for
sensor localization,” in Proc. ICANN-2007, Limassol, Cyprus, 2007, pp.
767–775.
[13] S. Van de Velde, H. Wymeersch, and H. Steendam, “Comparison of
message passing algorithms for cooperative localization under NLOS
conditions,” in Proc. IEEE WPNC-12, Dresden, Germany, 2012.
[14] C. Pedersen, T. Pedersen, and B. H. Fleury, “A variational message
passing algorithm for sensor self-localization in wireless networks,” in
Proc. IEEE ISIT-11, Saint Petersburg, Russia, Aug. 2011, pp. 2158–
2162.
[15] E. Riegler, G. E. Kirkelund, C. N. Manchon, M. Badiu, and B. H. Fleury,
“Merging belief propagation and the mean field approximation: A free
energy approach,” IEEE Trans. Inf. Theory, vol. 59, no. 1, pp. 588–602,
Jan. 2013.
[16] D. Koller and N. Friedman, Probabilistic Graphical Models: Principles
and Techniques. Cambridge, MA, USA: MIT Press, 2009.
[17] J. Nocedal and S. Wright, Numerical Optimization. New York, NY:
Springer, 2006.
[18] Y. Bar-Shalom, X.-R. Li, and T. Kirubarajan, Estimation with Applications to Tracking and Navigation. New York, NY, USA: Wiley, 2001.
[19] B. Çakmak, D. N. Urup, F. Meyer, T. Pedersen, B. H. Fleury,
and F. Hlawatsch, “Cooperative localization for mobile networks:
A distributed belief propagation – mean field message passing algorithm (supplementary material),” 2016, Available online:
http://arxiv.org/abs/1512.07782/anc/supplementaryMaterial.pdf.
[20] K. B. Petersen and M. S. Pedersen, The Matrix Cookbook. Copenhagen,
Denmark: Technical University of Denmark, 2008.
[21] S. O. Rice, “Mathematical analysis of random noise,” Bell System
Technical Journal, vol. 23, no. 3, pp. 282–332, 1944.
[22] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions
with Formulas, Graphs, and Mathematical Tables. Washington D.C.,
USA: National Bureau of Standards, 1964.
[23] D. N. Urup, “Distributed localization in dynamic cooperative wireless
sensor networks using the mean field approximation,” Master’s thesis,
Aalborg University, Denmark, Oct. 2014.
[24] E. Jones, T. Oliphant, and P. Peterson, “Open source scientific tools for
Python,” 2016, Available online: http://www.scipy.org/.
| 3cs.SY
|
Energy-efficient neuromorphic classifiers
Daniel Martı́ ∗ † , Mattia Rigotti ‡
†
Mingoo Seok § and Stefano Fusi †
∗
arXiv:1507.00235v1 [q-bio.NC] 1 Jul 2015
Département d’Études Cognitives, École Normale Supérieure - PSL Research University, Paris, France. Institut nationale de la santé et de la recherche médicale,
France.,‡ Physical Sciences Department, IBM T. J. Watson Research Center, Yorktown Heights, NY 10598,† Center for Theoretical Neuroscience, Columbia University, New
York, USA, and § Department of Electrical Engineering, Columbia University, New York, USA
Neuromorphic engineering combines the architectural and computational
principles of systems neuroscience with semiconductor electronics, with the
aim of building efficient and compact devices that mimic the synaptic and
neural machinery of the brain. Neuromorphic engineering promises extremely low energy consumptions, comparable to those of the nervous system. However, until now the neuromorphic approach has been restricted to
relatively simple circuits and specialized functions, rendering elusive a direct
comparison of their energy consumption to that used by conventional von
Neumann digital machines solving real-world tasks. Here we show that a recent technology developed by IBM can be leveraged to realize neuromorphic
circuits that operate as classifiers of complex real-world stimuli. These circuits emulate enough neurons to compete with state-of-the-art classifiers.
We also show that the energy consumption of the IBM chip is typically 2 or
more orders of magnitude lower than that of conventional digital machines
when implementing classifiers with comparable performance. Moreover,
the spike-based dynamics display a trade-off between integration time and
accuracy, which naturally translates into algorithms that can be flexibly
deployed for either fast and approximate classifications, or more accurate
classifications at the mere expense of longer running times and higher energy costs. This work finally proves that the neuromorphic approach can be
efficiently used in real-world applications and it has significant advantages
over conventional digital devices when energy consumption is considered.
neuromorphic electronic hardware
tion
|
VLSI technology
|
neural networks
|
classifica-
Abbreviations: SVM: support vector machine — SV: support vector — RCN: randomly connected neuron
Introduction Recent developments in digital technology and
machine learning are enabling computers to perform an increasing number of tasks that were once solely the domain
of human expertise, such as recognizing a face in a picture
or driving a car in city traffic. These are impressive achievements, but we should keep in mind that the human brain
carries out tasks of such complexity using only a small fraction of the energy needed by conventional computers, the difference in energy consumption being often of several orders
of magnitude. This suggests that one way to reduce energy
consumption is to design machines whose architecture takes
inspiration from the biological brain, an approach that was
proposed by Carver Mead in the late 1980s [1] and that is
now known as “neuromorphic engineering”. Mead’s idea was
to use very-large-scale integration (VLSI) technology to build
electronic circuits that mimic the architecture of the nervous
system. The first electronic devices inspired by this concept
were analog circuits that exploited the subthreshold properties
of transistors to emulate the biophysics of real neurons. Nowadays the term “neuromorphic” refers to any analog, digital,
or hybrid VLSI system whose design principles are inspired
by those of biological neural systems [2].
Neuromorphic hardware has convincingly demonstrated
its potential for energy efficiency, as proven by devices that
consume as little as a few picojoules per neural event (spike)
[3, 4, 5]. These devices contain however a relatively small
number of elements (neurons and synapses) and they can typically perform only simple and specialized tasks, making it
1–12
difficult to directly compare their energy consumption to that
of conventional digital machines.
The situation has changed recently with the development
by IBM of the TrueNorth processor, a neuromorphic device
that implements enough artificial neurons to perform complex real-world tasks, like large-scale pattern classification [6].
Here we show that a pattern classifier implemented on the
IBM chip can achieve performances comparable to those of
state-of-the-art conventional devices based on the von Neumann architecture. More importantly, our chip-implemented
classifier uses 2 or more orders of magnitude less energy than
current digital machines performing the same classification
tasks. These results show for the first time the deployment
of a neuromorphic device able to solve a complex task, while
meeting the claims of energy efficiency contented by the neuromorphic engineering community for the last few decades.
Results
We chose pattern classification as an example of a complex
task because of the availability of well-established benchmarks. A classifier takes an input, like the image of a handwritten character, and assigns it to one among a set of discrete
classes, like the set of digits. To train and evaluate our classifiers we used three different datasets consisting of images of
different complexity (see Fig. 1a).
We start by describing the architecture of the classifier
that we plan to implement on the neuromorphic chip. The
classifier is a feed-forward neural network with three layers of
neurons, and it can be simulated on a traditional digital computers. We will call this network the ‘neural classifier’ to distinguish it from its final chip implementation, which requires
adapting the architecture to the connectivity constraints imposed by the hardware. The neural classifier also differs from
the final hardware implementation in that it employs neurons with a continuous activation function, whereas the IBM
neuromorphic chip emulates spiking neurons. Despite the differences, the functionality of the neural classifier and its final
chip implementation is approximately the same, as we show
below. We list the procedure for adapting the architecture of
the neural classifier into its chip implementation as a contribution in its own right, since it can be directly extended for the
implementation of generic neural systems on other hardware
substrates.
Architecture of the neural classifier Figure 1b illustrates the
three-layer neural classifier. The first layer encodes the preprocessed input and projects to the neurons in the intermediate layer through connections with random weights. Each
of these Randomly Connected Neurons (RCNs) receives therefore a synaptic current given by a randomly weighted sum of
the inputs, which the RCNs transform into activation levels
in a non-linear way—in our case, through a linear rectification function: f (x) = x if x > 0, and 0 otherwise. The
combination of a random mixing of the inputs together with
a non-linear input-output transformation efficiently expands
the dimensionality of the resulting signal (see e.g. [7, 8, 9]),
thereby increasing the chances that downstream neurons can
MNIST
random
connections
readout
connections
0
1
2
3
4
5
6
7
8
9
input
MNISTback-image
LATEX
c
output
2
nonlinear units
(RCN layer)
n neurons
(outputs)
d
n axons
(inputs)
b
a
+
+
+
output
readout
input
RCN
layer
Figure 1. Datasets, architecture of the classifier, architecture of a single core, and chip implementation. a Samples of the three
datasets used to evaluate the performance of our classifier. MNIST contains handwritten digits (10 classes); MNIST-back-image contains the digits of the MNIST dataset
on a background patch extracted randomly from a set of 20 images downloaded from the Internet; LATEX contains distorted versions of 293 characters used in the LATEX
document preparation system. For more details about the datasets, see Methods. b Architecture of the neural network classifier. The images to classify are preprocessed (see
Methods) and represented as patterns of activity of a population of Nin input neurons (left, black dots). These input neurons send random projections to a second layer of
N Randomly Connected Neurons (RCNs) (green circles), which transform nonlinearly their synaptic inputs into firing activity. The activity of the RCNs is then read
out by a third layer of neurons, each of which is trained to respond to only one class (red circles). c Architecture of a single core in the chip. Horizontal lines represent inputs,
provided by the axons of neurons that project to the core. Vertical lines represent the dendrites of the neurons in the core (one dendrite per neuron). Active synapses are shown
as dots in a particular axon-dendrite junction. The synaptic input collected by the dendrites is integrated and transduced into spike activity at the soma (filled squares on top).
The spikes emitted by the neuron are sent via its axon to a particular input line, not necessarily on the same core. Blue lines represent the flow of input and output signals.
The panel includes an example of internal connection: the upmost axon carries the output activity of the leftmost neuron in the core (other connections are left unspecified).
d Implementation of the neural network classifier in a chip with connectivity constraints. The input is fed into all the cores in the RCN layer (shaded blue), whose neurons
project to the input lines of readout cores (shaded yellow) in a one-to-one manner (green curves). The outputs of the readout units are combined together off-line to generate
the response of the output neuron (shaded red). See the main text for the description of the different modules.
discriminate signals belonging to distinct classes. This discrimination is carried out by a set of output units in the last
layer, which compute a weighted sum of the RCNs activity.
The weights are trained so that each output unit responds to
one separate class (one-vs-all code). Details are given in the
Methods. Once the network is trained, a class is assigned to
each input patterns according to which output unit exhibits
the highest activation.
Chip implementation of the neural classifier We implemented
the neural classifier on the IBM neuromorphic chip described
in [10, 6]. The first step of the conversion of the abstract neural classifier to an explicit chip implementation is the transformation of the input patterns into a format that is compatible
with the spike-based coding of the TrueNorth system. For this
we simply employ a firing rate coding and convert the integer
value of every input component to a spike train with a proportional number of spikes, a prescription that is commonly used
in neurocomputational models such as the Neural Engineering
framework [11]. Specifically, input patterns are preprocessed
and formatted into 256-dimensional vectors representing the
firing activity of the input layer (the same preprocessing step
was applied in the neural classifier, see Methods). This vector
of activities is then used to generate 256 regular-firing spike
trains that are fed into a set of cores with random and sparse
connectivity. This set of cores constitutes the RCN layer. Like
in the neural classifier, the neurons in the RCN layer receive
synaptic inputs that consist of randomly weighted combinations of the input, and transform their synaptic inputs into
firing activity according to a nonlinear function. On the chip
this function is given by the neuronal current-to-rate transduction, which approximates a linear-rectification function [12].
Discriminating the inputs coming from the RCN layer requires each output unit to read from the whole layer of RCNs,
which in our implementation contains a number of neurons N
that can be as large as 214 . Moreover, all the readout connectitions have to be set at the weights computed by the training
procedure. These requirements exceed the constraints set by
the chip design, in terms of the maximal number of both incoming and outgoing connections per neuron, as well as the
resolution and the freedom with which synaptic weights can
be set. In this paragraph we will present a set of prescrip2
tions that will allow us to circumvent these limitations, and
successfully instantiate our neural classifier on the IBM system. The prescriptions we are presenting are specific to the
TrueNorth architecture, but the types of constraints that they
solve are shared by any physical implementation of neural
systems, whether it is biological or electronic. It is therefore
instructive to discuss the constraints and the prescriptions to
obviate them in detail, as they can be easily extended to other
more generic settings.
1. Constraints on connectivity. The IBM chip is organized in
cores, each of which contains 256 integrate-and-fire neurons
and 256 input lines that intersect with one another forming
a crossbar matrix of programmable synapses (Fig. 1c). Each
neuron can connect to other neurons by projecting its axon
(output) to a single input line, either on the same core or
on a different core. With this hardware design the maximum
number of incoming connections per neuron, or fan in, is 256.
Likewise, the maximal number of outgoing connections per
neuron, or fan out, is 256, each of which are restricted to
target neurons within a single core.
2. Constraints on synaptic weight precision. Synapses can be
either inactive or active. The weight of an active synapse can
be selected from a set of four values given by signed integers
of 9-bit precision. These values can differ from neuron to neuron. Which of the four values is assigned to an active synapse
depends on the input line: all synapses on the same input line
are assigned an index that determines which of the four values is taken by each synapse (e.g. if the index assigned to the
input line is 2, all synapses on the input line take the second
value of the set of four available synaptic weights, which may
differ from neuron to neuron).
The design constraints that we just described can be overcome with the following set of architectural prescriptions.
P1. Overcoming the constraints on connectivity. We introduced an intermediate layer of neurons, each of which integrates the inputs from 256 out of the total N RCNs. Accordingly, the firing rates of these intermediate neurons represent
a 256/N portion of the total input to an output unit. These
partial inputs can then be combined by a downstream neuron, which will have the same activity as the original output
P2. Overcoming the constraints on synaptic weight precision. Reducing the weight precision after learning usually
only causes moderate drops in classification performance. For
example, in the case of random uncorrelated inputs, the scaling properties of the capacity of the classifier (i.e., number
of classes that can be correctly classified) remain unchanged,
even when the number of states of the synaptic weights is
reduced to two [13]. Instead, the performance drop is catastrophically larger when the weight precision is limited also
during learning [14, 15] and in some situations the learning
problem becomes NP-complete [16] In our case the readout
weights are determined off-chip, using digital conventional
computers that operate on 64 bit numbers, and then quantized in the chip implementation. The performance drop is
almost negligible for a sufficient number of synaptic levels.
In our case we quantized the readout weights of the original
classifier on an integer scale between −28 and 28. Each quantized weight was then implemented as the sum of four groups
of 6 synaptic contacts, where each contact in the group can
either be inactive (value 0) or activated at one of the 6 values:
±1, ±2, ±4. The multiplicity of this decomposition (19 can be
for instance decomposed as (4) + (1 + 4) + (1 + 4) + (1 + 4) or
(2)+(2+4)+(2+4)+(1+4)) is resolved by choosing the decomposition that is closest to a balanced assignment of the weights
across the 4 groups (e.g. 19 = (4) + (1 + 4) + (1 + 4) + (1 + 4)).
This strategy requires that each original synapse be represented by 24 synapses. We implemented this strategy by
replicating each readout neuron 24 times and by distributing
each original weight across 24 different dendritic trees. These
synaptic inputs are then summed together by the off-line summation of all readout neuron activities that correspond to the
partial inputs to a specific output unit (see Methods for details). A similar strategy can be used to implement networks
with synaptic weights that have even a larger number of levels and the number of additional synapses would scale only
logarithmically with the total number of synaptic levels that
is required. However, it is crucial to limit individual synapses
to low values, in order to avoid synchronization between neurons. This is why we limited to 4 the maximum synaptic value
of individual synapses of the chip.
Classification performance and speed-accuracy trade-off Our
neuromorphic classifier implemented on the TrueNorth chip
was emulated on a simulator developed by IBM. As the
TrueNorth chip is entirely digital, the simulator reproduces
exactly the behavior of the chip [10]. In Fig. 2a we show
the dynamics of two typical runs of the simulator classifying
images from the MNIST-back-image dataset. Upon image
presentation, the RCNs in the intermediate layer start inte-
0
250
0
− 250
c
0
1
2
3
4
5
6
7
8
9
500
0
time (ms)
500
b
time (ms)
0
performance (%)
emitted spikes, detrended
a
performance (%)
unit. If the total number of the partial inputs is larger than
the total number of incoming connections of the neurons that
represent the output units (in our case 256), the procedure
can be iterated by introducing additional intermediate layers.
The final tree will contain a number of layers that scales only
logarithmically with the total number of RCNs. For simplicity we did not implement this tree on chip and we summed
off-chip the partial inputs represented by the firing activity
of the readout neurons. Notice also that this configuration
requires readout neurons to respond approximately linearly
to their inputs, which can be easily achieved by tuning readout neurons to operate in the linear regime of their currentto-rate transduction function (i.e., the regime in which their
average input current is positive). This procedure strongly
relies on the assumption that information is encoded in the
firing rates of neurons; if the spiking inputs happen to be
highly synchronized and synchronization encodes important
information, this approach would not work.
250
100
MNIST
MNISTback-image
50
0
500
0
0.5
1
energy consumption (mJ)
MNIST
MNIST-back-image
97.2
78.4
75.0
87.9
77.9
69.9
0
200
400
classification time
spike
difference
∞
80
40
20
10
67.6
59.6
0
200
400
classification time
Figure 2. The neuromorphic classifier in action. a Spikes emitted by readout neurons during an easy (top) and a difficult (bottom) classification,
after removing the trend caused by the intrinsic constant currents. Each curve corresponds to the readout output associated with the digit indicated by the color code.
Samples are drawn from the MNIST-back-image dataset. b Test error as a function
of classification time (i.e., the time over which spikes are integrated) and energy. The
error is averaged over the first 1000 test samples of the MNIST (red) and MNISTback-image (blue) datasets. Each dashed horizontal line indicates the best test error
achieved with support vector classifiers for a given dataset, based on the evaluation
of the whole test set. c Classification times for different thresholds in spike difference (as indicated in the legend), for the MNIST and MNIST-back-image datasets.
For each threshold we plot all classification times (thin lines) as well as the sample
mean (shorter ticks on top). The performances associated with each threshold are
indicated in the y -axis. When the threshold in spike difference is infinite (black), the
classification is assessed at t = 500 ms (i.e., there is no stopping criterion). In all
panels the chip uses N = 16384 RCNs.
grating the input signal (not shown) and, a few tens of milliseconds later, they start emitting spikes, which are passed
to the readout neurons. The figure shows the total number of
spikes emitted by the readout neurons since input activation,
after subtracting the overall activity trend caused by baseline
activity.
For simple classifications, in which the input is easily recognizable, the readout neuron associated with the correct class
is activated in less than 100 ms (Fig. 2a, top). More difficult cases require the integration of spikes over longer time
intervals, as the average synaptic inputs to different readout
neurons can be very similar (Fig. 2a, bottom). This suggests that the performance of the classifier, as measured by
the classification error rate on the test set, should improve
with longer integration intervals. This trade-off between speed
and performance is illustrated in Fig. 2b, which shows the
classification performance versus elapsed time for the MNIST
and MNIST-back-image datasets. The performance increases
monotonically with time until it saturates in about half a
second, with a highest performance of 97.27% for MNIST
(98.2% with 10-fold bagging), and 77.30% for MNIST-backimage. These performances are not too far from the best
classification results achieved so far: 99.06% for MNIST (using maxout networks on the permutation invariant version of
the MNIST dataset, which does not exploit any prior knowledge about the two-dimensional structure of the patterns [17])
and 77.39% for MNIST-back-image (with support vector classifiers [18], although methods combining deep nets, feature
3
learning, and feature selection can achieve performances as
high as 87.75% [19]).
Energy-speed-accuracy trade-off As just discussed, accuracy
has a cost in term of energy because longer integration times
entail more emitted spikes per classification and a larger baseline energy costs, which in our case is the dominant contribution to the total energy consumption. We estimated the
energy consumption as described in section 5 and we found
that the energy per classification never exceeds 1 mJ for our
network configuration. With the energy needed to keep lit a
100 W light bulb for a second, one could perform 105 classifications, which is equivalent to around one classification per
second uninterruptedly for almost one day. Notice that this
estimate is based on a classification that lasts 0.5 s and, therefore, does not take into account the fact that most patterns are
correctly classified in a significantly shorter time (see Fig. 2a,
top). If the integration and emission of spikes is stopped as
soon as one of the output units is significantly more active
than the others, then the average energy consumption can be
strongly reduced. The criterion we used to decide when to
stop the integration of spikes (and thus the classification) was
based on the spikes emitted by the readout units. Specifically,
we monitored the cumulative activity of each output unit by
counting all the spikes emitted by the corresponding readout
neurons. We stopped the classification when the accumulated
activity of the leading unit exceeded that of the second unit
by some threshold. The decision was the class associated with
the leading output unit.
In Fig. 2c we show the performances and the corresponding classification times for several thresholds. Low thresholds
allow for faster yet less accurate classifications. In both the
MNIST and MNIST-back-image datasets, the patterns that
require long classifications times are rare. While the performance barely changes for large enough thresholds, the average
classification time can be substantially reduced by lowering
the threshold. For example, for the MNIST dataset the classification time drops by a factor of 5 (from 500 ms to 100 ms)
and, accordingly, so does the energy consumption (from 1 mJ
to 0.2 mJ). Faster classifications are also possible by increasing either the average firing rate or the total number of RCNs,
both of which entail an increase in energy consumption, which
might be partially or entirely compensated by the decrease in
the classification time. These expedients will speed up the
integration of spike-counts and, as a result, the output class
will be determined faster.
In all cases both the energy cost and the classification performance increase with the total number of emitted spikes or,
equivalently, with integration time, if the average firing rate
is fixed. This is a simple form of a more general energy-speedaccuracy trade-off, a phenomenon that has been described in
several biological information-processing systems (e.g. [20]),
and that can confer great functional flexibility to our classifier. One advantage of basing the computation on a temporal
accumulation of spikes is that the classifier can be interrupted
at any time at the cost of reduced performance, but without compromising its function. This is in stark contrast to
some conventional clock-based centralized architectures whose
mode of computation crucially relies on the completion of entire monolithic sets of instructions. We can then envisage utilization scenarios where a spiking-based chip implementation
of our classifier is required to flexibly switch between precise
long-latency classifications (like, e.g., those involving the identification of targets of interest) and rapid responses of limited
accuracy (like the quick avoidance of imminent danger).
Notice that both the simulated and implemented networks, although entirely feed-forward, exhibit complex dy4
namics leading to classification times that depend on the difficulty associated with the input. This is because neurons are
spiking and the final decision requires some sort of accumulation of evidence. When a stimulus is ambiguous, the units
representing the different decisions receive similar inputs and
the competition becomes harder and longer. This type of behavior is also observed in human brains [21].
We will now focus on the comparison of energy consumption and performance between the neuromorphic classifier and
more conventional digital machines.
Energy consumption and performance: comparison with conventional digital machines We compared both the classification performance and the energy consumption of our neuromorphic classifier to those obtained with conventional digital machines implementing Support Vector Machines (SVMs).
SVMs offer a reasonable comparison because they are among
the most successful and widespread techniques for solving
machine-learning problems involving classification and regression [22, 23, 24], and because they can be efficiently implemented on digital machines.
To better understand how the energy consumption scales
with the complexity of the classification problem, it is useful
to summarize how SVMs work. After training, SVMs classify
an input pattern according to its similarity to a set of templates, called the support vectors, which are determined by the
learning algorithm to define the boundaries between classes.
The similarity is expressed in terms of the scalar product between the input vectors and the support vectors. As argued
above, we can improve classification performance by embedding the input vectors in a higher-dimensional space before
classifying them. In this case SVMs evaluate similarities by
computing classical scalar products in the higher-dimensional
space. One of the appealing properties of SVMs is that there
is no need to compute explicitly the transformation of inputs
into high-dimensional representations. Indeed, one can skip
this step and compute directly the scalar product between the
transformed vectors and templates, provided that one knows
how the distances are distorted by the transformation. This is
known as the “kernel trick” because the similarities in a highdimensional space can be computed and optimized over with
a kernel function applied to the inputs. Interestingly, the kernel associated with the transformation induced by the RCNs
of our neural classifier can be computed explicitly in the limit
of a large number of RCNs [18]. This is also the kernel that
we used to compare the performance of SVMs against that of
our neural classifier.
Unfortunately, classifying a test input by computing its
similarity to all support vectors becomes unwieldy and computationally inefficient for large datasets, as the number of
support vectors typically scales linearly with the size of the
training set in many estimation problems [25]. This means
that the number of operations to perform, and hence the energy consumption per classification, also scales with the size of
the training set. This makes SVMs and kernel methods computationally and energetically expensive in many large-scale
tasks. In contrast, our neural network algorithm evaluates a
test sample by means of the transformation carried out by
the RCNs. If the RCN layer comprises N neurons and the
input dimension is Nin , evaluating the output of a test sample requires O(Nin · N ) synaptic events. Thus for large sample sizes, evaluating a test sample in the network requires far
fewer operations than when using the “kernel trick”, because
the number is effectively independent of the size of training
set (cfr. [26, 27]). Systems such as ours may therefore display
considerable energy advantages over SVMs when datasets are
large.
a
performance (%)
MNIST
b
100
100
network
chip
libsvm
SVCperf
80
network
chip
primalSVC
60
10
1
10
2
10
3
10
60
4
0.01
number of RCNs or support vectors
MNISTback-image
performance (%)
c
d
75
network
chip
50
libsvm
SVCperf
25
10
1
10
2
primalSVC
0.1
10
3
10
4
number of RCNs or support vectors
1
10
average energy consumption (mJ)
100
75
25
libsvm
SVCperf
50
network
chip
primalSVC
libsvm
SVCperf
80
0.01
primalSVC
0.1
1
10
average energy consumption (mJ)
Figure 3. Energy-accuracy trade-off. a,c
Dependence of the classification accuracy on the number of Randomly Connected Neurons (RCNs) in the neural
classifier and on the number of support vectors (SVs) in the SVC. Panel a shows this dependence for the MNIST dataset, and panel b for the MNIST-back-image. As the
number of RCNs increases, the classifier becomes more accurate at the cost of higher energy consumptions (b,d). The energy consumption is based on the average time it
takes to the neural classifier to perform the classification (see Fig. 2c). We also show the performance achieved by three different implementations of support vector classifiers
(legend code: SVC, libsvm; rSVC, reduced primal; SVC; SVCperf, cutting plane subspace pursuit). The algorithms rSVC, SVCperf minimize the number of support vectors
(SVs) with respect to the optimal value and reduce, therefore, the energy consumption levels at test time. The number of SVs used by the standard algorithm (libsvm), on
the other hand, can go beyond the optimal value by reducing sufficiently the soft-margin parameter and pushing the classifier to overfit the data. In all cases, the energy
consumption increases linearly with the number of SVs, as the number of operations per classification at test time scales linearly with the number of SVs. The vertical thin
lines indicate the abscissa at best performance for the IBM chip (red) and SVM implementations (black). For reference we indicate the best performance achieved by the
chip with a horizontal dashed line. The horizontal arrow indicates the reduction in energy consumption that would be attained if the efficiency of digital machines reached the
theoretical lower bound estimated by [28]. The relation between number of SVs and energy consumption was determined by simulating the i7 Intel chip running a program that
implements an SVM at test time. b Same as a, but on the MNIST-back-image dataset. In both cases our neuromorphic classifier exhibits an energy cost per classification
that is orders of magnitude smaller.
In Fig. 3 we compare the energy consumption and performance of the neuromorphic classifier to those of an SVM
implemented on a conventional digital machine. More specifically, we estimated the energy expenditure of a digital SVM
using a simulator of the Intel i7 processor, which was the machine with the best energy performance among those that we
simulated (see Methods section 5 and Discussion). The energy cost per support vector per pattern was estimated to be
around 5.2 µJ, a quantity that is not far above what is considered as a lower bound on energy consumption for digital
machines [28]. For both the neuromorphic classifier and the
digital SVM we progressively increased the performance of
the classifiers by increasing the number of RCNs (in the case
of the neuromorphic classifier), and by varying the number of
support vectors (in the case of the SVM), see Figures 4a,c. For
the SVM we tried three different algorithms to minimize the
number of support vectors and hence the energy consumption (for more details, see caption of Fig. 3 and Methods).
For the IBM chip we estimated the energy consumption both
in the case in which we stopped the classifications with the
criterion described in the previous section and in the case in
which the classification time was fixed at 500 ms (see Fig. 7 in
Suppl. Info.). In both cases the energy consumption is significantly lower for the neuromorphic classifier, being in the former case approximately 2 orders of magnitude smaller for both
the MNIST and the MNIST-back-image datasets, while still
achieving comparable maximal performances (Fig. 4b,d).
Scalability The MNIST dataset only has 10 output classes. We
wondered whether the advantage of the neuromorphic classifier in terms of energy consumption is preserved when the
number of classes increases and the classification task becomes
more complex. To study how the energy consumption scales
with the number of classes we used the LATEX dataset, which
contains 293 classes of distorted characters. We progressively
increased the number of classes to be learned and classified
and we studied the performance and the energy consumption
of both the digital implementation of the SVM and the neuromorphic classifier. Specifically, given a number of classes that
was varied between 2 and 293, we selected a random subset of
all the available classes, and we trained both the SVM and the
neural classifier on the same subset. The results are averaged
over 10 repetitions, each one with a different sample of output
classes.
To make a meaningful comparison between the the energy consumed by a SVM and the neuromorphic classifier, we
equalized all the classification accuracies, as follows. For each
classification problem we varied the margin penalty parameter of the standard SVC using grid search and picked the
best performance achieved. We then varied the relevant parameters of the other two classifiers so that their classification
accuracy matched or exceeded the accuracy of the standard
SVC. Specifically, we progressively increased the number of
basis functions (in the primalSVC method) and the number
of RCNs (in the neural classifier) until both reached the target performance. For each classification problem we averaged
over 10 realizations of the random projections of the neural
classifier.
The results are summarized in Fig. 4. The energy consumption is about two orders of magnitude larger for the
SVM throughout the entire range of variation of the number of classes that we considered, although for a very small
(2–3) number of classes the advantage of the neuromorphic
classifier strongly reduces, most likely because the algorithms
to minimize the number of SVs work best when the number
5
of classes is low. This plot indicates that the energy advantage of the neuromorphic classifier over SVMs implemented
on conventional digital machines is maintained also for more
complex tasks involving a larger number of classes.
It is interesting to discuss the expected scaling for growing number of classes. Consider the case of generic C classes
multi-class problems solved through reduction with multiple
combined binary SVMs. In a one-vs-all reduction scheme,
each binary classifier is trained to respond to exactly one of
the C classes, and hence C SVMs are required. For each
SVM, one needs to compute the scalar products between the
test sample to be classified and the NSV support vectors. Each
scalar product requires Nin multiplications and sums. In the
favorable case in which all binary classifiers happen to share
the same support vectors, the scalar products can be computed only once and would require Nin ·NSV operations. These
NSV scalar products then need to be multiplied by the corresponding coefficients, which are different for the different
SVMs. This requires additional CNSV operations. If NSV
scales linearly with C, as in the cases we analyzed, then the
total energy E will scale as
E ∼ Nin C + C 2 .
When C is small compared to Nin , the first term dominates,
and the expected scaling is linear. However, for C > Nin
the scaling is expected to be at least quadratic. It can grow
more rapidly if the support vectors are different for different
classifiers.
Interestingly the expected scaling for the neural network
classifiers that we considered is the same. The energy consumption mostly depends on the number of needed cores. This
number will be proportional to the number of RCNs, N , multiplied by the number of classes. Indeed, each core can receive
up to 256 inputs, so the total number of needed cores will be
proportional to dN/256e, with d·e denoting the ceiling function. Moreover, the number of readout units, which are the
output lines of these cores, will be proportional to the number of classes. Hence the N C dependence. In the cases we
analyzed N depends linearly on the number of classes, and
hence the energy depends quadratically on C, as in the case
of the SVMs when C is large enough. Notice that the there
is a second term which also scales quadratically with C that
contributes to the energy. The second term comes from the
necessity of replicating the RCNs C times, due to the limited
fan out of the RCNs. Again, under the assumption of N ∼ C,
also this term will scale quadratically with C.
Given that the scaling with the number of classes is basically the same for the neuromorhic classifier as for the SVMs,
it is not unreasonable to hypothesize that the energy consumption advantage of the neuromorphic implementation would be
preserved also for a much larger number of classes.
Discussion
Our results indicate that neuromorphic devices are mature
enough to achieve performances on a real-world machinelearning task that are comparable to those of state-of-the-art
conventional devices with von Neumann architecture, all just
by using a tiny fraction of their energy. Our conclusions are
based on a few significant tests, based on a comparison limited
to our neuromorphic classifier and a few digital implementations of SVMs. This clearly restricts the generality of our
results and does not preclude situations in which the advantage of the neuromorphic approach might be less prominent.
In any case, the merit of our study is to offer a solid comparison with implementations on current conventional digital
6
platforms that are energy-efficient themselves. In particular,
the algorithm we used on conventional digital machines involves only multiplications between matrices and vectors, the
efficiency of which has been dramatically increased in the last
decades thanks to optimized parallelization. Furthermore, not
only we tried to match the classification performance of the
competitors, but we also considered two additional SVM algorithms that minimize the number of support vectors, and
hence the final number of operations. Other choices for SVM
algorithms would certainly lead to different estimates for energy consumption, but it is rather unlikely that they would
change across 2 orders of magnitude. It is possible that full
custom unconventional digital machines based, e.g., on field
programmable gate arrays (FPGAs) would be more energyefficient, but it is hard to imagine that they would break the
predicted energy wall discussed in [28]. If this assumption is
right, neuromorphic hardware would always be more efficient
when performing the type of tasks that we considered. Moreover, analog neuromorphic VLSI or unreliable digital technologies might allow for a further reduction of energy consumption, probably by another order of magnitude [5, 29, 30].
The current energy consumption levels achieved by analog systems are very close to those of biological brains in terms of energy per spike, although many of these systems are relatively
small and it is unclear whether they can ever be extended to
brain-scale architectures.
Other custom chips that can solve real-world tasks have
been designed. An example is the FPGA chip NeuFlow, designed to implement convolutional networks for visual recognition. The chip is digital and uses as little as 4.9 ×
1011 operations/W or, equivalently, 2 pJ/operation.
It is also interesting to discuss the performance of other
conventional digital processors in the benchmarks we examined. Let us consider for example the implementations of
SVMs classifying the MNIST digits with about 104 support
vectors, which is roughly the number of vectors we need to
achieve the best classification accuracy. As we have shown,
the Intel i7 takes about 10 ms to perform a classification, at
an approximate cost of 50 mJ. The IBM chip, in contrast, required 1 mJ for the longest classification times (500 ms), and
0.2 mJ for the average classification time (100 ms). We also
quantified the energy cost of the ARMv7, which is a more
energy-efficient yet slower microprocessor often used in mobile technologies. Its energy consumption per classification
was substantially higher, around 700 mJ. The main reason for
this high consumption is that it takes more than 0.6 seconds to
perform a single classification. And the baseline consumption,
which increases linearly with the classification time, is a large
portion of the total energy needed for a classification. Finally,
we considered the recent Xeon Phi, which has a massively parallel architecture and is employed in high performance computing applications. As we do not have a simulator for the
Phi, we could only indirectly estimate a lower bound for the
energy consumption (see Methods for more details). According to our estimate, a single classification requires only 0.2 µs
and uses about 16 mJ, which would be significantly lower than
the energy cost of the i7 and very close to the estimated lower
limit of energy consumption [28], but still larger than the consumption of the IBM chip. Notice however that both the classification time and the energy consumption of the Xeon Phi
processor are very likely to be grossly underestimated, as they
are simply derived from the peak performance of 100 Tflop/s.
The estimates for the i7 and the ARMv7 are significantly more
reliable, because we derived them by simulating the processors.
To summarize, our results compellingly suggest that the
neuromorphic approach is finally competitive in terms of en-
RCN
libsvm
primalSVC
0.5
1
0.99
10
0
100
10
100
number of classes
10
c
RCN
primalSVC
libsvm
number SVs, RCNs
b
1
energy consumption (mJ)
performance
a
1
0.1
0.01
10
100
number of classes
10
10
10
10
4
RCN
libsvm
primalSVC
3
2
1
10
100
number of classes
Figure 4. Dependence of the energy consumption on the number of classes a Classification accuracy for the neural classifier and for two SVM
algorithms, as a function of the number C of classes for the LATEX dataset. The parameters of the different classifiers are tuned to have approximately the same classification
accuracy. b Energy consumption as a function of the number of classes, for the LATEX dataset. Given a number C of classes, every point in the plot is obtained by training
a given classifier on C randomly sampled classes among the 293 available ones. This procedure is repeated 10 times for every value of C and every type of classifier. Each
datapoint associated with the neural classifier (’RCN’) was in turn estimated from a sample of 10 realizations of the random connections (squares indicate sample means,
errorbars indicate the 0.1 and 0.9 fractile of the sample). c As in b, but number of support vectors and RCNs as a function of the number of classes.
Dataset
MNIST
MNIST-back-image
LATEX
image
size
28 × 28
28 × 28
32 × 32
num.
classes
size training set
10
10
293
60000
12000
14650
size
test set
10000
50000
9376
ergy consumption in useful real-world machine learning tasks
and constitutes a promising direction for future scalable technologies. The recent success of deep networks for large-scale
machine learning [31, 32] makes neuromorphic approaches
particularly relevant and valuable. This will be certainly true
for neuromorphic systems with synaptic plasticity, which will
enable these devices to learn autonomously from experience.
Learning is now available only in small neuromorphic systems
[33, 34, 35], but hopefully new VLSI technologies will allow
us to implement it also in large-scale neural systems.
ACKNOWLEDGMENTS. This work was supported by DARPA SyNAPSE, Gatsby
Charitable Foundation, Swartz Foundation and Kavli Foundation. We are grateful to
the IBM team led by Dr. D. Modha for their assistance with the IBM chip simulator.
In particular we thank John Arthur and Paul Merolla for their help with the estimate
of the chip energy consumption. DM acknowledges the support from the FP7 Marie
Curie Actions of the European Commission and the ANR-10-LABX-0087 IEC and
ANR-10-IDEX-0001-02 PSL grants.
Materials and Methods
Images sets for classification benchmarksWe used three
datasets in our study: MNIST, MNIST-back-image, and
LATEX. The MNIST dataset consists of images of handrwitten digits (10 classes) [36]. The MNIST-back-image dataset
contains the same digits of MNIST, but in this case the background of each pattern is a random patch extracted from a
set of 20 black and white images downloaded from the Internet [37]. Patches with low pixel variance (i.e. containing
little texture) are discarded. The LATEX dataset consists of
distorted versions of 293 characters used in the LATEX document preparation system [38, 39]. All datasets consist of l × l
pixel gray-scale images, and each of such pixel images is associated with one out of C possible classes. The size of the pixel
images, the number of classes, and the sizes of the training
and test sets depend on the data set (see table below).
Preprocessing Every sample image was reshaped as a l2 dimensional vector, and the average gray level of each component was subtracted from the data. The dimensionality of the
resulting image vector was then reduced to 256 using PCA. To
guarantee that all the selected components contributed uniformly to the patterns, we applied a random rotation to the
principal subspace (see, e.g., [40]). We denote by Nin = 256
the dimension of that subspace.
The architecture of the network and the training algorithm
We map the preprocessed Nin -dimensional vector image, s,
into a higher dimensional space through the transformation
xi = f (wi · s),
i = 1, . . . , N,
where wi is an Nin -dimensional sparse random vector and f (·)
is a nonlinear function. This is the transformation induced by
a neural network with Nin input units and N output units
with activation function f (·). More succinctly,
x = f (WT s),
[1]
where W is a weight matrix of dimensions Nin × N formed
by adjoining all the column weight vectors wi , and where
f (·) acts componentwise, i.e., f (x) ≡ (f (x1 ), . . . , f (xNin ))T .
The output of the random nonlinear transformation, x, is
used as the input to a linear NC -class discriminant,
consistP
ing of NC linear functions of the type yj = N
J
x
jk
k , with
k=1
j = 1, . . . , C. More compactly,
y = Jx,
[2]
where y = (y1 , . . . , yC )T , J is a C×N matrix, and x is given by
Eq. [ 1 ]. A pattern x is assigned to class Cj if yj (x) > yk (x)
for all j 6= k. The elements of J are learned offline by imposing a 1-of-NC coding scheme on the output: if the target class
is j then the target output t is a vector of length NC where
all components are zero except component tj , which is 1. For
the offline training of weights we use the pseudoinverse, which
minimizes the mean squared error of the outputs. This technique has been shown to be a good replacement for empirical
minimization problems when the dataset is embedded in a random high-dimensional space, which is our case [41, 26, 42, 27].
Neuromorphic chip implementation
The chip is composed of multiple identical cores, each of which
consists of a neuromorphic circuit that comprises n = 256 axons, n neurons, and n2 adjustable synapses ([43, 10, 6], see
also Fig. 1c). Each axon provides the inputs by feeding the
spiking activity of one given neuron that may or not reside in
the core. The incoming spiking activity to all n axons in a core
is represented by a vector of activity bits (A1 (t), . . . , An (t))
whose elements indicate whether or not the neurons associated with the incoming axons emitted a spike in the previous time step. The intersection of the the n axons with the
n neurons forms a matrix of programmable synapses. The
weight of active synapses is determined by the type of axon
and the type of neuron the synapse lies on. Specifically, each
core can contain up to four different types of axon, labeled
Gj = {1, 2, 3, 4}, whereas it can accommodate an unlimited
7
number of neuron types, each of which having four associated
synaptic weights Si = (Si1, . . . , Si4 ). The strength of an active
G
synapse connecting axon j with neuron i is Si j , that is, the
axon type determines which weight to pick among the weights
associated with neuron i. The net input received by neuron i
P
Gj
at time step t is therefore hi (t) = n
j=1 Si Bij Aj (t), where
Bij is 1 or 0 depending on whether the synapse between axon
j and dendrite i is active or inactive.
At each time step the membrane potential Vi (t) of neuron i receiving input h(t) is updated according to Vi (t + 1) =
Vi (t) − β + hi (t), where β is a constant leak. If Vi (t) becomes
negative after an update, it is clipped to 0. Conversely, when
Vi (t) reaches the threshold Vthr , the potential is reset to Vreset
and the neuron emits a spike, which is sent through the neuron’s axon to the target core and neuron. This design implies
that each neuron can connect to at most n neurons, which
are necessarily in the same core. The initial voltage of each
neuron was initialized by drawing randomly and with equal
probability from a set of 4 evenly spaced values from Vreset to
Vthr .
Signal-to-rate transduction The input to the neuromorphic chip consists of a set of spike trains fed to the neurons
of the input layer. To transform the vector signal s into spike
trains, we first shifted the signal by s̄ = 3σ, where σ is the
standard deviation across all signal components of all patterns. The shifted signal was then scaled by a factor νsc chosen to ensure moderate output rates in the RCN layer, and
the result was linear-rectified to positive values. In short, the
input rate νi associated with signal si is νi = νsc [si + s̄]+ ,
i = 1, . . . , Nin , where [x]+ is x if x > 0, or 0 otherwise. The
values νi were then used to generate regular spike trains with
fixed inter-spike-interval 1/νi .
Basic architecture The circuit is divided in two functional
groups, or layers, each of which comprises several cores. The
first functional group is the RCN layer, which computes the
random nonlinear expansion in Eq. [ 1 ]. The second functional group computes the C-class discriminant y = Jx. The
output of the classifier is just argmaxj yj , where j runs over
the C possible categories. The argmax operation was not computed by the chip, but was determined off-line by comparing
the accumulated spike counts across all outputs. In the following, we describe the implementation of the two layers in
more detail.
RCN layer We first set the dimensionality of the input to
the number of available axons per core, i.e., Nin = n = 256. A
convenient choice for W is a n × N matrix where each column
is vector of zeros except for exactly m < n nonzero entries,
which are randomly placed and take a fixed integer value w.
We took m = 26, which corresponds to a connectivity level
of around 0.1. Lowering the connectivity has the advantage
of decreasing energy costs by reducing the number of total
spikes and active synapses, without impacting the classification performance. The random expansion was mapped in the
chip by splitting the matrix WT into dN/ne submatrices of
size n × n, and using each submatrix as the (boolean) connectivity matrix Bij of a core.
With this arrangement, each of the N neurons distributed
among the dN/ne cores receives a sparse and random linear
combination of signals. Specifically, the average current received by each RCN is
hi =
n
X
j=1
8
Wji νj ,
i = 1, . . . , N.
A zero-th order approximation of the firing rate of a general vlsi neuron receiving a current hi is
ri =
[hi − β]+
,
Vthr − Vreset
[3]
where Vthr is the threshold for spike emission and Vreset is the
reset potential [12].
We chose the parameters w and β to meet two criteria.
First, we required the fraction of RCNs showing any firing
activity (i.e., the coding level f ) to be around 0.25. This
coding level is a good compromise between the need for discrimination and generalization, and it keeps finite-size effects
at bay [9]. Second, we required the distribution of activities
across active RCNs to be sufficiently wide. Otherwise the information carried by the spiking activity of the RCNs is too
imprecise to discriminate among patterns.
All the cores in the RCN layer receive exactly the same
n-dimensional input signal.
Readout
The readout matrix J was trained offline and mapped to the
chip architecture as follows.
Weight quantization Because the chip can hold only
integer-valued synapses, we need to map the set of all components of J into an appropriate finite set of integers. We started
clipping the synaptic weights within the bounds (−4σ, 4σ),
where σ is the standard deviation of the sample composed of
all the components of J. We then rescaled the weights to a
convenient magnitude Jmax = 28 (see below), and rounded
the weight values to the nearest integer.
Weight assignment The TrueNorth connectivity constraints dictate that each RCN can project to only one axon,
meaning that there are at most n = 256 synaptic contacts
available to encode the C = 10 weights, J0i , . . . , J9i associated with the i-th RCN. We allocated 24 contacts per class
and per axon (see Fig. 5). Each of these 24 contacts were
divided in four groups comprising 6 weights each, with values 1, 2, 4, −1, −2, −4. This allowed us to represent any integer weight from −28 to 28 (each of the 4 groups encodes a
maximum weight of 7, sign aside). To distribute any weight
value w across the available synaptic contacts, we decomposed w in a sum of four terms, given by the integer division of w by 4 with the remainder spread evenly across terms
(Ex: 19 = 4 + 5 + 5 + 5). Each of such values was assigned
to one group, represented in base 2, and mapped to a pattern of active-inactive synapses according to the weight associated with each axon-dendrite intersection. Positive and
negative weights, as well as strong and week weights, were
balanced along a dendrite by changing the sign and order of
the weights in the crossbar (see alternating colors and saturations in Fig. 5).
Negative threshold For the readout to work properly, the
firing activity of readout neurons must be proportional to the
linear sum of the inputs from the RCNs. This requires neurons to operate in the linear regime of their dynamic range,
a regime that can be enforced by lowering the threshold βout
of readout neurons. We set βout < 0, which is equivalent
to adding a constant positive current to each neuron. If
the current-to-rate transduction function were the thresholdlinear function of Eq. [ 3 ], the baseline activity induced by
this constant current would be |βout |/(Vthr − Vreset ) per readout neuron. The contribution of this background signal should
0
1
+
+
from RCN 1
2
3
4
5
6
7
8
J13 = −9
Figure 5. Implementation of the readout matrix in a core.
The diagram represents the first 8 input lines and first 48 dendrites (two output units)
of a typical readout core. Under each axon-dendrite contact is a square that indicates the potential synaptic strength at the site: color indicates whether the connection is
excitatory (red) or inhibitory (blue), while the saturation level represents the absolute value of the synaptic strength, which can be 1, 2, or 4 (low, medium, and high saturation,
respectively). Only the sites marked with a dot are active. The green frame highlights all the synaptic contacts allocated for an arbitrary weight of the readout matrix, in this
case J13 = −9, which is decomposed as the 4-term sum −9 = −2 − 3 − 2 − 2 = −0102 − 0112 − 0102 − 0102 . Note that in this particular axon the ordering of
weights is 20 , 21 , 22 (rightmost bit is the most significant).
a
b
5 · 10
2.5 · 10
8
0.1
8
0.05
0
0
2500
5000
number of SVs
7500
0
running time (s)
Estimation of the IBM chip energy consumptionThe energy consumption of the IBM chip was estimated from the
TrueNorth specifications [6]. The total energy consumption
comprises the baseline energy (15.9 µW per core), the energy
to emit spikes (109 pJ per spike), the energy needed to read
active synapses (10.7 pJ per active synapse), and the energy
necessary to update membrane potentials (1.2 pJ per neuron).
We ignored the input-output energy needed to transmit spikes
off chip and receive spikes on chip. These numbers provide
a reasonable estimate of the energy consumption of systems
with a conservative supply voltage of 0.775 V; most chips operate near or below this estimate. For a setup with 214 RCNs,
26 dendrites per class, and 10 classes, the power was about
2.08 mW, 95% of which corresponds to the baseline power.
energy consumption (J)
Support Vector Machines. We trained SVMs to perform multiclass classifications based on a one-vs-all scheme, so that
the number of output units coincides with the number of
classes (as in the neural classifier). SVMs were evaluated using arc-cosine kernels, whic mimic the computation of large
feedforward networks with one or more layers of hidden nonlinear units [18]. For our particular architecture, based on
one hidden layer built with threshold-linear units, the kernel
is k(x, y) = kxkkykJ1 (θ), where J1 (θ) = sin θ + (π − θ) cos θ
and θ is the angle between the inputs x and y.
We considered three types of SVM. For the standard SVM
we used the open library libsvm [44], which we patched to include the arccos kernel. The other two SVMs reduce the number of support vectors without sacrificing performance substantially. One of such algorithms is primalSVC, which selects
greedily the basis functions by optimizing the primal objective function [45]. The other method is based on the so-called
Cutting-Plane Subspace Pursuit algorithm, which reduces the
number of support vectors by using basis functions that, unlike standard SVMs, are not necessarily training vectors [46].
Such method is implemented in the library SVMperf. Unlike
the other two classifiers, SVMperf used RBF kernels instead of
arccos kernels.
number of classes C increases, so does the number of readout
neurons necessary to perform a classification and, therefore,
so does the required number of readout cores. Specifically, if
we assign sc synaptic contacts per axon and per class, we will
need a total of sc C output lines. These output lines need to
be connected to all the N neurons through the input lines of
the readout cores . Because each readout core can accomodate
256 output lines, connected to 256 input lines, the total number of readout cores will be dN/256edsc C/256e (d·e indicates
the ceiling function). In principle the number of RCN cores
will be simply dN/256e. However, each RCN should project to
dsc C/256e cores, which implies that each RCN core must be
cloned dsc C/256e times due to the fan-out constraint—each
RCN can project to only one core. The total number of cores
is therefore Ncores = 2dN/256edsc C/256e, where the factor 2
accounts for the contributions of both the readout and the
RCN cores.
The total number of spikes emitted was estimated from
the reference value we got from the chip simulation (for 10
classes, N = 214 , sc = 24, and 500 ms of classification time),
scaled appropriately for the new Ncores . More concretely, if
we denote by n0sp the number of spikes emitted during our
reference simulation, the number of emitted spikes in a general case is nsp = n0sp dsc C/256e(T /500)(N/214 ), where T is
the duration of the simulation in milliseconds. We chose this
duration to be T = 108 ms, which is the average classification
time of the chip implementation the MNIST dataset, when
the spike difference is 80 spikes and which yields only 0.1%
less in performance than in the fixed-duration case (97.2%
vs 97.3%). With T and the estimated values of Ncores and
number of operations
be subtracted from the readout outputs if one wants to get the
equivalent to Eq. [ 2 ], although the step is unnecessary if one
only wishes to compare output magnitudes (as we implicitly
do in order to find the maximal output).
0.4
V
r S s)
pe tion
J
m ifica
3
05 ss
0. cla
0
(1
0.2
0
0
2500
5000
number of SVs
7500
Figure 6. Simulation of a digital support vector machine. a
Scaling of the energy with the number of classes The
estimation was based on the energy cost of the simulated classifications of the MNIST dataset, and extrapolated to the designs required by an increasing number of classes. As the
Number of operations (black circles, left ordinate) and runtime (blue dots, right ordinate) required by a digital SVM to classify 10 test patterns from the MNIST dataset,
as a function of the number of support vectors. The SVM performance was estimated
with a simulator of the Intel i7 processor. b Energy consumption associated to the
datapoints shown in a (squares). The straight line is a least-square fit.
9
b
MNIST
100
network
chip
60
libsvm
SVCperf
80
primalSVC
0.1
1
10
average energy consumption (mJ)
100
performance (%)
performance (%)
a
MNIST- back-image
75
network
chip
25
libsvm
SVCperf
50
primalSVC
0.1
1
10
average energy consumption (mJ)
Figure 7. Performance versus energy consumption at a fixed classification time.
Panels are like in Figs. 3b,d, classification time is now
fixed at 500 ms, rather than determined by a stopping criterion.
nsp , it is straightforward to compute the energy consumption
according to the values given in the previous paragraph.
Energy consumption in von Neumann digital machines
Configuration The runtime and power of microprocessors
with von Neumann architectures were estimated with the recently developed simulators GEM5 (gem5.opt 2.0) [47] and
McPAT (ver. 1.2) [48]. For the estimation we used an architecture configuration similar to that of the recent Intel CoreTM
i7 processors [49], which incorporate state-of-the-art CMOS
technology. Specifically, we used an x86 64, O3, single core
architecture at 2.66 GHz clock frequency, with 32KB 8-way
L1-i and 32KB 8-way L1-d caches, 256KB 8-way L2 cache,
64B cache line size, and 8GB DDR3 1600 DRAM. Channel
length was 22 nm, HP type, using long channel if appropriate.
VDD was 0.9V, so slightly higher than the 0.775 V used for
the IBM chip. However, could we use the same voltage in
Intel i7 simulator, the energy consumption would be lower by
a factor (0.775/0.9)2 = 0.74. This 26% reduction would not
change the main conclusions about the energy consumption
gap between the IBM chip and the conventional von Neumann
digital machines, which is 2–3 orders of magnitude.
Simulations The benchmark was the test phase of the
SVMs, already trained. Simulations showed that a modern
microprocessor based on a von-Neumann architecture takes
115.5 ms to evaluate the test set with 8087 SVs, while consuming 424.6 mJ (DRAM energy consumption not included).
When we varied the number of support vectors from 9 to
8087, both the runtime and energy consumption grew proportionally to the number of SVs, while the power was roughly
constant due to the fixed hardware configuration (see Fig 6).
To estimate how the energy used by von Neumann digital
SVMs scales with the number of classes, we ran another set
of simulations with Intel i7 simulator, this time varying both
the number of support vectors and the number of classes in
the classification problem. This step was necessary to determine the overhead incurred when we increase the number of
output units. For a given number of classes, the energy cost
per support vector was estimated from the least-square fit of
the energies against the number of support vectors.
Mobile processor We also investigated the runtime and
energy consumption of a more energy-efficient but slower
10
mobile microprocessor performing the same target workload.
The architecture configuration was: ARMv7, O3, single core,
1GHz CPU clock frequency, 32kB 4-way L1i and 32kB 4-way
L1d caches, and 128kB 8-way L2 cache, which is similar to the
architecture of ARM Cortex-A9 [50]. The technology node
(22 nm) and simulators were the same as in the experiment
with the microprocessor mimicking Intel Core i7. For the
benchmark code with the largest number of SVs, the task required 1.2 · 1010 operations that took 6.35 s at a cost of 7.34 J.
Discussion on Intel Xeon Phi Massively parallel architectures have gained a significant amount of attention to
improve the throughput and power efficiency of the highperformance computing (HPC) technology, in response to the
relatively stagnated improvement in clock frequency. The
Xeon Phi coprocessor, recently developed by Intel, is one of
such efforts [51]. It integrates more than 50 CPU cores together with L1/L2 caches, network-on-chips, GDDR memory
controller, and PCIe interface. Each core supports up-to 4thread in-order operation and the 512b SIMD VPU (Vector
processing unit). While the runtime and energy-consumption
of the coprocessor are highly dependent on the target workloads, several recent investigations quantified the performance
and energy-efficiency. In the high-performance configuration,
the system integrating Xeon and Xeon Phi shows the throughput of 100 Tera floating-point operations (flop) per second, the
power consumption of 72.9 kW, marking the energy efficiency
of 0.74 nJ/flop [51]. The classification benchmark codes (with
the largest number of SVs) require 0.02235 Gigaflop on the
desktop processor configuration similar to Intel Core i7. At
a first order approximation, therefore, the Xeon and Xeon
Phi-based system takes 0.2235 µs and uses 16.5 mJ per classification. This energy consumption seems significantly lower
than the one of the Intel Core i7, and very close to its lower
bound, which is approximately 3 mJ. However, one should
keep in mind that the energy is grossly underestimated, as
not only we ignored the energy needed for the RAM, but we
also neglected the cost of the non floating point operations,
which are approximately twice as many as the floating point
operations. For all these reasons it is difficult to compare the
energy consumption for the Xean Phi to the Intel Core i7.
In any case, even for our very conservative energy consumption estimate, the IBM chip remains significantly more energy
efficient.
1. Mead C (1989) Analog VLSI implementation of neural systems (Addison Wesley Publishing
Company).
2. Indiveri G, et al. (2011) Neuromorphic silicon neuron circuits. Front. Neurosci. 5.
3. Livi P, Indiveri G (2009) A current-mode conductance-based silicon neuron for address-event
neuromorphic systems pp 2898–2901.
4. Rangan V, Ghosh A, Aparin V, Cauwenberghs G (2010) A subthreshold aVLSI implementation
of the Izhikevich simple neuron model pp 4164–4167.
5. Chicca E, Stefanini F, Indiveri G (2014) Neuromorphic electronic circuits for building autonomous cognitive systems. Proceedings of the IEEE PP:1–22.
6. Merolla PA, et al. (2014) A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 345:668–673.
7. Jaeger H, Haas H (2004) Harnessing nonlinearity: predicting chaotic systems and saving energy
in wireless communication. Science 304:78–80.
8. Buonomano DV, Maass W (2009) State-dependent computations: spatiotemporal processing
in cortical networks. Nat Rev Neurosci 10:113–125.
9. Barak O, Rigotti M, Fusi S (2013) The sparseness of mixed selectivity neurons controls the
generalization–discrimination trade-off. J. Neurosci. 33:3844–3856.
10. Arthur JV, et al. (2012) Building block of a programmable neuromorphic substrate: A digital
neurosynaptic core (IEEE), pp 1–8.
11. Eliasmith C, et al. (2012) A large-scale model of the functioning brain. Science 338:1202–1205.
12. Fusi S, Mattia M (1999) Collective behavior of networks with linear (VLSI) integrate-and-fire
neurons. Neural Comput. 11:633–652.
13. Sompolinsky H (1986) Neural networks with non-linear synapses and static noise. Phys. Rev.
A 34:2571.
14. Amit DJ, Fusi S (1994) Learning in neural networks with material synapses. Neural Computation 6:957–982.
15. Fusi S (2002) Hebbian spike-driven synaptic plasticity for learning patterns of mean firing rates.
Biological cybernetics 87:459–470.
16. Garey MR, Johnson DS (1979) Computers and intractability: a guide to NP-completeness
(WH Freeman New York).
17. Goodfellow I, Warde-farley D, Mirza M, Courville A, Bengio Y (2013) Maxout Networks pp
1319–1327.
18. Cho Y, Saul LK (2010) Large-margin classification in infinite neural networks. Neural Comput.
22:2678–2697.
19. Sohn K, Zhou G, Lee C, Lee H (2013) Learning and Selecting Features Jointly with Point-wise
Gated Boltzmann Machines pp 217–225.
20. Lan G, Sartori P, Neumann S, Sourjik V, Tu Y (2012) The energy-speed-accuracy trade-off in
sensory adaptation. Nature physics 8:422–428.
21. Tang H, et al. (2014) Spatiotemporal dynamics underlying object completion in human ventral
visual cortex. Neuron 83:736–748.
22. Boser BE, Guyon IM, Vapnik VN (1992) A training algorithm for optimal margin classifiers
(ACM), pp 144–152.
23. Cortes C, Vapnik V (1995) Support-vector networks. Mach. Learn. 20:273–297.
24. Vapnik V, Golowich SE, Smola A (1997) Support vector method for function approximation,
regression estimation, and signal processing pp 281–287.
25. Steinwart I, Christmann A (2008) Support vector machines (Springer).
26. Rahimi A, Recht B (2008) Weighted sums of random kitchen sinks: Replacing minimization
with randomization in learning pp 1313–1320.
27. Le Q, Sarlós T, Smola A (2013) Fastfoodapproximating kernel expansions in loglinear time.
28. Hasler J, Marr B (2013) Finding a roadmap to achieve large neuromorphic hardware systems.
Front. Neurosci. 7.
29. Arthur JV, Boahen K (2011) Silicon-neuron design: A dynamical systems approach. Circuits
and Systems I: Regular Papers, IEEE Transactions on 58:1034–1043.
30. Han J, Orshansky M (2013) Approximate computing: An emerging paradigm for energyefficient design (IEEE), pp 1–6.
31. Krizhevsky A, Sutskever I, Hinton GE (2012) ImageNet classification with deep convolutional
neural networks pp 1097–1105.
32. Deng L, Hinton G, Kingsbury B (2013) New types of deep neural network learning for speech
recognition and related applications: An overview (IEEE), pp 8599–8603.
33. Mitra S, Fusi S, Indiveri G (2009) Real-time classification of complex patterns using spikebased learning in neuromorphic vlsi. Biomedical Circuits and Systems, IEEE Transactions on
3:32–42.
34. Giulioni M, et al. (2011) Robust working memory in an asynchronously spiking neural network
realized with neuromorphic vlsi. Frontiers in neuroscience 5.
35. Arthur J, Boahen K (2006) Learning in silicon: timing is everything. Advances in neural
information processing systems 18:75.
36. LeCun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-based learning applied to document
recognition. Proc. IEEE 86:2278–2324.
37. Larochelle H, Erhan D, Courville A, Bergstra J, Bengio Y (2007) An empirical evaluation of
deep architectures on problems with many factors of variation, ICML ’07 (ACM, New York,
NY, USA), pp 473–480.
38. Amit Y, Geman D (1997) Shape quantization and recognition with randomized trees. Neural
Comput. 9:1545–1588.
39. Amit Y (2002) 2D Object Detection and Recognition: Models, Algorithms, and Networks
(MIT Press).
40. Raiko T, Valpola H, LeCun Y (2012) Deep Learning Made Easier by Linear Transformations
in Perceptrons Vol. 22, pp 924–932.
41. Huang GB, Zhu QY, Siew CK (2006) Extreme learning machine: theory and applications.
Neurocomputing 70:489–501.
42. Tapson J, van Schaik A (2013) Learning the pseudoinverse solution to network weights. Neural
Netw. 45:94–100.
43. Merolla P, et al. (2011) A digital neurosynaptic core using embedded crossbar memory with
45pJ per spike in 45nm (IEEE), pp 1–4.
44. Chang CC, Lin CJ (2011) LIBSVM: A library for support vector machines. ACM T. Intel. Sys.
Techn. 2:27:1–27:27 Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm.
45. Keerthi SS, Chapelle O, DeCoste D (2006) Building support vector machines with reduced
classifier complexity. J. Mach. Learn. Res. 7:1493–1515.
46. Joachims T, Yu CN (2009) Sparse kernel SVMs via cutting-plane training. Mach. Learn.
76:179–193.
47. Binkert N, et al. (2011) The GEM5 simulator. ACM SIGARCH Computer Architecture News
39:1–7.
48. Li S, et al. (2009) McPAT: an integrated power, area, and timing modeling framework for
multicore and manycore architectures (IEEE), pp 469–480.
49. (2014) Intel core i7 processor., Technical report http://www.intel.com/content/www/us/en/
processors/core- i7- processor.html.
50. (2014) Arm cortex-a9., Technical report http://www.arm.com/products/processors/cortex-a/
cortex- a9.php.
51. Chrysos G, Engineer SP (2012) Intel Xeon Phi coprocessor (codename knights corner).
11
| 9cs.NE
|
arXiv:1610.04920v1 [math.GT] 16 Oct 2016
ON THE MONODROMY GROUP OF THE FAMILY OF SMOOTH PLANE
CURVES
NICK SALTER
Abstract. We consider the space Pd of smooth complex projective plane curves of degree
d. There is the tautological family of plane curves defined over Pd , and hence there is a
monodromy representation ρd : π1 (Pd ) → Mod(Σg ) into the mapping class group of the
fiber. We show two results concerning the image of ρd . First, we show that the presence of
an invariant known as a “n-spin structure” constrains the image in ways not predicted by
previous work of Beauville [Bea86]. Second, we show that for d = 5, our invariant is the
only obstruction for a mapping class to be contained in the image. This requires combining
the algebro-geometric work of Lönne [Lön09] with Johnson’s theory [Joh83] of the Torelli
subgroup of Mod(Σg ).
1. Introduction
Let Pd denote the moduli space of smooth degree-d plane curves.1 The tautological family
of plane curves over Pd determines a monodromy representation
ρd : π1 (Pd ) → Mod(Σg ),
d−1
2
where g =
and Mod(Σg ) is the mapping class group of the surface Σg of genus g. This
note concerns the the problem of computing the image of ρd .
The first step towards determining the image of ρd has been carried out by A. Beauville
in [Bea86], building off of earlier work of W. Janssen [Jan83] and S. Chmutov [Chm82]. Let
Ψ : Mod(Σg ) → Sp2g (Z) denote the symplectic representation on H1 (Σg ; Z). Beauville has
determined Ψ ◦ ρd ; he shows that for d even it is a surjection, while for d odd it is the (finiteindex) stabilizer of a certain spin structure. A priori, it is therefore possible that ρd could
surject onto Mod(Σg ) or a spin mapping class group, depending on the parity of d.
The first theorem of the present paper is that in general, this does not happen. We show that
a so-called n-spin structure provides an obstruction for f ∈ Mod(Σg ) to be contained in Im(ρd ),
and that this obstruction is not detectable on the level of homology, i.e. that Beauville’s “upper
bound” on Im(ρd ) is not sharp.
Theorem A. For all d ≥ 4, there is a finite-index subgroup Mod(Σg )[φd ] ≤ Mod(Σg ) for which
Im(ρd ) ⊆ Mod(Σg )[φd ].
Date: October 16, 2016.
1See Section 2.1 for a review of these algebro-geometric notions.
1
2
NICK SALTER
For d ≥ 6, the containment
Mod(Σg )[φd ] $ Ψ−1 (Ψ(Mod(Σg )[φd ]))
is strict. Consequently, for d ≥ 6, the same is true for Im(ρd ):
Im(ρd ) $ Ψ−1 (Ψ(Im(ρd ))).
In the statement of the theorem, φd is a cohomology class in H 1 (T ∗,1 Σg ; Z/(d − 3)Z), where
T ∗,1 Σg denotes the unit cotangent bundle of Σg , and Mod(Σg )[φd ] denotes the stabilizer of φd
in the natural action of Mod(Σg ) on H 1 (Σg ; Z/(d − 3)Z). The class φd is an instance of an
n-spin structure for n = d − 3, and is constructed in a natural way from a (d − 3)rd root of the
canonical bundle of a plane curve. Such objects, and the subgroups of Mod(Σg ) fixing the set
of all n-spin structures, were studied by P. Sipe [Sip82, Sip86].
Theorem A will be proved by giving a construction of φd that makes the invariance of φd
under Im(ρd ) transparent. Using a topological interpretation of n-spin structures based on the
work of S. Humphries-D. Johnson [HJ89], it will then be possible to see how the invariance of
φd provides a strictly stronger constraint on Im(ρd ) than that of Beauville.
The second half of the paper concerns the problem of determining sufficient conditions for
an element f ∈ Mod(Σg ) to be contained in Im(ρd ).
Theorem B. For d = 5, there is an equality
Im(ρ5 ) = Mod(Σ6 )[φ5 ].
Here φ5 ∈ H 1 (T ∗,1 Σ6 ; Z/2Z) is a (classical) spin structure of odd parity, and Mod(Σ6 )[φ5 ]
denotes its stabilizer within Mod(Σ6 ).
Analogous theorems hold for d = 3, 4 as well. The case d = 3 (where g = 1) follows
immediately from Beauville’s computation, in light of the fact that Ψ is an isomorphism Ψ :
Mod(Σ1 ) → SL2 (Z) for g = 1. This case is also closely related to the work of I. Dolgachev - A.
Libgober [DL81]. The case d = 4 (asserting the surjectivity Im(ρ4 ) = Mod(Σ3 )) was established
by Y. Kuno [Kun08]. Kuno’s methods are very different from those of the present paper, and
make essential use of the fact that the generic curve of genus 3 is a plane curve of degree 4.
Theorem B thus treats the first case where planarity is an exceptional property for a curve to
possess, and shows that despite this, the monodromy of the family of plane curves of degree 5
is still very large.
Theorem B is obtained by a novel combination of techniques from algebraic geometry and
the theory of the mapping class group. The starting point is Beauville’s work, which allows
one to restrict attention to Im(ρ5 ) ∩ I6 , where I6 is the Torelli group.2
The bridge between algebraic geometry and mapping class groups arises from the work of M.
Lönne [Lön09]. The main theorem of [Lön09] gives an explicit presentation for the fundamental
group of the space Pn,d of smooth hypersurfaces in CP n of degree d. Picard-Lefschetz theory
2See Section 4.3 for the definition of the Torelli group.
MONODROMY OF PLANE CURVES
3
allows one to recognize Lönne’s generators as Dehn twists. Theorem B is then proved by
carrying out a careful examination of the configuration of vanishing cycles as simple closed
curves on a surface of genus 6. This analysis is used to exhibit the elements of Johnson’s
generating set for the Torelli group inside Im(ρ5 ).
In genus 6, Johnson’s generating set has 4470 elements. In order to make this computation
tractable, we find a new relation in Mod(Σg ) known as the “genus-g star relation”. Using this,
we reduce the problem to eight easily-verified cases. An implicit corollary of the proof is a
determination of a simple finite set of Dehn twist generators for the spin mapping class group
Mod(Σ6 )[φ5 ]. An alternative set of generators was obtained by Hirose [Hir05, Theorem 6.1].
Outline. Section 2 is devoted to the construction of φd . In Section 3, we recall some work of
S. Humphries and D. Johnson that relates H 1 (T ∗,1 Σg ; V ) for an abelian group V to the notion
of a “generalized winding number function”. We will use this perspective to show that the
invariance of φd under Im(ρd ) provides an obstruction to the surjectivity of ρd .
The proof of Theorem B is carried out in sections 4 through 7. Section 4 collects a number
of results from the theory of mapping class groups. Section 5 recalls Lönne’s presentation and
establishes some first properties of Im(ρd ). Section 6 continues the analysis of Im(ρd ). Finally
Section 7 collects these results together to prove Theorem B.
Acknowledgements. The author would like to thank Dan Margalit for a series of valuable
discussions concerning this work. He would also like to thank Benson Farb for alerting him to
Lönne’s work and for extensive comments on drafts of this paper, as well as ongoing support
in his role as advisor.
2. nth roots of the canonical bundle and generalized spin structures
2.1. Plane curves and Pd . By definition, a (projective) plane curve of degree d is the vanishing locus V (f ) in CP 2 of a nonzero homogeneous polynomial f (x, y, z) of degree d. The space
of all plane curves is identified with CP N , where N = d+2
− 1. A plane curve X of degree d
2
is smooth if X ∼
,
and
otherwise
X
is
said
to be singular.
= Σg with g = d−1
2
We define the discriminant as the set
Dd = {f ∈ CP N | V (f ) is singular.}.
The discriminant Dd is the vanishing locus of a polynomial pd known as the discriminant
polynomial, and is therefore a hypersurface in CP N . The space of smooth plane curves is then
defined as
Pd = CP N \ Dd .
The universal family of plane curves is the space Xd ⊂ Pd × CP2 defined via
Xd = {(f, [x : y : z]) ∈ Pd × CP2 | f (x, y, z) = 0}.
4
NICK SALTER
The projection π : Xd → Pd is the projection map for a C ∞ fiber bundle structure on Xd with
fibers diffeomorphic to Σg .
2.2. n-spin structures. Let X be a smooth projective algebraic curve over C and let K ∈
Pic(X) denote the canonical bundle.3 Recall that a spin structure on X is an element L ∈ Pic(X)
satisfying L⊗2 = K. This admits an obvious generalization.
Definition 2.1. An n-spin structure is a line bundle L ∈ Pic(X) satisfying L⊗n = K.
Let T ∗,1 X denote the unit cotangent bundle of X, relative to an arbitrary Riemannian
metric on X. Just as ordinary spin structures are closely related to H 1 (T ∗,1 X; Z/2Z), there is
an analogous picture of n-spin structures.
Proposition 2.2. Let L be an n-spin structure on X. Associated to L are
∗,1 X → T ∗,1 X with deck group Z/nZ, and
(1) a regular n-sheeted covering space T^
(2) a cohomology class φL ∈ H 1 (T ∗,1 X; Z/nZ) restricting to a generator of the cohomology
H 1 (S 1 ; Z/nZ) of the fiber of T ∗,1 X → X.
Proof. In view of the equality L⊗n = K in Pic(X), taking nth powers in the fiber induces a map
µ : L → K. Let L◦ denote the complement of the zero section in L, and define K ◦ similarly.
Then µ : L◦ → K ◦ is an n-sheeted covering space with deck group Z/nZ induced from the
∗,1 X → T ∗,1 X is obtained
multiplicative action of the nth roots of unity. The covering space T^
from L◦ → K ◦ by restriction.
∗,1 X → T ∗,1 X is a regular cover with deck group Z/nZ, the Galois correspondence for
As T^
∗,1 X is associated to some homomorphism φ : π (T ∗,1 X) →
covering spaces asserts that T^
L
1
Z/nZ. This gives rise to a class, also denoted φL , in H 1 (T 1,∗ X; Z/nZ). On a given fiber of
∗,1 X → T ∗,1 X restricts to an n-sheeted cover S 1 → S 1 ; this proves
T ∗,1 X → X, the covering T^
the assertion concerning the restriction of φL to H 1 (S 1 ; Z/nZ).
Our interest in n-spin structures arises from the fact that degree-d plane curves are equipped
with a canonical (d − 3)-spin structure.
Fact 2.3. Let X be a smooth degree-d plane curve, d ≥ 3. The canonical bundle K ∈ Pic(X)
is induced from the restriction of O(d − 3) ∈ Pic(CP2 ). Consequently, O(1) determines a
(d − 3)-spin structure on X for d ≥ 4.
Let ̟ : Xd → CP2 denote the projection onto the second factor. Then ̟∗ (O(d − 3)) ∈
Pic(Xd ) restricts to the canonical bundle on each fiber, and ̟∗ (O(1)) determines a (d − 3)-spin
structure. Let T ∗,1 Xd denote the S 1 -bundle over Xd for which the fiber over x ∈ X consists of
the unit cotangent vectors Tx∗,1 X.
3Recall that the canonical bundle is the line bundle whose underlying R2 bundle is T ∗ X, the cotangent
bundle.
MONODROMY OF PLANE CURVES
5
Definition 2.4. The cohomology class
φd ∈ H 1 (T ∗,1 Xd ; Z/(d − 3)Z)
is obtained by applying the construction of Proposition 2.2 to the pair of line bundles ̟∗ (O(1)),
̟∗ (O(d − 3)) ∈ Pic(Xd ).
3. Generalized winding numbers and obstructions to surjectivity
In this section, we show that the existence of φd gives rise to an obstruction for a mapping
class f ∈ Mod(Σg ) to be contained in Im(ρd ). For any system of coefficients V , there is a natural
action of Mod(Σg ) on H 1 (T ∗,1 Σg ; V ) which extends the action of Mod(Σg ) on H 1 (Σg ; V ) via
Ψ. To prove Theorem A, it therefore suffices to show that the stabilizer Mod(Σg )[φd ] of each
nonzero element of H 1 (T ∗,1 Σg ; Z/(d − 3)Z) is not the full group Ψ−1 (Ψ(Im(ρd ))).
The natural setting for what follows is in the unit tangent bundle of a surface, which we
write T 1 Σ. Of course, a choice of Riemannian metric on Σ identifies T 1 Σ with T ∗,1 Σ, and a
choice of metric in each fiber identifies T ∗,1 Xd with the “vertical unit tangent bundle” T 1 Xd ;
we will make no further comment on these matters.
The basis for our approach is the work of Humphries-Johnson [HJ89], which gives an interpretation of H 1 (T 1 Σg ; V ) as the space of “V -valued generalized winding number functions”.
A basic notion here is that of a Johnson lift. For our purposes, a simple closed curve is a
C 1 -embedded S 1 -submanifold.
Definition 3.1. Let a be a simple closed curve on the surface Σ given by a unit-speed C 1
embedding a : S 1 → Σ. A choice of orientation on S 1 induces an orientation on a, as well
as providing a coherent identification Tx1 S 1 = {−1, 1} for each x ∈ S 1 . The Johnson lift of a,
written ~a, is the map ~a : S 1 → T 1 Σ given by
~a(t) = (a(t), Dt a(1)).
That is, the Johnson lift of a is simply the curve in T 1 Σ induced from a by tracking the tangent
vector.
The Johnson lift allows for the evaluation of elements of H 1 (T 1 Σ; V ) on simple closed curves.
Let Σ be a surface, V an abelian group, and α ∈ H 1 (T 1 Σ; V ) a cohomology class. Let a be
a simple closed curve. By an abuse of notation, we write α(a) for the evaluation of α on the
1-cycle determined by the Johnson lift ~a. In this context we call α a “generalized winding
number function”.4 In [HJ89], it is shown that this pairing satisfies the following properties:
Theorem 3.2 (Humphries-Johnson).
(i) The evaluation α(a) ∈ V is well-defined on the isotopy class of a.
4The terminology “generalized winding number” is inspired by the fact that the twist-linearity property was
first encountered in the context of computing winding numbers of curves on surfaces relative to a vector field.
6
NICK SALTER
(ii) (Twist-linearity) If b is another simple closed curve and Tb denotes the Dehn twist about
b, then α is “twist-linear” in the following sense:
α(Tb (a)) = α(a) + ha, biα(b),
(1)
where ha, bi denotes the algebraic intersection pairing.
(iii) Let ζ be a curve enclosing a small null-homotopic disk on Σ, and let S ⊂ Σ be a subsurface
with boundary components b1 , . . . , bk . If each bi is oriented so that S is on the left and ζ
is oriented similarly, then
α(b1 ) + · · · + α(bk ) = χ(S)α(ζ),
(2)
where χ(S) is the Euler characteristic of S.
Remark 3.3. Humprhies-Johnson in fact establish much more: they show that every V -valued
twist-linear function arises as a class α ∈ H 1 (T 1 Σ; V ). For what follows we only need the results
of Theorem 3.2.
Proof of Theorem A. Consider the class φd ∈ H 1 (T ∗,1 Xd ; Z/(d − 3)Z). The above discussion
implies that on a given fiber X of Xd → Pd , the restriction of φd determines a generalized
winding number function; we write αd ∈ H 1 (T 1 X; Z/(d−3)Z) for this class. Since αd is induced
from the globally-defined form φd , it follows that αd is monodromy-invariant: if f ∈ Im(ρd ),
then for any simple closed curve a on X, the equation
αd (f (a)) = αd (a)
(3)
must hold. Consequently,
Im(ρd ) ⊆ Mod(Σg )[φd ]
as claimed.
We wish to exhibit a nonseparating simple closed curve b for which αd (b) 6= 0. Given such a
b, there is another simple closed curve a satisfying ha, bi = 1. Then the twist-linearity condition
(1) will show that
αd (Tb (a)) = αd (a) + αd (b) 6= αd (a);
this contradicts (3). It follows that the Dehn twist Tb for such a curve cannot be contained in
Mod(Σg )[φd ].
In the case when d is even, when Ψ−1 (Ψ(Im(ρd ))) = Mod(Σg ), this will prove Theorem A.
1
∗,1
For d odd, there is an additional complication. Here, the class d−3
Xd ; Z/2Z)
2 φd ∈ H (T
determines an ordinary spin structure, and according to Beauville, the group Ψ(Mod(Σg )[φd ])
is the stabilizer of d−3
2 φd in Sp(2g, Z). We must therefore exhibit a curve b for which αd (b) is
d−3
nonzero and 2 -torsion. Equation (1) shows that such a curve does stabilize the spin structure
d−3
2 φd , but not the refinement to a (d − 3)-spin structure φd .
It remains to exhibit a suitable curve b. It follows easily from the twist-linearity condition
(1) that given any subsurface S ⊂ X of genus 1 with one boundary component, there is some
(necessarily nonseparating) curve c contained in S with αd (c) = 0. Let S1 , S2 , S3 be a collection
MONODROMY OF PLANE CURVES
7
of mutually-disjoint subsurfaces of genus 1 with one boundary component, and let c1 , c2 , c3 be
curves satisfying αd (ci ) = 0, and for which ci is contained in Si (recall that d ≥ 6 and so the
genus of X is g ≥ 10). Choose b disjoint from all ci so that the collection of curves b, c1 , c2 , c3
encloses a subsurface Σ homeomorphic to a sphere with 4 boundary components. From (2) and
the construction of the ci , it follows that when b is suitably oriented, it satisfies
αd (b) = χ(Σ)αd (ζ) = −2αd (ζ).
Recall that by Proposition 2.2.2, the element αd (ζ) ∈ Z/(d − 3)Z is primitive. Thus αd (b) 6= 0
for any d, but is d−3
2 -torsion when d is odd, as required.
4. Results from the theory of the mapping class group
We turn now to the proof of Theorem B. From this section onwards, we adopt the conventions
and notations of the reference [FM12]. In particular, the left-handed Dehn twist about a curve
c is written Tc , and the geometric intersection number between curves a, b is written i(a, b). We
pause briefly to establish some further conventions. We will often refer to a simple closed curve
as simply a “curve”, and will often confuse the distinction between a curve and its isotopy class.
Unless otherwise specified, we will assume that all intersections between curves are essential.
4.1. The change-of-coordinates principle. The change-of-coordinates principle roughly asserts that if two configurations of simple closed curves and arcs on a surface have the same
intersection pattern, then there is a homeomorphism taking one configuration to the other.
There are many variants of the change-of-coordinates principle, all based on the classification
of surfaces. See the discussion in [FM12, Section 1.3.2].
Basic principle. Suppose c1 , . . . , cn and d1 , . . . , dn are configurations of curves on a surface
S all meeting transversely. The surface S \ {ci } has a labeling on segments of its boundary,
corresponding to the segments of the curves ci from which the boundary component arises.
Suppose there is a homeomorphism
f : S \ {ci } → S \ {di }
taking every boundary segment labeled by ci to the corresponding di segment. Then f can be
extended to a homeomorphsim f : S → S taking the configuration ci to di .
We illustrate this in the case of chains.
Definition 4.1. Let S be a surface. A chain on S of length k is a collection of curves (c1 , . . . , ck )
for which the geometric intersection number i(ci , cj ) is 1 if |i − j| = 1 and 0 otherwise. If
C = (c1 , . . . , ck ) is a chain, the boundary of C, written ∂C, is defined to be the boundary
of a small regular neighborhood of c1 ∪ · · · ∪ ck . When k is even, ∂C is a single (necessarily
separating) curve, and when k is odd, ∂C = d1 ∪ d2 consists of two curves d1 , d2 whose union
separates S.
8
NICK SALTER
Lemma 4.2 (Change-of-coordinates for chains). Let (c1 , . . . , ck ) and (d1 , . . . , dk ) be chains
of even length k on a surface S. Then there is a homeomorphism f : S → S for which
f (ci ) = di , 1 ≤ i ≤ k.
Proof. See [FM12, Section 1.3.2].
4.2. Some relations in the mapping class group.
Proposition 4.3 (Braid relation). Let S be a surface, and a, b curves on S satisfying i(a, b) = 1.
Then
Ta Tb Ta = Tb Ta Tb .
(4)
On the level of curves,
Ta Tb (a) = b.
Any such a, b are necessarily non-separating.
Conversely, if a, b are curves on S in distinct isotopy classes that satisfy the braid relation
(4), then i(a, b) = 1.
Proof. See [FM12, Proposition 3.11] for the proof of the first assertion, and [FM12, Proposition
3.13] for the second.
The chain relation. The chain relation relates Dehn twists about curves in a chain to Dehn
twists around the boundary. We will require a slightly less well-known form of the chain relation
for chains of odd length; see [FM12, Section 4.4.1] for details.
Proposition 4.4 (Chain relation). Let C = (c1 , . . . , ck ) be a chain with k odd. Let d1 , d2
denote the components of ∂C. Then the following relation holds:
(Tc21 Tc2 . . . Tck )k = Td1 Td2 .
The genus-g star relation. We will also need to make use of a novel relation generalizing
the star relation (setting g = 1 below recovers the classical star relation).
Proposition 4.5 (Genus-g star relation). With reference to the curves a1 , a2 , c1 , . . . , c2g , d1 , d2 , d3
on the surface Σg,3 of Figure 1, the following relation holds in Mod(Σg,3 ):
(Ta1 Ta2 Tc1 . . . Tc2g )2g+1 = Tdg1 Td2 Td3 .
(5)
Proof. We will derive the genus-g star relation from a more transparent relation in a braid group.
Figure 1 depicts a 2 : 1 covering Σg,3 → Σ0,2 ramified at 2g + 1 points. Number the ramification
points clockwise p1 , . . . , p2g+1 , and consider the mapping class group Mod(Σ0,2,2g+1 ) relative
to these points. Under the covering, the double-twist Tδ21 lifts to Td1 ∈ Mod(Σg,3 ), and the
twist Tδ2 lifts to Td2 Td3 . The twist Tα lifts to Ta1 Ta2 , and the half-twist σi lifts to Tci . Let f ∈
MONODROMY OF PLANE CURVES
9
δ2
p1
a1
d1
c1
c2
d2
p2g+1
c2g−1
p2
α
c2g
σ2
δ1
a2
d3
Figure 1. The genus-g star relation.
Mod(Σ0,2,2g+1 ) be the push map moving each pi clockwise to pi+1 , with subscripts interpreted
mod 2g + 1. One verifies the equality
.
f = Tα σ1 . . . σ2g Tδ−1
1
It follows that
−(2g+1)
f 2g+1 = (Tα σ1 . . . σ2g )2g+1 Tδ1
,
since Tδ1 is central. As f 2g+1 is the push map around the core of the annulus, there is an
equality
Tδ2 .
f 2g+1 = Tδ−1
1
Combining these results,
= (Tα σ1 . . . σ2g )2g+1 .
Tδ2 Tδ2g
1
(6)
Under the lifting described above, the relation (6) in Mod(Σ0,2,2g+1 ) lifts to the relation (5) in
Mod(Σg,3 ).
4.3. The Johnson generating set for Ig . There is a natural map
Ψ : Mod(Σg ) → Sp2g (Z)
taking a mapping class f to the induced automorphism f∗ of H1 (Σg ; Z). The Torelli group Ig
is defined to be the kernel of this map:
Ig = ker(Ψ).
In [Joh83], Johnson produced a finite set of generators for Ig , for all g ≥ 3. Elements of this
generating set are known as chain maps. Let C = (c1 , . . . , ck ) be a chain of odd length with
boundary ∂C = d1 ∪ d2 . There are exactly two ways to orient the collection of curves c1 , . . . , ck
in such a way that the algebraic intersection number ci · ci+1 = +1. Relative to such a choice,
, where d1 is distinguished as
the chain map associated to C is then the mapping class Td1 Td−1
2
10
NICK SALTER
−1
is
the boundary component to the left of the curves c1 , c3 , . . . , ck . The mapping class Td1 Td−2
also called the bounding pair map for d1 , d2 .
While a complete description of Johnson’s generating set is quite tidy and elegant, it has
the disadvantage of requiring several preliminary notions before it can be stated. We therefore
content ourselves with a distillation of his work that is more immediately applicable to our
situation.
β
c3
c1
c2g
c2g−2
c6
c4
c2
c2g−1
c5
Figure 2. Curves involved in the Johnson generating set.
Theorem 4.6 (Johnson). For g ≥ 3, let Γ ≤ Mod(Σg ) be a subgroup that contains the Dehn
twists about the curves c1 , . . . , c2g shown in Figure 2. Suppose that Γ contains all chain maps
for the odd-length chains of the form (c1 , . . . , ck ) and (β, c5 , . . . , ck ). Then Ig ≤ Γ.
Proof. The interested reader should have no trouble deducing Theorem 4.6 from the Main
Theorem and Lemma 1(f) of [Joh83].
5. The Lönne presentation
In this section, we recall Lönne’s work [Lön09] computing π1 (Pd ), and apply this to derive
some first properties of the monodromy map ρd : π1 (Pd ) → Mod(Σg ).
5.1. Picard-Lefschetz theory. Picard-Lefschetz theory concerns the problem of computing
monodromies attached to singular points of holomorphic functions f : Cn → C. This then
serves as the local theory underpinning more global monodromy computations. Our reference
is [AGZV12].
Let U ⊂ C2 and V ⊂ C be open sets for which 0 ∈ V . Let f (u, v) : U → V be a
holomorphic function. Suppose f has an isolated critical value at z = 0, and that there is a
single corresponding critical point p ∈ C2 . Suppose that p is of Morse type in the sense that
the Hessian
!
∂2f
∂2x
∂2f
∂y∂x
is non-singular at p.
∂2f
∂x∂y
∂2f
∂2y
MONODROMY OF PLANE CURVES
11
In such a situation, the fiber f −1 (z) for z 6= 0 is diffeomorphic to an open annulus. The core
curve of such an annulus is called a vanishing cycle. Let γ be a small circle in C enclosing only
the critical value at z = 0. Let z1 ∈ γ be a basepoint with corresponding core curve c ⊂ f −1 (z1 ).
The Picard-Lefschetz theorem describes the monodromy obtained by traversing γ.
Theorem 5.1 (Picard-Lefschetz for n = 2). With reference to the preceding discussion, the
monodromy µ ∈ Mod(f −1 (z1 )) attached to traversing γ counter-clockwise is given by a righthanded Dehn twist about the vanishing cycle:
µ = Tc−1 .
More generally, let D∗ denote the punctured unit disk
D∗ = {w ∈ C | 0 < |w| ≤ 1},
and write D = {w ∈ C | |w| ≤ 1} for the closed unit disk.
Let f (x, y, z) be a homogeneous polynomial of degree d with the following properties:
(1) For c ∈ D, the plane curve cz d − f (x, y, z) is singular only for c = 0.
(2) The only critical point for f of the form (x, y, 0) is the point (0, 0, 0).
(3) The function f (x, y, 1) has a single critical point of Morse type at (x, y) = (0, 0).
In this setting, the local theory of Theorem 5.1 can be used to analyze the monodromy of the
family
E ⊂ D∗ × CP 2 = {(c, [x : y : z]) | cz d = f (x, y, z)}
around the boundary ∂D∗ .
Theorem 5.2 (Picard-Lefschetz for plane curve families). Let f ∈ CP N satisfy the properties
(1), (2), (3) listed above. Let X = V (z d − f (x, y, z)) denote the fiber above 1 ∈ D∗ . Then there
is a vanishing cycle c ⊂ X so that the monodromy µ ∈ Mod(X) obtained by traversing ∂D∗
counter-clockwise is given by a right-handed Dehn twist about the vanishing cycle:
µ = Tc−1 .
Proof. Condition (2) asserts that the monodromy can be computed by restricting attention to
the affine subfamily obtained by setting z = 1. Define
E ◦ ⊂ D∗ × C2 = {(c, x, y) | c = f (x, y, 1)}.
Define U = {(x, y) ∈ C2 | |f (x, y, 1)| ≤ 1}, and consider f (x, y, 1) as a holomorphic function
f : U → D. The monodromy of this family then corresponds to the monodromy of the original
family E → D∗ . The result now follows from Condition (3) in combination with Theorem 5.1
as applied to f (x, y, 1).
12
NICK SALTER
5.2. Lönne’s theorem. There are some preliminary notions to establish before Lönne’s theorem can be stated. We begin by introducing the Lönne graphs Γd . Lönne obtains his presentation of π1 (Pd ) as a quotient a certain group constructed from Γd .
Definition 5.3. [Lönne graph] Let d ≥ 3 be given. The Lönne graph Γd has vertex set
Id = {(a, b) | 1 ≤ a, b ≤ d − 1}.
Vertices (a1 , b1 ) and (a2 , b2 ) are connected by an edge if and only if both of the following
conditions are met:
(1) |a1 − a2 | ≤ 1 and |b1 − b2 | ≤ 1.
(2) (a1 − a2 )(b1 − b2 ) ≤ 0.
The set of edges of Γd is denoted Ed .
Figure 3. The Lönne graph Γ5 .
Vertices i, j, k ∈ Γd are said to form a triangle when i, j, k are mutually adjacent. The
triangles in the Lönne graph are crucial to what follows. It will be necessary to endow them
with orientations.
Definition 5.4 (Orientation of triangles). Let i, j, k determine a triangle in Γd .
(1) If
i = (a, b),
j = (a, b + 1),
k = (a + 1, b),
then the triangle i, j, k is positively-oriented by traversing the boundary clockwise.
(2) If
i = (a, b), j = (a, b + 1), k = (a − 1, b + 1),
then the triangle i, j, k is positively-oriented by traversing the boundary counterclockwise.
We say that the ordered triple (i, j, k) of vertices determining a triangle is positively-oriented
if traversing the boundary from i to j to k agrees with the orientation specified above.
Definition 5.5 (Artin group). Let Γ be a graph with vertex set V and edge set E. The Artin
group A(Γ) is defined to be the group with generators
σi , i ∈ V,
subject to the following relations:
MONODROMY OF PLANE CURVES
13
(1) σi σj = σj σi for all (i, j) 6∈ E.
(2) σi σj σi = σj σi σj for all (i, j) ∈ E.
Theorem 5.6 (Lönne). For d ≥ 3, the group π1 (Pd ) is isomorphic to a quotient of the Artin
group A(Γd ), subject to the following additional relations:
(3) σi σj σk σi = σj σk σi σj if (i, j, k) forms a positively-oriented triangle in Γd .
(4) An additional family of relations Ri , i ∈ Id .
e
(5) An additional relation R.
Remark 5.7. Define the group B(Γd ) as the quotient of the Artin group A(Γd ) by the family
of relations (3) in Theorem 5.6. As our statement of Lönne’s theorem indicates, the additional
relations will be of no use to us, and our theorem really concerns the lift of the monodromy
representation ρ̃d : B(Γd ) → Mod(Σg ).
For the analysis to follow, it is essential to understand the mapping classes ρd (σi ), i ∈ Id .
Proposition 5.8. For each generator σi of Theorem 5.6, the image
ρd (σi ) = Tc−1
i
is a right-handed Dehn twist about some vanishing cycle ci on a fiber X ∈ Pd .
Proof. The result will follow from Theorem 5.2, once certain aspects of Lönne’s proof are
recalled.
The generators σi of Theorem 5.6 correspond to specific loops in Pd known as geometric
elements.
Definition 5.9 (Geometric element). Let D = V (p) be a hypersurface in Cn defined by some
polynomial p. An element x ∈ π1 (Cn \ D) that can be represented by a path isotopic to the
e is a projective
boundary of a small disk transversal to D is called a geometric element. If D
n
e is said to be a geometric element if it can be
hypersurface, an element x ∈ π1 (CP \ D)
represented by a geometric element in some affine chart.
In Lönne’s terminology, the generators σi , i ∈ Id arise as a “Hefez-Lazzeri basis” - this will
require some explanation. Consider the linearly-perturbed Fermat polynomial
f (x, y, z) = xd + y d + νx xz d−1 + νy yz d−1
for well-chosen constants νx , νy . Such an f satisfies the conditions (1)-(3) of Theorem 5.2 near
each critical point. Moreover, there is a bijection between the critical points of f (x, y, 1) and
the set Id of Definition 5.3. If νx , νy are chosen carefully, each critical point lies above a distinct
critical value - in this way we embed Id ⊂ C.
Each c ∈ C determines a plane curve V (cz d − f ). The values of c for which V (cz d − f ) is
not smooth are exactly the critical values of f (x, y, 1). The family
H = {V (cz d − f ) | V (cz d − f ) is smooth}
14
NICK SALTER
is a subfamily of Pd defined over C \ Id . The Hefez-Lazzeri basis {σi | i ∈ Id } is a carefullychosen set of paths in C \ Id with each σi encircling an individual i ∈ Id . Lönne shows that
the inclusions of these paths into Pd via the family H generate π1 (Pd ). The result now follows
from an application of Theorem 5.2.
5.3. First properties of ρd . Proposition 5.8 establishes the existence of a collection ci , i ∈ Id
of vanishing cycles on X. In this section, we derive some basic topological properties of this
must satisfy the relations (1)-(3)
configuration arising from the fact that the Dehn twists Tc−1
i
of Lönne’s presentation.
Lemma 5.10.
(1)
(2)
(3)
(4)
If the vertices vi , vj are adjacent, then the curves ci , cj satisfy i(ci , cj ) = 1.
For d ≥ 4, the curves ci , i ∈ Id are pairwise distinct, and all ci are non-separating.
If the vertices vi , vj in Γd are non-adjacent, then the curves ci and cj are disjoint.
For d ≥ 4, if the vertices vi , vj , vk form a triangle in Γd , then the curves ci , cj , ck are
supported on an essential subsurface5 Sijk ⊂ X homeomorphic to Σ1,2 . Moreover, if
(ck )) = 0.
the triangle determined by vi , vj , vk is positively oriented, then i(ci , Tc−1
j
satisfy the braid
and Tc−1
Proof. (1): If vi and vj are adjacent, then the Dehn twists Tc−1
j
i
relation. It follows from Proposition 4.3 that i(ci , cj ) = 1.
(2): Suppose vi and vj are distinct vertices. For d ≥ 4, no two vertices have the same set of
adjacent vertices, so that there is some vk adjacent to vi and not vj . By (1) above, it follows
and Tc−1
do not, showing that the
and Tc−1
satisfy the braid relation, while Tc−1
that Tc−1
j
i
k
k
isotopy classes of ci and cj are distinct. Since each ci satisfies a braid relation with some other
cj , Proposition 4.3 shows that ci is non-separating.
commute. According to
and Tc−1
(3): If vi and vj are non-adjacent, then the Dehn twists Tc−1
j
i
[FM12, Section 3.5.2], this implies that either ci = cj or else ci and cj are disjoint. By (2), the
former possibility cannot hold.
(4): Via the change-of-coordinates principle, it can be checked that if ci , cj , ck are curves that
pairwise intersect once, then ci ∪ cj ∪ ck is supported on an essential subsurface of the form
(cj ). It
Σ1,b for 1 ≤ b ≤ 3. In the case b = 1, the curve ck must be of the form ck = Tc±1
i
follows that if d is a curve such that i(d, ck ) 6= 0, then at least one of i(d, ci ) and i(d, cj ) must
also be nonzero. However, as d ≥ 4, there is always some vertex vl adjacent to exactly one of
ci , cj , ck . The corresponding curve cl would violate the condition required of d above (possibly
after permuting the indices i, j, k).
5A subsurface S ′ ⊂ S is essential if every component of ∂S ′ is not null-homotopic.
MONODROMY OF PLANE CURVES
15
It remains to eliminate the possibility b = 3. In this case, the change-of-coordinates principle
implies that up to homeomorphism, the curves ci , cj , ck must be arranged as in Figure 4. It
can be checked directly (e.g. by examining the action on H1 (Σ1,3 )) that for this configuration,
the relation
Tc−1
Tc−1
Tc−1
= Tc−1
Tc−1
Tc−1
Tc−1
Tc−1
j
i
j
i
j
i
k
k
does not hold. This violates relation (3) in Lönne’s presentation of π1 (Pd ). We conclude that
necessarily b = 2.
(ck )) = 0 for a positivelyHaving shown that b = 2, it remains to show the condition i(ci , Tc−1
j
oriented triangle. Let (x, y, z) denote a 3-chain on Σ1,2 . The change-of-coordinates principle
implies that without loss of generality, ci = x, cj = y, and ck = Ty±1 (z). We wish to show that
necessarily ck = Ty (z). It can be checked directly (e.g. by considering the action on H1 (Σ1,2 ))
that only in the case ck = Ty (z) does relation (3) in the Lönne presentation hold.
Figure 4. Lemma 5.10.4: the configuration of ci , cj , ck in the b = 3 case.
6. Configurations of vanishing cycles
The goal of this section is to derive an explicit picture of the configuration of vanishing cycles
on a plane curve of degree 5. The main result of the section is Lemma 6.1.
Lemma 6.1. There is a homeomorphism f : X → Σ6 such that with reference to Figure 5,
(1) The curves c1 , . . . , c12 are vanishing cycles; that is, Tci ∈ Im(ρ5 ) for 1 ≤ i ≤ 12. The
curves x, y, z are also vanishing cycles.
(2) The curve b satisfies Tb2 ∈ Im(ρ5 ).
Proof. Lemma 6.1 will be proved in three steps.
Step 1: Uniqueness of Lönne configurations.
16
NICK SALTER
x
c6
b
c5
c4
c2
c7
c3
c1
c8
z
c13
c9
c12
c11
c10
y
Figure 5. The curves of Lemma 6.1. The bottom halves of curves b, x, y, z,
and ci for i odd have been omitted for clarity; on the bottom half, each curve
follows its mirror image on the top.
Lemma 6.2. Suppose d ≥ 5 is odd. Up to homeomorphism, there is a unique configuration of
curves ci , i ∈ Id on Σg whose intersection pattern is prescribed by Γd and such that the twists
satisfy the relations (1),(2),(3) given by Lönne’s presentation.
Tc−1
i
A configuration of curves ci , i ∈ Id as in Lemma 6.2 will be referred to as a Lönne configuration.
Proof. Let a1,1 , . . . , ad−1,d−1 determine a Lönne configuration on Σg . We will exhibit a homeomorphism of Σg taking each ai,j to a corresponding bi,j in a “reference” configuration {bi,j }
to be constructed in the course of the proof. This will require three steps.
Step 1: A collection of disjoint chains. Each row in the Lönne graph determines a chain
of length d − 1. The change of coordinates principle for chains of even length (Lemma 4.2)
asserts that any two chains of length d − 1 are equivalent up to homeomorphism. Considering
the odd-numbered rows of Γd , it follows that there is a homeomorphism f1 of Σg that takes
each a2i−1,j for 1 ≤ i ≤ d − 1 to a curve b2i−1,j in a standard picture of a chain. We denote the
subsurface of Σg determined by the chain a2i−1,1 , . . . , a2i−1,d−1 as Ai , and similarly we define
the subsurfaces Bi of the reference configuration. Each Ai , Bi is homeomorphic to Σ(d−1)/2,1 .
Step 2: Arcs on Ai . The next step is to show that up to homeomorphism, there is a unique
S
picture of what the intersection of the remaining curves a2i,j with Ai looks like. Consider
a curve a2i,j . Up to isotopy, a2i,j intersects only the subsurfaces Ai and Ai+1 . We claim that
MONODROMY OF PLANE CURVES
17
a2i,j can be isotoped so that its intersection with Ai is a single arc, and similarly for Ai+1 . If
j = d − 1, then a2i,d−1 intersects only the curve a2i−1,d−1 , and i(a2i,d−1 , a2i−1,d−1 ) = 1. It
follows that if a2i,d−1 ∩ Ai has multiple components, exactly one is essential, and the remaining
components can be isotoped off of Ai .
In the general case where a2i,j intersects both a2i−1,j and a2i−1,j+1 , an analogous argument
shows that a2i,j ∩ Ai consists of one or two essential arcs. Consider the triangle in the Lönne
graph determined by a2i,j , a2i−1,j , a2i−1,j+1 . According to Lemma 5.10.4, the union a2i,j ∪
a2i−1,j ∪ a2i−1,j+1 is supported on an essential subsurface of the form Σ1,2 . Figure 6 shows that
if a2i,j ∩ Ai consists of two essential arcs, then a2i,j ∪ a2i−1,j ∪ a2i−1,j+1 is supported on an
essential subsurface of the form Σ1,3 , in contradiction with Lemma 5.10.4. Similar arguments
establish that a2i−2,j ∩ Ai is a single essential arc as well.
Figure 6. If a2i,j cannot be isotoped onto a single arc inside Ai , then the
curve enclosed by the inner strip (shaded) is essential in Σg , causing a2i,j ∪
a2i−1,j ∪ a2i−1,j+1 to be supported on a surface Σ1,3 .
We next show that all points of intersection a2i,j ∩ a2i,j+1 can be isotoped to occur on both
Ai and Ai+1 . This also follows from Lemma 5.10.4. If some point of intersection a2i,j ∩ a2i,j+1
could not be isotoped onto Ai , then the union a2i,j ∪ a2i,j+1 ∪ a2i−1,j+1 could not be supported
on a subsurface homeomorphic to Σ1,2 . An analogous argument applies with Ai+1 in place of
Ai . This is explained in Figure 7.
Figure 7. If the intersection a2i,j ∩ a2i,j+1 cannot be isotoped to occur on Ai ,
then both curves indicated by the shaded regions are essential in Σg , causing
a2i,j ∪ a2i,j+1 ∪ a2i−1,j+1 to be supported on a surface Σ1,3 .
18
NICK SALTER
It follows from this analysis that all crossings between curves in row 2i can be isotoped to
occur in a collar neighborhood of ∂Ai . We define A+
i to be a slight enlargement of Ai along
such a neighborhood, so that all crossings between curves in row 2i occur in A+
i \ Ai .
+
We can now understand what the collection of arcs a2i,1 ∩ A+
,
.
.
.
,
a
∩
A
2i,d−1
i
i looks like. To
begin with, the change-of-coordinates principle asserts that up to a homeomorphism of Ai fixing
the curves {a2i−1,j }, the arc a2i,1 ∩ Ai can be drawn in one of two ways. The first possibility
is shown in Figure 8(a), and the second is its mirror-image obtained by reflection through the
plane of the page (i.e. the curve with the dotted and solid portions exchanged). In fact, a2i,1 ∩Ai
must look as shown. This follows from Lemma 5.10.4. The vertices (a2i−1,1 , a2i−1,2 , a2i,1 ) form
(a2i,1 )) = 0. This condition precludes the
a positively-oriented triangle, and so i(a2i−1,1 , Ta−1
2i−1,2
other possibility.
The pictures for a2i,2 , . . . , a2i,d−1 are obtained by proceeding inductively. In each case, there
are exactly two ways to draw an arc satisfying the requisite intersection properties, and Lemma
5.10.4 precludes one of these possibilities. The result is shown in Figure 8(b).
It remains to understand how the crossings between curves in row 2i are organized on A+
i .
As shown, the arcs a2i,j ∩Ai and a2i,j+1 ∩Ai intersect ∂Ai twice each, and in both instances the
intersections are adjacent relative to the other arcs. There are thus apparently two possibilities
for where the crossing can occur. However, one can see from Figure 8(c) that once a choice
is made for one crossing, this enforces choices for the remaining crossings. Moreover, the two
apparently distinct configurations are in fact equivalent: the cyclic ordering of the arcs along
∂A+
i is the same in either case, and the combinatorial type of the cut-up surface
A◦i := A+
i \
[
{ak,j | 2i − 1 ≤ k ≤ 2i, 1 ≤ j ≤ d − 1}
is the same in either situation. The change-of-coordinates principle then asserts the existence
of a homeomorphism of A+
i sending each a2i−1,j to itself, and taking one configuration of arcs
to the other.
Having seen that the arcs a2i,j ∩A+
i can be put into standard form, it remains to examine the
+
other collection of arcs on Ai , namely those of the form a2i−2,j . It is easy to see by induction
on d that the cut-up surface A◦i is a union of polyhedral disks for which the edges correspond
+
to portions of the curves a2i−1,j , the arcs a2i,j ∩ A+
i , or else the boundary ∂Ai . It follows that
the isotopy class of an arc a2i−2,j ∩ A+
I is uniquely determined by its intersection data with the
curves a2i−1,j and a2i,j .
For j ≥ 2, the curve a2i−2,j intersects a2i−1,j−1 and a2i−1,j , and is disjoint from all curves
a2i,k . As a2i,j−1 has the same set of intersections as a2i−2,j , it follows that a2i−2,j ∩ A+
i must
run parallel to a2i,j−1 . The curve a2i−2,1 intersects only a2i,1 ; consequently a2i−2,1 ∩ A+
i is
uniquely determined. As can be seen from Figure 8(c), this forces each subsequent a2i−2,j onto
a particular side of a2i,j−1 .
MONODROMY OF PLANE CURVES
19
(a)
(b)
(c)
Figure 8. The surface A+
i . (a): The correct choice for a2i,1 ∩ Ai . (b): The
configuration a2i,j ∩ Ai . (c) The configuration a2i,j ∩ A+
i .
Step 3: Arcs on the remainder of Σg . Consider now the subsurface
Σ◦g := Σg \
[
Ai .
This has (d−1)/2 boundary components ∂k , indexed by the corresponding Ak . The intersection
S
a2i,j ∩ (Σg \ Ai ) consists of two arcs, each connecting ∂i with ∂i+1 . The strategy for the
remainder of the proof is to argue that when all these arcs are deleted from Σ◦g , the result is
a union of disks. The change-of-coordinates principle will then assert the uniqueness of such a
configuration of arcs, completing the proof.
For what follows, it will be convenient to refer to a product neighborhood [0, 1] × [0, 1] ⊂ Σ◦g
of some arc a2i,j ∩ Σ◦g as a strip. Our first objective is to compute the Euler characteristic χ
◦
of the surface Σ◦◦
g obtained by deleting strips for all arcs from Σg . Then an analysis of the
pattern by which strips are attached will determine the number of components of this surface.
To begin, we return to the setting of Figure 7. Above, it was argued that for 2i < (d − 1)/2,
the intersection a2i,j ∩ a2i,j+1 can be isotoped onto either Ai or Ai+1 . This means that there
is a strip that contains both a2i,j ∩ Σ◦g and a2i,j+1 ∩ Σ◦g . Grouping such strips together, it can
be seen that for 1 ≤ i ≤ (d − 3)/2, the 2ith row of the Lönne graph gives rise to d strips. In
the last row, there are d − 1 strips. So in total there are 1/2(d + 1)(d − 2) strips, and each strip
contributes −1 to the Euler characteristic.
Recall the relation g = (d − 1)(d − 2)/2: this means that
χ(Σg ) = 2 − (d − 1)(d − 2).
20
NICK SALTER
Each Ai has Euler characteristic χ(Ai ) = 2 − d. It follows that
(d−1)/2
χ(Σ◦g )
= χ(Σg ) −
X
χ(Ai ) = 2 − (d − 1)(d − 2) + (d − 1)(d − 2)/2
i=1
=
2 − (d − 1)(d − 2)/2.
Therefore,
χ(Σ◦◦
g )=
χ(Σ◦g ) + 1/2(d + 1)(d − 2)
=
d.
We claim that Σ◦◦
g has d boundary components. This will finish the proof, as a surface of
Euler characteristic d and b = d boundary components must be a union of d disks. The claim
can easily be checked directly in the case d = 5 of immediate relevance. For general d, this
follows from a straightforward, if notationally tedious, verification, proceeding by an analysis
of the cyclic ordering of the arcs ai,j around the boundary components ∂A+
k.
Step 2: A convenient configuration.
Figure 9 presents a picture of a Lönne configuration in the case of interest d = 5. This
was obtained by “building the surface” curve by curve, attaching one-handles in the sequence
indicated by the numbering of the curves a1 , . . . , a16 . There are other, more uniform depictions
of Lönne configurations which arise from Akbulut-Kirby’s picture of a plane curve of degree d
derived from a Seifert surface of the (d, d) torus link (see [AK80] or [GS99, Section 6.2.7]), but
the analysis to follow is easier to carry out using the model of Figure 9.
Step 3: Producing vanishing cycles. The bulk of this step will establish claim (1); claim
(2) follows as an immediate porism. The principle is to exploit the fact that if a and b are
vanishing cycles, then so is Ta (b). To begin with, curves c1 , c2 , c4 , c8 , and c12 are elements of
the Lönne configuration and so are already vanishing cycles. The curve c3 is obtained as
(a3 );
c3 = Ta−1
2
similarly,
(a4 ).
c13 = Ta−1
2
Curve c10 is obtained as
c10 = Ta15 (a13 );
c6 is obtained from a14 and a16 analogously.
The curve c9 is obtained as
(a11 );
c9 = Tc10 Ta−1
10
c7 is obtained from a10 , a12 , and c6 analogously.
To obtain c5 , twist a13 along the chain c6 , . . . , c10 :
(a13 ).
Tc−1
Tc−1
Tc−1
Tc−1
c5 = Tc−1
10
9
8
7
6
MONODROMY OF PLANE CURVES
21
a9
a13
a1
a5
ι
a10
a2
a15
a3
a7
a11
a15
a5
a3
a1
a13
a7
a2
a4
a11
a9
a8
a6
a10
a12
a14
a16
Figure 9. A Lönne configuration on Σ6 . Only a portion of the figure has
been drawn: the omitted curves are obtained by applying the involution ι to
the depicted curves.
c11 is obtained by an analogous procedure on a14 .
The sequence of twists used to exhibit x as a vanishing cycle is illustrated in Figure 10.
Symbolically,
(a7 ).
Ta−1
Tc−1
Tc−1
Tc5 Tc4 Tc−1
Tc−1
Tc−1
Tc−1
x = Tc−1
9
8
7
6
9
8
7
6
y is produced in an analogous fashion, starting with a8 in place of a6 .
To produce z, we appeal to the genus-2 star relation. Applied to the surface bounded by
b, y, z, it shows that Tb2 Ty Tz ∈ Im(ρ5 ), and hence Tb2 Tz ∈ Im(ρ5 ) since Ty ∈ Im(ρ5 ) by above.
Observe that i(c10 , z) = 1, and that Tc10 ∈ Im(ρ5 ). Making use of the fact that b is disjoint
22
NICK SALTER
from both z and c10 , the braid relation gives
Tc10 Tb2 Tz (c10 ) = Tc10 Tz (c10 ) = z.
This exhibits z as a vanishing cycle, establishing claim (1) of Lemma 6.1. As Tb2 Tz and Tz are
now both known to be elements of Im(ρ5 ), it follows that Tb2 ∈ Im(ρ5 ) as well, completing claim
(2).
Ta−1
9
Tc−1
Tc−1
Tc−1
6
7
8
T c5 T c4
Tc−1
Tc−1
Tc−1
Tc−1
8
9
7
6
Figure 10. The sequence of twists used to obtain x.
7. Proof of Theorem B
In this final section we assemble the work we have done so far in order to prove Theorem B.
Step 1: Reduction to the Torelli group. The first step is to reduce the problem of
determining Im(ρ5 ) to the determination of Im(ρ5 ) ∩ I6 . This will follow from [Bea86]. Recall
that Beauville establishes that Im(Ψ ◦ ρ5 ) is the entire stabilizer of an odd-parity spin structure
on H1 (Σ6 ; Z). This spin structure was identified as φ5 in Section 2. Therefore Im(Ψ ◦ ρ5 ) =
Im(Ψ ◦ Mod(Σ6 )[φ5 ]). It therefore suffices to show that
Im(ρ5 ) ∩ ker Ψ = Mod(Σ6 )[φ5 ] ∩ ker Ψ = I6 .
(7)
Step 2: Enumeration of cases. Equation (7) will be derived as a consequence of Theorem
4.6. Lemma 6.1.1 asserts that the curves c1 , . . . , c12 in the Johnson generating set are contained
in Im(ρ5 ), so that the first hypothesis of Theorem 4.6 is satisfied. There are then eight cases
MONODROMY OF PLANE CURVES
23
γ
k=3
k=6
k=5
k=8
k=7
k = 10
k=9
k = 12
Figure 11. The cases of Step 2
to check: the four straight chain maps of the form (c1 , . . . , ck ) for k = 3, 5, 7, 9, and the four
β-chain maps of the form (β, c5 , . . . , ck ) for k = 6, 8, 10, 12. See Figure 11.
The verification of the β-chain cases will be easier to accomplish after conjugating by the
∈ Im(ρ5 ). This has the following effect on the curves in the β-chains (the
Tc−1
class g = Tx Tc−1
4
5
24
NICK SALTER
curve γ is indicated in Figure 11 in the picture for k = 6):
g(β) = b,
g(c5 ) = c4 ,
g(c6 ) = γ,
g(ck ) = ck for k ≥ 7.
Step 3: Producing bounding-pair maps. In this step, we explain the method by which
we will obtain the necessary bounding-pair maps. This is an easy consequence of the chain
relation.
Lemma 7.1. Let C = (c1 , . . . , ck ) be a chain of odd length k and boundary ∂C = d1 ∪ d2 .
Suppose that the mapping classes
Tc21 , Tc2 , . . . , Tck , Td21
are all contained in some subgroup Γ ≤ Mod(Σg ). Then the chain map associated to C (i.e.
) is also contained in Γ.
the bounding pair map Td1 Td−1
2
Proof. The chain relation (Proposition 4.4) implies that Td1 Td2 ∈ Γ. By hypothesis, Td21 ∈ Γ,
∈ Γ as well.
so the bounding pair map Td1 Td−1
2
Step 4: Verification of cases. Lemma 6.1 asserts that the classes Tci , 1 ≤ i ≤ 12, as well
as Tb2 are all contained in Im(ρ5 ). The class γ is obtained from c6 by the element g ∈ Im(ρ5 ),
so γ is a vanishing cycle as well. Via Lemma 7.1, it remains only to show that in each of the
cases in Step 2, one of the boundary components d1 satisfies Td21 ∈ Im(ρ5 ).
The straight chain maps are depicted in the left-hand column of Figure 11. For k = 3, one
boundary component is b; we have already remarked how Tb2 ∈ Im(ρ5 ). For k = 5, one of the
boundary components is x. For k = 7, one uses the methods of Lemma 6.1 to show that the
right-hand boundary component c satisfies Tc2 ∈ Im(ρ5 ) (the proof is identical to that for b).
Finally, for k = 9, one of the boundary components is y.
We turn to the β-chains. The images of the β-chains under the map g are depicted in the
right-hand column of Figure 11. For k = 6, 8, 10, 12, let dk denote the boundary component
depicted there for the chain (b, c4 , γ, c7 , . . . , ck ). Observe that dk is also a boundary component
of the chain map for (c6 , . . . , ck ) (in the case k = 6, the boundary component d6 is just c6 ).
Moreover, the chain map for (c6 , . . . , ck ) is conjugate to the chain map for (c1 , . . . , ck−5 ) by an
element of Im(ρ5 ) (this is easy to see using the isomorphism between the group generated by
c1 , . . . , c12 and the braid group B13 on 13 strands). Via the verification of the straight-chain
cases, it follows that Td2k ∈ Im(ρ5 ), and so by Lemma 7.1 the β-chain maps are also contained
in Im(ρ5 ).
References
[AGZV12] V. I. Arnold, S. M. Gusein-Zade, and A. N. Varchenko. Singularities of differentiable maps. Volume
2. Modern Birkhäuser Classics. Birkhäuser/Springer, New York, 2012. Monodromy and asymptotics
of integrals, Translated from the Russian by Hugh Porteous and revised by the authors and James
Montaldi, Reprint of the 1988 translation.
MONODROMY OF PLANE CURVES
25
[AK80]
S. Akbulut and R. Kirby. Branched covers of surfaces in 4-manifolds. Math. Ann., 252(2):111–131,
[Bea86]
1979/80.
A. Beauville. Le groupe de monodromie des familles universelles d’hypersurfaces et d’intersections
[Chm82]
Notes in Math., pages 8–18. Springer, Berlin, 1986.
S. V. Chmutov. Monodromy groups of critical point of functions. Invent. Math., 67(1):123–131, 1982.
complètes. In Complex analysis and algebraic geometry (Göttingen, 1985), volume 1194 of Lecture
[DL81]
I. Dolgachev and A. Libgober. On the fundamental group of the complement to a discriminant
variety. In Algebraic geometry (Chicago, Ill., 1980), volume 862 of Lecture Notes in Math., pages
1–25. Springer, Berlin-New York, 1981.
[FM12]
B. Farb and D. Margalit. A primer on mapping class groups, volume 49 of Princeton Mathematical
Series. Princeton University Press, Princeton, NJ, 2012.
[GS99]
R. Gompf and A. Stipsicz. 4-manifolds and Kirby calculus, volume 20 of Graduate Studies in Math-
[Hir05]
ematics. American Mathematical Society, Providence, RI, 1999.
S. Hirose. Surfaces in the complex projective plane and their mapping class groups. Algebr. Geom.
[HJ89]
Topol., 5:577–613 (electronic), 2005.
S. Humphries and D. Johnson. A generalization of winding number functions on surfaces. Proc.
London Math. Soc. (3), 58(2):366–386, 1989.
[Jan83]
W. A. M. Janssen. Skew-symmetric vanishing lattices and their monodromy groups. Math. Ann.,
266(1):115–133, 1983.
[Joh83]
D. Johnson. The structure of the Torelli group. I. A finite set of generators for I. Ann. of Math. (2),
[Kun08]
118(3):423–442, 1983.
Y. Kuno. The mapping class group and the Meyer function for plane curves. Math. Ann., 342(4):923–
[Lön09]
949, 2008.
M. Lönne. Fundamental groups of projective discriminant complements. Duke Math. J., 150(2):357–
405, 2009.
[Sip82]
[Sip86]
P. Sipe. Roots of the canonical bundle of the universal Teichmüller curve and certain subgroups of
the mapping class group. Math. Ann., 260(1):67–92, 1982.
P. Sipe. Some finite quotients of the mapping class group of a surface. Proc. Amer. Math. Soc.,
97(3):515–524, 1986.
E-mail address: [email protected]
Department of Mathematics, University of Chicago, 5734 S. University Ave., Chicago, IL 60637
| 4math.GR
|
Unrestricted Facial Geometry Reconstruction Using Image-to-Image Translation
Matan Sela
Elad Richardson
Ron Kimmel
Department of Computer Science, Technion - Israel Institute of Technology
arXiv:1703.10131v2 [cs.CV] 15 Sep 2017
{matansel,eladrich,ron}@cs.technion.ac.il
Figure 1: Results of the proposed method. Reconstructed geometries are shown next to the corresponding input images.
Abstract
It has been recently shown that neural networks can recover the geometric structure of a face from a single given
image. A common denominator of most existing face geometry reconstruction methods is the restriction of the solution space to some low-dimensional subspace. While such
a model significantly simplifies the reconstruction problem,
it is inherently limited in its expressiveness. As an alternative, we propose an Image-to-Image translation network
that jointly maps the input image to a depth image and a
facial correspondence map. This explicit pixel-based mapping can then be utilized to provide high quality reconstructions of diverse faces under extreme expressions, using a
purely geometric refinement process. In the spirit of recent
approaches, the network is trained only with synthetic data,
and is then evaluated on “in-the-wild” facial images. Both
qualitative and quantitative analyses demonstrate the accuracy and the robustness of our approach.
1. Introduction
Recovering the geometric structure of a face is a fundamental task in computer vision with numerous applications. For example, facial characteristics of actors in realistic movies can be manually edited with facial rigs that
are carefully designed for manipulating the expression [42].
While producing animation movies, tracking the geometry
of an actor across multiple frames allows transferring the
expression to an animated avatar [14, 8, 7]. Image-based
face recognition methods deform the recovered geometry
for producing a neutralized and frontal version of the input face in a given image, reducing the variations between
images of the same subject [49, 19]. As for medical applications, acquiring the structure of a face allows for fine
planning of aesthetic operations and plastic surgeries, designing of personalized masks [2, 37] and even bio-printing
facial organs.
Here, we focus on the recovery of the geometric structure
of a face from a single facial image under a wide range of
expressions and poses. This problem has been investigated
for decades and most existing solutions involve one or more
of the following components.
• Facial landmarks [25, 46, 32, 47] - a set of automatically detected key points on the face such as the tip of
the nose and the corners of the eyes, which can guide
the reconstruction process [49, 26, 1, 12, 29].
• A reference facial model - an average neutral face that
is used as an initialization of optical flow or shape from
shading procedures [19, 26].
• A three-dimensional morphable model - a prior lowdimensional linear subspace of plausible facial geometries which allows an efficient, yet rough, recovery of
a facial structure [4, 6, 49, 36, 23, 33, 43],
While using these components can simplify the reconstruction problem, they introduce some inherent limitations.
Methods that rely only on landmarks are limited to a sparse
set of constrained points. Classical techniques that use a
Non-Rigid
Registration
Fine Detail
Reconstruction
Image-to-Image
Network
Figure 2: The algorithmic reconstruction pipeline.
reference facial model might fail to recover extreme expressions and non-frontal poses, as optical flows restrict the deformation to the image plane. The morphable model, while
providing some robustness, limits the reconstruction as it
can express only coarse geometries. Integrating some of
these components together could mitigate the problems, yet,
the underlying limitations are still manifested in the final reconstruction.
Alternatively, we propose an unrestricted approach
which involves a fully convolutional network that learns to
translate an input facial image to a representation containing
two maps. The first map is an estimation of a depth image,
while the second is an embedding of a facial template mesh
in the image domain. This network is trained following the
Image-to-Image translation framework of [22], where an
additional normal-based loss is introduced to enhance the
depth result. Similar to previous approaches, we use synthetic images for training, where the images are sampled
from a wide range of facial identities, poses, expressions,
lighting conditions, backgrounds and material parameters.
Surprisingly, even though the network is still trained with
faces that are drawn from a limited generative model, it
can generalize and produce structures far and beyond the
limited scope of that model. To process the raw network
results, an iterative facial deformation procedure is used
which combines the representations into a full facial mesh.
Finally, a refinement step is applied to produce a detailed reconstruction. This novel blending of neural networks with
purely geometric techniques allows us to reconstruct highquality meshes with wrinkles and details at a mesoscopiclevel from only a single image.
While using a neural network for face reconstruction was
proposed in the past [33, 34, 43, 48, 24], previous methods
were still limited by the expressiveness of the linear model.
In [34], a second network was proposed to refine the coarse
facial reconstruction, yet, it could not compensate for large
geometric variations beyond the given subspace. For example, the structure of the nose was still limited by the span
of a facial morphable model. By learning the unconstrained
geometry directly in the image domain, we overcome this
limitation, as demonstrated by both quantitative and qualitative experimental results. To further analyze the potential of the proposed representation we devise an application
for translating images from one domain to another. As a
case study, we transform synthetic facial images into realistic ones, enforcing our network as a loss function to preserve the geometry throughout the cross domain mapping.
The main contributions of this paper are:
• A novel formulation for predicting a geometric representation of a face from a single image, which is not
restricted to a linear model.
• A purely geometric deformation and refinement procedure that utilizes the network representation to produce
high quality facial reconstructions.
• A novel application of the proposed network which allows translating synthetic facial images into realistic
ones, while keeping the geometric structure intact.
2. Overview
The algorithmic pipeline is presented in Figure 2. The
input of the network is a facial image, and the network
produces two outputs: The first is an estimated depth map
aligned with the input image. The second output is a dense
map from each pixel to a corresponding vertex on a reference facial mesh. To bring the results into full vertex correspondence and complete occluded parts of the face, we
warp a template mesh in the three-dimensional space by an
iterative non-rigid deformation procedure. Finally, a fine
detail reconstruction algorithm guided by the input image
recovers the subtle geometric structure of the face. Code
for evaluation is available at https://github.com/
matansel/pix2vertex.
3. Learning the Geometric Representation
There are several design choices to consider when working with neural networks. First and foremost is the training
data, including the input channels, their labels, and how to
gather the samples. Second is the choice of the architecture.
A common approach is to start from an existing architecture [27, 39, 40, 20] and to adapt it to the problem at hand.
Finally, there is the choice of the training process, including
the loss criteria and the optimization technique. Next, we
describe our choices for each of these elements.
3.1. The Data and its Representation
The purpose of the suggested network is to regress a geometric representation from a given facial image. This representation is composed of the following two components:
Depth Image A depth profile of the facial geometry. Indeed, for many facial reconstruction tasks providing only
the depth profile is sufficient [18, 26].
Correspondence Map An embedding which allows mapping image pixels to points on a template facial model,
given as a triangulated mesh. To compute this signature
for any facial geometry, we paint each vertex with the x, y,
and z coordinates of the corresponding point on a normalized canonical face. Then, we paint each pixel in the map
Figure 3: A reference template face presented alongside the
dense correspondence signature from different viewpoints.
Figure 4: Training data samples alongside their representations.
with the color value of the corresponding projected vertex,
see Figure 3. This feature map is a deformation agnostic
representation, which is useful for applications such as facial motion capture [44], face normalization [49] and texture mapping [50]. While a similar representation was used
in [34, 48] as feedback channel for an iterative network, the
facial recovery was still restricted to the span of a facial
morphable model.
For training the network, we adopt the same synthetic
data generation procedure proposed in [33]. Each random
face is generated by drawing random mesh coordinates S
and texture T from a facial morphable model [4]. In practice, we draw a pair of Gaussian random vectors, αg and αt ,
and recover the synthetic face as follows
S = µg + Ag αg
T = µt + At αt .
where µg and µt are the stacked average facial geometry
and texture of the model, respectively. Ag and At are matrices whose columns are the bases of low-dimensional linear subspaces spanning plausible facial geometries and textures, respectively. Notice that geometry basis Ag is composed to both identity and expression basis elements, as proposed in [10]. Next, we render the random textured meshes
under various illumination conditions and poses, generating a dataset of synthetic facial images. As the ground-truth
geometry is known for each synthetic image, one readily
has the matching depth and correspondence maps to use as
labels. Some examples of input images alongside their desired outputs are shown in Figure 4.
Working with synthetic data can still present some gaps
when generalizing to “in-the-wild” images [9, 33], however
it provides much-needed flexibility in the generation process and ensures a deterministic connection from an image
to its label. Alternatively, other methods [16, 43] proposed
to generate training data by employing existing reconstruction algorithms and regarding their results as ground-truth
labels. For example, Güler et al. [16], used a framework
similar to that of [48] to match dense correspondence maps
to a dataset of facial images, starting from only a sparse set
of landmarks. These correspondence maps were then used
as training labels for their method. Notice that such data
can also be used for training our network without requiring
any other modification.
3.2. Image to Geometry Translation
Pixel-wise prediction requires a proper network architecture [30, 17]. The proposed structure is inspired by
the recent Image-to-Image translation framework proposed
in [22], where a network was trained to map the input image to output images of various types. The architecture used
there is based on the U-net [35] layout, where skip connections are used between corresponding layers in the encoder
and the decoder. Additional considerations as to the network implementation are given in the supplementary.
While in [22] a combination of L1 and adversarial loss
functions were used, in the proposed framework, we chose
to omit the adversarial loss. That is because unlike the problems explored in [22], our setup includes less ambiguity in
the mapping. Hence, a distributional loss function is less
effective, and mainly introduces artifacts. Still, since the
basic L1 loss function favors sparse errors in the depth prediction and does not account for differences between pixel
neighborhoods, it is insufficient for producing fine geometric structures, see Figure 5b. Hence, we propose to augment the loss function with an additional L1 term, which
penalizes the discrepancy between the normals of the reconstructed depth and ground truth.
LN (ẑ, z) = k~n (ẑ) − ~n (z)k1 ,
(1)
where ẑ is the recovered depth, and z denotes the groundtruth depth image. During training we set λL1 = 100 and
λN = 10, where λL1 and λN are the matching loss weights.
Note that for the correspondence image only the L1 loss was
applied. Figure 5 demonstrates the contribution of the LN
to the quality of the depth reconstruction provided by the
network.
4. From Representations to a Mesh
Based on the resulting depth and correspondence we introduce an approach to translate the 2.5D representation to
a 3D facial mesh. The procedure is composed of an iterative
elastic deformation algorithm (4.1) followed by a fine detail
recovery step driven by the input image (4.2). The resulting
output is an accurate reconstructed facial mesh with a full
vertex correspondence to a template mesh with fixed triangulation. This type of data is helpful for various dynamic
facial processing applications, such as facial rigs, which allows creating and editing photo-realistic animations of actors. As a byproduct, this process also corrects the prediction of the network by completing domains in the face
which are mistakenly classified as part of the background.
4.1. Non-Rigid Registration
Next, we describe the iterative deformation-based registration pipeline. First, we turn the depth map from the network into a mesh, by connecting neighboring pixels. Based
on the correspondence map from the network, we compute
the affine transformation from a template face to the mesh.
This operation is done by minimizing the squared Euclidean
distances between corresponding vertex pairs. Next, similar to [28], an iterative non-rigid registration process deforms the transformed template, aligning it with the mesh.
Note that throughout the registration, only the template is
warped, while the target mesh remains fixed. Each iteration
involves the following four steps.
1. Each vertex in the template mesh, vi ∈ V, is associated with a vertex, ci , on the target mesh, by evaluating the nearest neighbor in the correspondence embedding space. This step is different from the method
described in [28], which computes the nearest neighbor in the Euclidean space. As a result, the proposed
step allows registering a single template face to different facial identities with arbitrary expressions.
2. Pairs, (vi , ci ), which are physically distant and those
whose normal directions disagree are detected and ignored in the next step.
3. The template mesh is deformed by minimizing the following energy
X
E(V, C) = αp2point
kvi − ci k22
(vi ,ci )∈J
+αp2plane
(a)
(b)
(c)
+αmemb
X
2
|~n(ci )(vi − ci )|
(vi ,ci )∈J
X
X
wi,j kvi − vj k22 ,
i∈V vj ∈N (vi )
Figure 5: (a) the input image, (b) the result with only the L1
loss function and (c) the result with the additional normals
loss function. Note the artifacts in (b).
(2)
where, wi,j is the weight corresponding to the biharmonic Laplacian operator (see [21, 5]), ~n(ci ) is the
normal of the corresponding vertex at the target mesh
ci , J is the set of the remaining associated vertex pairs
(vi , ci ), and N (vi ) is the set 1-ring neighboring vertices about the vertex vi . Notice that the first term
above is the sum of squared Euclidean distances between matches. The second term is the distance from
the point vi to the tangent plane at the corresponding
point of the target mesh. The third term quantifies the
stiffness of the mesh.
4. If the motion of the template mesh between the current
iteration and the previous one is below a fixed threshold, we divide the weight αmemb by two. This relaxes
the stiffness term and allows a greater deformation in
the next iteration.
This iterative process terminates when the stiffness weight
is below a given threshold. Further implementation information and parameters of the registration process are provided in the supplementary material. The resulting output
of this phase is a deformed template with fixed triangulation, which contains the overall facial structure recovered
by the network, yet, is smoother and complete, see the third
column of Figure 9.
4.2. Fine Detail Reconstruction
Although the network already recovers some fine geometric details, such as wrinkles and moles, across parts
of the face, a geometric approach can reconstruct details
at a finer level, on the entire face, independently of the
resolution. Here, we propose an approach motivated by
the passive-stereo facial reconstruction method suggested
in [3]. The underlying assumption here is that subtle geometric structures can be explained by local variations in the
image domain. For some skin tissues, such as nevi, this assumption is inaccurate as the intensity variation results from
the albedo. In such cases, the geometric structure would be
wrongly modified. Still, for most parts of the face, the reconstructed details are consistent with the actual variations
in depth.
The method begins from an interpolated version of the
deformed template. Each vertex v ∈ VD is painted with the
intensity value of the nearest pixel in the image plane. Since
we are interested in recovering small details, only the high
spatial frequencies, µ(v), of the texture, τ (v), are taken into
consideration in this phase. For computing this frequency
band, we subtract the synthesized low frequencies from the
original intensity values. This low-pass filtered part can be
computed by convolving the texture with a spatially varying Gaussian kernel in the image domain, as originally proposed. In contrast, since this convolution is equivalent to
computing the heat distribution upon the shape after time
dt, where the initial heat profile is the original texture, we
Figure 6: Mesoscopic displacement. From left to right:
an input image, the shape after the iterative registration,
the high-frequency part of the texture - µ(v), and the final
shape.
propose to compute µ(v) as
µ(v) = τ (v) − (I − dt · ∆g )−1 τ (v),
(3)
where I is the identity matrix, ∆g is the cotangent weight
discrete Laplacian operator for triangulated meshes [31],
and dt is a scalar proportional to the cut-off frequency of
the filter.
Next, we displace each vertex along its normal direction
such that v 0 = v + δ(v)~n(v). The step size of the displacement, δ(v), is a combination of a data-driven term, δµ (v),
and a regularization one, δs (v). The data-driven term is
guided by the high-pass filtered part of the texture, µ(v). In
practice, we require the local differences in the geometry to
be proportional to the local variation in the high frequency
band of the texture. For each vertex v, with a normal ~n(v),
and a neighboring vertex vi , the data-driven term is given
by
P
n(v)i|
i ,~
α(v,vi ) (µ(v) − µ(vi )) 1 − |hv−v
kv−vi k
δµ (v) =
vi ∈N (v)
P
,
α(v,vi )
vi ∈N (v)
(4)
where α(v,vi ) = exp (−kv − vi k). For further explanation
of Equation 4, we refer the reader to the supplementary
material of this paper or the implementation details of [3].
Since we move each vertex along the normal direction,
triangles could intersect each other, particularly in domains
of high curvature. To reduce the probability of such collisions, a regularizing displacement field, δs (v), is added.
This term is proportional to the mean curvature of the original surface, and is equivalent to a single explicit mesh fairing step [11]. The final surface modification is given by
v 0 = v + (ηδµ (v) + (1 − η)δs (v)) · ~n(v),
(5)
for some constant η ∈ [0, 1]. A demonstration of the results
before and after this step is presented in Figure 6
5. Experiments
Next, we present evaluations on both the proposed network and the pipeline as a whole, and comparison to different prominent methods of single image based facial reconstruction [26, 49, 34].
5.2. Quantitative Evaluation
Figure 7: Network Output.
For a quantitative comparison, we used the first 200 subjects from the BU-3DFE dataset [45], which contains facial images aligned with ground truth depth images. Each
method provides its own estimation for the depth image
alongside a binary mask, representing the valid pixels to
be taken into account in the evaluation. Obviously, since
the problem of reconstructing depth from a single image is
ill-posed, the estimation needs to be judged up to global
scaling and transition along the depth axis. Thus, we compute these paramters using the Random Sample Concensus
(RANSAC) approach [13], for normalizing the estimation
according to the ground truth depth. This significantly reduces the absolute error of each method as the global parameter estimation is robust to outliers. Note that the parameters of the RANSAC were identical for all the methods
and samples. The results of this comparison are given in Table 1, where the units are given in terms of the percentile of
the ground-truth depth range. As a further analysis of the
reconstruction accuracy, we computed the mean absolute
error of each method based on expressions, see Table 2.
Figure 8: Texture mapping via the embedding.
5.1. Qualitative Evaluation
The first component of our algorithm is an Image-toImage network. In Figure 7, we show samples of output
maps produced by the proposed network. Although the
network was trained with synthetic data, with simple random backgrounds (see Figure 4), it successfully separates
the hair and background from the face itself and learns the
corresponding representations. To qualitatively assess the
accuracy of the correspondence, we present a visualization where an average facial texture is mapped to the image plane via the predicted embedding, see Figure 8, this
shows how the network successfully learns to represent the
facial structure. Next, in Figure 9 we show the reconstruction of the network, alongside the registered template and
the final shape. Notice how the structural information retrieved by the network is preserved through the geometric
stages. Figure 10 shows a qualitative comparison between
the proposed method and others. One can see that our
method better matches the global structure, as well as the
facial details. To better perceive these differences, see Figure 11. Finally, to demonstrate the limited expressiveness
of the 3DMM space compared to our method, Figure 12
presents our registered template next to its projection onto
the 3DMM space. This clearly shows that our network is
able to learn structures which are not spanned by the 3DMM
model.
Figure 9: The reconstruction stages. From left to right: the
input image, the reconstruction of the network, the registered template and the final shape.
Input
Proposed
[34]
[26]
[49]
Proposed
[34]
[26]
[49]
Figure 10: Qualitative comparison. Input images are presented alongside the reconstructions of the different methods.
[26]
[49]
[34]
Ours
Mean Err.
3.89
3.85
3.61
3.51
Std Err.
4.14
3.23
2.99
2.69
Median Err.
2.94
2.93
2.72
2.65
90% Err.
7.34
7.91
6.82
6.59
Table 1: Quantitative evaluation on the BU-3DFE Dataset.
From left to right: the absolute depth errors evaluated by
mean, standard deviation, median and the average ninety
percent largest error.
Input
Proposed
[34]
[26]
[49]
Figure 11: Zoomed qualitative result of first and fourth subjects from Figure 10.
5.3. The Network as a Geometric Constraint
As demonstrated by the results, the proposed network
successfully learns both the depth and the embedding representations for a variety of images. This representation is
the key part behind the reconstruction pipeline. However,
it can also be helpful for other face-related tasks. As an
example, we show that the network can be used as a geometric constraint for facial image manipulations, such as
transforming synthetic images into realistic ones. This idea
[26]
[49]
[34]
Ours
AN
3.47
4.00
3.42
3.67
DI
4.03
3.93
3.46
3.34
FE
3.94
3.91
3.64
3.36
HA
4.30
3.70
3.41
3.01
NE
3.43
3.76
4.22
3.17
SA
3.52
3.61
3.59
3.37
SU
4.19
3.96
4.00
4.41
Table 2: The mean error by expression. From left to right:
Anger, Disgust, Fear, Happy, Neutral, Sad, Surprise.
is based on recent advances in applying Generative Adversarial Networks (GAN) [15] for domain adaption tasks [41].
In the basic GAN framework, a Generator Network (G)
learns to map from the source domain, DS , to the target domain DT , where a Discriminator Network (D) tries to dis-
Figure 12: 3DMM Projection. From left to right: the input
image, the registered template, the projected mesh and the
projection error.
tinguish between generated images and samples from the
target domain, by optimizing the following objective
minmaxV (D, G) =
G
D
Ey∼DT [log D (y)]
(6)
+ Ex∼DS [log (1 − D (G (x)))] .
Theoretically, this framework could also translate images from the synthetic domain into the realistic one. However, it does not guarantee that the underlying geometry of
the synthetic data is preserved throughout that transformation. That is, the generated image might look realistic, but
have a completely different facial structure from the synthetic input. To solve that potential inconsistency, we suggest to involve the proposed network as an additional loss
function on the output of the generator.
LGeom (x) = kN et (x) − N et (G (x))k1 ,
(7)
where N et(·) represents the operation of the introduced network. Note that this is feasible, thanks to the fact that the
proposed network is fully differentiable. The additional geometric fidelity term forces the generator to learn a mapping that makes a synthetic image more realistic while keeping the underlying geometry intact. This translation process could potentially be useful for data generation procedures, similarly to [38]. Some successful translations are
visualized in Figure 13. Notice that the network implicitly
learns to add facial hair and teeth, and modify the texture
the and shading, without changing the facial structure. As
demonstrated by this analysis, the proposed network learns
a strong representation that has merit not only for reconstruction, but for other tasks as well.
Figure 13: Translation results. From top to bottom: synthetic input images, the correspondence and the depth maps
recovered by the network, and the transformed result.
network represents the solution in the extremely highdimensional image domain. This structure is learned from
synthetic examples, and shown to successfully generalize to
“in-the-wild” images. Still, facial images that significantly
deviate from our training domain are challenging, resulting
in missing areas and errors inside the representation maps.
More specifically, our network has difficulty handling extreme occlusions such as sunglasses, hands or beards, as
these were not seen in the training data. Similarly to other
methods, reconstructions under strong rotations are also not
well handled. Reconstructions under such scenarios are
shown in the supplementary material. Another limiting factor of our pipeline is speed. While the suggested network
by itself can be applied efficiently, our template registration
step is currently not optimized for speed and can take a few
minutes to converge.
7. Conclusion
We presented an unrestricted approach for recovering the
geometric structure of a face from a single image. Our algorithm employs an Image-to-Image network which maps the
input image to a pixel-based geometric representation, followed by geometric deformation and refinement steps. The
network is trained only by synthetic facial images, yet, is
capable of reconstructing real faces. Using the network as a
loss function, we propose a framework for translating synthetic facial images into realistic ones while preserving the
geometric structure.
6. Limitations
One of the core ideas of this work was a model-free
approach, where the solution space is not restricted by a
low dimensional subspace. Instead, the Image-to-Image
Acknowledgments
We would like to thank Roy Or-El for the helpful discussions and comments.
References
[1] O. Aldrian and W. A. Smith. A linear approach of 3D face
shape and texture recovery using a 3d morphable model.
In Proceedings of the British Machine Vision Conference,
pages, pages 75–1, 2010.
[2] I. Amirav, A. S. Luder, A. Halamish, D. Raviv, R. Kimmel, D. Waisman, and M. T. Newhouse. Design of aerosol
face masks for children using computerized 3d face analysis.
Journal of aerosol medicine and pulmonary drug delivery,
27(4):272–278, 2014.
[3] T. Beeler, B. Bickel, P. Beardsley, B. Sumner, and M. Gross.
High-quality single-shot capture of facial geometry. In ACM
SIGGRAPH 2010 Papers, SIGGRAPH ’10, pages 40:1–
40:9, New York, NY, USA, 2010. ACM.
[4] V. Blanz and T. Vetter. A morphable model for the synthesis
of 3D faces. In Proceedings of the 26th annual conference on
Computer graphics and interactive techniques, pages 187–
194. ACM Press/Addison-Wesley Publishing Co., 1999.
[5] M. Botsch and O. Sorkine. On linear variational surface deformation methods. IEEE Transactions on Visualization and
Computer Graphics, 14(1):213–230, Jan 2008.
[6] P. Breuer, K.-I. Kim, W. Kienzle, B. Scholkopf, and V. Blanz.
Automatic 3D face reconstruction from single images or
video. In Automatic Face & Gesture Recognition, 2008.
FG’08. 8th IEEE International Conference on, pages 1–8.
IEEE, 2008.
[7] C. Cao, D. Bradley, K. Zhou, and T. Beeler. Real-time highfidelity facial performance capture. ACM Transactions on
Graphics (TOG), 34(4):46, 2015.
[8] C. Cao, Y. Weng, S. Lin, and K. Zhou. 3D shape regression
for real-time facial animation. ACM Transactions on Graphics (TOG), 32(4):41, 2013.
[9] W. Chen, H. Wang, Y. Li, H. Su, D. Lischinsk, D. Cohen-Or,
B. Chen, et al. Synthesizing training images for boosting human 3D pose estimation. arXiv preprint arXiv:1604.02703,
2016.
[10] B. Chu, S. Romdhani, and L. Chen. 3d-aided face recognition robust to expression and pose variations. In 2014 IEEE
Conference on Computer Vision and Pattern Recognition,
pages 1907–1914. IEEE, 2014.
[11] M. Desbrun, M. Meyer, P. Schröder, and A. H. Barr. Implicit fairing of irregular meshes using diffusion and curvature flow. In Proceedings of the 26th Annual Conference
on Computer Graphics and Interactive Techniques, SIGGRAPH ’99, pages 317–324, New York, NY, USA, 1999.
ACM Press/Addison-Wesley Publishing Co.
[12] P. Dou, Y. Wu, S. K. Shah, and I. A. Kakadiaris. Robust 3D
face shape reconstruction from single images via two-fold
coupled structure learning. In Proc. British Machine Vision
Conference, pages 1–13, 2014.
[13] M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image
analysis and automated cartography. Communications of the
ACM, 24(6):381–395, 1981.
[14] P. Garrido, M. Zollhöfer, D. Casas, L. Valgaerts, K. Varanasi,
P. Pérez, and C. Theobalt. Reconstruction of personalized
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
3D face rigs from monocular video. ACM Transactions on
Graphics (TOG), 35(3):28, 2016.
I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in Neural Information
Processing Systems, pages 2672–2680, 2014.
R. A. Güler, G. Trigeorgis, E. Antonakos, P. Snape,
S. Zafeiriou, and I. Kokkinos. Densereg: Fully convolutional dense shape regression in-the-wild. arXiv preprint
arXiv:1612.01202, 2016.
B. Hariharan, P. Arbeláez, R. Girshick, and J. Malik. Hypercolumns for object segmentation and fine-grained localization. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pages 447–456, 2015.
T. Hassner. Viewing real-world faces in 3d. In Proceedings
of the IEEE International Conference on Computer Vision,
pages 3607–3614, 2013.
T. Hassner, S. Harel, E. Paz, and R. Enbar. Effective face
frontalization in unconstrained images. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 4295–4304, 2015.
K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In The IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), June 2016.
B. T. Helenbrook. Mesh deformation using the biharmonic
operator. International journal for numerical methods in engineering, 56(7):1007–1021, 2003.
P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Imageto-image translation with conditional adversarial networks.
arXiv preprint arXiv:1611.07004, 2016.
L. Jiang, J. Zhang, B. Deng, H. Li, and L. Liu. 3d face reconstruction with geometry details from a single image. arXiv
preprint arXiv:1702.05619, 2017.
A. Jourabloo and X. Liu. Large-pose face alignment via
cnn-based dense 3D model fitting. In The IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), June
2016.
V. Kazemi and J. Sullivan. One millisecond face alignment
with an ensemble of regression trees. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 1867–1874, 2014.
I. Kemelmacher-Shlizerman and R. Basri. 3D face reconstruction from a single image using a single reference face
shape. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 33(2):394–405, 2011.
A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
Advances in neural information processing systems, pages
1097–1105, 2012.
H. Li. Animation Reconstruction of Deformable Surfaces.
PhD thesis, ETH Zurich, November 2010.
F. Liu, D. Zeng, J. Li, and Q. Zhao. Cascaded regressor based
3D face reconstruction from a single arbitrary view image.
arXiv preprint arXiv:1509.06161, 2015.
J. Long, E. Shelhamer, and T. Darrell. Fully convolutional
networks for semantic segmentation. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 3431–3440, 2015.
[31] M. Meyer, M. Desbrun, P. Schröder, A. H. Barr, et al.
Discrete differential-geometry operators for triangulated 2manifolds. Visualization and mathematics, 3(2):52–58,
2002.
[32] S. Ren, X. Cao, Y. Wei, and J. Sun. Face alignment at 3000
fps via regressing local binary features. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 1685–1692, 2014.
[33] E. Richardson, M. Sela, and R. Kimmel. 3D face reconstruction by learning from synthetic data. In 3D Vision (3DV),
2016 International Conference on, pages 460–469. IEEE,
2016.
[34] E. Richardson, M. Sela, R. Or-El, and R. Kimmel. Learning detailed face reconstruction from a single image. arXiv
preprint arXiv:1611.05053, 2016.
[35] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and
Computer-Assisted Intervention, pages 234–241. Springer,
2015.
[36] S. Saito, L. Wei, L. Hu, K. Nagano, and H. Li. Photorealistic
facial texture inference using deep neural networks. arXiv
preprint arXiv:1612.00523, 2016.
[37] M. Sela, N. Toledo, Y. Honen, and R. Kimmel. Customized
facial constant positive air pressure (cpap) masks. arXiv
preprint arXiv:1609.07049, 2016.
[38] A. Shrivastava, T. Pfister, O. Tuzel, J. Susskind, W. Wang,
and R. Webb. Learning from simulated and unsupervised images through adversarial training. arXiv preprint
arXiv:1612.07828, 2016.
[39] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556, 2014.
[40] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 1–9, 2015.
[41] Y. Taigman, A. Polyak, and L. Wolf. Unsupervised crossdomain image generation. arXiv preprint arXiv:1611.02200,
2016.
[42] J. Thies, M. Zollhofer, M. Stamminger, C. Theobalt, and
M. Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pages 2387–
2395, 2016.
[43] A. T. Tran, T. Hassner, I. Masi, and G. Medioni. Regressing robust and discriminative 3d morphable models with a
very deep neural network. arXiv preprint arXiv:1612.04904,
2016.
[44] T. Weise, S. Bouaziz, H. Li, and M. Pauly. Realtime
performance-based facial animation. In ACM Transactions
on Graphics (TOG), volume 30, page 77. ACM, 2011.
[45] L. Yin, X. Wei, Y. Sun, J. Wang, and M. J. Rosato. A 3d facial expression database for facial behavior research. In Automatic face and gesture recognition, 2006. FGR 2006. 7th
international conference on, pages 211–216. IEEE, 2006.
[46] Z. Zhang, P. Luo, C. C. Loy, and X. Tang. Facial landmark
detection by deep multi-task learning. In European Conference on Computer Vision, pages 94–108. Springer, 2014.
[47] E. Zhou, H. Fan, Z. Cao, Y. Jiang, and Q. Yin. Extensive facial landmark localization with coarse-to-fine convolutional
network cascade. In Proceedings of the IEEE International
Conference on Computer Vision Workshops, pages 386–391,
2013.
[48] X. Zhu, Z. Lei, X. Liu, H. Shi, and S. Z. Li. Face alignment
across large poses: A 3d solution. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 146–155, 2016.
[49] X. Zhu, Z. Lei, J. Yan, D. Yi, and S. Z. Li. High-fidelity
pose and expression normalization for face recognition in the
wild. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pages 787–796, 2015.
[50] G. Zigelman, R. Kimmel, and N. Kiryati. Texture mapping using surface flattening via multidimensional scaling.
IEEE Transactions on Visualization and Computer Graphics, 8(2):198–207, 2002.
Supplementary Material
A. Additional Network Details
Here, we summarize additional considerations concerning the network and its training procedure.
• The proposed architecture is based on the one introduced in [22]. For allowing further refinement of the results, three
additional convolution layers with a kernel of size 1 × 1 were concatenated at the end. Following the notations of [22],
the encoder architecture is given as
C64 − C128 − C256 − C512 − C512 − C512 − C512 − C512,
while the decoder is given by
CD512 − CD512 − CD512 − C512 − C512 − C256 − C128 − C64 − C ∗ 64 − C ∗ 32 − C ∗ 4,
where C ∗ represents a 1 × 1 convolution with stride 1.
• The resolution of the input and output training images was 512 × 512 pixels. While this is a relatively large input size for
training, the Image-to-Image architecture was able to process it successfully, and provided accurate results. Although,
one could train a network on smaller resolutions and then evaluate it on larger images, as shown in [22], we found that
our network did not successfully scale up for unseen resolutions.
• While a single network was successfully trained to retrieve both depth and correspondence representations, our experiments show that training separated networks to recover the representations is preferable. Note that the architectures
of both networks were identical. This can be justified by the observation that during training, a network allocates its
resources for a specific translation task and the representation maps we used have different characteristics.
• A necessary parameter for the registration step is the scale of the face with respect to the image dimensions. While
this can be estimated based on global features, such as the distance between the eyes, we opted to retrieve it directly by
training the network to predict the x and y coordinates of each pixel in the image alongside the z coordinate.
B. Additional Registration and Refinement Details
Next, we provide a detailed version of the iterative deformation-based registration phase, including implementation details
of the fine detail reconstruction.
B.1. Non-Rigid Registration
First, we turn the x,y and z maps from the network into a mesh, by connecting four neighboring pixels, for which the
coordinates are known, with a couple of triangles. This step yields a target mesh that might have holes but has dense map
to our template model. Based on the correspondence given by the network, we compute the affine transformation from a
template face to the mesh. This operation is done by minimizing the squared Euclidean distances between corresponding
vertex pairs. To handle outliers, a RANSAC approach is used [13] with 1, 000 iterations and a threshold of 3 millimeters for
detecting inliers. Next, similar to [28], an iterative non-rigid registration process deforms the transformed template, aligning
it with the mesh. Note, that throughout the registration, only the template is warped, while the target mesh remains fixed.
Each iteration involves the following four steps.
11
1. Each vertex in the template mesh, vi ∈ V, is associated with a vertex, ci , on the target mesh, by evaluating the nearest
neighbor in the embedding space. This step is different from the method described in [28], which computes the nearest
neighbor in the Euclidean space. As a result, the proposed step allows registering a single template face to different
facial identities with arbitrary expressions.
2. Pairs, (vi , ci ), which are physically distant by more than 1 millimeter and those with normal direction disagreement of
more than 5 degrees are detected and ignored in the next step.
3. The template mesh is deformed by minimizing the following energy
X
E(V, C) = αp2point
kvi − ci k22
(vi ,ci )∈J
+αp2plane
+αmemb
X
2
|~n(ci )(vi − ci )|
(vi ,ci )∈J
X
X
wi,j kvi − vj k22 ,
i∈V vj ∈N (vi )
(8)
where, wi,j is the weight corresponding to the biharmonic Laplacian operator (see [21, 5]), ~n(ci ) is the normal of the
corresponding vertex at the target mesh ci , J is the set of the remaining associated vertex pairs (vi , ci ), and N (vi ) is
the set 1-ring neighboring vertices about the vertex vi . Notice that the first term above is the sum of squared Euclidean
distances between matches and its weight αp2point is set to 0.1. The second term is the distance from the point vi to the
tangent plane at the corresponding point on the target mesh, and its weight αp2plane is set to 1. The third term quantifies
the stiffness of the mesh and its weight αmemb is initialized to 108 . In practice, the energy term given in Equation 8 is
minimized iteratively by an inner loop which contains a linear system of equations. We run this loop until the norm of
the difference between the vertex positions of the current iteration and the previous one is below 0.01.
4. If the motion of the template mesh between the current outer iteration and the previous one is below 0.1, we divide the
weight αmemb by two. This relaxes the stiffness term and allows a greater deformation in the next outer iteration. In
addition, we evaluate the difference between the number of remaining pairwise matches in the current iteration versus
the previous one. If the difference is below 500, we modify the vertex association step to estimate the physical nearest
neighbor vertex, instead of the the nearest neighbor in the space of the embedding given by the network.
This iterative process terminates when the stiffness weight αmemb is below 106 . The resulting output of this phase is a
deformed template with fixed triangulation, which contains the overall facial structure recovered by the network, yet, is
smoother and complete.
B.2. Fine Detail Reconstruction
Although the network already recovers fine geometric details, such as wrinkles and moles, across parts of the face, a geometric approach can reconstruct details at a finer level, on the entire face, independently of the resolution. Here, we propose
an approach motivated by the passive-stereo facial reconstruction method suggested in [3]. The underlying assumption here
is that subtle geometric structures can be explained by local variations in the image domain. For some skin tissues, such as
nevi, this assumption is inaccurate as the intensity variation results from the albedo. In such cases, the geometric structure
would be wrongly modified. Still, for most parts of the face, the reconstructed details are consistent with the actual variations
in depth.
The method begins from an interpolated version of the deformed template, provided by a surface subdivision technique.
Each vertex v ∈ VD is painted with the intensity value of the nearest pixel in the image plane. Since we are interested in
recovering small details, only the high spatial frequencies, µ(v), of the texture, τ (v), are taken into consideration in this
phase. For computing this frequency band, we subtract the synthesized low frequencies from the original intensity values.
This low-pass filtered part can be computed by convolving the texture with a spatially varying Gaussian kernel in the image
domain, as originally proposed. In contrast, since this convolution is equivalent to computing the heat distribution upon the
shape after time dt, where the initial heat profile is the original texture, we propose to compute µ(v) as
µ(v) = τ (v) − (I − dt · ∆g )−1 τ (v),
(9)
where I is the identity matrix, ∆g is the cotangent weight discrete Laplacian operator for triangulated meshes [31], and
dt = 0.2 is a scalar proportional to the cut-off frequency of the filter.
Next, we displace each vertex along its normal direction such that v 0 = v + δ(v)~n(v). The step size of the displacement,
δ(v), is a combination of a data-driven term, δµ (v), and a regularization one, δs (v). The data-driven term is guided by the
high-pass filtered part of the texture, µ(v). In practice, we require the local differences in the geometry to be proportional to
the local variation in the high frequency band of the texture. That is for each vertex v, with a normal ~n(v), and a neighboring
vertex vi , the data-driven term is given by
(µ(v) − µ(vi )) = hv + δµ (v)~n(v) − vi , ~n(v)i.
(10)
Thus, the step size assuming a single neighboring vertex can be calculated by
δµ (v) = γ(µ(v) − µ(vi )) − hv − vi , ~n(v)i.
(11)
In the presence of any number of neighboring vertices of v, we compute the weighted average of its 1-ring neighborhood
P
n(v)i]
vi ∈N (v) α(v, vi )γ [(µ(v) − µ(vi )) − hv − vi , ~
P
,
(12)
δµ (v) =
vi ∈N (v) α(v, vi )
An alternative term can spatially attenuate the contribution of the data-driven term in curved regions for regularizing the
reconstruction by
P
n(v)i|
i ,~
α(v,vi ) (µ(v) − µ(vi )) 1 − |hv−v
kv−vi k
δµ (v) =
vi ∈N (v)
P
α(v,vi )
,
(13)
vi ∈N (v)
where α(v,vi ) = exp (−kv − vi k). where N (v) is the set 1-ring neighboring vertices about the vertex v, and ~n(v) is the unit
normal at the vertex v.
Since we move each vertex along the normal direction, triangles could intersect each other, particularly in regions with
high curvature. To reduce the probability of such collisions, a regularizing displacement field, δs (v), is added. This term is
proportional to the mean curvature of the original surface, and is equivalent to a single explicit mesh fairing step [11]. The
final surface modification is given by
v 0 = v + (ηδµ (v) + (1 − η)δs (v)) · ~n(v),
for a constant η = 0.2.
(14)
C. Additional Experimental Results
We present additional qualitative results of our method. Figure 14 shows the output representations of the proposed
network for a variety of different faces, notice the failure cases presented in the last two rows. One can see that the network
generalizes well, but is still limited by the synthetic data. Specifically, the network might fail in presence of occlusions, facial
hair or extreme poses. This is also visualized in Figure 15 where the correspondence error is visualized using the texture
mapping. Additional reconstruction results of our method are presented in Figure 16. For analyzing the distribution of the
error along the face, we present an additional comparison in Figure 17, where the absolute error, given in percents of the
ground truth depth, is shown for several facial images.
Figure 14: Network Output.
Figure 15: Results under occlusions and rotations. Input images are shown next to the matching correspondence result,
visualized using the texture mapping to better show the errors.
Figure 16: Additional reconstruction results.
Figure 16: Additional reconstruction results.
Figure 16: Additional reconstruction results.
Input
Proposed
[34]
[26]
Figure 17: Error heat maps in percentile of ground truth depth.
[49]
Err. % Scale
| 1cs.CV
|
HILBERT SERIES OF SYMMETRIC IDEALS
IN INFINITE POLYNOMIAL RINGS
VIA FORMAL LANGUAGES
arXiv:1606.07956v1 [math.AC] 25 Jun 2016
ROBERT KRONE, ANTON LEYKIN, AND ANDREW SNOWDEN
Abstract. Let R be the polynomial ring K[xi,j ] where 1 ≤ i ≤ r and j ∈ N, and let I be
an ideal of R stable under the natural action of the infinite symmetric group S∞ . Nagel–
Römer recently defined a Hilbert series HI (s, t) of I and proved that it is rational. We give
a much shorter proof of this theorem using tools from the theory of formal languages and a
simple algorithm that computes the series.
Contents
1. Introduction
2. Background on regular languages
3. Monomial ideals
4. General ideals
5. An algorithm for Hilbert series
6. Hilbert series of modules
References
1
2
3
5
6
7
8
1. Introduction
1.1. Statement of results. Let R be the polynomial ring over the field K in variables xi,j ,
where i ∈ {1, . . . , r} and j ∈ N. The infinite symmetric group S∞ acts on R (by fixing the
first index and moving the second), and a fundamental result, proved originally by Cohen
[Co] but subsequently rediscovered [AH, HS], is that R is S∞ -noetherian: that is, any S∞ ideal in R is generated by the S∞ -oribts of finitely many elements. Given this, one can begin
to study finer properties of ideals. In this paper, we investigate their Hilbert series.
Let I ⊂ R be a homogeneous S∞ -ideal. For n ≥ 1, let Rn ⊂ R be the subalgebra
generated by the variables xi,j with 1 ≤ i ≤ r and j ≤ n, and put In = I ∩ Rn . Then In
is a finitely generated graded Rn -module, and so its Hilbert series HIn (t) is a well-defined
rational function. We define the Hilbert series of I by
X
HI (s, t) =
HIn (t)sn .
n≥0
This series was introduced by Nagel–Römer [NR], who proved the following theorem:
Theorem 1.1. The series HI (s, t) is a rational function of s and t.
Date: June 28, 2016.
AL was supported by NSF grant DMS-1151297.
AS was supported by NSF grants DMS-1303082 and DMS-1453893.
1
2
ROBERT KRONE, ANTON LEYKIN, AND ANDREW SNOWDEN
The purpose of this paper is to give a new proof of this theorem. Our proof is shorter and
(in our opinion) conceptually clearer than the one given in [NR].
Remark 1.2. In fact, [NR] work with what we would call HR/I (s, t), but it is a trivial matter
to pass between this and our HI (s, t).
Remark 1.3. The result of [NR] gives information about the denominator of HI (s, t). Our
method gives some information as well, though we have not carefully traced through everything to see exactly what it yields. In particular, we do not know which method will
ultimately say more about the denominator.
1.2. Overview of proof. We now describe the idea of our proof. First, passing to the initial
ideal one can reduce to the case where I is a monomial ideal. One then has what is essentially
a complicated bookkeeping problem: one must understand which of the monomials in the
infinitely many variables xi,j appear in I. Our main idea is to use a sort of encoding scheme to
make the problem more finite: more precisely, we establish a bijection between the monomials
in R and a certain set of words in a finite alphabet. Thus, in a sense, we trade the infinitely
many commuting variables of R for finitely many non-commuting variables. We show that,
under this encoding scheme, I (or rather, the set of monomials it contains) corresponds to
a regular language. The theorem then follows from standard results on generating functions
of regular languages.
The idea of using formal languages was motivated by the approach to Hilbert series in
[SS2]. However, the result and methods of this paper do not appear to fit into the general
setup of [SS2].
1.3. Outline. In §2 we review background material on regular languages. In §3 we prove
the main theorem in the case of monomial ideals; this is really the bulk of the work. In §4
we complete the proof of the theorem by reducing to the monomial case. In §5 we explicitly
describe an algorithm for computing HI (s, t), given a set of generators for I. Finally, in §6
we discuss the possibility of treating Hilbert series of R-modules.
1.4. Notation. We write N for the set of non-negative integers. We let Inc(N) be the socalled increasing monoid: this is the set of functions f : N → N satisfying f (n) < f (m) for
n < m, using composition as the monoidal operation. Throughout, K denotes an arbitrary
field.
2. Background on regular languages
In this section we review some well-known material on formal languages, especially regular
languages. We refer the reader to the text [HU] for more details.
Let Σ be a finite set and let Σ⋆ be the set of words in the alphabet Σ; alternatively, Σ⋆
is the free monoid on Σ. A formal language on Σ is simply a subset of Σ⋆ . Given a
formal language L on Σ⋆ , we define the Kleene star L⋆ of L to be the language consisting
of all words of the form w1 · · · wn where wi ∈ L; alternatively, L⋆ is the submonoid of Σ⋆
generated by L. Given two formal languages L1 and L2 , we define their concatenation
L1 L2 to be the formal language consisting of all words of the form w1 w2 with w1 ∈ L1 and
w2 ∈ L2 . We also make use of the standard set-theoretic operations of union, intersection,
and complement on formal languages.
The class of regular languages on Σ is the smallest class of languages containing the
singleton languages {σ} for each σ ∈ Σ, and closed under union, concatenation, and Kleene
HILBERT SERIES VIA FORMAL LANGUAGES
3
star. (Actually, the empty language and the language consisting only of the empty word are
also counted as regular languages, but do not fit the previous definition.) It turns out that
the class of regular languages is also closed under intersection and complement.
Let t1 , . . . , tk be a set of formal variables, let M be the set of monomials in these variables,
and let ρ : Σ⋆ → M be a monoid homomorphism, which we refer to as the weight function.
We note that ρ is determined by its restriction to Σ. Given a language L on Σ, we define its
generating function with respect to ρ by
X
ρ(w),
HL,ρ(t1 , . . . , tk ) =
w∈L
assuming this sum makes sense (i.e., there are only finitely many w ∈ L for which ρ(w) is a
given monomial). We consider this as a formal power series in the variables t1 , . . . , tk . For
example, suppose k = 1 and ρ is defined by ρ(σ) = t for all σ ∈ Σ. Then for a word w we
have ρ(w) = tlen(w) , and so the coefficient of tn in HL,ρ(t) is the number of words in L of
length n. We require the following standard result (see, e.g., [St, Theorem 4.7.2], though the
terminology there is somewhat different):
Proposition 2.1. If L is a regular language then HL,ρ (t1 , . . . , tk ) is a rational function of
the ti ’s, for any weight function ρ (for which the series makes sense).
3. Monomial ideals
Let R = K[xi,j ] where 1 ≤ i ≤ r and j ∈ N, and let M be the set of monomials in R.
Let Σ be the alphabet {τ, ξ1, . . . , ξr }. Let T : M → M be the shift operator, defined by
T (xi,j ) = xi,j+1 and extended multiplicatively. We define a function m : Σ⋆ → M inductively
using the following three rules: (a) m(∅) = 1; (b) m(ξi w) = xi,0 · m(w); and (c) m(τ w) =
T (m(w)). Thus, concretely, to compute m(w) simply change each ξi in w to xi,0 and each τ
to T applied to the string following it.
Example 3.1. We have m(τ ξ1 τ ξ2 τ ) = T (x1,0 T (x2,0 T (1))) = T (x1,0 x2,1 ) = x1,1 x2,2 .
It is clear that the map m : Σ⋆ → M is surjective, though it is not injective since the
variables xi,j commute, e.g., m(ξ1 ξ2 ) = m(ξ2 ξ1 ). We therefore introduce a subset of Σ⋆ to
obtain a bijection. We say that a word w in Σ⋆ is standard if it satisfies the condition that
every substring ξi ξj of w has i ≤ j. Let Σ⋆std be the set of standard words, and let Σ⋆std,n be
the set of standard words in which τ occurs exactly n times. Let Mn be the set of monomials
in the variables xi,j with 1 ≤ i ≤ r and 0 ≤ j ≤ n.
Proposition 3.2. For each n the map m : Σ⋆std,n → Mn is a bijection.
Proof. Let u and w be words in Σ⋆std,n such that m(u) = m(w), and let us prove u = w.
Let u′ be the segment of u appearing before the first τ in u, and write u = u′u′′ ; similarly
decompose w = w ′ w ′′. Note that u′ , u′′ , w ′ , and w ′′ are all standard. Every variable in m(u′ )
has second index equal to 0, while every variable in m(u′′ ) has second index greater than 0,
and similarly for m(w ′ ) and m(w ′′ ). We have
m(u′ )m(u′′ ) = m(u) = m(w) = m(w ′ )m(w ′′ )
and so m(u′ ) = m(w ′ ) and m(u′′ ) = m(w ′′ ). Since u′ and w ′ are standard, it is clear that
u′ = w ′. If n = 0 then u′′ and w ′′ are empty and thus equal. If n > 0 then u′′ = τ u′′′ and
w ′′ = τ w ′′′ and u′′′ , w ′′′ ∈ Σ⋆std,n−1 . Since T is injective on M, we have m(u′′′ ) = m(w ′′′ ). By
4
ROBERT KRONE, ANTON LEYKIN, AND ANDREW SNOWDEN
induction on n, u′′′ = w ′′′ , thus u = w. We have thus shown that m : Σ⋆std → M is injective;
it is clearly surjective.
We let w : M → Σ⋆std be the right-inverse to the map m which sends monomial m to the
minimal length word w such that m(w) = m. The image of w is the set of words in Σ⋆std
that do not end in τ . On the other hand m−1 (m) = wτ ∗ , the set of words consisting of w
followed by any number of trailing τ s.
Given a monomial m ∈ M, let hmi be the set of monomials m′ ∈ M such that σ(m) | m′
for some σ ∈ Inc(N). Given monomials m1 , . . . , mn , let hm1 , . . . , mn i be the union of the
hmi i’s.
Proposition 3.3. Let m1 , . . . , mn be monomials in R and let I the monomial ideal generated
by the Inc(N)-orbits of m1 , . . . , mn . Let ρ be the weight function defined by ρ(τ ) = s and
ρ(ξi ) = t for all i = 1, . . . , r. Then
HI (s, t) = Hm−1 (hm1 ,...,mn i),ρ (s, t).
Proof. Let I ⊂ M be the set of monomials in I. Then I = hm1 , . . . , mn i. The coefficient of
sn tm in HI (s, t) is the number of monomials in I ∩ Mn of degree m. This equals the number
of words in m−1 (I) in which τ appears exactly n times and which contain exactly m non-τ
letters. But this is just the coefficient of sn tm in Hm−1 (I),ρ (s, t) as defined at the end of §2.
Thus HI (s, t) = Hm−1 (I),ρ (s, t), and so the result follows from Proposition 2.1.
We say that a word in Σ⋆ is simple if it contains no τ .
Proposition 3.4. The set Σ⋆std is a regular language on Σ.
Proof. Let L be the language of simple standard words. The identity
L = {ξ1 }⋆ · {ξ2 }⋆ · · · {ξn }⋆
shows that L is regular. The identity Σ⋆std = L · (τ L)⋆ now shows that Σ⋆std is regular.
Proposition 3.5. Let m ∈ M. Then m−1 (hmi) is a regular language on Σ.
Proof. Write w(m) = w0 τ w1 τ · · · τ wn , where each wi is simple. Let Li be the language
consisting of simple standard words w ′ such that m(wi ) | m(w ′ ). One easily sees that Li is
a regular language. Let L be the regular language on Σ defined by
(3.6)
Σ⋆ L0 Σ⋆ τ L1 Σ⋆ τ L2 · · · Σ⋆ τ Ln Σ⋆ .
We claim that a monomial m′ belongs to hmi if and only if w(m′ ) ∈ L. This will prove the
proposition, as then m−1 (hmi) will coincide with L ∩ Σ⋆std , and Σ⋆std is also regular.
First suppose m′ ∈ hmi, so that σ(m) | m′ for some σ ∈ Inc(N). Write m = m0 · · · mn
where mj uses only the variables xi,j , and similarly write m′ = m′0 · · · m′t . Then σ(mj ) | m′σ(j)
for 0 ≤ j ≤ n. We have w(m′ ) = w0′ τ w1′ τ · · · τ wt′ where τ j wj′ = w(m′j ). We can regroup this
expression as
′
′
′
(· · · ) · · · (· · · )τ wσ(n)
(· · · )
(· · · )τ wσ(1)
w(m′ ) = (· · · )wσ(0)
′
′
) and so wσ(j)
∈ Lj .
Since τ σ(j) wj = w(σ(mj )) and σ(mj ) | m′σ(j) , we see that m(wj ) | m(wσ(j)
′
′
Thus the above expression shows that w(m ) ∈ L. Finally, if w(m ) is in L then so is the
set w(m′ )τ ⋆ = m−1 (m′ ).
Now suppose w ′ ∈ L. Write m(w ′ ) = m′0 · · · m′t and w ′ = w0′ τ · · · τ wt′ τ k as above. Since
′
∈ Lj for 0 ≤ j ≤ n.
w(m′ ) ∈ L, we can find σ(0) < σ(1) < · · · < σ(n) such that wσ(j)
HILBERT SERIES VIA FORMAL LANGUAGES
5
Extend σ arbitrarily to an element of Inc(N). Then it is clear that σ(m) | m(w ′ ), and so
m(w ′ ) ∈ hmi.
Corollary 3.7. Let m1 , . . . , mn ∈ M. Then m−1 (hm1 , . . . , mn i) is a regular language on Σ.
Theorem 3.8. Let I ⊂ R be an Inc(N)-stable monomial ideal. Then HI (s, t) is a rational
function.
Proof. It is known (see [Co] or [HS]) that I is finitely generated up to the action of Inc(N):
that is, there exist m1 , . . . , mn ∈ I, which can be taken to be monomials, such that I is the
ideal generated by the Inc(N)-orbits of m1 , . . . , mn . By Propositions 3.3 and 2.1, HI (s, t) is
rational if m−1 (hm1 , . . . , mn i) is a regular language, which is the result of Corollary 3.7.
Remark 3.9. The above construction can be generalized from the total degree grading to
arbitrary Inc(N)-stable (multi-) grading. An Inc(N)-stable multi-grading, deg : M → Zk , is
determined by the values of deg(xi,0 ) for i = 1, . . . , r. The series HI is then given by
HI (s, t1 , . . . , tk ) = HΣ⋆std ,ρ (s, t1 , . . . , tk )
for weight function ρ with ρ(τ ) = s and ρ(ξi ) = deg(xi,0 ) for i = 1, . . . , r.
4. General ideals
Let R be as in the previous section. We define an order ≤ on the monomials in R as follows.
First, we order the variables xi,j lexicographically by comparing the second index first: that
is, xi,j < xk,ℓ if j < ℓ or j = ℓ and i < k. We then order monomials by lexicographically
comparing their exponents. This is a well-ordering of the monomials and compatible with
multiplication. We write in(f ) for the initial term of a non-zero element f ∈ R and in(I) for
the initial ideal associated to an ideal I ⊂ R.
Lemma 4.1. We have in(I) ∩ Rn = in(I ∩ Rn ).
Proof. It is clear that in(I ∩ Rn ) ⊂ in(I) ∩ Rn , so let us prove the reverse containment. The
ideal in(I) ∩ Rn is monomial, so it suffices to show that if f ∈ I and in(f ) ∈ Rn then f ∈ Rn .
But this is clear from how we ordered the variables: indeed, if in(f ) = m ∈ Rn then no
monomial appearing in f can contain a variable of the form xi,j with j > n, for then that
monomial would exceed m in our ordering and contradict m being the initial term, and so
it follows that f ∈ Rn .
Lemma 4.2. We have HI (s, t) = Hin(I) (s, t).
Proof. The coefficient of sn in HI (s, t) is equal to HI∩Rn (t). It is a standard fact that passing
to the initial ideal does not affect Hilbert series, and so this is equal to Hin(I∩Rn ) (t). By the
lemma, this is equal to Hin(I)∩Rn (t), which is the coefficient of sn in Hin(I) (s, t).
Theorem 4.3. Let I be an Inc(N)-stable ideal in R. Then HI (s, t) is a rational function.
Proof. This follows from the previous lemma and Theorem 3.8. (Note that our monomial
ordering is compatible with the action of Inc(N), and so in(I) is still Inc(N)-stable.)
6
ROBERT KRONE, ANTON LEYKIN, AND ANDREW SNOWDEN
5. An algorithm for Hilbert series
We now describe an algorithm for computing HI (s, t) for an Inc(N)-stable ideal I as above.
We first recall some additional background material. Suppose that L is a regular language.
Then there is a finite-state automaton A that accepts precisely the words in L, see [HU,
Ch. 2]. Fix such an A, and suppose that it has N states. For ℓ ∈ Σ let MA,ℓ be the associated
transition matrix for A. This is the 0-1, left-stochastic N × N matrix with 1 in entry (i, j)
if there is edge labeled byPℓ from state j to state i. Let e1 ∈ K n be the basis vector for the
initial state, and let u = i∈F ei ∈ K n be the sum of the basis vectors corresponding to the
accept states F. Then for a word w = w1 · · · wn , we have
(
1 if A accepts w
t
uMA,wn · · · MA,w1 e1 =
0 if A rejects w.
Let ρ : Σ⋆ → M be a weight function, where M is the set of monomials in t1 , . . . , tk . Summing
the above expression over all words, we find
HL,ρ(t1 , . . . , tk ) =
(5.1)
X
w∈L
ρ(w) =
X
n≥0
t
X
n
u
ρ(ℓ)MA,ℓ e1
ℓ∈Σ
−1
X
t
ρ(ℓ)MA,ℓ
e1 .
= u Id −
ℓ∈Σ
Thus the generating function for L can be computed directly from the automaton A.
The following is our algorithm for computing HI (s, t), given as input a set of elements
f1 , . . . , fr of I whose Inc(N)-orbits generate I:
(1) First compute the initial ideal of I. This can be done using standard equivariant
Gröbner basis techniques. We suppose that m1 , . . . , ms are monomials whose Inc(N)
orbits generate the initial ideal.
(2) Next construct a regular expression for the language L = m−1 (hm1 , . . . , ms i). We
note that (3.6) is essentially a regular expression for m−1 (hmi) (and is obviously
constructed algorithmically from m), and a regular expression for L can be obtained
by “or-ing” the regular expressions for the various m−1 (hmi i).
(3) From the regular expression for L, construct an automaton A that accepts L. It is
well-understood how to algorithmically pass from a regular expression to an automaton, see [HU, Ch. 2].
(4) Finally, compute the Hilbert series from the automaton via (5.1), using the weight
function from Proposition 3.3. This really computes the Hilbert series of the initial
ideal, but this coincides with the Hilbert series of the original ideal I by Lemma 4.2.
Example 5.2. Let r = 1 and I = hx21,0 i. The language m−1 (I) is detected by the regular
expression
(ξ1 |τ )⋆ ξ1 ξ1 (ξ1 |τ )⋆
and by the automaton
HILBERT SERIES VIA FORMAL LANGUAGES
ξ1 , τ
τ
start
1
7
ξ1
2
ξ1
3
τ
where the first two states are rejecting and the last accepting. The automaton has transition
matrices
1 1 0
0 0 0
MA,τ = 0 0 0 ,
MA,ξ1 = 1 0 0 .
0 0 1
0 1 1
We have e1 = (1, 0, 0) and u = (0, 0, 1), and so
HI (s, t) = t u(Id −sMA,τ − tMA,ξ1 )−1 e1 =
t2
.
(1 − s − t)(1 − s − st)
We implemented functions constructing automata corresponding to monomial ideals in R
and computing their Hilbert series in Macaulay2 [M2]. These along with some examples are
posted at http://rckr.one/eHilbert.html.
6. Hilbert series of modules
Let M be a graded R-module equipped with a compatible action of S∞ that is generated
by the S∞ orbits of finitely many elements.1 A natural problem is to define a notion of
Hilbert series for M and extend Theorem 1.1 to this setting.
One can generalize the definition of HI as follows. Let G(n) ⊂ S∞ be the subgroup
consisting of permutations that fix each of the elements 0, . . . , n. Then I ∩ Rn is identified
with the invariants I G(n) . Thus in the definition of HI we can simply replace I ∩ Rn with
M G(n) to obtain a definition for HM , i.e.:
X
HM (s, t) =
HM G(n) (t)sn .
n≥0
This is a perfectly well-defined series, and so one can certainly study it and investigate
its rationality properties. However, as a definition of Hilbert series it is fatally flawed:
formation of G(n) invariants is not exact, and so the above quantity is not additive in short
exact sequences of R-modules. (For example, if I = hx1,1 − x1,0 i ⊂ R then RG(−1) is the
set of constants where G(−1) = S∞ . Meanwhile (R/I)G(−1) = R/I ∼
= K[y], and therefore
(R/I)G(−1) 6= RG(−1) /I G(−1) .)
There are various ways one could try to fix this problem: one could substitute invariants
with derived invariants, which is known to be well-behaved by [SS, §6.4.4], or with coinvariants, which is known to be exact by [SS, §6.2.11]. However, the best series to study is
probably
X
HM =
[Mn ]tn ,
n≥0
where [Mn ] is the class of the S∞ -representation Mn in the Grothendieck group of finitely generated algebraic representations (in the sense of [SS, §6]). Any reasonable notion of Hilbert
1For
technical reasons related to our uses of [SS] below, we assume that every element of M is stabilized
by a subgroup of S∞ of the form Aut({n, n + 1, . . .}). This is automatic if M is an ideal in R.
8
ROBERT KRONE, ANTON LEYKIN, AND ANDREW SNOWDEN
series for M should factor through the above definition. We note that the Grothendieck
group in question is identified with the ring of symmetric functions Λ, so that above series
can be considered as a power series in t with coefficients in Λ. We believe there should be
some sort of rationality theorem for HM , but leave this as an open problem.
References
[AH]
Matthias Aschenbrenner, Christopher J. Hillar, Finite generation of symmetric ideals, Trans. Amer.
Math. Soc., 359 (2007), 5171–5192; erratum, ibid. 361 (2009), 5627–5627. arXiv:math/0411514
[Co] D. E. Cohen, On the laws of a metabelian variety, J. Algebra 5 (1967), 267–273.
[HS] Christopher J. Hillar, Seth Sullivant, Finite Gröbner bases in infinite dimensional polynomial rings
and applications, Advances in Mathematics, 229 (2012), no. 1, 1–25. arXiv:0908.1777
[HU] John E. Hopcroft, Jeffrey D. Ullman, Introduction to Automata Theory, Languages, and Computation,
Addison–Wesley Series in Computer Science, Addison-Wesley Publishing Co., 1979.
[M2] D.R. Grayson and M.E. Stillman, Macaulay2, a software system for research in algebraic geometry,
Available at http://www.math.uiuc.edu/Macaulay2/
[NR] Uwe Nagel, Tim Roemer, Equivariant Hilbert series in non-noetherian polynomial rings, preprint,
2015. arXiv:1510.02757
[SS] Steven V Sam, Andrew Snowden, Stability patterns in representation theory, Forum. Math. Sigma 3
(2015), e11, 108 pp. arXiv:1302.5859v2
[SS2] Steven V Sam, Andrew Snowden, Gröbner methods for representations of combinatorial categories,
J. Amer. Math. Soc., to appear. arXiv:1409.1670v3
[St]
Richard P. Stanley, Enumerative Combinatorics. Volume 1, second edition, Cambridge Studies in
Advanced Mathematics 49, Cambridge University Press, Cambridge, 2012.
Department of Mathematics and Statistics, Queen’s University, Kingston, ON
E-mail address: [email protected]
URL: http://rckr.one/
School of Mathematics, Georgia Institute of Technology, Atlanta, GA
E-mail address: [email protected]
URL: http://people.math.gatech.edu/~aleykin3/
Department of Mathematics, University of Michigan, Ann Arbor, MI
E-mail address: [email protected]
URL: http://www-personal.umich.edu/~asnowden/
| 0math.AC
|
arXiv:1609.08900v2 [math.GR] 12 May 2017
GRADIENTS OF SEQUENCES OF SUBGROUPS IN A
DIRECT PRODUCT
NIKOLAY NIKOLOV, ZVI SHEMTOV, AND MARK SHUSTERMAN
Abstract. For a sequence {Un }∞
n=1 of finite index subgroups of a direct
product G ··= A × B of finitely generated groups, we show that
min{|X| : hXi = Un }
=0
lim
n→∞
[G : Un ]
once [A : A ∩ Un ], [B : B ∩ Un ] → ∞ as n → ∞. Our proof relies on the
classification of finite simple groups. For A, B that are finitely presented
we show that
log |Torsion(Unab )|
lim
= 0.
n→∞
[G : Un ]
1. Introduction
With motivation coming from the theory of 3-manifolds, Lackenby defined
in [31] the rank gradient of a sequence {Un }∞
n=1 of finite index subgroups of
a finitely generated group G to be the following combinatorial invariant
(1.1)
inf
n
d(Un ) − 1
[G : Un ]
where d(K) is the least cardinality of a generating set for the group K. As
can be seen from [2, 3, 4, 5, 19, 20, 23, 30, 34, 40, 41, 48, 50, 51, 52] the
rank gradient has been extensively studied, calculated in various cases, and
related to the notion of cost from ergodic theory. Furthermore, the rank
gradient controls other interesting combinatorial invariants such as the pgradient, and the rate of growth of the first Betti number in finite index
subgroups, that were studied, for instance, in [1, 3, 7, 13, 15, 18, 27, 32, 33,
35, 36, 37, 41, 46]. It is one of the main goals of asymptotic and measured
group theory to calculate these gradients, and if possible, to show that they
coincide for various sequences, and equal to the appropriate analytic and
ergodic invariants (such as the first ℓ2 -Betti number and the cost of G).
Here we accomplish this task for all the aforementioned gradients and all
sequences in case that G is a direct product.
Theorem 1.1. Let A, B be finitely generated groups, set G ··= A × B, and
let {Un }∞
n=1 be a sequence of finite index subgroups of G. Assume that
(1.2)
[A : A ∩ Un ], [B : B ∩ Un ] → ∞
1
2
NIKOLAY NIKOLOV, ZVI SHEMTOV, AND MARK SHUSTERMAN
as n → ∞. Then
(1.3)
lim
n→∞
d(Un ) − 1
= 0.
[G : Un ]
We should note that even the vanishing of the p-gradient (or the Betti
gradient) was not previously known for these sequences.
An open problem in ergodic theory analogous to Theorem 1.1 is to determine whether the direct product of every pair of finitely generated groups
has fixed price. Indeed, the special case of Theorem 1.1 considering normal
(or even Farber) chains (nested sequences), is at the focus of [47]. Another
special case that was already known is when A, B are torsion-free (and thus
G is right-angled, see [2, 26]) and the sequence is Farber.
Additional motivation for Theorem 1.1 comes directly from (arithmetic)
3-manifolds. As recent breakthroughs by Agol, Wise (and others) tell us that
many 3-manifolds virtually fiber over the circle, we know that their fundamental groups are semidirect products (up to a finite index). Calculating the
rank gradient (or the p-gradient) of semidirect products is both notoriously
difficult, and interesting from a topological and number-theoretic point of
view, so it makes sense to examine the special case of direct products first.
Our assumption (1.2) is necessary, since otherwise the gradients may become positive once this is the case for one of the factors A, B. Indeed the
sequences we consider are very general, and this allows us to obtain vanishing of gradients for Farber chains in groups (virtually) generated by a pair
of commuting infinite subgroups. These are the groups ‘presentable by a
product’ that were studied in [16, 17, 28, 29].
Having considered the Betti gradient (measuring the free part of abelianizations in sequences), it is tempting to study the torsion gradient
(1.4)
lim inf
n→∞
log |Torsion(Unab )|
[G : Un ]
for which we assume that G is finitely presented (otherwise, (1.4) may be
infinite). This gradient, having connections to number theory, geometry,
and topology, became an object of intensive study, as witnessed by [2, 6, 8,
9, 10, 11, 12, 14, 24, 35, 36, 39, 42, 43, 49, 53, 54].
Theorem 1.2. Let A, B be finitely presented groups, set G ··= A × B, and
let {Un }∞
n=1 be a sequence of finite index subgroups of G. Assume that
(1.5)
[A : A ∩ Un ], [B : B ∩ Un ] → ∞
as n → ∞. Then
(1.6)
log |Torsion(Unab )|
= 0.
[G : Un ]
Theorem 1.2 provides the first large family of groups for which the torsion
gradient vanishes in (almost) all sequences, including some that are not even
Farber.
GRADIENTS OF SEQUENCES OF SUBGROUPS IN A DIRECT PRODUCT
3
2. Preliminaries
Let G be a finitely generated group and let H ≤ G be a finite index
subgroup of G. We will be frequently using the bound
(2.1)
d(H) ≤ d(G)[G : H].
that follows from [45, Theorem 11.44].
Proposition 2.1. Let A, B be groups, set G ··= A × B and let K ≤ G be a
subgroup for which BK = G. Then A ∩ K ⊳ A.
Proof. Take a ∈ A, x ∈ A ∩ K and note that since A ⊆ BK, there exist
b ∈ B, k ∈ K such that a = bk. As A ⊳ G we see that kxk−1 ∈ A ∩ K, and
[A, B] = {1} implies that axa−1 = bkxk−1 b−1 = kxk−1 ∈ A ∩ K.
Definition 2.2. Let G be a group and let S, X ⊆ G be subsets. We denote
by hSiX the subgroup of G generated by the conjugates of S by elements of
X. For a normal subgroup N ⊳G we define dG (N ) to be the least cardinality
of a subset S ⊆ N for which hSiG = N .
Definition 2.3. Let K be a finite group, let d ∈ N, and let T = {t1 , . . . , td }
be a generating multiset of K. Take a free group F on X = {x1 , . . . , xd } and
let ϕ : F → K be the unique surjection with ϕ(xi ) = ti for 1 ≤ i ≤ d. We
set r(K, T ) ··= dF (Ker(ϕ)) and think of this quantity as the least number of
relations needed to present K using the generating multiset T .
Proposition 2.4. Let G be a group generated by a finite subset T ⊆ G,
let N ⊳ G be a normal subgroup of finite index in G, and define a multiset
T ⊆ G/N by T ··= {tN | t ∈ T }. Then dG (N ) ≤ r(G/N, T ).
Proof. Let F be the free group on T , and let ϕ : F → G/N be the unique
surjection for which ϕ(t) = tN for all t ∈ T . By 2.3 there exists a subset
S ⊆ Ker(ϕ) of cardinality r(G/N, T ) such that hSiF = Ker(ϕ). Define a
surjection ψ : F → G by ψ|T = IdT , and observe that
(2.2)
N = ψ(Ker(ϕ)) = ψ(hSiF ) = hψ(S)iG
2.2
so that dG (N ) ≤ |ψ(S)| ≤ |S| = r(G/N, T ) as required.
Let K be a finite group generated by a multiset T = {t1 , . . . , td }, let N be
a minimal normal subgroup of K, let S be a finite simple group onto which
N surjects, and define the multiset T ··= {t1 N, . . . , td N }. The following
bound comes from the argument appearing in the proof of Theorem 1 in
[38] and in the discussion following the proof of Theorem 2 therein.
(2.3)
r(K, T ) ≤ r(K/N, T ) + 6d log2 |N | + r(S, W )
where W is any pair of elements generating S. It follows from [21, Corollary
A’, Lemma 2.1, Theorem 4.34] and [38, Theorem 1] that
(2.4)
r(S, W ) ≤ 8|S|3/7 .
4
NIKOLAY NIKOLOV, ZVI SHEMTOV, AND MARK SHUSTERMAN
Corollary 2.5. Let K be a finite group generated by a multiset T . Then
(2.5)
r(K, T ) ≤ 128|T ||K|3/7 .
Proof. Let N be a minimal normal subgroup of K, and let S be a finite
simple factor of N . By induction, (2.3), and (2.4), we have
(2.6)
r(K, T ) ≤ r(K/N, T ) + 6|T | log2 |N | + 8|S|3/7
|K| 3/7
+ 6|T | log2 |N | + 8|N |3/7
≤ 128|T |
|N |
|K| 3/7
+ 24|T ||N |3/7 + 8|T ||N |3/7
≤ 128|T |
|N |
!
|K| 3/7
3
3/7
3/7
= 128|T |
+ |N | − |N |
|N |
4
3
≤ 128|T | |K|3/7 + 1 − |N |3/7
4
3
≤ 128|T | |K|3/7 + 1 − 23/7 ≤ 128|T ||K|3/7 .
4
Let E be a finite group with a presentation
(2.7)
1 −→ N −→ F −→ E −→ 1
where F is free. The Schur multiplier (see [44, Theorem 10.12]) of E is
(2.8)
M (E) = N ∩ [F, F ] /[F, N ].
This is a finite abelian group whose order admits the following bound.
Corollary 2.6. Let E be a finite group. Then
|M (E)| ≤ |E|log |E| .
(2.9)
Proof. Let S be the set of primes dividing |E|, and for each p ∈ S fix a
p-Sylow subgroup P of E. It follows from [25, Proposition 2.1.1] that
Y
(2.10)
|M (E)| =
|M (E) ⊗Z Zp |.
p∈S
By [25, Theorem 2.1.2], for every p ∈ S we have
(2.11)
|M (E) ⊗Z Zp | ≤ |M (P )|
and by [25, Corollary 3.1.5]
(2.12)
|M (P )| ≤ |P |log |P | .
GRADIENTS OF SEQUENCES OF SUBGROUPS IN A DIRECT PRODUCT
5
Combining everything, we get that
(2.13)
2.10 Y
2.11 Y
|M (E)| ≤
|M (E) ⊗Z Zp | ≤
|M (P )|
p∈S
p∈S
Y
2.12 Y
≤
|P |log |P | ≤
|P |log |E| = |E|log |E| .
p∈S
p∈S
In terms of our presentation for E we thus have
h
i h
i h
i
[F, F ] : [F, N ] ≤ [F, F ] : N ∩ [F, F ] · N ∩ [F, F ] : [F, N ]
(2.14)
≤ [F : N ] · |M (E)| ≤ |E|1+log |E| .
In a manner similar to Proposition 2.4 one can deduce the following.
Corollary 2.7. For a finite index normal subgroup A0 of a group A we have
h
i
(2.15)
[A, A] : [A, A0 ] ≤ [A : A0 ]1+log[A:A0] .
3. Upper bounds on the number of generators
We establish several bounds on the number of generators of a finite index
subgroup of a direct product, and conclude that the rank gradient vanishes.
Theorem 3.1. Let A, B be finitely generated groups, and let H be a finite
index subgroup of G ··= A × B. Then d(H) is bounded by:
(1) d(G)([G : AH] + [AH : H])
(2) d(G)([G : BH] + [BH : H])
(3) d(G)([G : AH] + 130[G : BH][G : H]3/7 ).
Proof. For (1) note that
(3.1)
2.1
d(H) ≤ d(H/H ∩ A) + d(H ∩ A) ≤ d(AH/A) + d(A)[A : H ∩ A]
2.1
≤ d(AH) + d(A)[AH : H] ≤ d(G)[G : AH] + d(G)[AH : H]
as required. To get (2) just replace A with B in (3.1).
For (3) let πA , πB be the projections from G onto A, B and observe
that H ≤ πA (H) × πB (H) is a subgroup that complements πB (H) (that
is, πB (H)H = πA (H) × πB (H)). By Proposition 2.1, πA (H) ∩ H ⊳ πA (H).
We can thus take a subset S ⊆ πA (H) ∩ H of least cardinality, with
(3.2)
hSiπA (H) = πA (H) ∩ H.
Furthermore, take RA (respectively, RB ) to be a subset of H mapped bijectively by πA (respectively, πB ) onto a generating set of πA (H) (respectively,
6
NIKOLAY NIKOLOV, ZVI SHEMTOV, AND MARK SHUSTERMAN
πB (H)) of least cardinality. Set L ··= hS ∪ RA ∪ RB i and note that L ≤ H.
Since S ⊆ A, conjugation by B does not affect it, so we find that
3.2
Ker(πB |H ) = πA (H) ∩ H = hSiπA (H)
(3.3)
2.2
= hSiBπA (H) = hSiBhRA i = hSihRA i ≤ L.
On the other hand, πB (H) = πB (hRB i) ≤ πB (L) so in conjunction with
(3.3) we conclude that L = H. Thus
d(H) = d(L) ≤ |S| + |RA | + |RB |
= dπA (H) (πA (H) ∩ H) + d(πA (H)) + d(πB (H))
2.4
≤ r(πA (H)/πA (H) ∩ H, πA (RA )) + d(πA (H)) + d(πB (H))
(3.4)
2.5
≤ 128d(πA (H))[G : H]3/7 + d(πA (H)) + d(πB (H)).
Moreover,
(3.5)
2.1
d(πA (H)) ≤ d(A)[A : πA (H)] = d(A)[G : BH] ≤ d(G)[G : BH]
and similarly, we have d(πB (H)) ≤ d(G)[G : AH]. Combining this inequality, (3.4), and (3.5) we obtain (3).
Let us now deduce Theorem 1.1.
Proof. Suppose that our claim is false, so (after passing to a subsequence)
we may assume that the limit in (1.6) is positive. By Theorem 3.1 (1),
1
1
d(Un )
≤ d(G)
+
(3.6)
[G : Un ]
[AUn : Un ] [G : AUn ]
and the first summand on the right hand side tends to 0 in view of our
assumption that [A : A ∩ Un ] → ∞ as n → ∞. Since the left hand side of
(3.6) tends to some c > 0, we conclude that [G : AUn ] is bounded as n → ∞.
Similarly, [G : BUn ] is bounded. Finally, apply Theorem 3.1 (3) to Un .
4. Lower bounds on the number of generators
To which extent are the bounds in Theorem 3.1 tight? Suppose that A
and B are isomorphic to a free group F on two generators. Clearly, (1) and
(2) are tight up to a constant once H ··= A0 × B0 where A0 , B0 ≤ A, B are
subgroups with [A : A0 ] = [B : B0 ].
It is conjectured ([38, Conjecture 2]) that every finite group has a presentation with a logarithmic number of relations. If this improvement of
Corollary 2.5 holds, then the argument from the proof of Theorem 3.1 (3)
gives a logarithmic bound on the number of generators as a function of the
index of H in G. Let us show that such a bound is tight (up to a constant).
Fix a prime p, and let {Pn }∞
n=1 be the lower p-central series defined by
(4.1)
P1 = F, Pn+1 = Pnp [F, Pn ].
GRADIENTS OF SEQUENCES OF SUBGROUPS IN A DIRECT PRODUCT
7
Set
(4.2)
pan ··= |Pn /Pn+1 |,
bn ··=
n
X
ai .
i=1
By [22, Corollary 3.4], an =
by
(4.3)
Pn
i=1 ri
where ri are the Witt numbers given
1X
µ
ri =
i
j|i
i
2j .
j
2n /n,
It is thus easy
and in particular
radius of conP∞to see nthat rn ∼
P∞ n the
2
vergence
of n=1 rn x is 1/2. Multiplying by ( n=0 x ) we deduce that
P∞
n has the same radius of convergence, and thus
b
x
n
n=1
(4.4)
lim sup
n→∞
bn
bn−1
≥ 2.
Set Un ··= {(x, y) ∈ F × F | xPn = yPn }. We have
(4.5)
[F × F : Un ] = pbn−1 = [F × {1} : F × {1} ∩ Un ].
We claim that Un maps onto the elementary abelian group Pn /Pn+1 . Let
∆ ≤ (F/Pn+1 ) × (F/Pn+1 ) be the image of Un mod Pn+1 . That is
(4.6)
∆ ··= {(xPn+1 , yPn+1 ) ∈ (F/Pn+1 ) × (F/Pn+1 ) | xPn = yPn }.
Let L ··= {(xPn+1 , xPn+1 ) | x ∈ F } be the diagonal subgroup of ∆. Clearly,
L is a normal subgroup of ∆ that commutes
with (Pn /Pn+1 ) × (Pn /Pn+1 ).
Moreover, L∩ (Pn /Pn+1 )×(Pn /Pn+1 ) is the diagonal subgroup isomorphic
to Pn /Pn+1 . It follows that
(4.7)
(Pn /Pn+1 ) × (Pn /Pn+1 )
∆∼
∼
=
= Pn /Pn+1 .
L
L ∩ ((Pn /Pn+1 ) × (Pn /Pn+1 ))
Hence Un has Pn /Pn+1 as a homomorphic image and therefore d(Un ) ≥ an .
At last, note that for any ǫ > 0 we have
(4.8)
4.4
d(Un )
an
bn
≥
=
−1 ≥ 1−ǫ
logp [F × F : Un ]
bn−1
bn−1
where the last inequality holds for infinitely many values of n (that is, (4.8)
holds for a subsequence of {Un }∞
n=1 ).
5. Upper bounds on torsion in the abelianization
In the following we give the bound needed to establish Theorem 1.2. The
torsion subgroup of an abelian group M is denoted by t(M ).
Lemma 5.1. Let A, B be finitely generated groups, and let H be a finite
index subgroup of G ··= A × B. Then
(5.1)
|t(H ab )| ≤ |t(πA (H)ab )| · |t(πB (H)ab )| · [G : H]2(1+log[G:H]).
8
NIKOLAY NIKOLOV, ZVI SHEMTOV, AND MARK SHUSTERMAN
Proof. We use the shorthand K ′ to denote the commutator subgroup [K, K]
of a group K, and set
(5.2)
A0 ··= πA (H) ∩ H = A ∩ H,
B0 ··= πB (H) ∩ H = B ∩ H.
Applying Proposition 2.1, we conclude that A0 ⊳ πA (H), B0 ⊳ πB (H) are
normal subgroups of finite index. One shows easily that
(5.3)
[πA (H), A0 ] × [πB (H), B0 ] ≤ H ′ ≤ H ∩ πA (H)′ × πB (H)′
so by Corollary 2.7 we have
i
h
H ∩ πA (H)′ × πB (H)′ : H ′ ≤
h
(5.4)
h
h
πA (H)′ × πB (H)′ : H ′
i 5.3
≤
i
πA (H)′ × πB (H)′ : [πA (H), A0 ] × [πB (H), B0 ] =
i h
i 2.7
πA (H)′ : [πA (H), A0 ] · πB (H)′ : [πB (H), B0 ] ≤
i1+log[πA (H):A0 ] h
i1+log[πB (H):B0 ] 5.2
πA (H) : A0
· πB (H) : B0
≤
h
i2(1+log[G:H])
G:H
.
h
In particular, by (5.4), the subgroup
(5.5)
M ··= H ∩ πA (H)′ × πB (H)′ /H ′
of H ab is finite, so
|t(H/H ′ )| = |M | · |H ab /M | ≤ |M | · t
(5.6)
ab
πA (H) × πB (H)
5.4
≤ [G : H]2(1+log[G:H]) · |t(πA (H)ab )| · |t(πB (H)ab )|.
In order to obtain Theorem 1.2 from Lemma 5.1, one just has to note that
our assumption (1.5) on the growth of the indices to infinity in Theorem 1.2,
implies that the first two factors in (5.1) grow subexponentially with the index of a subgroup in a sequence, since the torsion in the abelianization of
finite index subgroups of a finitely presented group grows at most exponentially with the index.
Acknowledgments
This work is partially supported by the ERC grant ANALYTIC no.
259527 of Goulnara Arzhantseva. Mark Shusterman is grateful to the Azrieli
Foundation for the award of an Azrieli Fellowship. The third author was
partially supported by a grant of the Israel Science Foundation with cooperation of UGC no. 40/14.
GRADIENTS OF SEQUENCES OF SUBGROUPS IN A DIRECT PRODUCT
9
References
[1] M. Abert, N. Bergeron, I. Biringer, T. Gelander, N. Nikolov, J. Raimbault, I. Samet,
On the growth of L2 -invariants for sequences of lattices in Lie groups, To appear in
Ann. Math.
[2] M. Abert, T. Gelander, N. Nikolov, Rank, combinatorial cost and homology torsion
growth in higher rank lattices, preprint, 2015.
[3] M. Abert, A. Jaikin-Zapirain, N. Nikolov, The rank gradient from a combinatorial
viewpoint, Groups, Geom., and Dyn. 5, 2, 213-230, 2011.
[4] M. Abert, N. Nikolov, Rank gradient, cost of groups and the rank versus Heegaard
genus problem, J. of the Europ. Math. Soc. 14, 5, 1657-1677, 2012.
[5] D. J. Allums, R. I. Grigorchuk, The rank gradient and the Lamplighter Group, Involve
J. of Math. 4, 3, 297-305, 2011.
[6] U. Bader, T. Gelander, R. Sauer, Homology and homotopy complexity in negative
curvature, preprint, arXiv:1612.04871, 2016.
[7] N. Bergeron, P. Linnell, W. Luck, R. Sauer, On the growth of Betti numbers in p-adic
analytic towers, preprint, arXiv:1204.3298.
[8] N. Bergeron, M. Lipnowski, Twisted limit formula for torsion and cyclic base change,
preprint, arXiv:1409.6749, 2014.
[9] N. Bergeron, A. Venkatesh, The asymptotic growth of torsion homology for arithmetic
groups, J. Inst. Math. Jussieu, 12(02), 391-447, 2013.
[10] N. Bergeron, M. H. Sengun, A. Venkatesh, Torsion homology growth and cycle complexity of arithmetic manifolds, Duke Math. J. 165(9), 1629-1693, 2016.
[11] O. Braunling,
Torsion homology growth beyond asymptotics,
preprint,
arXiv:1702.06243, 2017.
[12] J. Brock, N. Dunfield, Injectivity radii of hyperbolic integer homology 3-spheres, Geom.
& Top. 19(1), 497-523, 2015.
[13] F. Calegari, M. Emerton, Mod-p cohomology growth in p-adic analytic towers of 3manifolds, Groups, Geom. Dyn. 5, 2, 355-366, 2011.
[14] F. Calegari, A. Venkatesh, A torsion Jacquet–Langlands correspondence, preprint,
arXiv:1212.3847, 2012.
[15] B. Clair, K. Whyte, Growth of Betti numbers, Topology, 42, 5, 1125-1142, 2003.
[16] P. de la Harpe, Brouwer degree, domination of manifolds, and groups presentable by
products, preprint arXiv:1609.06637, 2016.
[17] P. de la Harpe, D. Kotschick, Presentability by products for some classes of groups,
J.Top. Anal. 1-23, 2014.
[18] M. Ershov, W. Luck, The First L2 -Betti Number and Approximation in Arbitrary
Characteristic, Doc. Math. 19, 313-331, 2014.
[19] D. Girao, Rank gradient in co-final towers of certain Kleinian groups, Groups Geom.
Dyn. 8, 143-155, 2014.
[20] R. Grigorchuk, R. Kravchenko, On the lattice of subgroups of the lamplighter group,
Int. J. Alg. Comput. 24, 837, 2014.
[21] R. Guralnick, W. Kantor, M. Kassabov, A. Lubotzky, Presentations of finite simple
groups: a quantitative approach, J. Amer. Math. Soc. 21, 3, 711-774, 2008.
[22] G.T. Helleloid, U. Martin, The automorphism group of a finite p-group is almost always a p-group, Formal Power Series and Algebraic Combinatorics, Nankai University,
Tianjin, China, 2007.
[23] A. Kar, N. Nikolov, Cost and rank gradient of Artin groups and their relatives,
Groups, Geom., and Dyn. 8, 4, 1195-1205, 2014.
[24] A. Kar, P. Kropholler, N. Nikolov, On growth of homology torsion in amenable groups,
Math. Proc. Camb. Phil. Soc. 162, 2, 2017.
[25] G. Karpilovsky, The Schur multiplier, Oxford, 1987.
10
NIKOLAY NIKOLOV, ZVI SHEMTOV, AND MARK SHUSTERMAN
[26] A. Kechris, B. Miller, Topics in orbit equivalence, No. 1852. Springer Science & Business Media, 2004.
[27] S. Kionke, J. Schwermer, On the growth of the first Betti number of arithmetic hyperbolic 3-manifolds, preprint, arXiv : 1204.3750, 2012.
[28] D. Kotschick, C. Loh, Fundamental classes not representable by products, J. London
Math. Soc. 79, 545-561, 2009.
[29] D. Kotschick, C. Loh, Groups not presentable by products, Groups Geom. Dyn. 7,
181-204, 2013.
[30] M. Lackenby, A characterisation of large finitely presented groups, J. Algebra, 287,
458-473, 2005.
[31] M. Lackenby, Expanders, rank and graphs of groups, Isr. J. Math. 146, 357-370, 2005.
[32] M. Lackenby, Finite covering spaces of 3-manifolds, Proc. ICM, 2, 2010.
[33] P. Linnell, W. Luck, R. Sauer, The limit of Fp -Betti numbers of a tower of finite
covers with amenable fundamental groups, Proc. Amer. Math. Soc. 139 (2) : 421-434,
2011.
[34] C. Loh, Rank gradient vs. stable integral simplicial volume, preprint,
arXiv:1704.05222, 2017.
[35] W. Luck, L2 -invariants: theory and applications to geometry and K-theory, Vol. 44.
Springer Science & Business Media, 2013.
[36] W. Luck, Survey on approximating L2 -invariants by their classical counterparts:
Betti numbers, torsion invariants and homological growth, preprint, arXiv:1501.07446,
2015.
[37] W. Luck, D. Osin, Approximating the first L2 -Betti number of residually finite groups,
J. Top. Anal. 3, 2, 153-160, 2011.
[38] A. Mann, Enumerating finite groups and their defining relations, J. Group Theory,
59-64, 1998.
[39] N. Nikolov, Growth of homology torsion of metabelian groups, preprint,
arXiv:1610.05944, 2016.
[40] D. Osin, Rank gradient and torsion groups, Bull. Lond. Math. Soc. 43, 1, 10-16, 2011.
[41] N. Pappas, Rank Gradient and p-gradient of Amalgamated Free Products and HNN
Extensions, Comm. in Alg. 43, 10, 4515-4527, 2015.
[42] J. Raimbault, Exponential growth of torsion in abelian coverings, Alg. & Geom. Top.
12(3), 1331-1372, 2012.
[43] I. Rivin, Statistics of Random 3-Manifolds occasionally fibering over the circle,
preprint, arXiv:1401.5736, 2014.
[44] J. Rotman, An introduction to homological algebra, Springer Science & Business Media, 2008.
[45] J. Rotman, An introduction to the theory of groups, Vol. 148. Springer Science &
Business Media, 2012.
[46] R. Sauer, Volume and homology growth of aspherical manifolds, Geom. Top. 20, 2,
1035-1059, 2016.
[47] B. Seward, R. Tucker-Drob, Rank gradient of Farber chains of subgroups in a direct
product, Unpublished note.
[48] J.C. Schlage-Puchta, A p-group with positive Rank Gradient, J. of Group Theory. 15,
2, 261-270, 2012.
[49] M. H. Sengun, On the torsion homology of non-arithmetic hyperbolic tetrahedral
groups, Inter. J. Number Theory, 8 (02), 311-320, 2012.
[50] M. Shusterman, Ranks of subgroups in boundedly generated groups, Bull. Lond. Math.
Soc. 48, 3, 539-547, 2016.
[51] M. Shusterman, Groups with positive rank gradient and their actions, To appear in
Mathematica Slovaca.
[52] M. Shusterman, Ascending chains of finitely generated subgroups, J. Alg. 471, 240-250,
2017.
GRADIENTS OF SEQUENCES OF SUBGROUPS IN A DIRECT PRODUCT
11
[53] D. Silver, S. Williams, Mahler measure, links and homology growth, Topology, 41(5),
979-991, 2002.
[54] H. Sun, Virtual homological torsion of closed hyperbolic 3-manifolds, J. Diff. Geom.
100 (3), 547-583, 2015.
University of Oxford, OX2 6GG Oxford, UK
E-mail address: [email protected]
Raymond and Beverly Sackler School of Mathematical Sciences, Tel-Aviv
University, Tel-Aviv, Israel
E-mail address: [email protected]
Raymond and Beverly Sackler School of Mathematical Sciences, Tel-Aviv
University, Tel-Aviv, Israel
E-mail address: [email protected]
| 4math.GR
|
ARES: Adaptive Receding-Horizon
Synthesis of Optimal Plans
Anna Lukina1 , Lukas Esterle1 , Christian Hirsch1 , Ezio Bartocci1 ,
Junxing Yang2 , Ashish Tiwari3 , Scott A. Smolka2 , and Radu Grosu1,2
arXiv:1612.07059v1 [cs.AI] 21 Dec 2016
1
Cyber-Physical Systems Group, Technische Universität Wien, Austria
2
Department of Computer Science, Stony Brook University, USA
3
SRI International, USA
Abstract. We introduce ARES, an efficient approximation algorithm
for generating optimal plans (action sequences) that take an initial state
of a Markov Decision Process (MDP) to a state whose cost is below a
specified (convergence) threshold. ARES uses Particle Swarm Optimization, with adaptive sizing for both the receding horizon and the particle
swarm. Inspired by Importance Splitting, the length of the horizon and
the number of particles are chosen such that at least one particle reaches
a next-level state, that is, a state where the cost decreases by a required
delta from the previous-level state. The level relation on states and the
plans constructed by ARES implicitly define a Lyapunov function and an
optimal policy, respectively, both of which could be explicitly generated
by applying ARES to all states of the MDP, up to some topological equivalence relation. We also assess the effectiveness of ARES by statistically
evaluating its rate of success in generating optimal plans. The ARES
algorithm resulted from our desire to clarify if flying in V-formation is
a flocking policy that optimizes energy conservation, clear view, and velocity alignment. That is, we were interested to see if one could find
optimal plans that bring a flock from an arbitrary initial state to a state
exhibiting a single connected V-formation. For flocks with 7 birds, ARES
is able to generate a plan that leads to a V-formation in 95% of the 8,000
random initial configurations within 63 seconds, on average. ARES can
also be easily customized into a model-predictive controller (MPC) with
an adaptive receding horizon and statistical guarantees of convergence.
To the best of our knowledge, our adaptive-sizing approach is the first
to provide convergence guarantees in receding-horizon techniques.
1
Introduction
Flocking or swarming in groups of social animals (birds, fish, ants, bees, etc.)
that results in a particular global formation is an emergent collective behavior
that continues to fascinate researchers [1, 8]. One would like to know if such a
formation serves a higher purpose, and, if so, what that purpose is.
One well-studied flight-formation behavior is V-formation. Most of the work
in this area has concentrated on devising simple dynamical rules that, when followed by each bird, eventually stabilize the flock to the desired V-formation [12,
2
Lukina, Esterle, Hirsch, Bartocci, Yang, Tiwari, Smolka, Grosu
13, 26]. This approach, however, does not shed very much light on the overall
purpose of this emergent behavior.
In previous work [35,36], we hypothesized that flying in V-formation is nothing but an optimal policy for a flocking-based Markov Decision Process (MDP)
M. States of M, at discrete time t, are of the form (xi (t), v i (t)), 1 6 i 6 N ,
where xi (t) and v i (t) are N -vectors (for an N -bird flock) of 2-dimensional positions and velocities, respectively. M’s transition relation, shown here for bird i
is simply and generically given by
xi (t + 1) = xi (t) + v i (t + 1),
v i (t + 1) = v i (t) + ai (t),
where ai (t) is an action, a 2-dimensional acceleration in this case, that bird i
can take at time t. M’s cost function reflects the energy-conservation, velocityalignment and clear-view benefits enjoyed by a state of M (see Section 2).
In this paper, we not only confirm this hypothesis, but we also devise a very
general adaptive, receding-horizon synthesis algorithm (ARES) that, given an
MDP and one of its initial states, generates an optimal plan (action sequence)
taking that state to a state whose cost is below a desired threshold. In fact,
ARES implicitly defines an optimal, online-policy, synthesis algorithm that could
be used in practice if plan generation can be performed in real-time.
ARES makes repeated use of Particle Swarm Optimization (PSO) [22] to effectively generate a plan. This was in principle unnecessary, as one could generate
an optimal plan by calling PSO only once, with a maximum plan-length horizon.
Such an approach, however, is in most cases impractical, as every unfolding of
the MDP adds a number of new dimensions to the search space. Consequently,
to obtain an adequate coverage of this space, one needs a very large number of
particles, a number that is either going to exhaust available memory or require
a prohibitive amount of time to find an optimal plan.
A simple solution to this problem would be to use a short horizon, typically of
size two or three. This is indeed the current practice in Model Predictive Control
(MPC) [14]. This approach, however, has at least three major drawbacks. First,
and most importantly, it does not guarantee convergence and optimality, as one
may oscillate or become stuck in a local optimum. Second, in some of the steps,
the window size is unnecessarily large thereby negatively impacting performance.
Third, in other steps, the window size may be not large enough to guide the
optimizer out of a local minimum (see Fig. 1 (left)). One would therefore like to
find the proper window size adaptively, but the question is how one can do it.
Inspired by Importance Splitting (IS), a sequential Monte-Carlo technique for
estimating the probability of rare events, we introduce the notion of a level-based
horizon (see Fig. 1 (right)). Level `0 is the cost of the initial state, and level `m is
the desired threshold. By using a state function, asymptotically converging to the
desired threshold, we can determine a sequence of levels, ensuring convergence
of ARES towards the desired optimal state(s) having a cost below `m = ϕ.
The levels serve two purposes. First, they implicitly define a Lyapunov function, which guarantees convergence. If desired, this function can be explicitly
ARES: Adaptive Receding-Horizon Synthesis of Optimal Plans
Cost
3
Level
`0
`0
`1
`1
.
.
.
`i
`i
`i+1
`i+1
.
.
.
ϕ
s0 s1 . . . si
...
si+3
State
`m
s0s1 si si+3 . . .
s∗
State
Fig. 1. Left: If state s0 has cost `0 , and its successor-state s1 has cost less than `1 , then
a horizon of length 1 is appropriate. However, if si has a local-minimum cost `i , one
has to pass over the cost ridge in order to reach level `i+1 , and therefore ARES has to
adaptively increase the horizon to 3. Right: The cost of the initial state defines `0 and
the given threshold ϕ defines `m . By choosing m equal segments on an asympthotically
converging (Lyapunov) function (where the number m is empirically determined), one
obtains on the vertical cost-axis the levels required for ARES to converge.
generated for all states, up to some topological equivalence. Second, the levels
help PSO overcome local minima (see Fig. 1 (left)). If reaching a next level
requires PSO to temporarily pass over a state-cost ridge, ARES incrementally
increases the size of the horizon, up to a maximum length.
Another idea imported from IS is to maintain n clones of the initial state at
a time, and run PSO on each of them (see Fig. 3). This allows us to call PSO
for each clone and desired horizon, with a very small number of particles per
clone. Clones that do not reach the next level are discarded, and the successful
ones are resampled. The number of particles is increased if no clone reaches a
next level, for all horizons chosen. Once this happens, we reset the horizon to
one, and repeat the process. In this way, we adaptively focus our resources on
escaping from local minima. At the last level, we choose the optimal particle (a
V-formation in case of flocking) and traverse its predecessors to find a plan.
We asses the rate of success in generating optimal plans in form of an (ε, δ)approximation scheme, for a desired error margin ε, and confidence ratio 1−δ.
Moreover, we can use the state-action pairs generated during the assessment (and
possibly some additional new plans) to construct an explicit (tabled) optimal
policy, modulo some topological equivalence. Given enough memory, one can
use this policy in real time, as it only requires a table look-up.
To experimentally validate our approach, we have applied ARES to the problem of V-formation in bird flocking (with a deterministic MDP). The cost function to be optimized is defined as a weighted sum of the (flock-wide) clear-view,
velocity-alignment, and upwash-benefit metrics. Clear view and velocity alignment are more or less obvious goals. Upwash optimizes energy savings. By flapping its wings, a bird generates a trailing upwash region off its wing tips; by
using this upwash, a bird flying in this region (left or right) can save energy.
4
Lukina, Esterle, Hirsch, Bartocci, Yang, Tiwari, Smolka, Grosu
Note that by requiring that at most one bird does not feel its effect, upwash can
be used to define an analog version of a connected graph.
We ran ARES on 8,000 initial states chosen uniformly and at random, such
that they are packed closely enough to feel upwash, but not too close to collide.
We succeeded to generate a V-formation 95% of the time, with an error margin
of 0.05 and a confidence ratio of 0.99. These error margin and confidence ratio
dramatically improve if we consider all generated states and the fact that each
state within a plan is independent from the states in all other plans.
The rest of this paper is organized as follows. Section 2 reviews our work on
bird flocking and V-formation, and defines the manner in which we measure the
cost of a flock (formation). Section 3 revisits the swarm optimization algorithm
used in this paper, and Section 4 examines the main characteristics of importance
splitting. Section 5 states the definition of the problem we are trying to solve.
Section 6 introduces ARES, our adaptive receding-horizon synthesis algorithm
for optimal plans, and discusses how we can extend this algorithm to explicitly
generate policies. Section 7 measures the efficiency of ARES in terms of an (ε, δ)approximation scheme. Section 8 compares our algorithm to related work, and
Section 9 draws our conclusions and discusses future work.
2
V-Formation MDP
We represent a flock of birds as a dynamically evolving system. Every bird in
our model [17] moves in 2-dimensional space performing acceleration actions
determined by a global controller. Let xi (t), v i (t) and ai (t) be 2-dimensional
vectors of positions, velocities, and accelerations, respectively, of bird i at time
t, where i ∈ {1, . . . , b}, for a fixed b. The discrete-time behavior of bird i is then
xi (t + 1) = xi (t) + v i (t + 1),
v i (t + 1) = v i (t) + ai (t).
(1)
The controller detects the positions and velocities of all birds through sensors,
and uses this information to compute an optimal acceleration for the entire flock.
A bird uses its own component of the solution to update its velocity and position.
We extend this discrete-time dynamical model to a (deterministic) MDP by
adding a cost (fitness) function4 based on the following metrics inspired by [35]:
– Clear View (CV ). A bird’s visual field is a cone with angle θ that can be
blocked by the wings of other birds. We define the clear-view metric by
accumulating the percentage of a bird’s visual field that is blocked by other
birds. Fig. 2 (left) illustrates the calculation of the clear-view metric. The
optimal value in a V-formation is CV ∗ = 0, as all birds have a clear view.
– Velocity Matching (VM ). The accumulated differences between the velocity
of each bird and all other birds, summed up over all birds in the flock defines
VM . Fig. 2 (middle) depicts the values of VM in a velocity-unmatched flock.
4
A classic MDP [28] is obtained by adding sensor/actuator or wind-gust noise.
ARES: Adaptive Receding-Horizon Synthesis of Optimal Plans
5
Fig. 2. Illustration of the clear view (CV ), velocity matching (VM ), and upwash
benefit (UB ) metrics. Left: Bird i’s view is partially blocked by birds j and k. Hence,
its clear view is CV = (α + β)/θ. Middle: A flock and its unaligned bird velocities results
in a velocity-matching metric VM = 6.2805. In contrast, VM = 0 when the velocities
of all birds are aligned. Right: Illustration of the (right-wing) upwash benefit bird i
receives from bird j depending on how it is positioned behind bird j. Note that bird
j’s downwash region is directly behind it.
The optimal value in a V-formation is VM ∗ = 0, as all birds will have the
same velocity (thus maintaining the V-formation).
– Upwash Benefit (UB ). The trailing upwash is generated near the wingtips of
a bird, while downwash is generated near the center of a bird. We accumulate
all birds’ upwash benefits using a Gaussian-like model of the upwash and
downwash region, as shown in Fig. 2 (right) for the right wing. The maximum
upwash a bird can obtain has an upper bound of 1. For bird i with UB i , we
use 1 −UB i as its upwash-benefit metric, because the optimization algorithm
performs minimization of the fitness metrics. The optimal value in a Vformation is UB ∗ = 1, as the leader does not receive any upwash.
Finding smooth and continuous formulations of the fitness metrics is a key element of solving optimization problems. The PSO algorithm has a very low probability of finding an optimal solution if the fitness metric is not well-designed.
Let c(t) = {ci (t)}bi=1 = {xi (t), v i (t)}bi=1 be a flock configuration at time-step
t. Given the above metrics, the overall fitness (cost) metric J is of a sum-ofsquares combination of VM , CV , and UB defined as follows:
J(c(t), ah (t), h) = (CV (cha (t)) − CV ∗ )2 + (VM (cha (t)) − VM ∗ )2
+ (UB (cha (t)) − UB ∗ )2 ,
(2)
where h is the receding prediction horizon (RPH), ah (t) is a sequence of accelerations of length h, and cha (t) is the configuration reached after applying ah (t)
to c(t). Formally, we have
cha (t) = {xha (t), v ha (t)} = {x(t) +
h(t)
X
τ =1
v(t + τ ), v(t) +
h(t)
X
aτ (t)},
(3)
τ =1
where aτ (t) is the τ th acceleration of ah (t). A novelty of this paper is that, as
described in Section 6, we allow RPH h(t) to be adaptive in nature.
6
Lukina, Esterle, Hirsch, Bartocci, Yang, Tiwari, Smolka, Grosu
The fitness function J has an optimal value of 0 in a perfect V-formation.
The main goal of ARES is to compute the sequence of acceleration actions
that lead the flock from a random initial configuration towards a controlled Vformation characterized by optimal fitness in order to conserve energy during
flight including optimal combination of a clear visual field along with visibility
of lateral neighbors. Similar to the centralized version of the approach given
in [35], ARES performs a single flock-wide minimization of J at each time-step
t to obtain an optimal plan of length h of acceleration actions:
opt-ah (t) = {opt-ahi (t)}bi=1 = arg min J(c(t), ah (t), h).
(4)
ah (t)
The optimization is subject to the following constraints on the maximum
velocities and accelerations: ||v i (t)|| 6 v max , ||ahi (t)|| 6 ρ||v i (t)|| ∀ i ∈ {1, . . . , b},
where v max is a constant and ρ ∈ (0, 1). The initial positions and velocities of
each bird are selected at random within certain ranges, and limited such that
the distance between any two birds is greater than a (collision) constant dmin ,
and small enough for all birds, except for at most one, to feel the UB . In the
following sections, we demonstrate how to generate optimal plans taking the
initial state to a stable state with optimal fitness.
3
Particle Swarm Optimization
Particle Swarm Optimization (PSO) is a randomized approximation algorithm
for computing the value of a parameter minimizing a possibly nonlinear cost
(fitness) function. Interestingly, PSO itself is inspired by bird flocking [22]. Hence,
PSO assumes that it works with a flock of birds.
Note, however, that in our running example, these birds are “acceleration
birds” (or particles), and not the actual birds in the flock. Each bird has the
same goal, finding food (reward), but none of them knows the location of the
food. However, every bird knows the distance (horizon) to the food location.
PSO works by moving each bird preferentially toward the bird closest to food.
ARES uses Matlab-Toolbox particleswarm, which performs the classical
version of PSO. This PSO creates a swarm of particles, of size say p, uniformly
at random within a given bound on their positions and velocities. Note that in
our example, each particle represents itself a flock of bird-acceleration sequences
{ahi }bi=1 , where h is the current length of the receding horizon. PSO further
chooses a neighborhood of a random size for each particle j, j = {1, . . . , p}, and
computes the fitness of each particle. Based on the fitness values, PSO stores two
vectors for j: its so-far personal-best position xjP (t), and its fittest neighbor’s
position xjG (t). The positions and velocities of each particle j in the particle
swarm 1 6 j 6 p are updated according to the following rule:
vj (t + 1) = ω · vj (t) + y1 · u1 (t + 1) ⊗ (xjP (t) − xj (t))
+ y2 · u2 (t + 1) ⊗ (xjG (t) − xj (t)),
(5)
ARES: Adaptive Receding-Horizon Synthesis of Optimal Plans
7
where ω is inertia weight, which determines the trade-off between global and
local exploration of the swarm (the value of ω is proportional to the exploration range); y1 and y2 are self adjustment and social adjustment, respectively;
u1 , u2 ∈ Uniform(0, 1) are randomization factors; and ⊗ is the vector dot product, that is, ∀ random vector z: (z1 , . . . , zb ) ⊗ (xj1 , . . . , xjb ) = (z1 xj1 , . . . , zb xjb ).
If the fitness value for xj (t + 1) = xj (t) + vj (t + 1) is lower than the one
for xjP (t), then xj (t + 1) is assigned to xjP (t + 1). The particle with the best
fitness over the whole swarm becomes a global best for the next iteration. The
procedure is repeated until the number of iterations reaches its maximum, the
time elapses, or the minimum criteria is satisfied. For our bird-flock example we
obtain in this way the best acceleration.
4
Importance Splitting
Importance Splitting (IS) is a sequential Monte-Carlo approximation technique
for estimating the probability of rare events in a Markov process [7]. The algorithm uses a sequence S0 , S1 , S2 , . . . , Sm of sets of states (of increasing “importance”) such that S0 is the set of initial states and Sm is the set of states defining
the rare event. The probability p, computed as P(Sm | S0 ) of reaching Sm from
the initial set of states S0 , is assumed to be extremely low (thus, a rare event),
and one desires to estimate this probability [16]. Random sampling approaches,
such as the additive-error approximation algorithm described in Section 7, are
bound to fail (are intractable) in this case, as they would require an enormous
number of samples to estimate p with low-variance.
Importance splitting is a way of decomposing the estimation of p. In IS, the
sequence S0 , S1 , . . . of sets of states is defined so that the conditional probabilities pi = P(Si | Si−1 ) of going from one level, Si−1 , to the next one, Si , are
considerably larger than p, and essentially equal to one another. Q
The resulting
k
probability of the rare event is then calculated as the product p = i=1 pi of the
intermediate probabilities. The levels can be defined adaptively [23].
To estimate pi , IS uses a swarm of particles of size N , with a given initial
distribution over the states of the stochastic process. During stage i of the algorithm, each particle starts at level Si−1 and traverses the states of the stochastic
process, checking if it reaches Si . If, at the end of the stage, the particle fails to
reach Si , the particle is discarded. Suppose that Ki particles survive. In this case,
pi = Ki /N . Before starting the next stage, the surviving particles are resampled,
such that IS once again has N particles. Whereas IS is used for estimating probability of a rare event in a Markov process, we use it here for synthesizing a plan
for a controllable Markov process, by combining it with ideas from controller
synthesis (receding-horizon control) and nonlinear optimization (PSO).
5
Problem Definition
Definition 1. A Markov decision process (MDP) M is a sequential decision problem that consists of a set of states S (with an initial state s0 ), a set of
8
Lukina, Esterle, Hirsch, Bartocci, Yang, Tiwari, Smolka, Grosu
actions A, a transition model T , and a cost function J. An MDP is deterministic if for each state and action, T : S × A → S specifies a unique state.
Definition 2. The optimal plan synthesis problem for an MDP M, an
arbitrary initial state s0 of M, and a threshold ϕ is to synthesize a sequence of
actions ai of length 1 6 i 6 m taking s0 to a state s∗ such that cost J(s∗ ) 6 ϕ.
Section 6 presents our adaptive receding-horizon synthesis algorithm (ARES)
for the optimal plan synthesis problem. In our flocking example (Section 2),
ARES is used to synthesize a sequence of acceleration-actions bringing an arbitrary bird flock s0 to an optimal state of V-formation s∗ . We assume that we
can easily extend such an optimal plan to maintain the cost of successor states
below ϕ ad infinitum (optimal stability).
6
The ARES Algorithm for Plan Synthesis
As mentioned in Section 1, one could in principle solve the optimization problem
defined in Section 5 by calling the PSO only once, with a horizon h in M equaling
the maximum length m allowed for a plan. This approach, however, tends to
explode the search space, and is therefore in most cases intractable. Indeed,
preliminary experiments with this technique applied to our running example
could not generate any convergent plan.
A more tractable approach is to make repeated calls to PSO with a small
horizon length h. The question is how small h can be. The current practice
in model-predictive control (MPC) is to use a fixed h, 1 6 h 6 3 (see the outer
loop of Fig. 3, where resampling and conditional branches are disregarded).
Unfortunately, this forces the selection of locally-optimal plans (of size less than
three) in each call, and there is no guarantee of convergence when joining them
together. In fact, in our running example, we were able to find plans leading to
a V-formation in only 45% of the time for 10, 000 random initial flocks.
Inspired by IS (see Fig. 1 (right) and Fig. 3), we introduce the notion of a
level-based horizon, where level `0 equals the cost of the initial state, and level
`m equals the threshold ϕ. Intuitively, by using an asymptotic cost-convergence
function ranging from `0 to `m , and dividing its graph in m equal segments, we
can determine on the vertical axis a sequence of levels ensuring convergence.
The asymptotic function ARES implements is essentially `i = `0 (m − i)/ m,
but specifically tuned for each particle. Formally, if particle k has previously
reached level equaling Jk (si−1 ), then its next target level is within the distance
∆k = Jk (si−1 )/(m − i + 1). In Fig. 3, after passing the thresholds assigned to
them, values of the cost function in the current state si are sorted in ascending
order {Jbk }nk=1 . The lowest cost Jb1 should be apart from the previous level `i−1
at least on its ∆1 for the algorithm to proceed to the next level `i := Jb1 .
The levels serve two purposes. First, they implicitly define a Lyapunov function, which guarantees convergence. If desired, this function can be explicitly
generated for all states, up to some topological equivalence. Second, the levels
`i help PSO overcome local minima (see Fig. 1 (left)). If reaching a next level
ARES: Adaptive Receding-Horizon Synthesis of Optimal Plans
9
next states after
n clones
s0
n
applying {ah
k }k=1
h si
a1
J1
PSO
PSO
s0
PSO
..
.
ah
2
I ...
J2
ah
3
J3
..
.
..
.
PSO
Resampling
ah
n
Sort
{Jbk }n
k=1
`i−1 − Jb1 > ∆1
Yes
..
.
..
.
Jn
No
h ++
h := 1;
p += pinc ;
i ++
Yes
Yes
p < pmax
h < hmax
No
No
Particle exhaustion
Yes
i<m
No
Timeout
Yes
`i > ϕ
`i := Jb1
No
Stable state
Fig. 3. Graphical representation of ARES.
requires PSO to temporarily pass over a state-cost ridge, then ARES incrementally increases the size of the horizon h, up to a maximum size hmax . For particle
k, passing the thresholds ∆k means that it reaches a new level, and the definition
of ∆k ensures a smooth degradation of its threshold.
Another idea imported from IS and shown in Fig. 3, is to maintain n clones
{Mk }nk=1 of the MDP M (and its initial state) at any time t, and run PSO, for
a horizon h, on each h-unfolding Mhk of them. This results in an action sequence
ahk of length h (see Algo. 1). This approach allows us to call PSO for each clone
and desired horizon, with a very small number of particles p per clone.
Algorithm 1: Simulate (M, h, i, {∆k , Jk (si−1 )}nk=1 )
1
2
3
4
5
6
7
foreach Mk ∈ M do
[ahk , Mhk ] ← particleswarm(Mk , p, h); // use PSO in order to determine
best next action for the MDP Mk with RPH h
Jk (si ) ← Cost(Mhk , ahk , h); // calculate cost function if applying the
sequence of optimal actions of length h
if Jk (si−1 ) − Jk (si ) > ∆k then
∆k ← Jk (si )/(m − i); // new level-threshold
end
end
10
Lukina, Esterle, Hirsch, Bartocci, Yang, Tiwari, Smolka, Grosu
Algorithm 2: Resample ({Mhk , Jk (si )}nk=1 )
1
2
3
4
5
6
7
8
I ← Sort ascending Mhk by their current costs; // find indexes of MDPs whose
costs are below the median among all the clones
for k = 1 to n do
if k ∈
/ I then
Sample r uniformly at random from I; Mk ← Mhr ;
else
Mk ← Mhk ; // Keep more successful MDPs unchanged
end
end
To check which particles have overcome their associated thresholds, we sort
the particles according to their current cost, and split them in two sets: the
successful set, having the indexes I and whose costs are lower than the median
among all clones; and the unsuccessful set with indexes in {1, . . ., n} \I, which are
discarded. The unsuccessful ones are further replenished, by sampling uniformly
at random from the successful set I (see Algo. 2).
The number of particles is increased p = p + pinc if no clone reaches a next
level, for all horizons chosen. Once this happens, we reset the horizon to one, and
repeat the process. In this way, we adaptively focus our resources on escaping
from local minima. From the last level, we choose the state s∗ with the minimal
cost, and traverse all of its predecessor states to find an optimal plan comprised
of actions {ai }16i6m that led MDP M to the optimal state s∗ . In our running
example, we select a flock in V-formation, and traverse all its predecessor flocks.
The overall procedure of ARES is shown in Algo. 3.
Proposition 1 (Optimality and Minimality). (1) Let M be an MDP. For
any initial state s0 of M, ARES is able to solve the optimal-plan synthesis
problem for M and s0 . (2) An optimal choice of m in function ∆k , for some
particle k, ensures that ARES also generates the shortest optimal plan.
Proof (Sketch). (1) The dynamic-threshold function ∆k ensures that the initial
cost in s0 is continuously decreased until it falls below ϕ. Moreover, for an appropriate number of clones, by adaptively determining the horizon and the number
of particles needed to overcome ∆k , ARES always converges, with probability
1, to an optimal state, given enough time and memory. (2) This follows from
convergence property (1), and from the fact that ARES always gives preference
to the shortest horizon while trying to overcome ∆k .
The optimality referred to in the title of the paper is in the sense of (1).
One, however, can do even better than (1), in the sense of (2), by empirically
determining parameter m in the dynamic-threshold function ∆k . Also note that
ARES is an approximation algorithm. As a consequence, it might return nonminimal plans. Even in these circumstances, however, the plans will still lead to
an optimal state. This is a V-formation in our flocking example.
ARES: Adaptive Receding-Horizon Synthesis of Optimal Plans
11
Algorithm 3: ARES
Input : M, ϕ, pstart , pinc , pmax , hmax , m, n
Output: {ai }16i6 m // synthesized optimal plans
1
Initialize `0 ← inf; {Jk (s0 )}n
k=1 ← inf; p ← pstart ; i ← 1; h ← 1; ∆k ← 0;
2
while (`i > ϕ) ∨ (i < m) do
// find and apply best actions with RPH h
n
[{ahk , Jk (si ), Mhk }n
k=1 ] ←Simulate(M, h, i, {∆k , Jk (si−1 )}k=1 );
Jb1 ← sort(J1 (si ), . . . , Jn (si )); // find minimum cost among all the clones
if `i−1 − Jb1 > ∆1 then
`i ← Jb1 ; // new level has been reached
i ← i + 1; h ← 1; p ← pstart ; // reset adaptive parameters
h
n
{Mk }n
k=1 ← Resample({Mk , Jk (si )}k=1 );
else
if h < hmax then
h ← h + 1; // improve time exploration
else
if p < pmax then
h ← 1; p ← p + pinc ; // improve space exploration
else
break;
end
end
end
end
Take a clone in the state with minimum cost `i = J(s∗i ) 6 ϕ at the last level i;
foreach i do
{s∗i−1 , ai } ← P re(s∗i ); // find predecessor and corresponding action
end
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
7
Experimental Results
To assess the performance of our approach, we developed a simple simulation
environment in Matlab. All experiments were run on an Intel Core i7-5820K
CPU with 3.30 GHz and with 32GB RAM available.
We performed numerous experiments with a varying number of birds. Unless
stated otherwise, results refer to 8,000 experiments with 7 birds with the following parameters: pstart = 10, pinc = 5, pmax = 40, `max = 20, hmax = 5, ϕ = 10−3 ,
and n = 20. The initial configurations were generated independently uniformly
at random subject to the following constraints:
1. Position constraints: ∀ i ∈ {1, . . ., 7}. xi (0) ∈ [0, 3] × [0, 3].
2. Velocity constraints: ∀ i ∈ {1, . . ., 7}. v i (0) ∈ [0.25, 0.75] × [0.25, 0.75].
Table 1 gives an overview of the results with respect to the 8,000 experiments
we performed with 7 birds for a maximum of 20 levels. The average fitness
across all experiments is at 0.0282 with a standard deviation of 0.1654. We
12
Lukina, Esterle, Hirsch, Bartocci, Yang, Tiwari, Smolka, Grosu
Fig. 4. Left: Example of an arbitrary initial configuration of 7 birds. Right: The Vformation obtained by applying the plan generated by ARES. In the figures, we show
the wings of the birds, bird orientations, bird speeds (as scaled arrows), upwash regions
in yellow, and downwash regions in dark blue.
Table 1. Overview of the results for 8,000 experiments with 7 birds
Successful
No. Experiments
Min
Cost, J
Time, t
Plan Length, i
RPH, h
Total
7573
8000
Max
Avg
Std
Min
Max
Avg
2.88·10−7 9·10−4 4·10−4 3·10−4 2.88·10−7 1.4840 0.0282
23.14s 310.83s 63.55s 22.81s 23.14s 661.46s 64.85s
7
20
12.80 2.39
7
20
13.13
1
5
1.40
0.15
1
5
1.27
Std
0.1607
28.05s
2.71
0.17
achieved a success rate of 94.66% with fitness threshold ϕ = 10−3 . The average
fitness is higher than the threshold due to comparably high fitness of unsuccessful
experiments. When increasing the bound for the maximal plan length m to 30 we
achieved a 98.4% success rate in 1,000 experiments at the expense of a slightly
longer average execution time.
The left plot in Fig. 5 depicts the resulting distribution of execution times for
8,000 runs of our algorithm, where it is clear that, excluding only a few outliers
from the histogram, an arbitrary configuration of birds (Fig. 4 (left)) reaches
V-formation (Fig. 4 (right)) in around 1 minute. The execution time rises with
the number of birds as shown in Table 2.
In Fig. 5, we illustrate for how many experiments the algorithm had to increase RPH h (Fig. 5 (middle)) and the number of particles used by PSO p
(Fig. 5 (right)) to improve time and space exploration, respectively.
After achieving such a high success rate of ARES for an arbitrary initial
configuration, we would like to demonstrate that the number of experiments
Table 2. Average duration for 100 experiments with various number of birds
No. of birds
3
5
7
9
Avg. duration 4.58s 18.92s 64.85s 269.33s
ARES: Adaptive Receding-Horizon Synthesis of Optimal Plans
13
Fig. 5. Left: Distribution of execution times for 8,000 runs. Middle: Statistics of increasing RPH h. Right: Particles of PSO p for 8,000 experiments
performed is sufficient for high confidence in our results. This requires us to
determine the appropriate number N of random variables Z1 , ...ZN necessary
for the Monte-Carlo approximation scheme we apply to assess efficiency of our
approach. For this purpose, we use the additive approximation algorithm as
discussed in [17]. If the sample mean µZ = (Z1 + . . . + ZN )/N is expected to be
large, then one can exploit the Bernstein’s inequality and fix N to Υ ∝ ln(1/δ)/ε2 .
This results in an additive or absolute-error (ε, δ)-approximation scheme:
P[µZ − ε ≤ µ
eZ ≤ µZ + ε)] ≥ 1 − δ,
where µ
eZ approximates µZ with absolute error ε and probability 1 − δ.
In particular, we are interested in Z being a Bernoulli random variable:
1, if J(c(t), a(t), h(t)) 6 ϕ,
Z=
0, otherwise.
Therefore, we can use the Chernoff-Hoeffding instantiation of the Bernstein’s
inequality, and further fix the proportionality constant to Υ = 4 ln(2/δ)/ε2 , as
in [20]. Hence, for our performed 8,000 experiments, we achieve a success rate of
95% with absolute error of ε = 0.05 and confidence ratio 0.99.
Moreover, considering that the average length of a plan is 13, and that each
state in a plan is independent from all other plans, we can roughly consider
that our above estimation generated 80,000 independent states. For the same
confidence ratio of 0.99 we then obtain an approximation error ε = 0.016, and
for a confidence ratio of 0.999, we obtain an approximation error ε = 0.019.
8
Related Work
Organized flight in flocks of birds can be categorized in cluster flocking and line
formation [19]. In cluster flocking the individual birds in a large flock seem to
be uncoordinated in general. However, the flock moves, turns, and wheels as
if it were one organism. In 1987 Reynolds [27] defined his three famous rules
describing separation, alignment, and cohesion for individual birds in order to
have them flock together. This work has been great inspiration for research in
the area of collective behavior and self-organization.
14
Lukina, Esterle, Hirsch, Bartocci, Yang, Tiwari, Smolka, Grosu
In contrast, line formation flight requires the individual birds to fly in a very
specific formation. Line formation has two main benefits for the long-distance
migrating birds. First, exploiting the generated uplift by birds flying in front,
trailing birds are able to conserve energy [10, 24, 34]. Second, in a staggered
formation, all birds have a clear view in front as well as a view on their neighbors [1]. While there has been quite some effort to keep a certain formation for
multiple entities when traveling together [11, 15, 30], only little work deals with
a task of achieving this extremely important formation from a random starting
configuration [6]. The convergence of bird flocking into V-formation has been
also analyzed with the use of combinatorial techniques [8].
Compared to previous work, in [5] this question is addressed without using
any behavioral rules but as problem of optimal control. In [35] a cost function was proposed that reflects all major features of V-formation, namely, Clear
View (CV), Velocity Matching (VM), and Upwash Benefit (UB). The technique
of MPC is used to achieve V-formation starting from an arbitrary initial configuration of n birds. MPC solves the task by minimizing a functional defined
as squared distance from the optimal values of CV, VM, and UB, subject to
constraints on input and output. The approach is to choose an optimal velocity
adjustment, as a control input, at each time-step applied to the velocity of each
bird by predicting model behavior several time-steps ahead.
The controller synthesis problem has been widely studied [33]. The most popular and natural technique is Dynamic Programming (DP) [4] that improves the
approximation of the functional at each iteration, eventually converging to the
optimal one given a fixed asymptotic error. Compared to DP, which considers all
the possible states of the system and might suffer from state-space explosion in
case of environmental uncertainties, approximate algorithms [2, 3, 18, 25, 31, 32]
take into account only the paths leading to desired target. One of the most efficient ones is Particle Swarm Optimization (PSO) [22] that has been adopted
for finding the next best step of MPC in [35]. Although it is a very powerful
optimization technique, it has not yet been possible to achieve a high success
rate in solving the considered flocking problem. Sequential Monte-Carlo methods
proved to be efficient in tackling the question of control for linear stochastic systems [9], in particular, Importance Splitting (IS) [23]. The approach we propose
is, however, the first attempt to combine adaptive IS, PSO, and receding-horizon
technique for synthesis of optimal plans for controllable systems. We use MPC
to synthesize a plan, but use IS to determine the intermediate fitness-based waypoints. We use PSO to solve the multi-step optimization problem generated by
MPC, but choose the planning horizon and the number of particles adaptively.
These choices are governed by the difficulty to reach the next level.
9
Conclusion and Future Work
In this paper, we have presented ARES, a very general adaptive, receding-horizon
synthesis algorithm for MDP-based optimal plans. Additionally, ARES can be
readily converted into a model-predictive controller with an adaptive receding
ARES: Adaptive Receding-Horizon Synthesis of Optimal Plans
15
horizon and statistical guarantees of convergence. We have also conducted a very
thorough performance analysis of ARES based on the problem of V-formation
in a flock of birds. For flocks of 7 birds, ARES is able to generate an optimal
plan leading to a V-formation in 95% of the 8,000 random initial configurations
we considered, with an average execution time of only 63 seconds per plan.
The execution time of the ARES algorithm can be even further improved in
a number of ways. First, we currently do not parallelize our implementation of
the PSO algorithm. Recent work [21, 29, 37] has shown how Graphic Processing Units (GPUs) are very efficient at accelerating PSO computation. Modern
GPUs, by providing thousands of cores, are well-suited for implementing PSO
as they enable execution of a very large number of particles in parallel, which
can improve accuracy of the optimization procedure. Likewise, the calculation
of the fitness function can also be run in parallel. The parallelization of these
steps should significantly speed up our simulations.
Second, we are currently using a static approach to decide how to increase
our prediction horizon and the number of particles used in PSO. Specifically, we
first increase the prediction horizon from 1 to 5, while keeping the number of
c1 satisfying
particles unchanged at 10; if this fails to find a solution with fitness J
c
`i−1 − J1 > ∆1 , we then increase the number of particles by 5. Based on our
results, we speculate that in the initial stages, increasing the prediction horizon is
more beneficial (leading rapidly to the appearance of cost-effective formations),
whereas in the later stages, increasing the number of particles is more helpful. As
future work, we will use machine-learning approaches to decide on the prediction
horizon and the number of particles deployed at runtime given the current level
and state of the MDP.
Third, in our approach, we always calculate the number of clones for resampling based on the current state. An alternative approach would rely on statistics
built up over multiple levels in combination with the rank in the sorted list to
determine whether a configuration should be used for resampling or not.
Finally, we are currently using our approach to generate plans for a flock
to go from an initial configuration to a final V-formation. Our eventual goal
is to achieve formation flight for a robotic swarm of (bird-like) drones. A realworld example is parcel-delivering drones that follow the same route to their
destinations. Letting them fly together for a while could save energy and increase
flight time. To achieve this goal, we first need to investigate the wind dynamics
of multi-rotor drones. Then, the fitness function needs to be adopted to the new
wind dynamics. Lastly, a decentralized approach of this method needs to be
implemented and tested on the drone firmware.
Acknowledgments. The first author and the last author would like to thank
Jan Kr̆etı́nský for very valuable feedback. This work was partially supported
by the Doctoral Program Logical Methods in Computer Science funded by the
Austrian Science Fund (FWF) project W1255-N23, and the Austrian National
Research Network (nr. S 11405-N23 and S 11412-N23) SHiNE funded by FWF.
16
Lukina, Esterle, Hirsch, Bartocci, Yang, Tiwari, Smolka, Grosu
References
1. Bajec, I.L., Heppner, F.H.: Organized flight in birds. Animal Behaviour 78(4),
777–789 (2009)
2. Bartocci, E., Bortolussi, L., Brázdil, T., Milios, D., Sanguinetti, G.: Policy learning
for time-bounded reachability in continuous-time markov decision processes via
doubly-stochastic gradient ascent. In: Proc. of QEST 2016: the 13th International
Conference on Quantitative Evaluation of Systems. vol. 9826, pp. 244–259 (2016)
3. Baxter, J., Bartlett, P.L., Weaver, L.: Experiments with infinite-horizon, policygradient estimation. J. Artif. Int. Res. 15(1), 351–381 (2011)
4. Bellman, R.: Dynamic Programming. Princeton University Press (1957)
5. Camacho, E.F., Alba, C.B.: Model Predictive Control. Advanced Textbooks in
Control and Signal Processing, Springer (2007)
6. Cattivelli, F.S., Sayed, A.H.: Modeling bird flight formations using diffusion adaptation. IEEE Transactions on Signal Processing 59(5), 2038–2051 (2011)
7. Cérou, F., Guyader, A.: Adaptive multilevel splitting for rare event analysis.
Stochastic Analysis and Applications 25, 417–443 (2007)
8. Chazelle, B.: The Convergence of Bird Flocking. Journal of the ACM 61(4), 21:1–
21:35 (2014)
9. Chen, Y., Wu, B., Lai, T.L.: Fast Particle Filters and Their Applications to Adaptive Control in Change-Point ARX Models and Robotics. INTECH Open Access
Publisher (2009)
10. Cutts, C., Speakman, J.: Energy savings in formation flight of pink-footed geese.
Journal of Experimental Biology 189(1), 251–261 (1994)
11. Dang, A.D., Horn, J.: Formation control of autonomous robots following desired
formation during tracking a moving target. In: Proceedings of the International
Conference on Cybernetics. pp. 160–165. IEEE (2015)
12. Dimock, G., Selig, M.: The Aerodynamic Benefits of Self-Organization in Bird
Flocks. Urbana 51, 1–9 (2003)
13. Flake, G.W.: The Computational Beauty of Nature: Computer Explorations of
Fractals, Chaos, Complex Systems, and Adaptation. MIT Press (1998)
14. Garcı́a, C.E., Prett, D.M., Morari, M.: Model predictive control: Theory and practice – a survey. Automatica 25(3), 335–348 (1989)
15. Gennaro, M.C.D., Iannelli, L., Vasca, F.: Formation Control and Collision Avoidance in Mobile Agent Systems. In: Proceedings of the International Symposium on
Control and Automation Intelligent Control. pp. 796–801. IEEE (2005)
16. Glasserman, P., Heidelberger, P., Shahabuddin, P., Zajic, T.: Multilevel Splitting for Estimating Rare Event Probabilities. Operations Research 47(4), 585–600
(1999)
17. Grosu, R., Peled, D., Ramakrishnan, C.R., Smolka, S.A., Stoller, S.D., Yang, J.:
Using statistical model checking for measuring systems. In: Proceedings of the
International Symposium Leveraging Applications of Formal Methods, Verification
and Validation. LNCS, vol. 8803, pp. 223–238. Springer (2014)
18. Henriques, D., Martins, J.G., Zuliani, P., Platzer, A., Clarke, E.M.: Statistical model checking for markov decision processes. In: Proc. of QEST 2012: the
Ninth International Conference on Quantitative Evaluation of Systems. pp. 84–93.
QEST’12, IEEE Computer Society (2012)
19. Heppner, F.H.: Avian flight formations. Bird-Banding 45(2), 160–169 (1974)
20. Hérault, T., Lassaigne, R., Magniette, F., Peyronnet, S.: Approximate probabilistic
model checking. In: Proceedings of the International Conference on Verification,
Model Checking, and Abstract Interpretation (2004)
ARES: Adaptive Receding-Horizon Synthesis of Optimal Plans
17
21. Hung, Y., Wang, W.: Accelerating parallel particle swarm optimization via gpu.
Optimization Methods and Software 27(1), 33–51 (2012)
22. James, K., Russell, E.: Particle swarm optimization. In: Proceedings of 1995 IEEE
International Conference on Neural Networks. pp. 1942–1948 (1995)
23. Kalajdzic, K., Jégourel, C., Lukina, A., Bartocci, E., Legay, A., Smolka, S.A.,
Grosu, R.: Feedback Control for Statistical Model Checking of Cyber-Physical
Systems. In: Proceedings of the International Symposium Leveraging Applications
of Formal Methods, Verification and Validation: Foundational Techniques. pp. 46–
61. LNCS, Springer (2016)
24. Lissaman, P., Shollenberger, C.A.: Formation flight of birds. Science 168(3934),
1003–1005 (1970)
25. Mannor, S., Rubinstein, R.Y., Gat, Y.: The cross entropy method for fast policy
search. In: ICML. pp. 512–519 (2003)
26. Nathan, A., Barbosa, V.C.: V-like Formations in Flocks of Artificial Birds. Artificial
Life 14(2), 179–188 (2008)
27. Reynolds, C.W.: Flocks, herds and schools: A distributed behavioral model. SIGGRAPH Computer Graphics 21(4), 25–34 (1987)
28. Russell, S., Norvig, P.: Artificial Intelligence: A Modern Approach. Prentice-Hall,
3rd edn. (2010)
29. Rymut, B., Kwolek, B., Krzeszowski, T.: GPU-Accelerated Human Motion Tracking Using Particle Filter Combined with PSO. In: Proceedings. of the International
Conference on Advanced Concepts for Intelligent Vision Systems. LNCS, vol. 8192,
pp. 426–437. Springer (2013)
30. Seiler, P., Pant, A., Hedrick, K.: Analysis of bird formations. In: Proceedings of
the Conference on Decision and Control. vol. 1, pp. 118–123 vol.1. IEEE (2002)
31. Stulp, F., Sigaud, O.: Path integral policy improvement with covariance matrix
adaptation. arXiv preprint arXiv:1206.4621 (2012), http://arxiv.org/abs/1206.
4621
32. Stulp, F., Sigaud, O.: Policy improvement methods: Between black-box optimization and episodic reinforcement learning (2012), http://hal.upmc.fr/
hal-00738463/
33. Verfaillie, G., Pralet, C., Teichteil, F., Infantes, G., Lesire, C.: Synthesis of plans
or policies for controlling dynamic systems. AerospaceLab (4), p. 1–12 (2012)
34. Weimerskirch, H., Martin, J., Clerquin, Y., Alexandre, P., Jiraskova, S.: Energy
Saving in Flight Formation. Nature 413(6857), 697–698 (2001)
35. Yang, J., Grosu, R., Smolka, S.A., Tiwari, A.: Love Thy Neighbor: V-Formation
as a Problem of Model Predictive Control. In: LIPIcs-Leibniz International Proceedings in Informatics. vol. 59. Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik
(2016)
36. Yang, J., Grosu, R., Smolka, S.A., Tiwari, A.: V-Formation as Optimal Control.
In: Proceedings of the Biological Distributed Algorithms Workshop 2016 (2016)
37. Zhou, Y., Tan, Y.: GPU-based Parallel Particle Swarm Optimization. In: Proceedings of the Congress on Evolutionary Computation. pp. 1493–1500. IEEE (2009)
| 3cs.SY
|
Make Up Your Mind:
The Price of Online Queries in Differential Privacy
arXiv:1604.04618v1 [cs.CR] 15 Apr 2016
Mark Bun∗
Thomas Steinke†
Jonathan Ullman‡
April 18, 2016
Abstract
We consider the problem of answering queries about a sensitive dataset subject to
differential privacy. The queries may be chosen adversarially from a larger set Q of allowable
queries in one of three ways, which we list in order from easiest to hardest to answer:
• Offline: The queries are chosen all at once and the differentially private mechanism
answers the queries in a single batch.
• Online: The queries are chosen all at once, but the mechanism only receives the queries
in a streaming fashion and must answer each query before seeing the next query.
• Adaptive: The queries are chosen one at a time and the mechanism must answer each
query before the next query is chosen. In particular, each query may depend on the
answers given to previous queries.
Many differentially private mechanisms are just as efficient in the adaptive model as they
are in the offline model. Meanwhile, most lower bounds for differential privacy hold in the
offline setting. This suggests that the three models may be equivalent.
We prove that these models are all, in fact, distinct. Specifically, we show that there is a
family of statistical queries such that exponentially more queries from this family can be
answered in the offline model than in the online model. We also exhibit a family of search
queries such that exponentially more queries from this family can be answered in the online
model than in the adaptive model. We also investigate whether such separations might hold
for simple queries like threshold queries over the real line.
∗ Harvard University John A. Paulson School of Engineering and Applied Sciences. Supported by an NDSEG
Fellowship and NSF grant CNS-1237235. Part of this work was done while the author was visiting Yale University.
[email protected]
† Harvard University John A. Paulson School of Engineering and Applied Sciences. Supported by NSF grants
CCF-1116616, CCF-1420938, and CNS-1237235. [email protected]
‡ Northeastern University College of Computer and Information Science. [email protected]
Contents
1 Introduction
1.1 Our Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
2
3
2 Preliminaries
2.1 Datasets and Differential Privacy . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Models of Interactive Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
5
6
6
3 A Separation Between Offline and Online Queries
3.1 Answering Offline Prefix Queries . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 A Lower Bound for Online Prefix Queries . . . . . . . . . . . . . . . . . . . . . .
8
9
10
4 A Separation Between Adaptive and Non-Adaptive Online Queries
4.1 Answering Online Correlated Vector Queries . . . . . . . . . . . . . . . . . . . .
4.2 A Lower Bound for Adaptive Correlated Vector Queries . . . . . . . . . . . . . .
14
14
16
5 Threshold Queries
5.1 Separation for Pure Differential Privacy . . . . . . . . . . . . . . . . .
5.2 The BetweenThresholds Algorithm . . . . . . . . . . . . . . . . . . . . .
5.3 The Online Interior Point Problem . . . . . . . . . . . . . . . . . . . .
5.4 Releasing Adaptive Thresholds with Approximate Differential Privacy
18
18
20
26
27
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Acknowledgements
30
References
30
A The Fingerprinting Lemma
32
1
Introduction
Differential privacy [DMNS06] is a formal guarantee that an algorithm run on a sensitive dataset
does not reveal too much about any individual in that dataset. Since its introduction, a rich
literature has developed to determine what statistics can be computed accurately subject to
differential privacy. For example, suppose we wish to approximate a real-valued query q(x) on
some dataset x that consists of the private data of many individuals. Then, this question has a
clean answer—we can compute a differentially private estimate of q(x) with error proportional
to the global sensitivity of q, and we cannot have smaller error in the worst case.
But how much error do we need to answer a large set of queries q1 , . . . , qk ? Before we can
answer this question, we have to define a model of how the queries are asked and answered.
The literature on differential privacy has considered three different interactive models1 for
specifying the queries:
• The Offline Model: The sequence of queries q1 , . . . , qk are given to the algorithm together in
a batch and the mechanism answers them together.
• The Online Model: The sequence of queries q1 , . . . , qk is chosen in advance and then the
mechanism must answer each query qj before seeing qj+1 .
• The Adaptive Model: The queries are not fixed in advance, each query qj+1 may depend on
the answers to queries q1 , . . . , qj .
In all three cases, we assume that q1 , · · · , qk are chosen from some family of allowable queries Q,
but may be chosen adversarially from this family.
Differential privacy seems well-suited to the adaptive model. Arguably its signature property
is that any adaptively-chosen sequence of differentially private algorithms remains collectively
differentially private, with a graceful degradation of the privacy parameters [DMNS06, DRV10].
As a consequence, there is a simple differentially private algorithm that takes a dataset
√ of
n individuals and answers Ω̃(n) statistical queries in the adaptive model with error o(1/ n),
simply by perturbing each answer independently with carefully calibrated noise. In contrast,
the seminal lower bound of Dinur and Nissim and its later refinements [DN03, DY08] shows
that there exists a fixed set of O(n) queries that cannot be answered by any differentially private
algorithm with such little error, even in the easiest offline model. For an even more surprising
example, the private multiplicative weights algorithm of Hardt and Rothblum [HR10] can in
many cases answer an exponential number of arbitrary, adaptively-chosen statistical queries
with a strong accuracy guarantee, whereas [BUV14] show that the accuracy guarantee of private
multiplicative weights is nearly optimal even for a simple, fixed family of queries.
These examples might give the impression that answering adaptively-chosen queries comes
“for free” in differential privacy—that everything that can be achieved in the offline model
can be matched in the adaptive model. Beyond just the lack of any separation between the
models, many of the most powerful differentially private algorithms in all of these models use
techniques from no-regret learning, which are explicitly designed for adaptive models.
Another motivation for studying the relationship between these models is the recent line
of work connecting differential privacy to statistical validity for adaptive data analysis [HU14,
1 Usually, the “interactive model” refers only to what we call the “adaptive model.” We prefer to call all of these
models interactive, since they each require an interaction with a data analyst who issues the queries. We use the
term “interactive” to distinguish these models from one where the algorithm only answers a fixed set of queries.
1
DFH+ 15, SU15b, BNS+ 16], which shows that differentially private algorithms for adaptivelychosen queries in fact yield state-of-the-art algorithms for statistical problems unrelated to
privacy. This connection further motivates studying the adaptive model and its relationship to
the other models in differential privacy.
In this work, we show for the first time that these three models are actually distinct. In
fact, we show exponential separations between each of the three models. These are the first
separations between these models in differential privacy.
1.1
Our Results
Given a dataset x whose elements come from a data universe X, a statistical query on X is defined
by a predicate φ on X and asks “what fraction of elements in the dataset satisfy φ?” The
answer to a statistical query lies in [0, 1] and our goal is to answer these queries up to some
small additive error ±α, for a suitable choice of 0 < α < 1. If the mechanism is required to
answer arbitrary statistical queries, then the offline, online, and adaptive models are essentially
equivalent — the upper bounds in the adaptive model match the lower bounds in the offline
model [DRV10, HR10, BUV14, SU15a]. However, we show that when the predicate φ is required
to take a specific form, then it becomes strictly easier to answer a set of these queries in the
offline model than it is to answer a sequence of queries presented online.
Theorem 1.1 (Informal). There exists a data universe X and a family of statistical queries Q on X
such that for every n ∈ N,
1. there is√a differentially private algorithm that takes a dataset x ∈ X n and answers any set of
k = 2Ω( n) offline queries from Q up to error ±1/100 from Q, but
2. no differentially private algorithm can take a dataset x ∈ X n and answer an arbitrary sequence
of k = O(n2 ) online (but not adaptively-chosen) queries from Q up to error ±1/100.
This result establishes that the online model is strictly harder than the offline model. We
also demonstrate that the adaptive model is strictly harder than the online model. Here, the
family of queries we use in our separation is not a family of statistical queries, but is rather a
family of search queries with a specific definition of accuracy that we will define later.
Theorem 1.2 (Informal). For every n ∈ N, there is a family of “search” queries Q on datasets in X n
such that
1. there is a differentially private algorithm that takes a dataset x ∈ {±1}n and accurately answers
any online (but not adaptively-chosen) sequence of k = 2Ω(n) queries from Q, but
2. no differentially private algorithm can take a dataset x ∈ {±1}n and accurately answer an
adaptively-chosen sequence of k = O(1) queries from Q.
We leave it as an interesting open question to separate the online and adaptive models for
statistical queries, or to show that the models are equivalent for statistical queries.
Although Theorems 1.1 and 1.2 separate the three models, these results use somewhat
contrived families of queries. Thus, we also investigate whether the models are distinct for
natural families of queries that are of use in practical applications. One very well studied class
of queries is threshold queries. These are a family of statistical queries Qthresh defined on the
universe [0, 1] and each query is specified by a point τ ∈ [0, 1] and asks “what fraction of the
2
elements of the dataset are at most τ?” If we restrict our attention to so-called pure differential
privacy (i.e. (ε, δ)-differential privacy with δ = 0), then we obtain an exponential separation
between the offline and online models for answering threshold queries.
Theorem 1.3 (Informal). For every n ∈ N,
1. there is a pure differentially private algorithm that takes a dataset x ∈ [0, 1]n and answers any
set of k = 2Ω(n) offline queries from Qthresh up to error ±1/100, but
2. no pure differentially private algorithm takes a dataset x ∈ [0, 1]n and answers an arbitrary
sequence of k = O(n) online (but not adaptively-chosen) queries from Qthresh up to error ±1/100.
We also ask whether or not such a separation exists for arbitrary differentially private algorithms (i.e. (ε, δ)-differential privacy with δ > 0). Theorem 1.3 shows that, for pure differential
privacy, threshold queries have near-maximal sample complexity. That is, up to constants, the
lower bound for online threshold queries matches what is achieved by the Laplace mechanism,
which is applicable to arbitrary statistical queries. This may lead one to conjecture that adaptive threshold queries also require near-maximal sample complexity subject to approximate
differential privacy. However, we show that this is not the case:
Theorem 1.4. For every n ∈ N, there is a differentially private algorithm that takes a dataset x ∈ [0, 1]n
and answers any set of k = 2Ω(n) adaptively-chosen queries from Qthresh up to error ±1/100.
In contrast, for any offline set of k thresholds τ1 , . . . , τk , we can round each element of the
dataset up to an element in the finite universe X = {τ1 , . . . , τk , 1} without changing the answers to
any of the queries. Then we can use known algorithms for answering all threshold queries over
any finite, totally ordered domain [BNS13, BNSV15] to answer the queries using a very small
∗
dataset of size n = 2O(log (k)) . We leave it as an interesting open question to settle the complexity
of answering adaptively-chosen threshold queries in the adaptive model.
1.2
Techniques
Separating Offline and Online Queries
To prove Theorem 1.1, we construct a sequence of queries q1 , · · · , qk such that, for all j ∈ [k],
• qj “reveals” the answers to q1 , · · · , qj−1 , but
• q1 , · · · , qj−1 do not reveal the answer to qj .
Thus, given the sequence q1 , · · · , qk in the offline setting, the answers to q1 , · · · , qk−1 are revealed
by qk . So only qk needs to be answered and the remaining query answers can be inferred.
However, in the online setting, each query qj−1 must be answered before qj is presented and this
approach does not work. This is the intuition for our separation.
To prove the online lower bound, we build on a lower bound for marginal queries [BUV14],
which is based on the existence of short secure fingerprinting codes [BS98, Tar08]. Consider the
data universe {±1}k . Given a dataset x ∈ {±1}n×k , a marginal query is a specific type of statistical
query that asks for the mean of a given column of x. Bun et al. [BUV14] showed that unless
k n2 , there is no differentially private algorithm that answers all k marginal queries with
non-trivial accuracy. This was done by showing that such an algorithm would violate the
security of a short fingerprinting code due to Tardos [Tar08]. We are able to “embed” k marginal
3
queries into the sequence of online queries q1 , · · · , qk . Thus a modification of the lower bound
for marginal queries applies in the online setting.
To prove the offline upper bound, we use the fact that every query reveals information
about other queries. However, we must handle arbitrary sequences of queries, not just the
specially-constructed sequences used for the lower bound. The key property of our family of
queries is the following. Each element x of the data universe X requires k bits to specify. On the
other hand, for any set of queries q1 , · · · , qk , we can specify q1 (x), · · · , qk (x) using only O(log(nk))
bits. Thus the effective size of the data universe given the queries is poly(nk), rather than 2k .
Then we can apply a differentially private algorithm that gives good accuracy as long as the
data universe has subexponential size [BLR13]. Reducing the size of the data universe is only
possible once the queries have been specified; hence this approach only works in the offline
setting.
Separating Online and Adaptive Queries
To prove Theorem 1.2, we start with the classical randomized response algorithm [War65].
Specifically, given a dataset x ∈ {±1}n , randomized response produces a new dataset y ∈ {±1}n
where each coordinate yi is independently set to +xi with probability (1 + α)/2 and is set to −xi
with probability (1 − α)/2. It is easy to prove that this algorithm is (O(α), 0)-differentially private.
What accuracy guarantee does this algorithm satisfy? By design, it outputs a vector y that has
correlation approximately α with the dataset x — that is, hy, xi ≈ αn. On the other hand, it is
also easy to prove that there is no differentially private algorithm (for any reasonable privacy
parameters) that can output a vector that has correlation at least 1/2 with the sensitive dataset.
Our separation between the online and adaptive models is based on the observation that,
if we can obtain O(1/α 2 ) “independent” vectors y1 , . . . , yk that are each roughly α-correlated
with x, then we can obtain a vector z that is (1/2)-correlated with x, simply by letting z be the
coordinate-wise majority of the yj s. Thus, no differentially private algorithm can output such a
set of vectors. More precisely, we require that hyi , yj i ≈ α 2 n for i , j, which is achieved if each yj
is an independent sample from randomized response.
Based on this observation, we devise a class of queries such that, if we are allowed to choose
k of these queries adaptively, then we obtain a set of vectors y1 , . . . , yk satisfying the conditions
above. This rules out differential privacy for k = O(1/α 2 ) adaptive queries. The key is that
we can use adaptivity to ensure that each query asks for an “independent” yj by adding the
previous answers y1 , · · · , yj−1 as constraints in the search query.
On the other hand, randomized response can answer each such query with high probability.
If a number of these queries is fixed in advance, then, by a union bound, the vector y output by
randomized response is simultaneously an accurate answer to every query with high probability.
Since randomized response is oblivious to the queries, we can also answer the queries in the
online model, as long as they are not chosen adaptively.
At a high level, the queries that achieve this property are of the form “output a vector
y ∈ {±1}n that is approximately α-correlated with x and is approximately as uncorrelated as
possible with the vectors v1 , . . . , vm .” A standard concentration argument shows that randomized
response gives an accurate answer to all the queries simulatneously with high probability. On
the other hand, if we are allowed to choose the queries adaptively, then for each query qi , we can
ask for a vector yi that is correlated with x but is as uncorrelated as possible with the previous
answers y1 , . . . , yi−1 .
4
Threshold Queries
For pure differential privacy, our separation between offline and online threshold queries uses
a simple argument based on binary search. Our starting point is a lower bound showing
that any purely differentially private algorithm that takes a dataset of n points x1 , . . . , xn ∈
{1, . . . , T } and outputs an approximate median of these points requires n = Ω(log(T )). This
lower bound follows from a standard application of the “packing” technique of Hardt and
Talwar [HT10]. On the other hand, by using binary search, any algorithm that can answer
k = O(log(T )) adaptively-chosen threshold queries can be used to find an approximate median.
Thus, any purely differentially private algorithm for answering such queries requires a dataset
of size n = Ω(k). Using the structure of the lower bound argument, we show that the same
lower bound holds for online non-adaptive queries as well. In contrast, using the algorithms
of [DNPR10, CSS11, DNRR15], we can answer k offline threshold queries on a dataset with only
n = O(log(k)) elements, giving an exponential separation.
The basis of our improved algorithm for adaptive threshold queries under approximate
differential privacy is a generalization of the sparse vector technique [DNPR10, RR10, HR10]
(see [DR14, §3.6] for a textbook treatment). Our algorithm makes crucial use of a stability
argument similar to the propose-test-release techniques of Dwork and Lei [DL09]. To our
knowledge, this is the first use of a stabiltiy argument for any online or adaptive problem in
differential privacy and may be of independent interest. In particular, our algorithm is given an
input x ∈ X n , a threshold t ∈ (0, 1), and an adaptive sequence of statistical (or low-sensitivity)
queries q1 , · · · , qk : X n → [0, 1] and, for each query qj , it reports (i) qj (x) ≥ t, (ii) qj (x) ≤ t, or (iii)
√
t − α ≤ qj (x) ≤ t + α. The sample complexity of this algorithm is n = O( c log(k/εδ)/εα), where k
is the total number of queries, c is an upper bound on the number of times (iii) may be reported,
and (ε, δ)-differential privacy is provided. We call this the Between Thresholds algorithm.
Once we have this algorithm, we can use it to answer adaptively-chosen thresholds using
an approach inspired by Bun et al. [BNSV15]. The high-level ideal is to sort the dataset
x(1) < x(2) < · · · < x(n) and then partition it into chunks of consecutive sorted elements. For
any chunk, and a threshold τ, we can use the between thresholds algorithm to determine
(approximately) whether τ lies below all elements in the chunk, above all elements in the chunk,
or inside the chunk. Obtaining this information for every chunk is enough to accurately estimate
the answer to the threshold query τ up to an error proportional to the size of the chunks. The
sample complexity is dominated by the O(log k) sample complexity of our Between Thresholds
algorithm multiplied by the number of chunks needed, namely O(1/α).
2
2.1
Preliminaries
Datasets and Differential Privacy
A dataset x ∈ (x1 , . . . , xn ) ∈ X n is an ordered tuple of n elements from some data universe X. We
say that two datasets x, x0 are adjacent if they differ on only a single element and denote this
relation by x ∼ x0 .
Definition 2.1 (Differential Privacy [DMNS06]). A randomized algorithm M : X n → R is (ε, δ)differentially private if for every two adjacent datasets x ∼ x0 , and every R ⊆ R,
P [M(x) ∈ R] ≤ eε P M(x0 ) ∈ R + δ.
5
We also use the following well known group privacy property of (ε, 0)-differential privacy.
We say that two datasets x, x0 are c-adjacent if the differ on at most c-elements, and denote this
relation by x ∼c x0 .
Lemma 2.2 ([DMNS06]). If M : X n → R is (ε, 0)-differentially private, then for every c ∈ N and
every two c-adjacent datasets x ∼c x0 , and every R ⊆ R,
P [M(x) ∈ R] ≤ ecε P M(x0 ) ∈ R .
2.2
Queries
In this work we consider two general classes of queries on the dataset: statistical queries, and
search queries. Although statistical queries are a very special case of search queries, we will
present each of them independently to avoid having to use overly abstract notation to describe
statistical queries.
Statistical Queries. A statistical query on a data universe X is defined by a Boolean predicate
q : X → {0, 1}. Abusing notation, we define the evaluation of a statistical query q on a dataset
x = (x1 , . . . , xn ) to be the average of the predicate over the rows
n
1X
q(x) =
q(xi ) ∈ [0, 1].
n
i=1
For a dataset x, a statistical query q, and an answer a ∈ [0, 1], the answer is α-accurate for q on x if
|q(x) − a| ≤ α.
Search Queries. A search query q on X n is defined by a loss function Lq : X n × R → [0, ∞), where
R is an arbitrary set representing the range of possible outputs. For a dataset x ∈ X n and an
output y ∈ R, we will say that y is α-accurate for q on x if Lq (x, y) ≤ α. In some cases the value of
Lq will always be either 0 or 1. Thus we simply say that y is accurate for q on x if Lq (x, y) = 0. For
example, if X n = {±1}n , we can define a search query by R = {±1}n , and Lq (x, y) = 0 if hx, yi ≥ αn
and Lq (x, y) = 1 otherwise. In this case, the search query would ask for any vector y that has
correlation α with the dataset.
To see that statistical queries are a special case of search queries, given a statistical query
q on X n , we can define a search query Lq with R = [0, 1] and Lq (x, a) = |q(x) − a|. Then both
definitions of α-accurate align.
2.3
Models of Interactive Queries
The goal of this work is to understand the implications of different ways to allow an adversary
to query a sensitive dataset. In each of these models there is an algorithm M that holds a dataset
x ∈ X n , and a fixed family of (statistical or search) queries Q on X n , and a bound k on the number
of queries that M has to answer. There is also an adversary A that chooses the queries. The
models differ in how the queries chosen by A are given to M.
6
Offline
In the offline model, the queries q1 , . . . , qk ∈ Q are specified by the adversary A in advance and
the algorithm M is given all the queries at once and must provide answers. Formally, we define
the following function OfflineA →
: X n → Qk × Rk depending A and M.
←M
Input: x ∈ X n .
A chooses q1 , · · · , qk ∈ Q.
M is given x and q1 , · · · , qk and outputs a1 , · · · , ak ∈ R.
Output: (q1 , · · · , qk , a1 , · · · , ak ) ∈ Qk × Rk .
Figure 1: OfflineA →
: X n → Qk × Rk
←M
Online Non-Adaptive
In the online non-adaptive model, the queries q1 , . . . , qk ∈ Q are again fixed in advance by the
adversary, but are then given to the algorithm one at a time, and the algorithm must give an
answer to query qj before it is shown qj+1 . We define a function OnlineA →
: X n → Qk × Rk
←M
depending on the adversary A and the algorithm M as follows.
Input: x ∈ X n .
A chooses q1 , · · · , qk ∈ Q.
M is given x.
For j = 1, . . . , k:
M is given qj and outputs aj ∈ R.2
Output: (q1 , · · · , qk , a1 , · · · , ak ) ∈ Qk × Rk .
Figure 2: OnlineA →
: X n → Qk × Rk
←M
Online Adaptive
In the online adaptive model, the queries q1 , . . . , qk ∈ Q are not fixed, and the adversary may
choose each qj based on the answers that the algorithm gave to the previous queries. We define
a function AdaptiveA →
: X n → Qk × Rk depending on the adversary A and the algorithm M as
←M
follows.
Definition 2.3 (Differential Privacy for Interactive Mechanisms). In each of the three cases
— Offline, Online Non-Adaptive, or Online Adaptive — we say that M is (ε, δ)-differentially
private if, for all adversaries A, respectively OfflineA →
, OnlineA →
, or AdaptiveA →
is (ε, δ)←M
←M
←M
differentially private.
Definition 2.4 (Accuracy for Interactive Mechanisms). In each case — Offline, Online NonAdaptive, or Online Adaptive queries — we say that M is (α, β)-accurate if, for all adversaries A
and all inputs x ∈ X n ,
"
#
P
max Lqj (x, aj ) ≤ α ≥ 1 − β,
q1 ,··· ,qk ,a1 ,··· ,ak j∈[k]
7
(1)
Input: x ∈ X n .
M is given x.
For j = 1, . . . , k:
A chooses a query qj ∈ Q.
M is given qj and outputs aj ∈ R.
Output: (q1 , · · · , qk , a1 , · · · , ak ) ∈ Qk × Rk .
Figure 3: AdaptiveA →
: X n → Qk × Rk
←M
where (q1 , · · · , qk , a1 , · · · , ak ) is respectively drawn from one of OfflineA →
(x), OnlineA →
(x), or
←M
←M
AdaptiveA →
(x). We also say that M is α-accurate if the above holds with (1) replaced by
←M
E
"
#
max Lqj (x, aj ) ≤ α.
q1 ,··· ,qk ,a1 ,··· ,ak j∈[k]
3
A Separation Between Offline and Online Queries
In this section we prove that online accuracy is strictly harder to achieve than offline accuracy,
even for statistical queries. We prove our results by constructing a set of statistical queries
that we call prefix queries for which it is possible to take a dataset of size n and accurately
answer superpolynomially many offline prefix queries in a differentially private manner, but
it is impossible to answer more than O(n2 ) online prefix queries while satisfying differential
privacy.
We now define the family of prefix queries. These queries are defined on the universe
S
j
3
∗
X = {±1}∗ = ∞
j=0 {±1} consisting of all finite length binary strings. For x, y ∈ {±1} , we use y x
to denote that y is a prefix of x. Formally
yx
⇐⇒
|y| ≤ |x| and ∀i = 1, . . . , |y| xi = yi .
Definition 3.1. For any finite set S ⊆ {±1}∗ of finite-length binary strings, we define the prefix
query qS : {±1}∗ → {±1} by
qS (x) = 1 ⇐⇒ ∃y ∈ S y x.
We also define
Qprefix = {qS | S ⊂ {±1}∗ }
B
Qprefix
= {qS | S ⊂ {±1}∗ , |S| ≤ B}
to be the set of all prefix queries and the set of prefix queries with sizes bounded by B, respectively.
3 All of the arguments in this section hold if we restrict to strings of length at most k + log n. However, we allow
strings of arbitrary length to reduce notational clutter.
8
3.1
Answering Offline Prefix Queries
We now prove that there is a differentially private algorithm that answers superpolynomially
many prefix queries, provided that the queries are specified offline.
Theorem 3.2 (Answering Offline Prefix Queries). For every α, ε ∈ (0, 1/10), every B ∈ N, and every
n ∈ N, there exists a
√
3
3
k = min 2Ω( α εn) , 2Ω(α εn/ log(B))
B
and an (ε, 0)-differentially private algorithm Mprefix : X n × (Qprefix
)k → Rk that is (α, 1/100)-accurate
B
for k offline queries from Qprefix
.
We remark that it is possible to answer even more offline prefix queries by relaxing to
(ε, δ)-differential privacy for some negligibly small δ > 0. However, we chose to state the results
for (ε, 0)-differential privacy to emphasize the contrast with the lower bound, which applies
even when δ > 0, and to simplify the statement.
Our algorithm for answering offline queries relies on the existence of a good differentially private algorithm for answering arbitrary offline statistical queries. For concreteness, the so-called
“BLR mechanism” of Blum, Ligett, and Roth [BLR13] suffices, although different parameter
tradeoffs can be obtained using different mechanisms. Differentially private algorithms with
this type of guarantee exist only when the data universe is bounded, which is not the case for
prefix queries. However, as we show, when the queries are specified offline, we can replace the
infinite universe X = {±1}∗ with a finite, restricted universe X 0 and run the BLR mechanism.
Looking ahead, the key to our separation will be the fact that this universe restriction is only
possible in the offline setting. Before we proceed with the proof of Theorem 3.2, we will state
the guarantees of the BLR mechanism.
Theorem 3.3 ([BLR13]). For every 0 < α, ε ≤ 1/10 and every finite data universe X, if QSQ is the set
of all statistical queries on X, then for every n ∈ N, there is a
k = 2Ω(α
3
εn/ log |X|)
and an (ε, 0)-differentially private algorithm MBLR : X n × QkSQ → Rk that is (α, 1/100)-accurate for k
offline queries from QSQ .
We are now ready to prove Theorem 3.2.
B
Proof of Theorem 3.2. Suppose we are given a set of queries qS1 , . . . , qSk ∈ Qprefix
and a dataset
S
k
n
∗
x ∈ X where X = {±1} . Let S = j=1 Sj . We define the universe XS = S ∪ {∅} where ∅ denotes
the empty string of length 0. Note that this universe depends on the choice of queries, and that
|XS | ≤ kB + 1. Since XS ⊂ X, it will be well defined to restrict the domain of each query qSj to
elements of XS .
Next, given a dataset x = (x1 , . . . , xn ) ∈ X n , and a collection of sets S1 , . . . , Sk ⊂ X, we give
a procedure for mapping each element of x to an element of XS to obtain a new dataset
xS = (x1S , . . . , xnS ) ∈ XSn that is equivalent to x with respect to the queries qS1 , . . . , qSk . Specifically,
define rS : X → XS by
rS (x) = arg max |y|.
y∈XS ,yx
That is, rS (x) is the longest string in XS that is a prefix of x. We summarize the key property of
rS in the following claim
9
Claim 3.4. For every x ∈ X, and j = 1, . . . , k, qSj (rS (x)) = qSj (x).
Proof of Claim 3.4. First, we state a simple but important fact about prefixes: If y, y 0 are both
prefixes of a string x with |y| ≤ |y 0 |, then y is a prefix of y 0 . Formally,
∀x, y, y 0 ∈ {0, 1}∗
(y x ∧ y 0 x ∧ |y| ≤ |y 0 |) =⇒ y y 0 .
(2)
Now, fix any x ∈ X and any query qSj and suppose that qSj (x) = 1. Then there exists a string
y ∈ Sj such that y x. By construction, we have that rS (x) x and that |rS (x)| ≥ |y|. Thus, by (2),
we have that y rS (x). Thus, there exists y ∈ Sj such that y rS (x), which means qSj (rS (x)) = 1,
as required.
Next, suppose that qSj (rS (x)) = 1. Then, there exists y ∈ Sj such that y rS (x). By construction, rS (x) x, so by transitivity we have that y x. Therefore, qSj (x) = 1, as required.
xS
Given this lemma, we can replace every row xi of x with xiS = rS (xi ) to obtain a new dataset
such that for every j = 1, . . . , k,
n
n
i=1
i=1
1X
1X
qSj (x ) =
qSj (xiS ) =
qSj (xi ) = qSj (x).
n
n
S
Thus, we can answer qS1 , · · · , qSk on xS ∈ XSn , rather than on x ∈ X n . Note that each row of xS
depends only on the corresponding row of x. Hence, for every set of queries qS1 , . . . , qSk , if
x ∼ x0 are adjacent datasets, then xS ∼ x0S are also adjacent datasets. Consequently, applying
a (ε, δ)-differentially private algorithm to xS yields a (ε, δ)-differentially private algorithm as a
function of x.
In particular, we can give α-accurate answers to these queries using the algorithm MBLR as
long as
3
3
k ≤ 2Ω(α εn/ log |XS |) = 2Ω(α εn/ log(kB+1)) .
Rearranging terms gives the bound in Theorem 3.2. We specify the complete algorithm Mprefix
in Figure 4.
Mprefix (x; qS1 , . . . , qSk ):
S
Write x = (x1 , . . . , xn ) ∈ X n , S = kj=1 Sj , XS = S ∪ {∅}.
For i = 1, . . . , n, let xiS = rS (xi ) and let xS = (x1S , . . . , xnS ) ∈ XSn .
Let (a1 , . . . , ak ) = MBLR (xS ; qS1 , . . . , qSk ).
Output (a1 , . . . , ak ).
Figure 4: Mprefix
3.2
A Lower Bound for Online Prefix Queries
Next, we prove a lower bound for online queries. Our lower bound shows that the simple
approach of perturbing the answer to each query with independent noise is essentially optimal
for prefix queries. Since this approach is only able to answer k = O(n2 ) queries, we obtain
an exponential separation between online and offline statistical queries for a broad range of
parameters.
10
Theorem 3.5 (Lower Bound for Online Prefix Queries). There exists a function k = O(n2 ) such
that for every sufficiently large n ∈ N, there is no (1, 1/30n)-differentially private algorithm M that
n
takes a dataset x ∈ X n and is (1/100, 1/100)-accurate for k online queries from Qprefix
.
√
In this parameter regime, our algorithm from Section 3.1 answers k = exp(Ω̃( n)) offline
prefix queries, so we obtain an exponential separation.
Our lower bound relies on a connection between fingerprinting codes and differential privacy
[Ull13, BUV14, SU15a, DSS+ 15]. However, instead of using fingerprinting codes in a black-box
way, we will make a direct use of the main techniques. Specifically, we will rely heavily on the
following key lemma. The proof appears in Appendix A.
Lemma 3.6 (Fingerprinting Lemma). Let f : {±1}n → [−1, 1] be any function. Suppose p is sampled
from the uniform distribution over [−1, 1] and c ∈ {±1}n is a vector of n independent bits, where each
bit has expectation p. Letting c denote the coordinate-wise mean of c, we have
X
1
(ci − p) + 2 f (c) − c ≥ .
E f (c) ·
3
p,c
i∈[n]
Roughly the fingerprinting lemma says that if we sample a vector c ∈ {±1}n in a specific
fashion, then for any bounded function f (c), we either have that f (c) has “significant” correlation
with ci for some coordinate i, or that f (c) is “far” from c on average. In our lower bound, the
vector c will represent a column of the dataset, so each coordinate ci will correspond to the value
of some row of the dataset. The function f (c) will represent the answer to some prefix query. We
will use the accuracy of a mechanism for answering prefix queries to argue that f (c) is not far
from c, and therefore conclude that f (c) must be significantly correlated with some coordinate
ci . On the other hand, if ci were excluded from the dataset, then ci is sufficiently random that
the mechanism’s answers cannot be significantly correlated with ci . We will use this to derive a
contradiction to differential privacy.
Proof of Theorem 3.5. First we define the distribution on the input dataset x = (x1 , . . . , xn ) and
the queries qS1 , · · · , qSk .
Input dataset x:
• Sample p1 , · · · , pk ∈ [−1, 1] independently and uniformly at random.
• Sample c1 , · · · , ck ∈ {±1}n independently, where each cj is a vector of n independent bits,
each with expectation pj .
• For i ∈ [n], define
xi = (binary(i), ci1 , · · · , cik ) ∈ {±1}dlog2 ne+k ,
where binary(i) ∈ {±1}dlog2 ne is the binaryrepresentation
n of i where 1 is mapped to +1 and
4
dlog
ne+k
0 is mapped to −1. Let x = (x1 , . . . , xn ) ∈ {±1} 2
.
4 This choice is arbitrary, and is immaterial to our lower bound. The only property we need is that binary(i)
uniquely identifies i and, for notational consistency, we require binary(i) to be a string over the alphabet {±1}.
11
Queries qS1 , · · · , qSk :
• For i ∈ [n] and j ∈ [k], define
j−1
zi,j = (binary(i), ci1 , · · · , ci
, 1) ∈ {±1}dlog2 ne+j .
n
o
n
• For j ∈ [k], define qSj ∈ Qprefix
by Sj = zi,j | i ∈ [n] .
These queries are designed so that the correct answer to each query j ∈ [k] is given by
qSj (x) = cj :
Claim 3.7. For every j ∈ [k], if the dataset x and the queries qS1 , . . . , qSk are constructed as above, then
with probability 1,
n
n
1X
1X j
qSj (x) =
qSj (xi ) =
ci = cj
n
n
i=1
i=1
Proof of Claim 3.7. We have
qSj (xi ) = 1 ⇐⇒ ∃w ∈ Sj (w xi ) ⇐⇒ ∃` ∈ [n] (z`,j xi ).
j
j
By construction, we have z`,j xi if and only if ` = i and xi = ci = 1, as required. Here, we have
used the fact that the strings binary(i) are unique to ensure that z`,j xi if and only if ` = i.
We now show no differentially private algorithm M is capable of giving accurate answers to
n
these queries. Let M be an algorithm that answers k online queries from Qprefix
. Suppose we
generate an input dataset x and queries qS1 , . . . , qSk as above, and run M(x) on this sequence of
queries. Let a1 , . . . , ak ∈ [−1, 1] denote the answers given by M.
First, we claim that, if M(x) is accurate for the given queries, then each answer aj is close to
P
j
the corresponding value cj = n1 ni=1 ci .
n
Claim 3.8. If M is (1/100, 1/100)-accurate for k online queries from Qprefix
, then with probability 1
over the choice of x and qS1 , . . . , qSk above,
X
E
aj − c j
M
j∈[k]
k
.
≤
10
Proof of Claim 3.8. By Claim 3.7, for every j ∈ [k], qSj (x) = cj . Since, by assumption, M is
n
(1/100, 1/100)-accurate for k online queries from Qprefix
, we have that with probability at least
99/100,
1
1
∀j ∈ [k] aj − qSj (x) ≤
=⇒ ∀j ∈ [k] aj − cj ≤
100
100
By linearity of expectation, this case contributes at most k/100 to the expectation. On the other
hand, |aj − qSj (x)| ≤ 2, so by linearity of expectation the case where M is inaccurate contributes
at most 2k/100 to the expectation. This suffices to prove the claim.
The next claim shows how the fingerprinting lemma (Lemma 3.6) can be applied to M.
12
Claim 3.9.
X X j
j
E
(ci − pj ) + 2 aj − cj
a
p,x,q,M
j∈[k]
i∈[n]
k
≥ .
3
Proof. By linearity of expectation, it suffices to show that, for every j ∈ [k],
1
X j
E aj
(ci − pj ) + 2 aj − cj ≥ .
3
p,x,q,M
i∈[n]
Since each column cj is generated independently from the columns c1 , . . . , cj−1 , cj and pj are
independent from qS1 , · · · , qSj . Thus, at the time M produces the output aj , it does not have any
information about cj or pj apart from its private input. (Although M later learns cj when it is
asked qSj+1 .) For any fixed values of c1 , . . . , cj−1 and the internal randomness of M, the answer aj
is a deterministic function of cj . Thus we can apply Lemma 3.6 to this function to establish the
claim.
Combining Claims 3.8 and 3.9 gives
2k
X X j
aj
E
(ci − pj ) ≥
.
15
p,x,q,M
j∈[k]
i∈[n]
In particular, there exists some i ∗ ∈ [n] such that
X
2k
j
aj (ci ∗ − pj ) ≥
E
.
15n
p,x,q,M
(3)
j∈[k]
To complete
the proof, we show that (3) violates the differential privacy guarantee unless
√
n ≥ Ω( k).
k
1
1
k
h To
i this end, fix any p , . . . , p ∈ [−1, 1], whence ci ∗ , · · · , ci ∗ ∈h {±1}
i are independent bits with
j
j
1
k
j
E c = p . Let c̃ , · · · , c̃ ∈ {±1} be independent bits with E c̃ = pj . The random variables
ci1∗ , · · · , cik∗ have the same marginal distribution as c̃1 , · · · , c̃k . However, c̃1 , · · · , c̃k are independent
from a1 , · · · , ak , whereas a1 , · · · , ak depend on ci1∗ , · · · , cik∗ . Consider the quantities
X
X
j
Z=
aj (ci ∗ − pj )
and
Z̃ =
aj (c̃j − pj ).
j∈[k]
j∈[k]
Differential privacy implies that Z and Z̃ have similar distributions. Specifically, if M is
(1, 1/30n)-differentially private, then
Z
2k
E [|Z|] =
Z
2k
P [|Z| > z] dz ≤
0
0
h
i
h i
1
k
eP |Z̃| > z +
dz = eE |Z̃| +
,
30n
15n
as |Z|, |Z̃| ≤ 2k with probability 1.
13
Now E [|Z|] ≥ E [Z] ≥ 2k/15n, by (3). On the other hand, aj is independent from c̃j and
h
i
h i
E c̃j − pj = 0, so E Z̃ = 0. We now observe that
h i2
h i
h i X
h
i X h
i
E |Z̃| ≤ E Z̃ 2 = Var Z̃ =
Var aj (c̃j − pj ) ≤
E (c̃j − pj )2 ≤ k.
j∈[k]
Thus, we have
j∈[k]
h i
√
k
2k
k
≤ E [|Z|] ≤ eE |Z̃| +
≤e k+
.
15n
15n
15n
√
The condition 2k/15n ≤ e k+k/15n is a contradiction unless k ≤ 225e2 n2 . Thus, we can conclude
that there exists a k = O(n2 ) such that no (1, 1/30n)-differentially private algorithm is accurate
n
for more than k online queries from Qprefix
, as desired. This completes the proof.
4
A Separation Between Adaptive and Non-Adaptive Online Queries
In this section we prove that even among online queries, answering adaptively-chosen queries
can be strictly harder than answering non-adaptively-chosen queries. Our separation applies to
a family of search queries that we call correlated vector queries. We show that for a certain regime
of parameters, it is possible to take a dataset of size n and privately answer an exponential
number of fixed correlated vector queries, even if the queries are presented online, but it is
impossible to answer more than a constant number of adaptively-chosen correlated vector
queries under differential privacy.
The queries are defined on datasets x ∈ {±1}n (hence the data universe is X = {±1}). For
every query, the range R = {±1}n is the set of n-bit vectors.n We fix some
o parameters 0 < α < 1
and m ∈ N. A query q is specified by a set V where V = v 1 , . . . , v m ⊆ {±1}n is a set of n-bit
vectors. Roughly, an accurate answer to a given search query is any vector y ∈ {±1}n that is
approximately α-correlated with the input dataset x ∈ {±1}n and has nearly as little correlation
as possible with every v j . By “as little correlation as possible with v j ” we mean that v j may itself
be correlated with x, in which case y should be correlated with v j only insofar as this correlation
comes through the correlation between y and x. Formally, for a query qV , we define the loss
function LqV : X n × X n → {0, 1} by
LqV (x, y) = 0
⇐⇒
hy − αx, xi ≤
α2n
α2n
∧ ∀v j ∈ V hy − αx, v j i ≤
.
100
100
We remark√that the choice of α 2 n/100 is somewhat arbitrary, and we can replace this choice with
C for any n C n and obtain quantitatively different results. We chose to fix this particular
choice in order to reduce notational clutter. We let
n,α,m
Qcorr
= {qV | V ⊆ {±1}n , |V | ≤ m}
be the set of all correlated vector queries on {±1}n for parameters α, m.
4.1
Answering Online Correlated Vector Queries
Provided that all the queries are fixed in advance, we can privately answer correlated vector
queries using the randomized response algorithm. This algorithm simply takes the input vector
14
x ∈ {±1}n and outputs a new vector y ∈ {±1}n where each bit yi is independent and is set to xi
with probability 1/2 + ρ for a suitable choice of ρ > 0. The algorithm will then answer every
correlated vector query with this same vector y. The following theorem captures the parameters
that this mechanism achieves.
Theorem 4.1 (Answering Online Correlated Vector Queries). For every 0 < α < 1/2, there exists
4
k = 2Ω(α n) such that, for every sufficiently large n ∈ N, there is a (3α, 0)-differentially private
n,α,k
algorithm Mcorr that takes a dataset x ∈ {±1}n and is (1/k)-accurate for k online queries from Qcorr
.
Proof Theorem 4.1. Our algorithm based on randomized response is presented in Figure 5 below.
Mcorr :
Input: a dataset x ∈ {±1}n .
Parameters: 0 < α < 1/2.
For i = 1, . . . , n, independently set
+xi
yi =
−xi
with probability
with probability
1+α
2
1−α
2
.
Let y = (y1 , . . . , yn ) ∈ {±1}n , and answer each query with y.
Figure 5: Mcorr
To establish privacy, observe that by construction each output bit yi depends only on xi and
is independent of all xj , yj for j , i. Therefore, it suffices to observe that if 0 < α < 1/2,
1≤
P[yi = +1 | xi = +1] 1 + α
=
≤ e3α
P[yi = +1 | xi = −1] 1 − α
and similarly
1≥
P[yi = −1 | xi = +1] 1 − α
=
≥ e−3α .
P[yi = −1 | xi = −1] 1 + α
To prove accuracy, observe that since the output y does not depend on the sequence of
n,α,k
queries, we can analyze the mechanism as if the queries qV1 , . . . , qVk ∈ Qcorr
were fixed and given
Sk
2
all at once. Let V = j=1 Vj , and note that |V | ≤ k . First, observe that E [y] = αx. Thus we have
E [hy − αx, xi] = 0
y
and
∀v ∈ V E [hy − αx, vi] = 0
y
Since x and every vector in V is fixed independently of y, and the coordinates of y are independent by construction, the quantities hy, xi and hy, vi are each the sum of n independent
{±1}-valued random variables. Thus, we can apply Hoeffding’s inequality5 and a union bound
5 We use the following statement of Hoeffding’s Inequality: if Z , . . . , Z are independent {±1}-valued random
1
n
P
variables, and Z = ni=1 Zi , then
√
2
P Z − E [Z] > C n ≤ 2e−C /2
15
to conclude
#
!
−α 4 n
α2n
≤ 2 exp
P |hy − αx, xi| >
y
100
20000
"
#
!
2
α n
−α 4 n
2
P ∃v ∈ V s.t. |hy − αx, vi| >
≤ 2k exp
y
100
20000
"
Ω(α
Thetheorem
now follows by setting an appropriate choice of k = 2
4
−α n
≤ 1/k.
exp 20000
4.2
4
n)
such that 2(k 2 + 1) ·
A Lower Bound for Adaptive Correlated Vector Queries
We now prove a contrasting lower bound showing that if the queries may be chosen adaptively,
then no differentially private algorithm can answer more than a constant number of correlated
vector queries. The key to our lower bound is that fact that adaptively-chosen correlated
vector queries allow an adversary to obtain many vectors y 1 , . . . , y k that are correlated with x
but pairwise nearly orthogonal with each other. As we prove, if k is sufficiently large, this
information is enough to recover a vector x̃ that has much larger correlation with x than any
of the vectors y 1 , . . . , y k have with x. By setting the parameters appropriately, we will obtain a
contradiction to differential privacy.
Theorem 4.2 (Lower Bound for Correlated Vector Queries). For every 0 < α < 1/2, there is a
k = O(1/α 2 ) such that for every sufficiently large n ∈ N, there is no (1, 1/20)-differentially private
n,α,k
algorithm that takes a dataset x ∈ {±1}n and is 1/100-accurate for k adaptive queries from Qcorr
We remark that the value of k in our lower bound is optimal up to constants, as there
is a (1, 1/20)-differentially private algorithm that can answer k = Ω(1/α 2 ) adaptively-chosen
queries of this sort. The algorithm simply answers each query with an independent invocation
of randomized response. Randomized response is O(α)-differentially private for each query,
and we can invoke the adaptive composition theorem [DMNS06, DRV10] to argue differential
privacy for k = Ω(1/α 2 )-queries.
Before proving Theorem 4.2, we state and prove the combinatorial lemma that forms the
foundation of our lower bound.
Lemma 4.3 (Reconstruction Lemma). Fix parameters 0 ≤ a, b ≤ 1. Let x ∈ {±1}n and y 1 , · · · , y k ∈
{±1}n be vectors such that
hy j , xi ≥ an
∀1 ≤ j ≤ k
∀1 ≤ j < j 0 ≤ k
Then, if we let x̃ = sign(
Pk
j=1 y
j ) ∈ {±1}n
0
|hy j , y j i| ≤ bn.
be the coordinate-wise majority of y 1 , . . . , y k , we have
!
2(b − a2 )
2
hx̃, xi ≥ 1 − 2 −
n.
a k
a2
Proof of Lemma 4.3. Let
k
1X j
y=
y ∈ [−1, 1]n .
k
j=1
16
By linearity, hy, xi ≥ an and
kyk22 =
k
1
1 X j j0
1
2
hy
,
y
i
≤
kn
+
(k
−
k)bn
≤
+
b
n.
k
k2 0
k2
j,j =1
Define a random variable W ∈ [−1, 1] to be xi y i for a uniformly random i ∈ [n]. Then
n
h
i 1X
1
1
1
E [W ] = hx, yi ≥ a and E W 2 =
xi2 y 2i = kyk22 ≤ + b
n
n
n
k
i=1
By Chebyshev’s inequality,
Var[W ] E[W 2 ] − E[W ]2
P [W ≤ 0] ≤ P |W − E [W ] | ≥ a ≤
=
≤
a2
a2
1
k
+ b − a2
a2
.
Meanwhile,
n
n
i=1
i=1
1X
1X
1 1
P [W ≤ 0] =
I[xi y i ≤ 0] ≥
I[sign(y i ) , xi ] = − hsign(y), xi.
n
n
2 2n
Thus we conclude
1
k + b − a2
hsign(y), xi ≥ n − 2nP [W ≤ 0] ≥ n − 2n
a2
P
To complete the proof, we rearrange terms and note that sign(y) = sign( kj=1 y j ).
Now we are ready to prove our lower bound for algorithms that answer adaptively-chosen
correlated vector queries.
Proof of Theorem 4.2. We will show that the output y 1 , . . . , y k of any algorithm M that takes a
dataset x ∈ {±1}n and answers k = 100/α 2 adaptively-chosen correlated vector queries can be
used to find a vector x̃ ∈ {±1}n such that hx̃, xi > n/2. In light of Lemma 4.3, this vector will
P
simply be x̃ = sign( kj=1 y j ). We will then invoke the following elementary fact that differentially
private algorithms do not admit this sort of reconstruction of their input dataset.
Fact 4.4. For every sufficiently large n ∈ N, there is no (1, 1/20)-differentially private algorithm
M : {±1}n → {±1}n such that for every x ∈ {±1}n , with probability at least 99/100, hM(x), xi > n/2.
n
o
The attack works as follows. For j = 1, . . . , k, define the set Vj = y 1 , . . . , y j−1 and ask the query
n,α,k
qVj (x) ∈ Qcorr
to obtain some vector y j . Since M is assumed to be accurate for k adaptively-
chosen queries, with probability 99/100, we obtain vectors y 1 , . . . , y k ∈ {±1}n such that
∀1 ≤ j ≤ k
hy j , xi ≥ hαx, xi − |hy − αx, xi|
≥ αn −
≥ an,
17
α2n
100
∀1 ≤ j < j 0 ≤ k
0
0
|hy j , y j i| ≤ |hαx, y j i| + |hy j − αx, y j i|
≤ α|hy j , xi| +
α2n
100
α2n
≤ α |hαx, xi| + |hy j − αx, xi| +
100
3
2
α n α n
≤ α2n +
+
100 100
51 2
≤
α n
50
= bn,
where a = 99α/100 and b = 51α 2 /50. Thus, by Lemma 4.3, if x̃ = sign(
we have
!
2(b − a2 )
2
hx̃, xi ≥ 1 − 2 −
n
a k
a2
Pk
j=1 y
j ),
and k = 100/α 2 ,
!
2(51α 2 /50 − (99α/100)2 )
2
−
n
= 1−
(99α/100)2 k
(99α/100)2
!!
(51/50) − (99/100)2
2(100/99)2
n
= 1−
−2
100
(99/100)2
≥ 0.89n ≥ n/2.
By Fact 4.4, this proves that M cannot be (1, 1/20)-differentially private.
5
Threshold Queries
First we define threshold queries, which are a family of statistical queries.
Definition 5.1. Let ThreshX denote the class of threshold queries over a totally ordered domain
X. That is, ThreshX = {cx : x ∈ X} where cx : X → {0, 1} is defined by cx (y) = 1 iff y ≤ x.
5.1
Separation for Pure Differential Privacy
In this section, we show that the sample complexity of answering adaptively-chosen thresholds
can be exponentially larger than that of answering thresholds offline.
Proposition 5.2 ([DNPR10, CSS11, DNRR15]). Let X be any totally ordered domain. Then there
exists a (ε, 0)-differentially private mechanism M that, given x ∈ X n , gives α-accurate answers to k
offline queries from ThreshX for
(
)!
log k + log2 (1/α) log2 k
n = O min
,
αε
αε
On the other hand, we show that answering k adaptively-chosen threshold queries can
require sample complexity as large as Ω(k) – an exponential gap. Note that this matches the
upper bound given by the Laplace mechanism [DMNS06].
18
Proposition 5.3. Answering k adaptively-chosen threshold queries on [2k−1 ] to accuracy α subject to
ε-differential privacy requires sample complexity n = Ω(k/αε).
The idea for the lower bound is that an analyst may adaptively choose k threshold queries to
binary search for an “approximate median” of the dataset. However, a packing argument shows
that locating an approximate median requires sample complexity Ω(k).
Definition 5.4 (Approximate Median). Let X be a totally ordered domain, α > 0, and x ∈ X n .
We call y ∈ X an α-approximate median of x if
1
1
{i ∈ [n] : xi ≤ y} ≥ − α
n
2
and
1
1
{i ∈ [n] : xi ≥ y} ≥ − α.
n
2
Proposition 5.3 is obtained by combining Lemmas 5.5 and 5.6 below.
Lemma 5.5. Suppose M answers k = d1 + log2 T e adaptively-chosen queries from Thresh[T ] with
ε-differential privacy and (α, β)-accuracy. Then there exists an ε-differentially private M 0 : [T ]n → [T ]
that computes an α-approximate median with probability at least 1 − β.
Proof. The algorithm M 0 , formalized in Figure 6, uses M to perform a binary search.
Input: x ∈ X n .
M is given x.
Initialize `1 = 0, u1 = T , and j = 1.
While uj − `j > 1 repeat:
Let mj = d(uj + `j )/2e.
Give M the query cmj ∈ Thresh[T ] and obtain the answer aj ∈ [0, 1].
If aj ≥ 21 , set (`j+1 , uj+1 ) = (`j , mj ); otherwise set (`j+1 , uj+1 ) = (mj , uj ).
Increment j.
Output uj .
Figure 6: M 0 : X n → X
We have u1 − `1 = T and, after every query j, uj+1 − `j+1 ≤ d(uj − `j )/2e. Since the process
stops when uj − `j = 1, it is easy to verify that M 0 makes at most d1 + log2 (T − 1)e queries to M.
Suppose all of the answers given by M are α-accurate. This happens with probability at least
1 − β. We will show that, given this, M 0 outputs an α-approximate median, which completes the
proof.
We claim that cuj (x) ≥ 12 − α for all j. This is easily shown by induction. The base case is
cT (x) = 1 ≥ 12 − α. At each step either uj+1 = uj (in which case the induction hypothesis can be
applied) or uj+1 = mj ; in the latter case our accuracy assumption gives
cuj+1 (x) = cmj (x) ≥ aj − α ≥
1
− α.
2
We also claim that c`j (x) < 12 +α for all j. This follows from a similar induction and completes
the proof.
19
Lemma 5.6. Let M : [T ]n → [T ] be an ε-differentially private algorithm that computes an αapproximate median with confidence 1 − β. Then
!
log T + log(1/β)
n≥Ω
.
αε
Proof. Let m = d( 12 − α)ne − 1. For each t ∈ [T ], let xt ∈ [T ]n denote the dataset containing m
copies of 1, m copies of T , and n − 2m copies of t. Then for each t ∈ [T ],
h
i
P M(xt ) = t ≥ 1 − β.
On the other hand, by the pigeonhole principle, there must exist t∗ ∈ [T − 1] such that
h
i
T
h
i P M(x ) ∈ [T − 1]
β
P M(xT ) = t∗ ≤
≤
.
T −1
T −1
The inputs xT and xt∗ differ in at most n − 2m ≤ 2αn + 2 entries. By group privacy,
h
i
h
i
β
1 − β ≤ P M(xt∗ ) = t∗ ≤ eε(2αn+2) P M(xT ) = t∗ ≤ eε(2αn+2)
.
T −1
Rearranging these inequalities gives
!
(1 − β)(T − 1)
≥ Ω(log(T /β)),
O(εαn) ≥ ε(2αn + 2) ≥ log
β
which yields the result.
Remark 5.7. Proposition 5.3 can be extended to online non-adaptive queries, which yields a
separation between the online non-adaptive and offline models for pure differential privacy and
threshold queries.
The key observation behind remark 5.7 is that, while Lemma 5.5 in general requires making
adaptive queries, for the inputs xt ∈ [T ]n (t ∈ [T ]) used in Lemma 5.6 the queries are “predictable.” In particular, on input xt , the algorithm M 0 from the proof of Lemma 5.5 will (with
probability at least 1 − β) always make the same sequence queries. This allows the queries to be
specified in advance in a non-adaptive manner. More precisely, we can produce an algorithm
Mt0 that produces non-adaptive online queries by simulating M 0 on input xt and using those
queries. Given the answers to these online non-adaptive queries, Mt0 can either accept or reject
its input depending on whether the answers are consistent with the input xt ; Mt0 will accept xt
0
with high probability and reject xt for t 0 , t with high probabiliy. The proof of Lemma 5.6 can
be carried out using Mt0∗ instead of M 0 at the end.
5.2
The BetweenThresholds Algorithm
The key technical novelty behind our algorithm for answering adaptively-chosen threshold
queries is a refinement of the “Above Threshold” algorithm [DR14, §3.6], which underlies the
ubiquitous “sparse vector” technique [DNR+ 09, RR10, DNPR10, HR10].
The sparse vector technique addresses a setting where we have a stream of k (adaptivelychosen) low-sensitivity queries and a threshold parameter t. Instead of answering all k queries
20
accurately, we are interested in answering only the ones that are above the threshold t – for the
remaining queries, we only require a signal that they are below the threshold. Intuitively, one
would expect to only pay in privacy for the queries that are actually above the threshold. And
indeed, one can get away with sample complexity proportional to the number of queries that
are above the threshold, and to the logarithm of the total number of queries.
We extend the sparse vector technique to settings where we demand slightly more information about each query beyond whether it is below a single threshold. In particular, we set two
thresholds t` < tu , and for each query, release a signal as to whether the query is below the lower
threshold, above the upper threshold, or between the two thresholds.
As long as the thresholds are sufficiently far apart, whether (the noisy answer to) a query
is below the lower threshold or above the upper threshold is stable, in that it is extremely
unlikely to change on neighboring datasets. As a result, we obtain an (ε, δ)-differentially private
algorithm that achieves the same accuracy guarantees as the traditional sparse vector technique,
i.e. sample complexity proportional to log k.
Our algorithm is summarised by the following theorem.6
Theorem 5.8. Let α, β, ε, δ, t ∈ (0, 1) and n, k ∈ N satisfy
n≥
1
max {12 log(30/εδ), 16 log((k + 1)/β)} .
αε
Then there exists a (ε, δ)-differentially private algorithm that takes as input x ∈ X n and answers
a sequence of adaptively-chosen queries q1 , · · · , qk : X n → [0, 1] of sensitivity 1/n with a1 , · · · , a≤k ∈
{L, R, >} such that, with probability at least 1 − β,
• aj = L =⇒ qj (x) ≤ t,
• aj = R =⇒ qj (x) ≥ t, and
• aj = > =⇒ t − α ≤ qj (x) ≤ t + α.
The algorithm may halt before answering all k queries; however, it only halts after outputting >.
Our algorithm is given in Figure 7. The analysis is split into Lemmas 5.9 and 5.10.
Input: x ∈ X n .
Parameters: ε, t` , tu ∈ (0, 1) and n, k ∈ N.
Sample µ ∼ Lap(2/εn) and initialize noisy thresholds t̂` = t` + µ and t̂u = tu − µ.
For j = 1, 2, · · · , k:
Receive query qj : X n → [0, 1].
Set cj = qj (x) + νj where νj ∼ Lap(6/εn).
If cj < t̂` , output L and continue.
If cj > t̂u , output R and continue.
If cj ∈ [t̂` , t̂u ], output > and halt.
Figure 7: BetweenThresholds
6 In Theorem 5.8, only one threshold is allowed. However, our algorithm is more general and permits the setting
of two thresholds. We have chosen this statement for simplicity.
21
Lemma 5.9 (Privacy for BetweenThresholds). Let ε, δ ∈ (0, 1) and n ∈ N. Then BetweenThresholds
(Figure 7) is (ε, δ)-differentially private for any adaptively-chosen sequence of queries as long as the
gap between the thresholds t` , tu satisfies
tu − t` ≥
12
(log(10/ε) + log(1/δ) + 1) .
εn
Lemma 5.10 (Accuracy for BetweenThresholds). Let α, β, ε, t` , tu ∈ (0, 1) and n, k ∈ N satisfy
n≥
8
(log(k + 1) + log(1/β)) .
αε
Then, for any input x ∈ X n and any adaptively-chosen sequence of queries q1 , q2 , · · · , qk , the answers a1 , a2 , · · · a≤k produced by BetweenThresholds (Figure 7) on input x satisfy the following with
probability at least 1 − β. For any j ∈ [k] such that aj is returned before BetweenThresholds halts,
• aj = L =⇒ qj (x) ≤ t` + α,
• aj = R =⇒ qj (x) ≥ tu − α, and
• aj = > =⇒ t` − α ≤ qj (x) ≤ tu + α.
Combining Lemmas 5.9 and 5.10 and setting t` = t − α/2 and tu = t + α/2 yields Theorem 5.8.
Proof of Lemma 5.9. Our analysis is an adaptation of Dwork and Roth’s [DR14, §3.6] analysis of
the AboveThreshold algorithm. Recall that a transcript of the execution of BetweenThresholds is
given by a ∈ {L, R, >}∗ . Let M : X n → {L, R, >}∗ denote the function that simulates BetweenThresholds
interacting with a given adaptive adversary (cf. Figure 3) and returns the transcript.
Let S ⊂ {L, R, >}∗ be a set of transcripts. Our goal is to show that for adjacent datasets x ∼ x0 ,
P [M(x) ∈ S] ≤ eε P M(x0 ) ∈ S + δ.
Let
1
6
2
z∗ = (tu − t` ) −
log(10/ε) − 1/n ≥
log(1/δ).
2
εn
εn
Our strategy will be to show that as long as the noise value µ is under control, in particular if
µ ≤ z∗ , then the algorithm behaves in essentially the same way as the standard AboveThreshold
algorithm. Meanwhile, the event µ > z∗ which corresponds to the (catastrophic) event where the
upper and lower thresholds are too close or overlap, happens with probability at most δ.
The following claim reduces the privacy analysis to examining the probability of obtaining
any single transcript a:
Claim 5.11. Suppose that for any transcript a ∈ {L, R, >}∗ , and any z ≤ z∗ , that
P [M(x) = a|µ = z] ≤ eε/2 P M(x0 ) = a|µ = z + 1/n .
Then M is (ε, δ)-differentially private.
Proof. By properties of the Laplace distribution, since µ ∼ Lap(2/εn), for any z ∈ R, we have
P [µ = z] ≤ eε/2 P [µ = z + 1/n] ,
22
and
∗
1
P [µ > z∗ ] = e−εnz /2 ≤ δ.
2
Fix a set of transcripts S. Combining these properties allows us to write
Z
P [M(x) ∈ S] =
P [M(x) ∈ S|µ = z] P [µ = z] dz
R
!
Z z∗
≤
P [M(x) ∈ S|µ = z] P [µ = z] dz + P [µ > z∗ ]
−∞
≤ e
≤ e
Z
ε/2
ε
z∗
!
P M(x ) ∈ S|µ = z + 1/n P [µ = z] dz + δ
Z
0
−∞
z∗
!
0
P M(x ) ∈ S|µ = z + 1/n P [µ = z + 1/n] dz + δ
−∞
ε
≤ e P M(x0 ) ∈ S + δ
Returning to the proof of Lemma 5.9, fix a transcript a ∈ {L, R, >}∗ . Our goal is now to show
that M satisfies the hypotheses of Claim 5.11, namely that for any z ≤ z∗ ,
P [M(x) = a|µ = z] ≤ eε/2 P M(x0 ) = a|µ = z + 1/n .
(4)
For some k ≥ 1, we can write the transcript a as (a1 , a2 , . . . , ak ), where aj ∈ {L, R} for each j < k,
and ak = >.
For convenience, let A = M(x) and A0 = M(x0 ). We may decompose
h
i
P [M(x) = a|µ = z] = P (∀j < k, Aj = aj ) ∧ qk (x) + νk ∈ [t̂` , t̂u ]|µ = z
h
i
h
i
= P (∀j < k, Aj = aj )|µ = z · P qk (x) + νk ∈ [t̂` , t̂u ]|µ = z ∧ (∀j < k, Aj = aj ) .
(5)
We upper bound each factor on the right-hand side separately.
Claim 5.12.
h
i
P [(∀i < k, Ai = ai )|µ = z] ≤ P (∀i < k, A0i = ai )|µ = z + 1/n
Proof. For fixed z, let Az (x) denote the set of noise vectors (ν1 , . . . , νk−1 ) for which (A1 , . . . , Ak−1 ) =
(a1 , . . . , ak−1 ) when ν = z. We claim that as long as z ≤ z∗ , then Az (x) ⊆ Az+1/n (x0 ). To argue this,
let (ν1 , . . . , νk−1 ) ∈ Az (x). Fix an index j ∈ {1, . . . , k − 1} and suppose aj = L. Then qj (x) + νj < t` + z,
but since qj has sensitivity 1/n, we also have qj (x0 ) + νj < t` + (z + 1/n). Likewise, if aj = R, then
qj (x) + νj > tu − z, so
qj (x0 ) + νj > tu − z − 1/n ≥ t` + (z + 1/n)
as long as z ≤ z∗ ≤ 12 (tu − t` ) − 1/n. (This ensures that M(x0 ) does not output L on the first branch
of the “if” statement, and proceeds to output R.)
23
Since Az (x) ⊆ Az+1/n (x0 ), this proves that
P [(∀i < k, Ai = ai )|µ = z] = P [(ν1 , . . . , νk−1 ) ∈ Az (x)]
≤ P (ν1 , . . . , νk−1 ) ∈ Az+1/n (x0 )
h
i
= P (∀i < k, A0i = ai )|µ = z + 1/n .
Given Claim 5.12, all that is needed to prove (4) and, thereby, prove Lemma 5.9 is to bound
the second factor in (5) — that is, we must only show that
h
i
h
i
P qk (x) + νk ∈ [t̂` , t̂u ]|µ = z ∧ (∀j < k, Aj = aj ) ≤ eε/2 P qk (x0 ) + νk ∈ [t̂` , t̂u ]|µ = z + 1/n ∧ (∀j < k, A0j = aj ) .
(6)
Let ∆ = (qk (x0 ) − qk (x)) ∈ [−1/n, 1/n]. Then
h
i
P qk (x) + νk ∈ [t̂` , t̂u ]|µ = z ∧ (∀j < k, Aj = aj )
= P [t` + z ≤ qk (x) + νk ≤ tu − z]
= P t` + z + ∆ ≤ qk (x0 ) + νk ≤ tu − z + ∆
= P t` + (z + 1/n) + (∆ − 1/n) ≤ qk (x0 ) + νk ≤ tu − (z + 1/n) + (∆ + 1/n)
h
i
= P qk (x0 ) + νk ∈ [t̂` + ∆ − 1/n, t̂u + ∆ + 1/n]|µ = z + 1/n
h
i
≤ eε/2 P qk (x0 ) + νk ∈ [t̂` , t̂u ]|µ = z + 1/n
h
i
= eε/2 P qk (x0 ) + νk ∈ [t̂` , t̂u ]|µ = z + 1/n ∧ (∀j < k, A0j = aj )
where the last inequality follows from Claim 5.13 below (setting η = 2/n, λ = 6/εn, [a, b] = [t̂` , t̂u ],
6
and [a0 , b0 ] = [t̂` + ∆ − 1/n, t̂u + ∆ + 1/n]) and the fact that z ≤ z∗ = 12 (tu − t` ) − εn
log(10/ε) − 1/n
implies
12
10
1
b − a = t̂u − t̂` = tu − t` − 2µ ≥
log
≥ 2λ log
εn
ε
1 − e−ε/6
whenever 0 ≤ ε ≤ 1.
Claim 5.13. Let ν ∼ Lap(λ) and let [a, b], [a0 , b0 ] ⊂ R be intervals satisfying [a, b] ⊂ [a0 , b0 ]. If
η ≥ (b0 − a0 ) − (b − a), then
P ν ∈ [a0 , b0 ] ≤
eη/λ
· P [ν ∈ [a, b]] .
1 − e−(b−a)/2λ
Proof. Recall that the probability density function of the Laplace distribution is given by
1 −|x|/λ
fλ (x) = 2λ
e
. There are four cases to consider: In the first case, a < b ≤ 0. In the second case,
a < 0 < b with |a| ≤ |b|. In the third case, 0 ≤ a < b. Finally, in the fourth case, a < 0 < b with
|a| ≥ |b|. Since the Laplace distribution is symmetric, it suffices to analyze the first two cases.
24
Case 1:
Suppose a < b ≤ 0. Then
P ν ∈ [a0 , b0 ] ≤ P [ν ∈ [a, b]] +
Z
b+η
b
1 x/λ
e dx
2λ
1
= (e(b+η)/λ − ea/λ )
2
!
1 eη/λ − e(a−b)/λ
= ·
· (eb/λ − ea/λ )
(a−b)/λ
2
1−e
!
η/λ
e − e−(b−a)/λ
· P [ν ∈ [a, b]] .
=
1 − e−(b−a)/λ
Case 2:
Suppose a < 0 < b and |a| ≤ |b|. Note that this implies b ≥ (b − a)/2. Then
1 a/λ
P ν ∈ [a0 , b0 ] ≤ P [ν ∈ [a, b]] + η ·
e
2λ
a/λ
η
e
≤ P [ν ∈ [a, b]] 1 +
2λ P [ν ∈ [0, b]]
η
= P [ν ∈ [a, b]]
1 − e−b/λ + λ ea/λ
1 − e−b/λ
1 + η/λ
≤ P [ν ∈ [a, b]]
1 − e−b/λ
eη/λ
.
≤ P [ν ∈ [a, b]]
1 − e−(b−a)/2λ
Proof of Lemma 5.10. We claim that it suffices to show that with probability at least 1 − β we
have
∀1 ≤ j ≤ k
|νj | + |µ| ≤ α.
To see this, suppose |νj | + |µ| ≤ α for every j. Then, if aj = L, we have
cj = qj (x) + νj < t̂` = t` + µ,
whence
qj (x) < t` + |µ| + |νj | ≤ t` + α.
Similarly, if aj = R, then
cj = qj (x) + νj > t̂u = tu − µ,
whence
qj (x) > tu − (|µ| + |νj |) ≥ tu − α.
Finally, if aj = >, then
cj = qj (x) + νj ∈ [t̂` , t̂u ] = [t` + µ, tu − µ],
whence
t` − α ≤ qj (x) ≤ tu + α.
We now show that indeed |νj | + |µ| ≤ α for every j with high probability. By tail bounds for
the Laplace distribution,
h
i
εαn
εαn
P [|µ| > α/4] = exp −
and
P |νj | > 3α/4 = exp −
8
8
25
for all j. By a union bound,
h
i
εαn
≤ β,
P |µ| > α/4 ∨ ∃j ∈ [k] |νj | > 3α/4 ≤ (k + 1) · exp −
8
as required.
5.3
The Online Interior Point Problem
Our algorithm extends a result of [BNSV15] showing how to reduce the problem of privately
releasing thresholds to the much simpler interior point problem. By analogy, our algorithm
for answering adaptively-chosen thresholds relies on solving multiple instances of an online
variant of the interior point problem in parallel. In this section, we present the OIP problem and
give an (ε, δ)-differentially private solution that can handle k adaptively-chosen queries with
sample complexity O(log k). Our OIP algorithm is a direct application of the BetweenThresholds
algorithm from Section 5.2.
Definition 5.14 (Online Interior Point Problem). An algorithm M solves the Online Interior
Point (OIP) Problem for k queries with confidence β if, when given as input any private dataset
x ∈ [0, 1]n and any adaptively-chosen sequence of real numbers y1 , · · · , yk ∈ [0, 1], with probability
at least 1 − β it produces a sequence of answers a1 , · · · , ak ∈ {L, R} such that
∀j ∈ {1, 2, · · · , k}
yj < min xi =⇒ aj = L,
i∈[n]
yj ≥ max xi =⇒ aj = R .
i∈[n]
(If mini∈[n] xi ≤ yj < maxi∈[n] xi , then M may output either symbol L or R.)
Input: Dataset x ∈ [0, 1]n .
Initialize a BetweenThresholds instance (Figure 7) B on dataset x with thresholds t` = 13 ,
tu = 23 .
For j = 1, 2, · · · , k:
Receive query yj ∈ [0, 1].
If B already halted on some query qy ∗ , output L if yj < y ∗ and output R if yj ≥ y ∗ .
Otherwise, give B the query cyj ∈ Thresh[0,1] .
If B returns >, output R. Otherwise, output the answer produced by B.
Figure 8: Online Interior Point Algorithm
Proposition 5.15. The algorithm in Figure 8 is (ε, δ)-differentially private and solves the OIP Problem
with confidence β as long as
n≥
36
(log(k + 1) + log(1/β) + log(10/ε) + log(1/δ) + 1) .
ε
Proof. Privacy follows immediately from Lemma 5.9, since Algorithm 8 is obtained by postprocessing Algorithm 7, run using thresholds with a gap of size 1/3.
To argue utility, let α = 1/3 so that
n≥
8
(log(k + 1) + log(1/β)).
εα
By Lemma 5.10, with probability at least 1 − β, the following events occur:
26
• If the BetweenThresholds instance B halts when it is queried on cy ∗ , then mini∈[n] xi ≤ y ∗ <
maxi∈[n] xi .
• If B has not yet halted and yj < mini∈[n] xi , its answer to cyj is L.
• If B has not yet halted and yj ≥ maxi∈[n] xi , its answer to cyj is R.
Thus, if B has not yet halted, the answers provided are accurate answers for the OIP Problem.
On the other hand, when B halts, it has successfully identified an “interior point” of the dataset
x, i.e. a y ∗ such that mini∈[n] xi ≤ y ∗ < maxi∈[n] xi . Thus, for any subsequent query y, we have that
y < min xi =⇒ y < y ∗ ,
i∈[n]
so Algorithm 8 correctly outputs L. Similarly,
y ≥ max xi =⇒ y ≥ y ∗ ,
i∈[n]
so Algorithm 8 correctly outputs R on such a query.
5.4
Releasing Adaptive Thresholds with Approximate Differential Privacy
We are now ready to state our reduction from releasing thresholds to solving the OIP Problem.
Theorem 5.16. If there exists an (ε, δ)-differentially private algorithm solving the OIP problem for k
queries with confidence αβ/8 and sample complexity n0 , then there is a (4ε, (1 + eε )δ)-differentially
private algorithm for releasing k threshold queries with (α, β)-accuracy and sample complexity
)
( 0
6n 24 log2.5 (4/α) · log(2/β)
,
.
n = max
α
αε
Combining this reduction with our algorithm for the OIP Problem (Proposition 5.15) yields:
Corollary 5.17. There is an (ε, δ)-differentially private algorithm for releasing k adaptively-chosen
threshold queries with (α, β)-accuracy for
!
log k + log2.5 (1/α) + log(1/βεδ)
n=O
.
αε
Proof of Theorem 5.16. Our algorithm and its analysis follow the reduction of Bun et al. [BNSV15]
for reducing the (offline) query release problem for thresholds to the offline interior point problem.
Let T be an (ε, δ)-differentially private algorithm solving the OIP Problem with confidence
αβ/8 and sample complexity n0 . Without loss of generality, we may assume that T is differentially
private in “add-or-remove-an-item sense”—i.e. if x ∈ [0, 1]∗ and x0 differs from x up to the
addition or removal of a row, then
h for every adversary
i A and
h set S of outcomes of
i the interaction
ε P Adaptive → (x0 ) ∈ S + δ. Moreover,
between A and T , we have P AdaptiveA →
(x)
∈
S
≤
e
A←T
←T
T provides accurate answers to the OIP Problem with probability at least 1 − αβ/8 whenever
its input is of size at least n0 . To force an algorithm T to have these properties, we may pad
any dataset of size less than n0 with an arbitrary fixed element. On the other hand, we may
subsample the first n0 elements from any dataset with more than this many elements.
27
Input: Dataset x ∈ [0, 1]n .
Parameter: α ∈ (0, 1).
Let (x(1) , . . . , x(M) ) ←R Partition(x1 , . . . , xn , α).
Initialize an instance of the OIP algorithm T (m) on each chunk x(m) ∈ [0, 1]∗ , for m ∈ [M].
For each j = 1, · · · , k:
Receive query cyj ∈ Thresh[0,1] .
(1)
(M)
Give query yj ∈ [0, 1] to every OIP instance T (m) , receiving answers aj , · · · , aj
(m)
1
· m ∈ [M] : aj = R .
Return aj = M
∈ {L, R}.
Figure 9: AdaptiveThresholdsT
Input: Dataset x ∈ [0, 1]n .
Parameter: α ∈ (0, 1).
Output: (Random) partition (x(1) , . . . , x(M) ) ∈ ([0, 1]∗ )M of x, where 2/α ≤ M < 4/α.
Let M = 2dlog2 (2/α)e .
Sort x in nondecreasing order x1 ≤ x2 ≤ · · · ≤ xn .
For each 0 ≤ ` ≤ log2 M and s ∈ {0, 1}` , sample νs ∼ Lap((log2 M)/ε) independently.
P
For each 1 ≤ m ≤ M − 1, let ηm = s∈P (m) νs , where P (m) is the set of all prefixes of the
binary representation
j
kof m.
j
k
n
Let t0 = 1, t1 = M + η1 , · · · , tm = m·n
M + ηm , · · · , tM = n + 1.
Let x(m) = (xtm−1 , . . . , xtm −1 ) for all m ∈ [M].
Figure 10: Partition
Consider the algorithm AdaptiveThresholdsT in Figures 9 and 10.
The proof of Theorem 5.16 relies on the following two claims about the Partition subroutine,
both of which are implicit in the work of Bun et al. [BNSV15, Appendix C] and are based on
ideas of Dwork et al. [DNPR10]. Claim 5.18 shows that for neighboring databases x ∼ x0 , the
behaviors of the Partition subroutine on x and x0 are “similar” the following sense: for any fixed
partition of x, one is roughly as likely (over the randomness of the partition algorithm) to obtain
a partition of x0 that differs on at most two chunks, where the different chunks themselves differ
only up to the addition or removal of a single item. This will allow us to show that running M
parallel copies of the OIP algorithm on the chunks remains roughly (ε, δ)-differentially private.
Claim 5.19 shows that, with high probability, each chunk is simultaneously large enough for
the corresponding OIP algorithm to succeed, but also small enough so that treating all of the
elements in a chunk as if they were the same element still permits us to get α-accurate answers
to arbitrary threshold queries.
Claim 5.18. Fix neighboring datasets x, x0 ∈ [0, 1]n . Then there exists a (measurable) bijection
ϕ : R2M → R2M with the following properties:
1. Let z ∈ R2M be any noise vector. Let x(1) , . . . , x(M) denote the partition of x obtained with random
noise set to ν = z. Similarly, let x0(1) , . . . , x0(M) denote the partition of x0 obtained under noise
28
ν = ϕ(z). Then there exist indices i1 , i2 such that: 1) For i ∈ {i1 , i2 }, the chunks x(i) and x0(i)
differ up to the addition or removal of at most one item and 2) For every index i < {i1 , i2 }, we
have x(i) = x0(i) .
2. For every noise vector z ∈ R2M , we have P [ν = ϕ(z)] ≤ e2ε P [ν = z].
Claim 5.19. With probability at least 1 − β/2, we have that |tm − m · n/M| ≤ αn/24 for all m ∈ [M].
Privacy of Algorithm 9. We first show how to use Claim 5.18 to show that Algorithm 9 is
differentially private. Fix an adversary A, and let B = AdaptiveA →
simulate the
← AdaptiveThresholdsT
interaction between A and Algorithm 9. Let S be a subset of the range of B. Then, by Property
(1) of Claim 5.18 and group privacy, we have that for any z ∈ R2M :
P [B(x) ∈ S|ν = z] ≤ e2ε P B(x0 ) ∈ S|ν = ϕ(z) + (1 + eε )δ.
By Property (2) of Claim 5.18, we also have Pr[ν = z] ≤ e2ε Pr[ν = ϕ(z)] for every z ∈ R2M .
Therefore,
Z
P [B(x) ∈ S] =
P [B(x) ∈ S|ν = z] · P [ν = z] dz
2M
R
Z
2ε
0
ε
≤
e P B(x ) ∈ S|ν = ϕ(z) + (1 + e )δ · P [ν = z] dz
R2M
Z
ε
≤ (1 + e )δ +
e2ε P B(x0 ) ∈ S|ν = ϕ(z) · e2ε P [ν = ϕ(z)] dz
ε
R2M
4ε
≤ (1 + e )δ + e P B(x0 ) ∈ S .
Hence, B is (e4ε , (1 + eε )δ)-differentially private, as claimed.
Accuracy of Algorithm 9. We now show how to use Claim 5.19 to show that Algorithm 9
produces (α, β)-accurate answers. By a union bound, the following three events occur with
probability at least 1 − β:
1. For all m ∈ [M],
m
M
− tnm ≤ α6 .
2. Every chunk x(m) has size |x(m) | = tm − tm−1 ∈ [αn/6, 2αn/3].
3. Every instance of T succeeds.
Now we need to show that if these three events occur, we can produce α-accurate answers
to every threshold query cy1 , . . . , cyk . Write the sorted input database as x1 ≤ x2 ≤ · · · ≤ xn . We
consider two cases for the j th query: As our first case, suppose xn ≤ yj . Then for every chunk
(m)
x(m) , we have max{x(m) } ≤ yj . Then the success condition of T (m) guarantees that aj = R. Thus,
the answer aj = 1 is (exactly) accurate for the query cj .
As our second case, let i be the smallest index for which xi > yj , and suppose the item xi is
in some chunk x(mi ) . Note that this means that the true answer to the query cyj is (i − 1)/n and
29
that tmi −1 ≤ i ≤ tmi − 1. Then again, for every m < mi we have max{x(m) } ≤ yj , so every such T (m)
(m)
instance yields aj
aj =
= R. Thus,
1
m − 1 tmi α α (i − 1)
(m)
· m ∈ [M] : aj = R ≥ i
≥
− − ≥
− α,
M
M
n
6 2
n
since M ≥ 2/α.
On the other hand, for every m > mi , we have min{x(m) } > yj , so every such T (m) instance
(m)
instead yields aj
= L.
aj ≤
mi tmi α tmi −1 + 2αn/3 α i 2α α i − 1
≤
+ ≤
+ ≤ +
+ ≤
+ α,
M
n
6
n
6 n
3
6
n
since n ≥ 6/α.
Acknowledgements
We thank Salil Vadhan for many helpful discussions.
References
[BLR13]
Avrim Blum, Katrina Ligett, and Aaron Roth. A learning theory approach to noninteractive database privacy. J. ACM, 60(2):12, 2013.
[BNS13]
Amos Beimel, Kobbi Nissim, and Uri Stemmer. Private learning and sanitization:
Pure vs. approximate differential privacy. In Approximation, Randomization, and
Combinatorial Optimization. Algorithms and Techniques - 16th International Workshop,
APPROX 2013, and 17th International Workshop, RANDOM 2013, Berkeley, CA, USA,
August 21-23, 2013. Proceedings, pages 363–378, 2013.
[BNS+ 16] Raef Bassily, Kobbi Nissim, Adam D. Smith, Thomas Steinke, Uri Stemmer, and
Jonathan Ullman. Algorithmic stability for adaptive data analysis. In Proceedings
of the Forty-Eighth Annual ACM on Symposium on Theory of Computing, STOC 2016,
Cambridge, MA, USA, 2016.
[BNSV15] Mark Bun, Kobbi Nissim, Uri Stemmer, and Salil P. Vadhan. Differentially private
release and learning of threshold functions. In IEEE 56th Annual Symposium on
Foundations of Computer Science, FOCS 2015, Berkeley, CA, USA, 17-20 October, 2015,
pages 634–649, 2015.
[BS98]
Dan Boneh and James Shaw. Collusion-secure fingerprinting for digital data. IEEE
Trans. Information Theory, 44(5):1897–1905, 1998.
[BUV14]
Mark Bun, Jonathan Ullman, and Salil P. Vadhan. Fingerprinting codes and the price
of approximate differential privacy. In Symposium on Theory of Computing, STOC
2014, New York, NY, USA, May 31 - June 03, 2014, pages 1–10, 2014.
30
[CSS11]
T.-H. Hubert Chan, Elaine Shi, and Dawn Song. Private and continual release of
statistics. ACM Trans. Inf. Syst. Secur., 14(3):26, 2011.
[DFH+ 15] Cynthia Dwork, Vitaly Feldman, Moritz Hardt, Toniann Pitassi, Omer Reingold,
and Aaron Leon Roth. Preserving statistical validity in adaptive data analysis. In
Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing,
STOC 2015, Portland, OR, USA, June 14-17, 2015, pages 117–126, 2015.
[DL09]
Cynthia Dwork and Jing Lei. Differential privacy and robust statistics. In Proceedings
of the 41st Annual ACM Symposium on Theory of Computing, STOC 2009, Bethesda,
MD, USA, May 31 - June 2, 2009, pages 371–380, 2009.
[DMNS06] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise
to sensitivity in private data analysis. In Theory of Cryptography, Third Theory of
Cryptography Conference, TCC 2006, New York, NY, USA, March 4-7, 2006, Proceedings,
pages 265–284, 2006.
[DN03]
Irit Dinur and Kobbi Nissim. Revealing information while preserving privacy. In
Proceedings of the Twenty-Second ACM SIGACT-SIGMOD-SIGART Symposium on
Principles of Database Systems, June 9-12, 2003, San Diego, CA, USA, pages 202–210,
2003.
[DNPR10] Cynthia Dwork, Moni Naor, Toniann Pitassi, and Guy N. Rothblum. Differential
privacy under continual observation. In Proceedings of the 42nd ACM Symposium
on Theory of Computing, STOC 2010, Cambridge, Massachusetts, USA, 5-8 June 2010,
pages 715–724, 2010.
[DNR+ 09] Cynthia Dwork, Moni Naor, Omer Reingold, Guy N. Rothblum, and Salil P. Vadhan.
On the complexity of differentially private data release: efficient algorithms and
hardness results. In Proceedings of the 41st Annual ACM Symposium on Theory of
Computing, STOC 2009, Bethesda, MD, USA, May 31 - June 2, 2009, pages 381–390,
2009.
[DNRR15] Cynthia Dwork, Moni Naor, Omer Reingold, and Guy N. Rothblum. Pure differential
privacy for rectangle queries via private partitions. In Advances in Cryptology ASIACRYPT 2015 - 21st International Conference on the Theory and Application of
Cryptology and Information Security, Auckland, New Zealand, November 29 - December
3, 2015, Proceedings, Part II, pages 735–751, 2015.
[DR14]
Cynthia Dwork and Aaron Roth. The algorithmic foundations of differential privacy.
Found. Trends Theor. Comput. Sci., 9(3–4):211–407, August 2014.
[DRV10]
Cynthia Dwork, Guy N. Rothblum, and Salil P. Vadhan. Boosting and differential
privacy. In IEEE Symposium on Foundations of Computer Science (FOCS ’10), pages
51–60. IEEE, 23–26 October 2010.
[DSS+ 15] Cynthia Dwork, Adam D. Smith, Thomas Steinke, Jonathan Ullman, and Salil P.
Vadhan. Robust traceability from trace amounts. In IEEE 56th Annual Symposium on
Foundations of Computer Science, FOCS 2015, Berkeley, CA, USA, 17-20 October, 2015,
pages 650–669, 2015.
31
[DY08]
Cynthia Dwork and Sergey Yekhanin. New efficient attacks on statistical disclosure control mechanisms. In Advances in Cryptology - CRYPTO 2008, 28th Annual
International Cryptology Conference, Santa Barbara, CA, USA, August 17-21, 2008.
Proceedings, pages 469–480, 2008.
[HR10]
Moritz Hardt and Guy N. Rothblum. A multiplicative weights mechanism for
privacy-preserving data analysis. In 51th Annual IEEE Symposium on Foundations of
Computer Science, FOCS 2010, October 23-26, 2010, Las Vegas, Nevada, USA, pages
61–70, 2010.
[HT10]
Moritz Hardt and Kunal Talwar. On the geometry of differential privacy. In Proceedings of the 42nd ACM Symposium on Theory of Computing, STOC 2010, Cambridge,
Massachusetts, USA, 5-8 June 2010, pages 705–714, 2010.
[HU14]
Moritz Hardt and Jonathan Ullman. Preventing false discovery in interactive data
analysis is hard. In 55th IEEE Annual Symposium on Foundations of Computer Science,
FOCS 2014, Philadelphia, PA, USA, October 18-21, 2014, pages 454–463, 2014.
[RR10]
Aaron Roth and Tim Roughgarden. Interactive privacy via the median mechanism.
In STOC, pages 765–774. ACM, June 5–8 2010.
[SU15a]
Thomas Steinke and Jonathan Ullman. Between pure and approximate differential
privacy. CoRR, abs/1501.06095, 2015.
[SU15b]
Thomas Steinke and Jonathan Ullman. Interactive fingerprinting codes and the
hardness of preventing false discovery. In Proceedings of The 28th Conference on
Learning Theory, COLT 2015, Paris, France, July 3-6, 2015, pages 1588–1628, 2015.
[Tar08]
Gábor Tardos. Optimal probabilistic fingerprint codes. J. ACM, 55(2), 2008.
[Ull13]
Jonathan Ullman. Answering n2+o(1) counting queries with differential privacy is
hard. In Symposium on Theory of Computing Conference, STOC’13, Palo Alto, CA, USA,
June 1-4, 2013, pages 361–370, 2013.
[War65]
Stanley L. Warner. Randomized response: A survey technique for eliminating evasive
answer bias. Journal of the American Statistical Association, 60(309):63–69, 1965.
A
The Fingerprinting Lemma
In this section we prove the fingerprinting lemma (Lemma 3.6). The proof is broken into several
lemmata.
Lemma A.1. Let f : {±1}n → R. Define g : [±1] → R by
g(p) =
Then
E [f (x)] .
x1···n ∼p
X
E f (x) ·
(xi − p) = g 0 (p) · (1 − p2 ).
x1···n ∼p
i∈[n]
32
A rescaling of this lemma appears in [SU15b]. The following proof is taken from [DSS+ 15].
Proof. We begin by establishing several identities.
Since x2 = 1 for x ∈ {±1}, we have the identity
d 1 + xp x 1 + xp x − p
= =
dp 2
2
2 1 − p2
for all x ∈ {±1} and p ∈ (−1, 1). By the product rule, we have
!
d Y 1 + xi p X d 1 + xi p Y 1 + xk p X xi − p Y 1 + xk p
=
=
dp
2
dp 2
2
2
1 − p2
i∈[n]
i∈[n]
k∈[n]\{i}
i∈[n]
k∈[n]
for all x ∈ {±1}n and p ∈ (−1, 1).
1+xp
Sampling x ∼ p samples each x ∈ {±1} with probability 2 . Thus sampling x1···n ∼ p,
Q
1+x p
samples each x ∈ {±1}n with probability i∈[n] 2 i .
Now we can write
X
Y 1+x p
i
g(p) = E [f (x)] =
f (x)
.
x1···n ∼p
2
n
x∈{±1}
i∈[n]
Using the above identities gives
g 0 (p) =
X
x∈{±1}n
f (x)
d Y 1 + xi p
dp
2
i∈[n]
X x −p Y 1+x p
i
k
=
f (x)
2
2
1−p
x∈{±1}n
i∈[n]
k∈[n]
X x − p
i
= E f (x)
x1···n ∼p
1 − p2
X
i∈[n]
Lemma A.2. Let g : [±1] → R be a polynomial. Then
h
i
E g 0 (p) · (1 − p2 ) = 2 E [g(p) · p] .
p∈[±1]
p∈[±1]
33
Proof. Let u(p) = 1 − p2 . By integration by parts and the fundamental theorem of calculus,
h
i 1Z 1
g 0 (p)(1 − p2 )dp
E g 0 (p) · (1 − p2 ) =
2 −1
p∈[±1]
Z
1 1 0
g (p)u(p)dp
=
2 −1
!
Z
1 1 d
=
g(p)u(p) − g(p)u 0 (p)dp
2 −1 dp
Z
1
1 1
g(p)(−2p)dp
= (g(1)u(1) − g(−1)u(−1)) −
2
2 −1
Z1
=0 +
g(p)pdp
−1
=2 E [g(p) · p] .
p∈[±1]
Proposition A.3. Let f : {±1}n → R. Then
X
1
(xi − p) + (f (x) − x)2 ≥ .
E
f (x) ·
3
p∈[±1],x1···n ∼p
i∈[n]
Proof. Define g : [±1] → R by
g(p) =
E [f (x)] .
x1···n ∼p
By Lemmas A.1 and A.2,
X
h
i
(xi − p) = E g 0 (p)(1 − p2 ) = E [2g(p)p] .
E
f (x) ·
p∈[±1]
p∈[±1]
p∈[±1],x1···n ∼p
i∈[n]
Moreover, by Jensen’s inequality,
!2
h
i
2
E
(f (x) − x) ≥ E E [f (x) − x]
p∈[±1],x1···n ∼p
p∈[±1] x1···n ∼p
h
i
= E (g(p) − p)2
p∈[±1]
h
i
= E g(p)2 − 2g(p)p + p2
p∈[±1]
h
i
= E g(p)2 −
p∈[±1]
X
1
E
(xi − p) + .
f (x) ·
3
p∈[±1],x1···n ∼p
i∈[n]
Rearranging yields the result:
X
h
i 1 1
E
(xi − p) + (f (x) − x)2 ≥ E g(p)2 + ≥ .
f (x) ·
p∈[±1]
3 3
p∈[±1],x1···n ∼p
i∈[n]
34
We also have an alternative version of Proposition A.3:
Proposition A.4. Let f : {±1}n → R. Then
X
1
2
E
(xi − p) + (f (x) − p) ≥ .
f (x) ·
3
p∈[±1],x1···n ∼p
i∈[n]
Proof. Define g : [±1] → R by
g(p) =
E [f (x)] .
x1···n ∼p
By Lemmas A.1 and A.2,
X
h
i
E
(xi − p) = E g 0 (p)(1 − p2 ) = E [2g(p)p] .
f (x) ·
p∈[±1]
p∈[±1],x1···n ∼p
p∈[±1]
i∈[n]
Moreover,
E
p∈[±1],x1···n ∼p
h
i
(f (x) − p)2 =
h
i
1
f (x)2 − 2g(p)p + p2 ≥ 0 − E [2g(p)p] + .
3
p∈[±1],x1···n ∼p
p∈[±1]
E
The result follows by combining the above equality and inequality.
Finally we restate and prove Lemma 3.6
Lemma A.5 (Fingerprinting Lemma). Let f : {±1}n → [±1]. Then
X
1
(xi − p) + 2 |f (x) − x| ≥ .
E
f (x) ·
3
p∈[±1],x1···n ∼p
i∈[n]
Proof. Since |f (x) − x| ≤ 2, we have |f (x) − x|2 ≤ 2 |f (x) − x|. The result thus follows from Proposition A.3.
35
| 8cs.DS
|
Frameworks for Designing In-place Graph Algorithms
Sankardeep Chakraborty1 , Anish Mukherjee2 , Venkatesh Raman3 , Srinivasa Rao Satti4
arXiv:1711.09859v1 [cs.DS] 27 Nov 2017
1
3
National Institute of Informatics, Tokyo, Japan, [email protected]
2
Chennai Mathematical Institute, Chennai, India, [email protected]
The Institute of Mathematical Sciences, HBNI, Chennai, India, [email protected]
4
Seoul National University, Seoul, South Korea, [email protected]
Read-only memory (ROM) model is a classical model of computation to study time-space
tradeoffs of algorithms. One of the classical results on the ROM model is that any sorting algorithm
that uses O(s) words of extra space requires Ω(n2 /s) comparisons for lg n ≤ s ≤ n/ lg n1 and the
bound has also been recently matched by an algorithm. However, if we relax the model (from
ROM), we do have sorting algorithms (say Heapsort) that can sort using O(n lg n) comparisons
using O(lg n) bits of extra space, even keeping a permutation of the given input sequence at any
point of time during the algorithm.
We address similar questions for graph algorithms. We show that a simple natural relaxation of
ROM model allows us to implement fundamental graph search methods like BFS and DFS more
space efficiently than in ROM. By simply allowing elements in the adjacency list of a vertex to be
permuted, we show that, on an undirected or directed connected graph G having n vertices and m
edges, the vertices of G can be output in a DFS or BFS order using O(lg n) bits of extra space and
O(n3 lg n) time. Thus we obtain similar bounds for reachability and shortest path distance (both for
undirected and directed graphs). With a little more (but still polynomial) time, we can also output
vertices in the lex-DFS order. As reachability in directed graphs (even in DAGs) and shortest path
distance (even in undirected graphs) are NL-complete, and lex-DFS is P-complete, our results show
that our model is more powerful than ROM if L 6= P.
En route, we also introduce and develop algorithms for another relaxation of ROM where the
adjacency lists of the vertices are circular lists and we can modify only the heads of the lists. Here
we first show a linear time DFS implementation using n + O(lg n) bits of extra space. Improving
the extra space further to only O(lg n) bits, we also obtain BFS and DFS albeit with a slightly
slower running time. Some of these algorithms also translate to improved algorithms for DFS and
its applications in ROM. Both the models we propose maintain the graph structure throughout the
algorithm, only the order of vertices in the adjacency list changes. In sharp contrast, for BFS and
DFS, to the best of our knowledge, there are no algorithms in ROM that use even O(n1− ) bits
of extra space; in fact, implementing DFS using cn bits for c < 1 has been mentioned as an open
problem. Furthermore, DFS (BFS) algorithms using n + o(n) (o(n)) bits of extra use Reingold’s
[JACM, 2008] or Barnes et al’s reachability algorithm [SICOMP, 1998] and hence have high runtime.
Our results can be contrasted with the recent result of Buhrman et al. [STOC, 2014] which gives an
algorithm for directed st-reachability on catalytic Turing machines using O(lg n) bits with catalytic
space O(n2 lg n) and time O(n9 ).
All our algorithms are simple but quite subtle, and we believe that these models are practical
enough to spur interest for other graph problems in these models.
1
We use lg to denote logarithm to the base 2
1
Contents
1 Introduction
1.1 Previous work in space efficient graph algorithms . .
1.2 In-place model for graph algorithms . . . . . . . . .
1.3 Definitions, computational complexity and notations
1.4 Our Results . . . . . . . . . . . . . . . . . . . . . . .
1.5 Techniques . . . . . . . . . . . . . . . . . . . . . . .
1.6 Consequences of our BFS and DFS results . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
3
5
5
7
8
9
2 DFS algorithms in the rotate model
2.1 Proof of Theorem 1(a) for undirected graphs
2.2 Proof of Theorem 1(a) for directed graphs . .
2.3 Proof of Theorem 1(b) for undirected graphs
2.4 Proof of Theorem 1(b) for directed graphs . .
2.5 Proof of Theorem 1(c) for undirected graphs
2.6 Proof of Theorem 1(c) for directed graphs . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
10
10
11
11
12
12
13
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3 BFS algorithms in the rotate model
15
3.1 BFS using n + O(lg n) bits–Proof of Theorem 2(a) . . . . . . . . . . . . . . . . . . . 15
3.2 BFS using O(lg n) bits–Proof of Theorem 2(b) . . . . . . . . . . . . . . . . . . . . . 16
4 Simulation of algorithms for rotate model in the implicit model
17
5 DFS algorithms in the implicit model—proof of Theorem 3
18
6 BFS algorithms in the implicit model—proof of Theorem 4
19
7 Minimum Spanning Tree
20
8 Consequences
8.1 Improved algorithm for DFS and applications in ROM
8.2 Space-efficient approximation algorithms using Baker’s
8.2.1 Baker’s Algorithm . . . . . . . . . . . . . . . .
8.3 Solving NP-hard problems in in-place models . . . . .
21
21
23
23
24
9 Concluding remarks
. . . . . .
approach
. . . . . .
. . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
24
2
1
Introduction
Motivated by the rapid growth of huge data set (“big data”), space efficient algorithms are
becoming increasingly important than ever before. The proliferation of specialized handheld devices
and embedded systems that have a limited supply of memory provide another motivation to
consider space efficient algorithms. To design space-efficient algorithms in general, several models
of computation have been proposed. Among them, the following two computational models have
received considerable attention in the literature.
• In the read-only memory (ROM) model, we assume that the input is given in a read-only
memory. The output of an algorithm is written on to a separate write-only memory, and
the output can not be read or modified again. In addition to the input and output media, a
limited random access workspace is available. Early work on this model was on designing lower
bounds [15, 17, 18], for designing algorithms for selection and sorting [27, 45, 52, 61, 62, 64]
and problems in computational geometry [6, 9, 13, 26, 32]. Recently there has been interest
on space-efficient graph algorithms [7, 11, 12, 23, 24, 44, 53, 54].
• In the in-place model, the input elements are given in an array, and the algorithm may use the
input array as working space. Hence the algorithm may modify the array during its execution.
After the execution, all the input elements should be present in the array (maybe in a permuted
order) and the output maybe put in the same array or sent to an output stream. The amount
of extra space usage during the entire execution of the algorithm is limited to O(lg n) bits.
A prominent example of an in-place algorithm is the classic heap-sort. Other than in-place
sorting [51], searching [49, 63] and selection [57] algorithms, many in-place algorithms have
been designed in areas such as computational geometry [19] and stringology [50].
Apart from these models, researchers have also considered (semi)-streaming models [3, 47, 61]
for designing space-efficient algorithms. Very recently the following two new models were introduced
in the literature with the same objective.
• Chan et al. [28] introduced the restore model which is a more relaxed version of read-only
memory (and a restricted version of the in-place model), where the input is allowed to be
modified, but at the end of the computation, the input has to be restored to its original form.
They also gave space efficient algorithms for selection and sorting on integer arrays in this
model. This has motivation, for example, in scenarios where the input (in its original form) is
required by some other application.
• Buhrman et al. [20, 21, 56] introduced and studied the catalytic-space model where a small
amount (typically O(lg n) bits) of clean space is provided along with some large additional
auxiliary space, with the condition that the additional space is initially in an arbitrary, possibly
incompressible, state and must be returned to this state when the computation is finished.
The input is assumed to be given in ROM. Thus this model can be thought of as having
an auxiliary storage that needs to be ‘restored’ in contrast to the model by Chan et al. [28]
where the input array has to be ‘restored’. They show various interesting complexity theoretic
consequences in this model and designed significantly better (in terms of space) algorithms in
comparison with the ROM model for a few combinatorial problems.
1.1
Previous work in space efficient graph algorithms
Even though these models were introduced in the literature with the aim of designing and/or
implementing various algorithms space efficiently, space efficient graph algorithms have been designed
only in the (semi)-streaming and the ROM model. In the streaming and semi-streaming models,
3
researchers have studied several basic and fundamental algorthmic problems such as connectivity,
minimum spanning tree, matching. See [59] for a comprehensive survey in this field. Research on
these two models (i.e., streaming and semi-streaming) is relatively new and has been going on for
last decade or so whereas the study in ROM could be traced back to almost 40 years. In fact there
is already a rich history of designing space efficient algorithms in the read-only memory model.
The complexity class L (also known as DLOGSPACE) is the class containing decision problems that
can be solved by a deterministic Turing machine using only logarithmic amount of work space for
computation. There are several important algorithmic results [35, 40, 41, 42] for this class, the
most celebrated being Reingold’s method [67] for checking st-reachability in an undirected graph,
i.e., to determine if there is a path between two given vertices s and t. NL is the non-deterministic
analogue of L and it is known that the st-reachability problem for directed graphs is NL-complete
(with respect to log space reductions). Using Savitch’s algorithm [5], this problem can be solved in
nO(lg n) time using O(lg2 n) bits of extra space. Savitch’s algorithm is very space efficient but its
running time is superpolynomial.
Among the deterministic algorithms running in polynomial time for directed st-reachability, the
most space
efficient algorithm is due to Barnes et al. [14] who gave a slightly sublinear space (using
√
n/2Θ( lg n) bits) algorithm for this problem running in polynomial time. We know of no better
polynomial time algorithm for this problem with better space bound. Moreover, the space used by
this algorithm matches a lower bound on space for solving directed st-reachability on a restricted
model of computation called Node Naming Jumping Automata on Graphs (NNJAGs) [29, 39].
This model was introduced especially for the study of directed st-reachability and most of the
known sublinear space algorithms for this problem can be implemented
on it. Thus, to design any
√
Θ(
lg
n)
polynomial time ROM algorithm taking space less than n/2
bits requires radically new ideas.
Recently there has been some improvement in the space bound for some special classes of graphs
like planar and H-minor free graphs [8, 22]. A drawback for all these algorithms using small space
i.e., sublinear number of bits, is that their running time is often some polynomial of high degree.
This is not surprising as Tompa [70] showed that for directed st-reachability, if the number of bits
available is o(n) then some natural algorithmic approaches to the problem require super-polynomial
time.
Motivated by these impossibility results from complexity theory and inspired by the practical
applications of these fundamental graph algorithms, recently there has been a surge of interest in
improving the space complexity of the fundamental graph algorithms without paying too much
penalty in the running time i.e., reducing the working space of the classical graph algorithms to O(n)
bits with little or no penalty in running time. Thus the goal is to design space-efficient yet reasonably
time-efficient graph algorithms on the ROM model. Generally most of the classical linear time graph
algorithms take O(n) words or equivalently O(n lg n) bits of space. Towards this recently Asano et
al. [7] gave an O(m lg n) time algorithm using O(n) bits, and another implementation taking n + o(n)
bits, using Reingold’s or Barnes et al’s reachability algorithm, and hence have high polynomial
running time. Later, time bound was improved to O(m lg lg n) still using O(n) bits in [44]. For
sparse graphs, the time bound is further improved in [11, 23] to optimal O(m) using still O(n) bits
of space. Improving on the classical linear time implementation of BFS which uses O(n lg n) bits of
space, recent space efficient algorithms [11, 44, 53] have resulted in a linear time algorithm using
n lg 3 + o(n) bits. We know of no algorithm for BFS using n + o(n) bits and O(m lgc n) (or even
O(mn))
√ time for some constant c in ROM. The only BFS algorithm taking sublinear space uses
n/2Θ( lg n) bits [14] and has a high polynomial runtime. A few other space efficient algorithms for
fundamental graph problems like checking strong connectivity [44], biconnectivity and performing
st-numbering [23], recognizing chordal and outerplanar graphs [25, 54] were also designed very
recently.
4
1.2
In-place model for graph algorithms
In order to break these inherent space bound barriers and obtain reasonably time and space efficient
graph algorithms, we want to relax the limitations of ROM. And the most natural and obvious
candidate in this regard is the classical in-place model. Thus our main objective is to initiate a
systematic study of efficient in-place (i.e., using O(lg n) bits of extra space) algorithms for graph
problems. To the best of our knowledge, this has not been done in the literature before. Our first
goal towards this is to properly define models for the in-place graph algorithms. As in the case of
standard in-place model, we need to ensure that the graph (adjacency) structure remains intact
throughout the algorithm. Let G = (V, E) be the input graph with n = |V |, m = |E|, and assume
that the vertex set V of G is the set V = {1, 2, · · · , n}. To describe these models, we assume that
the input graph representation consists of two parts: (i) an array V of length n, where V [i] stores
a pointer to the adjacency list of vertex i, and (ii) a list of singly linked lists, where the i-th list
consists of a singly linked list containing all the neighbors of vertex i with V [i] pointing to the head
of the list. In the ROM model, we assume that both these components cannot be modified. In our
relaxed models, we assume that one of these components can be modified in a limited way. This
gives rise to two different model which we define next.
Implicit model: The most natural analogue of in-place model allows any two elements in the
adjacency list of a vertex to be swapped (in constant time assuming that we have access to the nodes
storing those elements in the singly linked list). The adjacency “structure” of the representation
does not change; only the values stored can be swapped. (One may restrict this further to allow
only elements in adjacent nodes to be swapped. Most of our algorithms work with this restriction.)
We call it the implicit model inspired by the notion of implicit data structures [63]. We introduce
and develop algorithms for another relaxed model which we call the rotate model.
Rotate model: In this model, we assume that only the pointers stored in the array V can be
modified, that too in a limited way - to point to any node in the adjacency list, instead of always
pointing to the first node. In space-efficient setting, since we do not have additional space to store a
pointer to the beginning of the adjacency list explicitly, we assume that the second component of
the graph representation consists of a list of circular linked lists (instead of singly linked lists) – i.e.,
the last node in the adjacency list of each vertex points to the first node (instead of storing a null
pointer). See the Figure 1 to get a better visual description. We call the element pointed to by the
pointer as the front of the list, and a unit cost rotate operation changes the element pointed to by
the pointer to the next element in the list.
Thus the rotate model corresponds to keeping the adjacency lists in read-only memory and
allowing (limited) updates on the pointer array that points to these lists. And, the implicit model
corresponds to the reverse case, where we keep the pointer array in read-only memory and allow
swaps on the adjacency lists/arrays. A third alternative especially for the implicit model is to assume
that the input graph is represented as an adjacency array, i.e., adjacency lists are stored as arrays
instead of singly linked lists (see [23, 44, 54] for some results using this model); and we allow here
that any two elements in the adjacency array can be swapped. In this model, some of our algorithms
have improved performance in time.
1.3
Definitions, computational complexity and notations
We study some basic and fundamental graph problems in these models. In what follows we provide
the definitions and state the computational complexity of some these problems. For the DFS
problem, there have been two versions studied in the literature. In the lexicographically smallest
DFS or lex-DFS problem, when DFS looks for an unvisited vertex to visit in an adjacency list,
it picks the “first” unvisited vertex where the “first” is with respect to the appearance order in
5
1
1
2
4
2
1
3
5
4
5
2
3
3
5
4
(a)
2
4
4
1
2
5
1
2
3
5
4
(b)
Single Rotation
4
1
2
3
5
4
1
2
3
5
(c)
Figure 1: (a) An undirected graph G with 5 vertices and 8 edges. (b) A circular list representation of
G. To avoid cluttering the picture, we draw the vertices and the pointers to the next node separately
as opposed to a single node having two different fields in the circular list. (c) An illustration of a
single clockwise rotation in the circular list of vertex 4.
the adjacency list. The resulting DFS tree will be unique. In contrast to lex-DFS, an algorithm
that outputs some DFS numbering of a given graph, treats an adjacency list as a set, ignoring the
order of appearance of vertices in it, and outputs a vertex ordering T such that there exists some
adjacency ordering R such that T is the DFS numbering with respect to R. We say that such a
DFS algorithm performs general-DFS. Reif [66] has shown that lex-DFS is P-complete (with respect
to log-space reductions) implying that a logspace algorithm for lex-DFS results in the collapse
of complexity classes P and L. Anderson et al. [4] have shown that even computing the leftmost
root-to-leaf path of the lex-DFS tree is P-complete. For many years, these results seemed to imply
that the general-DFS problem, that is, the computation of any DFS tree is also inherently sequential.
However, Aggarwal et al. [1, 2] proved that the general-DFS problem can be solved much more
efficiently, and it is in RNC. Whether the general-DFS problem is in NC is still open.
As is standard in the design of space-efficient algorithms [11, 44], while working with directed
graphs, we assume that the graphs are given as in/out (circular) adjacency lists i.e., for a vertex v,
we have the (circular) lists of both in-neighbors and out-neighbors of v. We assume the word RAM
model of computation where the machine consists of words of size w in Ω(lg n) bits and any logical,
arithmetic and bitwise operation involving a constant number of words takes O(1) time. We count
space in terms of number of extra bits used by the algorithm other than the input, and this quantity
is referred as “extra space” and “space” interchangeably throughout the paper. By a path of length
d, we mean a simple path on d edges. By deg(x) we mean the degree of the vertex x. In directed
graphs, it should be clear from the context whether that denotes out-degree or in-degree. By a
BFS/DFS traversal of the input graph G, as in [7, 11, 23, 44, 53], we refer to reporting the vertices
of G in the BFS/DFS ordering, i.e., in the order in which the vertices are visited for the first time.
6
1.4
Our Results
Rotate Model:
For DFS, in the rotate model, we show the following in Sections 2.1, 2.3 and 2.5.
Theorem 1. Let G be a directed or an undirected graph, and ` ≤ n be the maximum depth of the
DFS tree starting at a source vertex s. Then in the rotate model, the vertices of G can be output in
(a) the lex-DFS order in O(m + n) time using n lg 3 + O(lg2 n) bits,
(b) a general-DFS order in O(m + n) time using n + O(lg n) bits, and
(c) a general-DFS order in O(m2 /n + m`) time for an undirected graph and in O(m(n + `2 )) time
for directed graphs using O(lg n) bits. For this algorithm, we assume that s can reach all other
vertices.
This is followed by the BFS algorithms where, in the rotate model, we show the following in
Sections 3.1 and 3.2.
Theorem 2. Let G be a directed or an undirected graph, and ` be the depth of the BFS tree starting
at the source vertex s. Then in the rotate model, the vertices of G can be output in a BFS order in
(a) O(m + n`2 ) time using n + O(lg n) bits, and
(b) O(m` + n`2 ) time using O(lg n) bits. Here we assume that the source vertex s can reach all
other vertices.
Implicit Model: In the implicit model, we obtain polynomial time implementations for lex-DFS
and general-DFS using O(lg n) bits. For lex-DFS, this is conjectured to be unlikely in ROM as the
problem is P-complete [66]. In particular, we show the following in Section 5.
Theorem 3. Let G be a directed or an undirected graph with a source vertex s and ` ≤ n be the
maximum depth of the DFS tree starting at s that can reach all other vertices. Then in the implicit
model, using O(lg n) bits the vertices of G can be output in
(a) the lex-DFS order in O(m3 /n2 +`m2 /n) time if G is given in adjacency list and in O(m2 lg n/n)
time if G is given in adjacency array for undirected graphs. For directed graphs our algorithm
takes O(m2 (n + `2 )/n) time if G is given in adjacency list and O(m lg n(n + `2 )) time if G is
given in adjacency array;
(b) a general-DFS traversal order in O(m2 /n) time if the input graph G is given in an adjacency
list and in O(m2 (lg n)/n + m` lg n)) time if it is given in an adjacency array.
In the implicit model, we can match the runtime of BFS from rotate model, and do better in
some special classes of graphs. In particular, we show the following in Section 6 .
Theorem 4. Let G be a directed or an undirected graph with a source vertex that can reach all
other vertices by a distance of at most `. Then in the implicit model, using O(lg n) bits the vertices
of G can be output in a BFS order in
(a) O(m + n`2 ) time;
(b) the runtime can be improved to O(m + n`) time if there are no degree 2 vertices;
(c) the runtime can be improved to O(m) if the degree of every vertex is at least 2 lg n + 3.
7
In sharp contrast, for space efficient algorithms for DFS in ROM, the landscape looks markedly
different. To the best of our knowledge, there are no DFS algorithms in general graphs in ROM that
use O(n1− ) bits. In fact, an implementation of DFS taking cn bits for c < 1 has been proposed as
an open problem by Asano et al. [7]. Similar to DFS, to the best of our knowledge, there are no
polynomial time BFS algorithms in ROM that use even O(n1− ) bits. On the other hand, we don’t
hope to have a BFS algorithm (for both undirected and directed graphs) using O(lg n) bits in ROM
as the problem is NL-complete [5].
Minimum Spanning Tree (MST). Moving on from DFS and BFS, we also study the problem
of reporting a minimum spanning tree (MST) of a given undirected connected graph G. We show
the following result in Section 7.
Theorem 5. A minimum spanning forest of a given undirected weighted graph G can be found
using O(lg n) bits and in
(a) O(mn) time in the rotate model,
(b) O(mn2 ) time in the implicit model if G is given in an adjacency list, and
(b) O(mn lg n) time in the implicit model when G is represented in an adjacency array.
Note that by the results of [65, 67], we already know logspace algorithms for MST in ROM
but again the drawback of those algorithms is their large time complexity. On the other hand, our
algorithms have relatively small polynomial running time, simplicity, making it an appealing choice
in applications with strict space constraints.
1.5
Techniques
Our implementations follow (variations of) the classical algorithms for BFS and DFS that use three
colors (white, gray and black), but avoid the use of stack (for DFS) and queue (for BFS). In the
rotate model, we first observe that in the usual search algorithms one can dispense with the extra
data structure space of pointers maintaining the search tree (while retaining the linear number
of bits and a single bit per vertex in place of the full unvisited/visited/explored array) simply by
rotating each circular adjacency lists to move the parent or a (typically the currently explored) child
to the beginning of the list to help navigate through the tree during the forward or the backtracking
step, i.e. by changing the pointer from the vertex to the list of its adjacencies by one node at a
time. This retains the basic efficiency of the search strategies. The nice part of this strategy is
that the total number of rotations also can be bounded. To reduce the extra space from linear
to logarithmic, it is noted that one can follow the vertices based on the current rotations at each
vertex to determine the visited status of a vertex, i.e. these algorithms use the rotate operation in
a non-trivial way to move elements within the lists to determine the color of the vertices as well.
However, the drawback is that to do so could require moving up (or down) the full height of the
implicit search tree. This yields super-linear rather than (near-) linear time algorithms.
In the implicit model, we use the classical bit encoding trick used in the development of the
implicit data structures [63]. We encode one (or two) bit(s) using a sequence of two (or three
respectively) distinct numbers. To encode a single bit b using two distinct values x and y with
x < y, we store the sequence x, y if b = 0, and y, x otherwise. Similarly, permuting three distinct
values x, y, z with x < y < z, we can represent six combinations. We can choose any of the four
combinations to represent up to 4 colors (i.e. two bits). Generalizing this further, we can encode
a pointer taking lg n bits using 2 lg n distinct elements where reading or updating a bit takes
constant time, and reading or updating a pointer takes O(lg n) time. This also is the reason for the
8
requirement of vertices with (high) degree at least 3 or 2 lg n + 3 for faster algorithms, which will
become clear in the description of the algorithms.
1.6
Consequences of our BFS and DFS results
There are many interesting and surprising consequences of our results for BFS and DFS in both the
rotate and implicit model. In what follows, we mention a few of them.
• For directed st-reachability,
√ as mentioned previously, the most space efficient polynomial time
algorithm [14] uses n/2Θ( lg n) bits. In sharp contrast, we obtain efficient (timewise) log-space
algorithms for this problem in both the rotate and implicit models (as a corollary of our directed
graph DFS/BFS results). In terms of workspace this is exponentially better than the best
known polynomial time algorithm [14] for this problem in ROM. For us, this provides one of
the main motivations to study this model. A somewhat incomparable result obtained recently
by Buhrman et al. [20, 56] where they designed an algorithm for directed st-reachability on
catalytic Turing machines in space O(lg n) with catalytic space O(n2 lg n) and time O(n9 ).
• Problems like directed st-reachability [5], distance [68] which asks whether a given G (directed,
undirected or even directed acyclic) contains a path of length at most k from s to t, are
NL-complete i.e., no deterministic log-space algorithm is known for these problems. But in our
(both the rotate and implicit) models, we design log-space algorithms for them. Assuming L 6=
NL, these results show that probably both our models with log-space are stronger than NL.
• The lex-DFS problem (both in undirected and directed graphs) is P-complete [66], and thus
polylogarithmic space algorithms are unlikely to exist in the ROM model. But we show
an O(lg n) space algorithm in the implicit model for lex-DFS. This implies that, probably
the implicit model is even more powerful than the rotate model. It could even be possible
that every problem in P can be computed using log-space in the implicit model. A result of
somewhat similar flavor is obtained recently Buhrman et al. [20, 56] where they showed that
any function in TC1 can be computed using catalytic log-space, i.e., TC1 ⊆ CSPACE(lg n).
Note that TC1 contains L, NL and even other classes that are conjectured to be different from
L.
• Our bounds for BFS and DFS in the rotate and implicit models immediately imply (with some
care) similar bounds, that are improvement over the best space bounds known so far in ROM,
for many applications of DFS/BFS. Moreover, as described before, any algorithm in the rotate
model can be implemented in ROM using extra O(n lg(m/n)) bits. Thus, our linear time DFS
algorithm in the rotate model can be implemented in ROM using O(n lg(m/n)) bits, matching
the bound of [23] for DFS. Using this DFS implementation, we can obtain improved space
efficient algorithms for various applications of DFS in ROM. This is discussed in Section 8.1.
• In Section 8.2 we present Logspace Approximation Scheme or LSAS ((1 ± ) approximation
algorithm running in logspace for every > 0) for a class of MSO-definable optimization
problems which are amenable to the Baker’s method [10] in locally bounded treewidth graphs
in both of our rotate and implicit models. No such algorithms are known in ROM as Baker’s
method requires to compute distance which is NL-complete. As BFS admits logspace algorithms
in our models, we can design such LSAS for these problems here.
• For a large number of NP-hard graph problems, the best algorithms in ROM run in exponential
time and polynomial space. We show that using just logarithmic amount of space, albeit using
exponential time, we can design algorithms for those NP-hard problems in both of our models
under some restrictions. This gives an exponential improvement over the ROM space bounds
9
for these problems. This is described in Section 8.3. In constrast, note that, no NP-hard
problem can be solved in the ROM model using O(lg n) bits unless P=NP.
2
DFS algorithms in the rotate model
In this section, we describe our space-efficient algorithms for DFS in the rotate model proving
Theorem 1.
2.1
Proof of Theorem 1(a) for undirected graphs
We begin by describing our algorithm for undirected graphs, and later mention the changes required
for directed graphs. In the normal exploration of DFS (see for example, Cormen et al. [30]) we use
three colors. Every vertex v is white initially while it has not been discovered yet, becomes gray
when DFS discovers v for the first time, and is colored black when it is finished i.e., all its neighbors
have been explored completely.
We maintain a color array C of length n that stores the color of each vertex at any point in
the algorithm. In the rest of the paper, when we say we scan the adjacency list of some vertex v,
what we mean is, we create a temporary pointer pointing to the current first element of the list
and move this temporary pointer untill we find the desired element. Once we get that element we
actually rotate the list so that the desired element now is at the front of the list. We start DFS at
the starting vertex, say s, changing its color from white to gray in the color array C. Then we scan
the adjacency list of s to find the first white neighbor, say w. We keep rotating the list to bring w
to the front of s’s adjacency list (as the one pointed to by the head V [s]), color w gray in the color
array C and proceed to the next step (i.e. to explore w’s adjacency list). This is the first forward
step of the algorithm. In general, at any step during the execution of the algorithm, whenever we
arrive at a gray vertex u (including the case when u’s color is changed from white to gray in the
current step), we scan u’s adjacency list to find the first white vertex. (i) If we find such a vertex,
say v, then we rotate u’s list to make v as the first element, and change the color of v to gray. (ii) If
we do not find any white vertex, then we change the color of u to black, and backtrack to its parent.
To identify u’s parent, we use the following lemma.
Lemma 1. Suppose w is a node that just became black. Then its parent p is the unique vertex in
w’s adjacency list which is (a) gray and (b) whose current adjacency list has w in the first position.
Proof. Among all the neighbors of w, some vertices are w’s children in the DFS tree, and the rest
of them are w’s ancestors, and among the ancestors, exactly one vertex is w’s parent in the DFS
tree. All the ancestors should have their currently explored (gray) child at the first position in their
adjacency list; and this current child would be different from w for all the ancestors except p (as
w was discovered from p). So, the second condition is violated for them. All of w’s children have
been fully processed earlier and have been colored black, and hence the first condition is violated
for them. Observe that, if w has a child, say k, which is a leaf in the DFS tree, it might happen
that k also has w at the first position in its current adjacency list, but, fortunately, k is black while
scanning w’s list. So for such vertices, the first condition gets violated. Only for w’s parent, which
is p here, both the conditions are satisfied.
So, the parent can be found by by scanning the w’s list, to find a neighbor p that is colored
gray such that the first element in p’s list is u. This completes the description of the backtracking
step. Once we backtrack to p, we find the next white vertex (as in the forward step) and continue
until all the vertices of G are explored. Other than some constant number of variables, clearly the
space usage is only for storing the color array C. Since C is of length n where each element has 3
possible values, C can be encoded using n lg 3 + O(lg2 n) bits, so that the i-th element in C can
10
be read and updated in O(1) time [38]. So overall space required is n lg 3 + O(lg2 n) bits. As the
algorithm systematically brings a white vertex to the front, makes it gray, and moves it to the end
after it becomes black, at most two full rotations of each of the list may happen (the second one to
determine that there are no more white vertices) resulting in a linear time lex-DFS algorithm. We
discuss the lex-DFS algorithm for the directed graphs below.
2.2
Proof of Theorem 1(a) for directed graphs
Recall that we have access to both the in-adjacency and the out-adjacency lists for each vertex w in
a directed graph G, hence we can use these lists separately for performing two steps of DFS. I.e.,
out-adjacency list is used for the exploration of DFS in the forward direction and the in-adjacency
list is used for finding parent of a node during the backtracking step. We provide the details below.
Similar to our algorithm for undirected graphs, in the forward direction, we scan the out-neighbor
list of w to find the next white neighbor and proceed. Once the out-neighbor list of w is fully
processed, we need to backtrack from w. Towards that we first have to identify w’s parent. In order
to do so we use the following lemma whose proof follows along the same lines as the Lemma 1 above.
Hence we omit the proof.
Lemma 2. Suppose w is a node that just became black. Then its parent p is the unique vertex in
w’s in-adjacency list which is (a) gray and (b) whose current out-adjacency list has w in the first
position.
Once we figure out w’s parent p, DFS backtracks to p, finds the next white neighbor (as done in
the forward step) and continues until all the vertices are exhausted. It is clear that this procedure
performs lex-DFS on a directed graph G correctly in linear time, and this completes the proof of
Theorem 1(a).
2.3
Proof of Theorem 1(b) for undirected graphs
To improve the space further, we replace the color array C with a bit array visited[1, . . . , n] which
stores a 0 for an unvisited vertex (white), and a 1 for a visited vertex (gray or black).
First we need a test similar to that in the statement of Lemma 1 without the distinction of gray
and black vertices to find the parent of a node. Due to the invariant we have maintained, every
internal vertex of the DFS tree will point to (i.e. have as first element in its list) its last child. So
the nodes that could potentially have a node w in its first position are its parent, and any leaf
vertex. Hence we modify the forward step in the following way.
Whenever we visit an unvisited vertex v for the first time from another vertex u (hence, u is the
parent of v in the DFS tree and u’s list has v in the first position), we, as before, mark v as visited
and in addition to that, we rotate v’s list to bring u to the front (during this rotation, we do not
mark any intermediate nodes as visited). Then we continue as before (by finding the first unvisited
vertex and bringing it to the front) in the forward step. Now the following invariants are easy to see
and are useful.
Invariants: During the exploration of DFS, in the (partial) DFS tree
1. any internal vertex has the first element in its list as its current last child; and
2. for any leaf vertex of the DFS tree, the first element in its list is its parent.
The first invariant is easy to see as we always keep the current explored vertex (child) as the first
element in the list. For leaves, the first time we encounter them, we make its parent as the first
element in the forward direction. Then we discover that it has no unvisited vertices in its list, and
so we make a full rotation and bring the parent to the front again. The following lemma provides a
test to find the parent of a node.
11
Lemma 3. Let w be a node that has just become black. Then its parent p is the first vertex x in
w’s list which is marked 1 in the visited array, and whose current adjacency list has w in the first
position.
Proof. From the invariants we observed, the nodes that can potentially have w in the first position
of their lists are its parent and its children that happen to be leaves. But in w’s list, as we began the
exploration of its neighbors starting from its parent, its parent will appear first before its children.
Hence the first node in w’s list which has w in the first position must be its parent.
Once we backtrack to p, we find the next white vertex, and continue until all the vertices of G
are explored. Overall this procedure takes linear time. As we rotate the list to bring the parent of a
node, before exploring its white neighbors, we are not guaranteed to explore the first white vertex in
its original list, and hence we loose the lexicographic property. We provide our DFS algorithm for
directed graphs below.
2.4
Proof of Theorem 1(b) for directed graphs
For performing DFS in directed graphs using n + O(lg n) bits, we don’t even need to apply the
modifications as we did for the undirected graphs during the forward direction, and we can essentially
use the same forward step idea as used for lex-DFS in undirected graphs of Section 2.1. We provide
the details below. When we arrive at a previously unvisited vertex v from the vertex u (hence u is
the parent of v in the DFS tree), we rotate the in-neighbor list of v to bring u to the front and u
stays there during the entire course of the exploration. Thus we maintain the invariant that for
any visited node v, the first element in its in-neighbor list is its parent in the DFS tree. Now the
algorithm scans v’s adjacency list to find its unvisited neighbor. (i) If we find such a vertex, say w,
then we rotate v’s list to make w as the first element, and mark w visited. (ii) If we do not find
any such unvisited neighbor of v, then DFS needs to backtrack to its parent. From the invariant
we maintain in the in-neighbor list of every visited vertex, this is easy. All we need to do is to see
the first entry in v’s in-neighbor list to retrieve its parent u and then continue from u. Overall this
procedure takes linear time. Upon closer inspection, it can be seen that, as we are not modifying
the ordering of the vertices in the out-neighbor lists in the forward direction (in contrast with the
undirected graph algorithm of Section 2.3), this procedure actually traverses the directed graph G
in lex-DFS ordering. This completes the proof of Theorem 1(b).
2.5
Proof of Theorem 1(c) for undirected graphs
Now to decrease the space to O(lg n), we dispense with the color/visited array, and give tests to
determine white, gray and black vertices. For now, assume that we can determine the color of a
vertex. The forward step is almost the same as before except performing the update in the color
array. I.e., whenever we visit a white vertex v for the first time from another vertex u (hence u is the
parent of v), we rotate vs list to bring u to the front. Then we continue to find the first white vertex
to explore. We maintain the following invariants. (i) any gray vertex has the first element in its list
as its last child in the (partial) DFS tree; (ii) any black vertex has its parent as the first element in
its list. We also store the depth of the current node in a variable d, which is incremented by 1 every
time we discover a white vertex and decremented by 1 whenever we backtrack. We maintain the
maximum depth the DFS has attained using a variable max. At a generic step during the execution
of the algorithm, assume that we are at a vertex x’s list, let p be x’s parent and let y be a vertex in
x’s list. We need to determine the color of y and continue the DFS based on the color of y. We use
the following characterization.
12
Lemma 4. Suppose the DFS has explored starting from a source vertex s, up to a vertex x at level
d. Let p be x’s parent. Note that both s and x are gray in the normal coloring procedure. Let max be
the maximum level of any vertex in the partial DFS exploration. Let y be a vertex in x’s list. Then,
1. y is gray (i.e., (x, y) is a back edge, and y is an ancestor of x) if and only if we can reach y
from s by following through the gray child (which is at the front of a gray node’s list) path in
at most d steps.
2. y is black (i.e., (x, y) is a back edge, and x is an ancestor of y) if and only if
• there is a path P of length at most (max − d) from y to x (obtained by following through
the first elements of the lists of every vertex in the path, starting from y), and
• let z be the node before x in the path P . The node z appears after p in x’s list.
3. y is white if y is not gray or black.
Proof. The test for gray and white vertices is easy to see. The vertex y is black implies that y is a
descendant of x in the partially explored DFS tree. This means that there is a path of length at
most (max − d) (obtained by following the parent which is in the first element of the adjacency list)
from x to y through an already explored child z . By the way we process x’s list, we first bring the
parent to the front of the list, and then explore the nodes in sequence, and hence z, the explored
neighbor of x must appear after p in x’s list. Conversely, the unexplored neighbors of x appear
before p in x’s list.
Now, if we use the above claim to test for colors of vertices, testing for gray takes at most d
steps. Testing for black takes at most (max − d) steps to find the path, and at most deg(x) steps to
determine whether p appears before. Thus for each vertexPin x’s list, we spend time proportional to
max + deg(x). So, the overall runtime of the algorithm is v∈V deg(v)(deg(v) + `) = O(m2 /n + m`),
where ` is the maximum depth of DFS tree. Maintaining the invariants for the gray and black
vertices are also straightforward. We provide the details of our log-space algorithm for directed
graphs below.
2.6
Proof of Theorem 1(c) for directed graphs
We describe our O(lg n) bits algorithm for directed graphs in the rotate model. More specifically, we
give a DFS algorithm to output all vertices reachable by a directed path from the source vertex s.
If we assume that s can reach all vertices, we get to output all vertices. In the preprocessing step,
the algorithm spends O(m) time to bring the minimum valued neighbor (denote it by min) in the
out-neighbor list of every vertex by rotation (hence we loose the lexicographic DFS property). For
now assume that we can determine the color of a vertex. Given this, in the forward direction, when
DFS arrives at a previously unvisited vertex v from the vertex u (hence u is the parent of v in the
DFS tree), we rotate the in-neighbor list of v to bring u to the front and u stays there during the
entire course of the exploration. Also in u’s out-neighbor list, v is made the first location. Hence we
maintain the following invariants.
Invariants: During the exploration of DFS, in the (partial) DFS tree
1. gray vertices have their current last child in the first location of their out-neighbor lists;
2. all the visited (i.e., gray and black) vertices have their parent in the first location of their
in-neighbor lists.
We also keep track of the depth of the current node (i.e., the last gray vertex in the gray path of
the DFS tree) in a variable d, which, as before, is incremented by 1 every time DFS visits a white
vertex and decremented by 1 whenever DFS backtracks. We also store the maximum depth the DFS
tree has attained so far in a variable max. At a generic step during the execution of the algorithm,
assume that we are at a vertex x’s list, let p be x’s parent (which can be found from the way x is
13
visited by a forward or a backtracking step using the invariants being maintained) and let y be a
vertex in x’s list. We need to determine the color of y and continue the DFS based on the color of y
and maintain the invariants. We use the following characterization.
s
s
y
z
c
x
s
s
z
x
p
y
x
p
x
y
y
(a)
d
(b)
(c)
(d)
Figure 2: Illustration of the different cases of the possible positions of the vertex y when DFS
considers the directed edge (x, y) at some intermediate step. Suppose the root of the DFS tree is
the vertex s and the curvy path starting from s and going straight below through x is the current
gray path in the DFS tree. Intuitively all the vertices on the left hand side of the path are black,
and right hand side are white and yet to be explored. From left to the right are cases when (x, y) is
(a) back edge, (b) cross edge, (c) forward edge, and (d) tree edge.
Lemma 5. Suppose the DFS has explored starting from a source vertex s up to a vertex x at level d.
Let p be x’s parent. Note that both s and x are gray in the normal coloring procedure. Let max be
the maximum level of any vertex in the partial DFS exploration. Let y be a vertex in x’s list. Then,
1. y is gray (i.e., (x, y) is a back edge and y is an ancestor of x) if and only if we can reach y
from s following the gray child (which is in the first location of each of the out-neighbor lists
of gray nodes) path in at most d steps.
2. y is black if and only if any of the following happens.
• There is a path P of length at most (max − d) from y to x obtained by following the
first elements of the in-neighbor lists of every vertex in the path P starting from y. This
happens when (x, y) is a forward edge, and x is an ancestor of y.
• There is a path P of length at most max from y to a gray vertex z 6= x (obtained by
following through the first elements of the in-neighbor lists of every vertex starting from
y) which is the first gray vertex in the path. Let c be the node before z in the path P ,
then c must appear after min in z’s list (this happens when (x, y) is a cross edge).
3. The vertex y is white if it is not black or gray(i.e., (x, y) is the next tree edge with x being the
parent of y in the DFS tree).
Proof. See Figure 2 for a picture of all the cases. The test for gray and white vertices is easy to see.
From a vertex x, there could be two types of outgoing edges to a black vertex y. When (x, y)
is a forward edge, y is a descendant of x and hence there must exist a path P of length at most
(max − d) (obtained by following the parent which is in the first location of the in-neighbor list of
every vertex in P , starting from y) from y to x through an already explored child t of x. In the
14
other case, when (x, y) is a cross edge, y has already been discovered and explored completely before
DFS reaches to x. Hence there must exist a gray ancestor z of x (z could be x) such that y belongs
to the subtree rooted at z in the DFS tree. Thus, from y’s in-neighbor list if we follow the path
starting with y’s parent for at most max steps, we must hit the gray path and the first vertex we
come across is z. Let c be the node before z in the path. By the way we process z’s out-neighbor
list, we first bring the min to the front of the list, and then explore the other neighbor nodes in
sequence, and hence c, the explored neighbor of z must appear after min in z’s list.
For the converse, suppose y is a white vertex. Either we never reach a gray vertex in max steps
(and we will correctly determin its color in this case) or we reach x or x’s ancestor z from y following
through the (spurious) first vertices of the in-neighbor list of a white vertex y. Note that the parent
of a white vertex is white or gray and it can never be black. Hence z’s child in the path is white.
Hence that child will appear before min in z’s list.
Given the above test, if y turns out to be white, the edge (x, y) is added to the DFS tree, and
y now becomes the current gray vertex. Note that maintaining the invariants are straightforward.
Also, when any vertex v has been completely explored by DFS, we retrieve its parent from the
in-neighbor list to complete the backtracking step. This procedure is continued until DFS comes
back to the source vertex s. We stop at this point. This is because, note that, our proof breaks
down in the case when DFS in a directed graph produces a forest and some cross edges go across
these trees. In that case, if we follow the path starting from y’s parent, we would reach the root
of the DFS tree containing y and this is different from the tree where x belongs to. As we cannot
maintain informations regarding all such roots of these previously explored DFS trees, we might
spuriously conclude that y is unvisited even though it is not the case. Thus our algorithm produces
the DFS tree containing only the vertices reachable from the source vertex s via some directed path
in G. We leave open the case for desigining such logspace algorithm for the general directed graphs.
Given the above lemma, if y turns out to be white, the edge (x, y) is added to the DFS tree,
and y now becomes the current gray vertex. Note that maintaining the invariants are easy. When
any vertex v has been completely explored by DFS, we retrieve its parent from the in-neighbor list
to complete the backtracking step. This procedure is continued until DFS comes back to the source
vertex s. We stop at this point and we have outputted all vertices reachable from s.
To analyse the running time of our algorithm observe that testing for gray takes at most d steps.
Testing for black takes, in the worst case, at most max steps to find the path, and at each step of
the path, we take d time to test whether the new node is gray. Once we reach a gray vertex, we
spend at most deg(z) steps to determine whether c appears before min. Thus for each vertex in x’s
list, we spend time proportional to d + (d.max) + deg(z) time. As z (which
P is independent of x)
can have degree at most n. Thus, the overall runtime of the algorithm is v∈V deg(v)(d + d` + n)
which is O(m(n + (1 + `)`) which is O(m(n + `2 )), where ` is the maximum depth of DFS tree.
3
3.1
BFS algorithms in the rotate model
BFS using n + O(lg n) bits–Proof of Theorem 2(a)
It is well-known that BFS actually computes the shortest path lengths in unweighted undirected or
directed graph G from a given source vertex s ∈ V to every vertex v ∈ V that is reachable from s.
I.e., if a vertex v belongs to the d-th level in the BFS tree (assuming the root s is at zero-th level),
then we know that the length of the shortest path from s to v is d. We use this crucially to design
our BFS algorithms. We use a bit array visited[1, · · · , n] that stores a 0 for an unvisited vertex,
and 1 for a visited vertex. We also maintain a counter dist which stores the level of the vertex that
is currently being explored in the BFS algorithm.
15
We start by setting visited[s] = 1, and initializing the counter dist to 0. At the next step, for
every unvisited neighbor v of s, we rotate their adjacency list so that s appears as the first element
in v’s list, set visited[v] = 1, and output v. This step ensures that for each visited vertex, its parent
is at the front of its adjacency list. We refer to this front element in the adjacency list of a visited
vertex as its parent pointer. (Also, once we set the parent pointer for a vertex, we will not rotate
its adjacency list in the remaining part of the algorithm.) Once the root s’s list is fully processed
as above, the dist is incremented to 1. The next step in the algorithm is to find all the vertices in
the first level and mark all their unvisited neighbors as visited. As we haven’t stored these vertices
(in the first level), the challenge is to find them first. We use the following claim, to find the level
number of a visited vertex. The proof easily follows from the fact that the parent pointers are set
for all the visited vertices, and that all the ancestors of a visited vertex are also visited.
Claim 1. If the BFS has been explored till distance d from the source, then for any k ≤ d, a vertex
x marked visited is in level k if and only if we can reach the source s in exactly k steps by following
through their parent pointers. Thus determining if a visited vertex is in level d takes at most d steps.
So now we continue the BFS by scanning through the vertices, finding those vertices in level d
(using the above claim by spending d steps for each vertex), and marking their unvisited neighbors
visited, and making in their adjacency lists, their parent vertex as the first element, and incrementing
dist. We stop our algorithm when we discover no new unvisited vertex while exploring any level.
The correctness of the procedure and the space used
P by the algorithm are clear. To analyze the
runtime, note that the time spent at level d is nd + i∈V (d) deg(i) where V (d) is the set of vertices
in level d and deg(i) is the degree of vertex i. Summing over all levels, we get a runtime O(m + n`2 ),
where ` is the depth of the BFS tree. To handle directed graphs, we follow the outneighbor list as
we go down, and we set up the parent at the first position in the in-neighbor list of every vertex
v. To verify if v belongs to the d-th level, we take d steps from v by following the parent pointers
in the in-neighbor lists of the (visited) vertices along the path, and check if we reach s at the end.
This completes the proof of Theorem 2(a).
3.2
BFS using O(lg n) bits–Proof of Theorem 2(b)
To reduce the space to O(lg n) bits, we dispense with the color array and explain how to determine
visited and unvisited vertices. Assume that we can determine this in constant time. Our first
observation is that Claim 1 is true even for unvisited vertices even though the first vertex in the
adjacency list of unvisited vertices can be an arbitrary vertex (not necessarily referring to their
parent in the BFS tree). However, we know (by the property of BFS) that no unvisited vertex is
adjacent to a vertex in level less than d, and hence they can not reach s by a path at most d. Using
the same argument, we can show that
Claim 2. If vertices up to level d have been explored and visited, then a vertex x is a visited vertex
if and only if by following through the parent pointers, x can reach s by a path of length at most d.
Furthermore, a vertex is in level d if and only if we can reach s by a path of length exactly d by
following through the parent pointers.
Thus, to check whether a vertex is visited, we spend O(d) time when exploring vertices at level
d instead of O(1) time P
when the visited array was stored explicitly. Hence, the total the time spent
at level d is O(nd + d i∈V (d) deg(i)), where the first term gives the time to find all the visited
vertices at level d, and the second term gives the time to explore those vertices (i.e., going through
their neighbor lists, identifying the unvisited neighbors and setting their parent pointers). Summing
over all levels, we get a total runtime of O(n`2 + m`). Note that this algorithm works only when the
16
input undirected graph is connected as Claim 2 breaks down if there are more than one component.
The modifications to handle the directed graphs are similar to those for the directed graph BFS
algorithm. This proves Theorem 2(b).
4
Simulation of algorithms for rotate model in the implicit model
The following result captures the overhead incurred while simulating any rotate model algorithm in
the implicit model. Most of our algorithms in the implicit model use these simulations often with
some enhancements and tricks to obtain better running time bounds for some specific problems.
Theorem 6. Let D be the maximum degree of a graph G. Then any algorithm running in t(m, n)
time in the rotate model can be simulated in the implicit model in (i) O(D · t(m, n)) time when
G is given in an adjacency list, and (ii) O(lg D · t(m, n)) time when G is given in an adjacency
array. Furthermore, let rv (m, n) denote the number of rotations made in v’s (whose degree is
dv ) list, and
P f (m, n) be the remaining number of operations. Then any algorithm running in
t(m, n) = v∈VPrv (m, n) + f (m, n) time in the rotate model can be simulated in the the implicit
model
P in (i) O( v∈V rv (m, n) · dv + f (m, n)) time when G is given in an adjacency list, and (ii)
O( v∈V rv (m, n) lg dv + f (m, n)) time when G is given in an adjacency array.
Proof. We can implement a single rotate operation that moves the head of the list by one position
(which is assumed to be a unit-cost operation in the rotate model) in the implicit model by moving
all the elements in the adjacency list circularly. If dv is the degree of a vertex v, then performing a
rotation of the adjacency list of v can be implemented in O(dv ) time in the implicit model. Thus, if
we have an algorithm in the rotate model that takes t(m, n) time, then it can be implemented in
O(D · t(m, n)) time in the implicit model, where D is the maximum degree of the graph. One can get
a better runtime by analysing the algorithm for the rotate model more carefully. In particular, if the
runtime of the algorithm in the rotate model can be expressed as r(m, n) + f (m, n), where r(m, n)
is the number of rotations performed and f (m, n) is the remaining number of operations, then the
algorithm canPbe implemented in the implicit model in O(D · r(m, n)) + f (m, n) time. Furthermore,
if r(m, n) ≤ v∈V rv (m, n) where rv (m, n) is the number of rotations
P made in v’s list, then the
runtime of the algorithm in the implicit model can be bounded by O( v∈V rv (m, n) · dv + f (m, n)).
If the input graph is given in the adjacency array representation and if the ordering of the
elements in the adjacency array is allowed to be changed, then one can simulate the rotate operation
even faster. The main idea is to simulate the algorithm for the rotate model after sorting the
adjacency arrays of each vertex. Using an in-place linear time radix sorting algorithm [51], sorting
all the adjacency arrays can be done in O(m) time. Next, we note that in the rotate model, the
head points to an element in an arbitrary position (called the front element) in the adjacency list of
any vertex and the unit operation moves it one position. Thus, it is enough to show how to access
the next element in sorted order. We maintain the following invariant: if the first element is the i-th
element in sorted order, then the adjacency array consists of the sorted array of all the adjacent
vertices, with the first element swapped with the i-th element. To bring the (i + 1)-st element to
the front, we first perform a binary search for the i-th element which is in the first position in
the ‘almost sorted’ adjacency array to find the position i, and move the elements appropriately
to maintain the invariant (with the (i + 1)-st element at the front). This takes O(lg dv ) time to
simulate the rotation of the adjacency array of a vertex v with degree dv . Thus, if we have an
algorithm in the rotate model that takes t(m, n) time, then it can be implemented in O(lg D · t(m, n))
time in the implicit
model. Moreover, if the runtime of the algorithm in the rotate model can be
P
expressed as v∈V rv (m, n) + f (m, n), where rv (m, n) is an upper bound on the number of rotations
performed on vertex v and f (m, n) is the remaining
number of operations, then the algorithm can
P
be implemented in the implicit model in O( v∈V rv (m, n) lg dv + f (m, n)) time.
17
5
DFS algorithms in the implicit model—proof of Theorem 3
To obtain a lex-DFS algorithm, we implement the O(lg n)-bit DFS algorithm in the rotate model,
described in Section 2.5, with a simple modification. First, note that in this algorithm (in the rotate
model), we bring the parent of a vertex to the front of its adjacency list (by performing rotations)
when we visit a vertex for the first time. Subsequently, we explore the remaining neighbors of the
vertex in the left-to-right order. Thus, for each vertex, if its parent in the DFS were at the beginning
of its adjacency list, then this algorithm would result in a lex-DFS algorithm. Now, to implement
this algorithm in the implicit model, whenever we need to bring the parent to the front, we simply
bring it to the front without changing the order of the other neighbors. Subsequently, we simulate
each rotation by moving all the elements in the adjacency
P list circularly. As mentioned in Section 4,
this results in an algorithm whose running time P
is O( v∈V dv (dv + `) · dv ) = O(m3 /n2 + `m2 /n) if
the graph is given in an adjancecy list and in O( v∈V dv (dv + `) · lg dv ) = O(m2 (lg n)/n + m` lg n))
when the graph is given in the form an adjacency array. This proves Theorem 3(a) for undirected
graphs. The results for the directed case follow from simulating the corresponding results for the
directed graphs.
To prove the result mentioned in Theorem 3(b), we implement the linear-time DFS algorithm of
Theorem
the rotate model that uses n + P
O(lg n) bits. This results in an algorithm that runs
P 1 for
2
2
in O( v∈V dv + n) = O(m /n) time (or in O( v∈V dv lg dv + n) = O(m lg m + n) time, when the
graph is given as an adjacency array representation), using n + O(lg n) bits. We reduce the space
usage of the algorithm to O(lg n) bits by encoding the visited/unvisited bit for each vertex with
degree at least 2 within its adjacency list (and not maintaining this bit for degree-1 vertices). We
describe the details below.
Whenever a node is visited for the first time in the algorithm for the rotated list model, we bring
its parent to the front of its adjacency list. In the remaining part of the algorithm, we process each
of its other adjacent vertices while rotating the adjacency list, untill the parent comes to the front
again. Thus, for each vertex v with degree dv , we need to rotate v’s adjacency list O(dv ) times. In
the implicit model, we also bring the parent to the front when a vertex is visited for the first time,
for any vertex with degree at least 3. We use the second and third elements in the adjacency list to
encode the visited/unvisited bit. But instead of rotating the adjacency list circularly, we simply
scan through the adjacency list from left to right everytime we need to find the next unvisited
vertex in its adjacency list. This requires O(dv ) time for a vertex v with degree dv . We show how
to handle vertices with degree at most 2 separately.
As before, we can deal with the degree-1 vertices without encoding visited/unvisited bit as we
encounter those vertices only once during the algorithm. For degree-2 vertices, we initially (at
preprocessing stage) encode the bit 0 using the two elements in their adjacency arrays - to indicate
that they are unvisited. When a degree-2 vertex is visited for the first time from a neighbor x, we
move to its other neighbor – continuing the process as long as we encounter degree-2 vertices until
we reach a vertex y with degree at least 3. If y is already visited, then we output the path consisting
of all the degree-2 vertices and backtrack to x. If y is not visited yet, then we output the path upto
y, and continue the search from y, and after marking y as visited. In both the cases, we also mark
all the degree-2 nodes as visited (by swapping the two elements in each of their adjacency arrays).
During the preprocessing, for each vertex with degree at least 3, we ensure that the second
and third elements in its adjacency list encode the bit 0 (to mark it unvisited). We maintain the
invariant that for any vertex with degree at least 3, as long as it is not visited, the second and
third elements in its adjacency array encode the bit 0; and after the vertex is visited, its parent
(in the DFS tree) is at the front of its adjacency array, and the second and third elements in its
adjacency array encode the bit 1. Thus, when we visit a node v with degree at least 3 for the
18
first time, we bring its parent to the front, and then swap the second and third elements in the
adjacency
list, if needed, to mark it as visited. The total running time of this algorithm is bounded
P
by v∈V d2v = O(m2 /n).
We can implement the above DFS algorithm even faster when the input graph is given in an
adjacency array representation. We deal with vertices with degree at most 2 exactly as before. For
a vertex v with degree at least 3, we bring its parent to the front and swap the second and third
elements to mark the node as visited (as before) whenever v is visited for the first time. We then sort
the remaining elements, if any, in the adjacency array, in-place (using the linear-time in-place radix
sort algorithm [51]), and implement the rotations on the remaining part
P of the array as described in
Section 4. The total running time of this algorithm is bounded by v∈V dv lg dv = O(m lg m + n).
This completes the proof of Theorem 3(b).
6
BFS algorithms in the implicit model—proof of Theorem 4
Before getting into the technical details, we first outline the main ideas involved in proving Theorem 4.
One can simulate the BFS algorithm of Theorem 2(b) (for the rotate model) in the implicit model
using the simulation described in Section 4. Since these BFS algorithms scan through each of the
adjacency lists/arrays at most twice during the algorithm, there won’t be any slowdown in the
runtime. This results in an algorithm running in O(m` + n`2 ) time, using O(lg n) bits. To improve
the running time, we simulate the algorithm of Theorem 2(a) using the trick of encoding the visited
bit of each vertex in its adjacency list instead of storing the visited array explicitly. This requires
special attention to degree-1 and degree-2 vertices along with few other technical issues which are
dealt in the proof given next.
Proof. Here we give the full proof of Theorem 4. In particular, we provide all the details of the
case when the degree of each vertex is at least 3, and in that case, we show that we can implement
the BFS algorithm using 4 colors of [11], by encoding the 4 color of a vertex using the first three
elements in its adjacency list, resulting in an algorithm that takes O(m + n`) time. Moreover, when
the degree of every vertex is at least 2 lg n + 3, then we show that the above algorithm can be
implemented more efficiently, resulting in an algorithm that takes O(m) time. Details follow. One
can simulate the BFS algorithm of in Item 2 of Theorem 2 (for the rotate model) in the implicit
model using the simulation described in Section 4. Since these BFS algorithms scan through each of
the adjacency lists/arrays at most twice during the algorithm, there won’t be any slowdown in the
runtime. This results in a BFS algorithm that runs in O(m` + n`2 ) time, using O(lg n) bits.
To improve the running time further, we simulate the algorithm of Theorem 2(a). But instead
of storing the visited array explicitly, we encode the visited bit of each vertex in its adjacency list,
as explained below, resulting in an algorithm that takes O(m + n`2 ) time, using O(lg n) bits. For a
vertex x with degree at least 3, we encode its visited bit using the second and third elements in its
adjacency list. To set the parent pointer for x (when it is visited for the first time), we bring its
parent to the front, and move the second and third elements, if necessary, to encode the visited bit.
We now describe how to deal with the degree-1 and degree-2 vertices.
First, observe that in the original BFS algorithm, we can simply output any degree-1 vertex,
when it is visited for the first time. Thus, we need not store the visited bit for degree-1 vertices. For
degree-2 vertices, we encode the visited bit using the two neighbors. We do not bring its parent to
the front, as we do for other vertices. Whenever we need to check whether a visited degree-2 vertex
is at depth d, we follow parent pointers from both the neighbors - if one of them does not exist,
then the other one is its parent - for a distance of length at most d. (While following the parent
pointers from a vertex to the root, it is easy to find its parent - since there is only one alternative to
follow.) Thus, the first part of the theorem follows from Theorem 2(a), with the above modification.
19
To improve the runtime further, we implement the BFS algorithm using 4 colors by [11], where
all the unvisited vertices are white, all the visited, yet unfinished vertices of the two consecutive
layers of BFS are gray1 and gray2 respectively, and all the vertices which are completely explored
are black. Suppose the degree of every vertex in the given graph is at least three. In this case, we
can encode the color of any vertex by permuting the first three elements in its adjacency array
appropriately. This enables us to retrieve (or modify) the color of any vertex in constant time by
reading (or modifying) the permuted order of the first elements in its adjacency array.
Since we haven’t stored the gray1 or gray2 vertices in a queue (as in the standard BFS), we scan
through the entire vertex set (in the increasing order of their labels), and when we find any vertex v
colored gray1, we color all its white neighbors with gray2, and color v itself with black. We call this
an exploration phase. The time taken for each exploration phase is equal to the sum of the degrees
of all the gray1 vertices at the beginning of the phase, plus O(n). At the end of each exploration
phase, we change the colors of all gray2 vertices to gray1 and also output them, using an additional
O(n) time. We call this a consolidation phase. We need to repeat the exploration and consolidation
phases for ` times before all the nodes are colored black, where ` is the height of the BFS tree. Thus
the overall time taken by this procedure can be bounded by O(m + n`). This proves the second
part of the theorem.
If every vertex has degree at least 2 lg n + 3, then we can perform BFS in O(m) time (note that
m ≥ n lg n in this case) – by encoding the current set of gray1 vertices as a linked list by using 2 lg n
vertices to encode (a pointer to) the next vertex in the list. The time to read all the gray1 vertices
in a phase when there are k vertices colored gray1 becomes O(k lg n) instead of O(n). This results
in O(m + n lg n) time which is O(m). This proves the third part of the theorem.
7
Minimum Spanning Tree
In this section, we start by giving an in-place implementation of the Prim’s algorithm [30] to find
a minimum spanning tree of a given weighted undirected graph in the rotate model. Here we are
given a weight function w : E → Z. We also assume that the weights of non-edges are ∞ and that
the weights can be represented using O(lg n) bits. The input representation also changes slightly to
accommodate these weights. Now each element of the circular linked list has three fields, (a) the
vertex label, (b) the weight, and (c) the pointer to the next vertex respectively. In what follows,
when we talk about computing a minimum spanning tree, what we mean is reporting the edges of
such a tree. Our result is the following,
Theorem 7. A minimum spanning tree of a given undirected weighted graph G can be found using
O(lg n) bits and in O(mn) time in the rotate model.
Proof. Our rotate model algorithm basically mimics Prim’s algorithm with a few tweaks. Prim’s
algorithm starts with initializing a set S with a vertex s. For every vertex v not in S, it finds
and maintains d[v] = min{w(v, x) : x ∈ S} and π[v] = x where w(v, x) is the minimum among
{w(v, y) : y ∈ S}. Then it repeatedly deletes the vertex with the smallest d value from V − S
adding it to S. Then the d values are updated by looking at the neighbors of the newly added
vertex. To implement this procedure using just extra O(lg n) bits of space, first we need to find a
way to mark/unmark a vertex v if it has been taken into S without using n bits explicitly. The
way we do this is as follows. In the preprocessing step, the algorithm spends O(m) time to bring
the minimum valued neighbor (denote it by min) in the neighbor list of every vertex v by rotation.
Subsequently we would attach the following meaning with the position of min in the list of any
vertex v. If the first element in the list of any vertex v is min, this means that v is not taken into S
so far during the execution of the algorithm, otherwise it belongs to S. This way we can store the
20
information regarding the status of any vertex v without using any extra space but actually figuring
out this information takes time proportional to the degree of v. Note that for vertices having degree
one, we cannot determine exactly its status correctly by this method. But a simple fact we can use
here. If a vertex z has degree one (say its neighbor is y), then the edge (y, z) is always a part of the
minimum spanning tree. Hence, after the preprocessing step, we can output all such edges at once
and embark on executing the rest of the algorithm.
The algorithm initializes the set S with the starting vertex s, and goes to its list to give a
rotation so that min does not stay in the first location in s’s list. We call this step as the marking
step. This is followed by finding the smallest weight neighbor (say u) of s. According to Prim’s
algorithm, u should now move to S. We achieve the same by marking u i.e., going to u’s list to give
a rotation to move min from the first location to indicate that u belongs to S now and subsequently
continue to in u’s list find its smallest weight neighbor and repeat. Thus at any generic step of
the algorithm, we go over the list of unmarked vertices (i.e., those vertices having min at the first
position of their respective lists) and collect the minimum weight vertex (say t), and t is then marked
and we continue until all the vertices are marked.
Clearly this method returns all the correct edges of a minimum spanning tree of G. Space bound
of this algorithm is easily seen to be O(lg n) bits for keeping a few variables. Total time spent by
algorithm can be bounded by O(mn) where at the preprocessing step, it spends O(m) time and
after the preprocessing, for reporting each edges of the minimum spanning tree, in the worst case,
the algorithm spends O(m) time, hence O(mn) is the bound on the running time.
As mentioned in Section 4, simulating this algorithm in the implicit model would result in an
algorithm having running time O(mn.dv )=O(mn2 ) if the graph is given in an adjacency list and in
O(mn. lg dv )=O(mn lg n) when the graph is represented in an adjacency array. Hence, we have the
following,
Theorem 8. In the implicit model a minimum spanning tree of a given undirected weighted graph
G can be found using O(lg n) bits and
1. O(mn2 ) time if the graph is given in an adjacency list, and
2. O(mn lg n) time when the graph is represented in an adjacency array.
8
Consequences
In this section, we provide some applications/consequences of our results that we derived in the
earlier sections.
8.1
Improved algorithm for DFS and applications in ROM
In this section, we show, using a little more space, how to simulate any rotate model algorithm in
the read-only model. This results in improved space-efficient algorithms for various fundamental
graph problems in the read-only model. Observe that, the only modification of the input that we
do in our rotate model algorithm is to make the head pointer point to an arbitrary element in the
adjacency list (instead of a fixed element) at various times. To simulate this in read-only memory,
we can simply maintain a pointer in each of the lists in the adjacency list. The resources required
to store and update such pointers is proven in the following lemma [23]. We provide the proof here
for completeness.
Lemma 6 ([23]). Given the adjacency list representation of a directed or an undirected graph G
on n vertices with m edges, using O(m) time, one can construct an auxiliary structure of size
O(n lg(m/n)) bits that can store a “pointer” into an arbitrary index within the adjacency list of
each vertex. Also, updating any of these pointers (within the adjacency list) takes O(1) time.
21
Proof. We first scan the adjacency list of each vertex and construct a bitvector B as follows: starting
with an empty bitvector B, for 1 ≤ i ≤ n, if di is the length of the adjacencyP
array of vertex vi
(i.e., its degree), then we append the string 0dlg di e−1 1 to B. The length of B is ni=1 dlg di e, which
is bounded by O(n lg(m/n)). We construct auxiliary structures to support select queries on B in
constant time [60]. We now construct another bitvector P of the same size as B, which stores the
pointers into the adjacency array of each vertex. The pointer into the adjacency array of vertex
vi is stored using the dlg di e bits in P from position select(i − 1, B) + 1 to position select(i, B),
where select(0, B) is defined to be 0. Now, using select operations on B and using constant time
word-level read/write operations, one can access and/or modify these pointers in constant time.
To actually get to the element in the list, we need the graph to be represented as what is referred
as adjacency array [44]. Here given an index in the list of a vertex, we can access the (adjacent)
vertex in that position of the vertex’s adjacency list in constant time. Now if we simulate our
rotate model algorithm of Theorem 1 in read-only memory using the auxiliary structure as stated in
Lemma 6 as additional storage, then we obtain,
Theorem 9. A DFS traversal of an undirected or a directed graph G, represented by an adjacency
array, on n vertices and m edges can be performed in O(m + n) time using O(n lg(m/n)) bits, in
the read-only model.
The above result improves the DFS tradeoff result of Elmasry et al. [44] for relatively sparse
graphs in the read-only model. In particular, they showed the following,
Theorem 10 ([44]). For every function t : N → N such that t(n) can be computed within the
resource bound of this theorem (e.g., in O(n) time using O(n) bits), the vertices of a directed or
undirected graph G, represented by adjacency arrays, with n vertices and m edges can be visited in
lg n
depth first order in O((m + n)t(n)) time with O(n + n lgt(n)
) bits.
Thus to achieve O(m + n) time for DFS, their algorithm (Theorem 10) uses O(n lg lg n) bits.
This is Ω(n lg(m/n)) for all values of m where m = O(n lg n). Banerjee et al. [11] and Kammer
et al. [54] recently provided another DFS implementation taking O(m + n) bits of space and runs
in O(m + n) time. Note that, Theorem 9 improves the space bound of the above mentioned DFS
implementations from O(m + n) space to O(n lg(m/n)), while maintaining the same linear running
time. Chakraborty et al. [23] obtained similar results by a slightly different technique. In what
follows, we show that using Theorem 9, we can improve the space bounds of some of the classical
applications of DFS in read-only model.
To illustrate this, note that, one of the many classical applications of DFS (see [30]) include (i)
topological sorting of the vertices of a directed acyclic graph [55], (ii) producing a sparse (having
O(n) edges) spanning biconnected subgraph of a undirected biconnected graph G [43], and (iii) given
an undirected 2-edge-connected graph G = (V, E), to orient each edge (u, v) ∈ E as an arc (u, v)
or (v, u) to obtain D = (V, A) such that D becomes strongly connected. For all of these problems,
classical algorithms [55, 43] take linear time and O(n lg n) bits of space. Recently, Banerjee et al.
[11] showed the following,
Theorem 11. [11] In the read-only model, if the DFS of G on n vertices and m edges, can be
performed in t(m, n) time using s(m, n), where s(m, n) = Ω(n), bits of space, then using O(s(m, n))
bits and in O(t(n, m)) time, we can output
1. the vertices of a directed acyclic graph in topologically sorted order,
2. the edges of a sparse spanning biconnected subgraph of a undirected biconnected graph G, and
3. a strongly connected orientation of a undirected 2-edge-connected graph G.
22
Now plugging the improved DFS algorithm of Theorem 9 in the above theorem, we obtain for
all the applications an improved (over the classical implementation) space-efficient implementations
taking O(m + n) time and O(n lg(m/n)) bits of space.
8.2
Space-efficient approximation algorithms using Baker’s approach
In this section we present in-place Logspace Approximation Scheme or LSAS ((1 ± ) approximation
algorithm running in logspace for every > 0) for a class of MSO-definable optimization problems
which are amenable to the Baker’s method [10], also known as shifting technique, in locally bounded
treewidth graphs.
There are two main general approaches for designing PTASs for problems on planar graphs.
The first approach is based on planar separators [58]. The approximation algorithms resulting
from this approach are impractical in general. To address this, Baker [10] introduced the second
approach for PTASs in planar graphs, based on decomposition into overlapping subgraphs of
bounded outerplanarity, which are of bounded treewidth. For a general account on these, see [37].
Baker’s method was originally designed to give PTASs for a host of NP-hard optimization problems on planar graphs like minimum vertex cover, minimum dominating set, maximum independent
set etc which are hard to approximate in general graphs. Many of these remain NP-hard even in
planar graphs. Later the technique was generalized to a broader class of graphs called graphs of
bounded local treewidth [46, 36]. For a vertex v of a graph G and integer k ≥ 0, by Gkv we denote
the subgraph of G induced by vertices within distance at most k from v in G. A class of graphs G
is of bounded local treewidth if there exists function f such that for every graph G ∈ G and every
vertex v of G, treewidth(Gkv ) ≤ f (k).
8.2.1
Baker’s Algorithm
The main two computational bottlenecks in Bakers approach are
1. decomposing the graph into bounded treewidth graphs, and
2. solving the optimization problem on bounded tree width graphs optimally and combining
these solutions.
Step (1) requires performing BFS on the graph G and considering the induced subgraphs Gi,j
between layers ki + j and k(i + 1) + j (which are of treewidth O(k)) for i ≥ 0 and offset 0 ≤ j ≤ k − 1
by deleting (or including in both the adjacent slices) the vertices/edges in every k = O(1/)-th BFS
layer (particular details differ depending on the problem). By choosing the right offset, we can make
sure this affects the optimum solution at most by factor. For Step (2), given an MSO-definable
problem, using Boadlander’s [16] and Courcelles theorem [31] on bounded treewidth graphs we can
get the optimal solution. Baker’s approach is highly efficient in terms of time complexity as this
gives linear time approximation schemes but the scenario changes when considering the efficiency of
the approach in terms of the space needed.
Though we can use the result of [40] for the second part of the algorithm which proves logspace
analogue of Boadlander’s and Courcelles theorem, in the the first part we need to compute distance.
BFS can be performed in NL in general graphs and in UL ∩ co-UL for planar graphs [69]. Since
UL ⊆ NL is not known (or even believed) to be inside L, the NL bound (UL ∩ co-UL for planar
graphs) for Baker’s algorithm is the best known in ROM [33]. Recently [34] has given an LSAS for
maximum matching in planar graphs and some more sparse graph classes (a problem amenable to
Baker’s method) but a logspace analogue of Baker’s algorithm in general is not yet known.
Since in our in-place models (both rotate and implicit ) we can overcome this hurdle by computing
distance in L, we obtain the following result. Notice that though our in-place algorithms change
the ordering of the vertices in an adjacency list (or array) the graph remains the same and so the
23
distances between vertices.
Theorem 12. In both the rotate and the implicit model MSO-definable optimization problems which
are amenable to Baker’s method has an LSAS in locally bounded treewidth graphs.
8.3
Solving NP-hard problems in in-place models
We show how to solve some NP-hard graph problems using logspace and exponential time in the
rotate and implicit models. In particular, this implies that problems such as vertex cover and
dominating set can be solved in exponential time using O(lg n) bits in both the models. In constrast,
note that, no NP-hard problem can be solved in the ROM model using O(lg n) bits unless P=NP.
Similar to Fomin et al. [48], we define a class of graph problems in NP which we call graph subset
problems where the goal is to find a subset of vertices satisfying some property. We show a meta
theorem showing that a restricted class of graph subset problems that are in NP admit log-space
exponential algorithms in the rotate and implicit models.
Given a graph G with its adjacency list, we encode a subset of the vertices as follows. For every
vertex in the subset, we bring in the minimum labelled vertex among its neighbors to the front of the
list, and for others, we keep a vertex with a higher label (than the minimum) at the front of the list.
So it takes a linear time to check whether a vertex is in the subset. The algorithm enumerates all
subsets (by this encoding) and simply verifies using the NP algorithm whether that subset satisfies
the required property until it finds a subset satisfying the property or it has enumerated all the
subsets. By a standard theorem in complexity theory [5], every problem in NP is actually verifiable
by a log-space ROM and hence the overall space taken by our algorithm is only logarithmic. Note
that our algorithm requires that the adjacency list of any vertex has at least two values, i.e. that
the degree of any vertex is at least two. Thus we have
Theorem 13. Any graph subset problem in NP can be solved using O(lg n) bits of extra space (and
exponential time) in the rotate and implicit models in graphs G having minimum degree 2.
Remark 1. We remark that the above idea can work for other graph problems that are not
necessarily subset problems. For example, for testing hamiltonicity, we can simply explore all
neighbors of a vertex (starting at the smallest labelled neighbor so we know when we have explored
them all) in a systematic fashion simply encoding them into the adjacency list by moving the current
neighbor to the front of the list, and test whether together they form a cycle of length n.
If the graphs have larger minimum degree (say at least 2 lg n), we can even encode pointers in
each adjacency list and using that we can even test for graph isomorphism in logarithmic space by
explicitly maintaining the vertex mapping by these encoding pointers.
Remark 2. The minimum degree 2 restriction in the above theorem is not a serious restriction
as for many problems (like vertex cover, dominating set and traveling salesperson problem), there
are preprocessing routines that can handle (not necessarily in our model) and eliminate degree 1
vertices.
9
Concluding remarks
Our initial motivation was to get around the limitations of ROM to obtain a reasonable model for
graphs in which we can obtain space efficient algorithms. We achieved that by introducing two new
frameworks and obtained efficient (of the order of O(n3 lg n)) algorithms using O(lg n) bits of space
for fundamental graph search procedures. We also discussed various applications of our DFS/BFS
results, and it is not surprising that many simple corollaries would follow as DFS/BFS being the
backbone of so many graph algorithms. We showed that some of these results also translate to
improved space efficient algorithms in ROM (by simulating the rotate model algorithms in ROM
24
with one pointer per list). With some effort, we can obtain log space algorithm for minimum
spanning tree. These results can be contrasted with the state of the art results in ROM that take
almost linear bits for some of these problems other than having large runtime bounds. All our
algorithms are conceptually simple, and as they don’t use any heavy data structures, we believe
that they are also practical to implement. Still, there are plenty of algorithmic graph problems
to be studied in these models. We believe that our work is the first step towards this and will
inspire further investigation into designing in-place algorithms for other graph problems. One future
direction would be to improve the running time of our algorithms to make them more practical.
Surprisingly we could design log-space algorithm for some P-complete problems, and so it is
important to understand the power of our models. Towards that we discovered that we can even
obtain log-space algorithms for some NP-hard graph problems. More specifically, we defined graph
subset problems and obtained log-space exponential time algorithms for problems belonging to this
class. One interesting future direction would be to determine the exact computational power of
these models along with exploring the horizon of interesting complexity theoretic consequences of
problems in these models.
References
[1] A. Aggarwal and R. J. Anderson. A random NC algorithm for depth first search. Combinatorica,
8(1):1–12, 1988.
[2] A. Aggarwal, R. J. Anderson, and M. Kao. Parallel depth-first search in general directed graphs.
SIAM J. Comput., 19(2):397–409, 1990.
[3] N. Alon, Y. Matias, and M. Szegedy. The space complexity of approximating the frequency
moments. J. Comput. Syst. Sci., 58(1):137–147, 1999.
[4] R. J. Anderson and E. W. Mayr. Parallelism and the maximal path problem. Inf. Process.
Lett., 24(2):121–126, 1987.
[5] S. Arora and B. Barak. Computational Complexity - A Modern Approach. Cambridge University
Press, 2009.
[6] T. Asano, K. Buchin, M. Buchin, M.Korman, W. Mulzer, G. Rote, and A. Schulz. Reprint of:
Memory-constrained algorithms for simple polygons. Comput. Geom., 47(3):469–479, 2014.
[7] T. Asano, T. Izumi, M. Kiyomi, M. Konagaya, H. Ono, Y. Otachi, P. Schweitzer, J. Tarui, and
R. Uehara. Depth-first search using O(n) bits. In 25th ISAAC, pages 553–564, 2014.
√
[8] T. Asano, D. G. Kirkpatrick, K. Nakagawa, and O. Watanabe. Õ( n)-space and polynomialtime algorithm for planar directed graph reachability. In 39th MFCS LNCS 8634, pages 45–56,
2014.
[9] T. Asano, W. Mulzer, G. Rote, and Y. Wang. Constant-work-space algorithms for geometric
problems. JoCG, 2(1):46–68, 2011.
[10] B. S. Baker. Approximation algorithms for np-complete problems on planar graphs. J. ACM,
41(1):153–180, 1994.
[11] N. Banerjee, S. Chakraborty, and V. Raman. Improved space efficient algorithms for BFS, DFS
and applications. In 22nd COCOON, 2016.
25
[12] N. Banerjee, S. Chakraborty, V. Raman, S. Roy, and S. Saurabh. Time-space tradeoffs for
dynamic programming in trees and bounded treewidth graphs. In 21st COCOON, volume 9198,
pages 349–360. springer, LNCS, 2015.
[13] L. Barba, M. Korman, S. Langerman, K. Sadakane, and R. I. Silveira. Space-time trade-offs
for stack-based algorithms. Algorithmica, 72(4):1097–1129, 2015.
[14] G. Barnes, J. Buss, W. Ruzzo, and B. Schieber. A sublinear space, polynomial time algorithm
for directed s-t connectivity. SIAM J. Comput., 27(5):1273–1282, 1998.
[15] Paul Beame. A general sequential time-space tradeoff for finding unique elements. SIAM J.
Comput., 20(2):270–277, 1991.
[16] H. L. Bodlaender. A linear-time algorithm for finding tree-decompositions of small treewidth.
SIAM J. Comput., 25(6):1305–1317, 1996.
[17] A. Borodin and S. A. Cook. A time-space tradeoff for sorting on a general sequential model of
computation. SIAM J. Comput., 11(2):287–297, 1982.
[18] A. Borodin, M. J. Fischer, D. G. Kirkpatrick, N. A. Lynch, and M. Tompa. A time-space
tradeoff for sorting on non-oblivious machines. J. Comput. Syst. Sci., 22(3):351–364, 1981.
[19] H. Brönnimann, T. M. Chan, and E. Y. Chen. Towards in-place geometric algorithms and data
structures. In Proceedings of the 20th ACM Symposium on Computational Geometry, Brooklyn,
New York, USA, June 8-11, 2004, pages 239–246, 2004.
[20] H. Buhrman, R. Cleve, M. Koucký, B. Loff, and F. Speelman. Computing with a full memory:
catalytic space. In Symposium on Theory of Computing, STOC 2014, New York, NY, USA,
May 31 - June 03, 2014, pages 857–866, 2014.
[21] H. Buhrman, M.l Koucký, B. Loff, and F. Speelman. Catalytic space: Non-determinism and
hierarchy. In 33rd STACS 2016, February 17-20, 2016, Orléans, France, pages 24:1–24:13,
2016.
[22] D. Chakraborty, A. Pavan, R. Tewari, N. V. Vinodchandran, and L. Yang. New time-space
upperbounds for directed reachability in high-genus and h-minor-free graphs. In FSTTCS,
pages 585–595, 2014.
[23] S. Chakraborty, V. Raman, and S. R. Satti. Biconnectivity, chain decomposition and stnumbering using O(n) bits. In 27th ISAAC, pages 22:1–22:13, 2016.
[24] S. Chakraborty, V. Raman, and S. R. Satti. Biconnectivity, st-numbering and other applications
of DFS using O(n) bits. J. Comput. Syst. Sci., 90:63–79, 2017.
[25] S. Chakraborty and S. R. Satti. Space-efficient algorithms for maximum cardinality search,
stack bfs, queue BFS and applications. In Computing and Combinatorics - 23rd International
Conference, COCOON 2017, Hong Kong, China, August 3-5, 2017, Proceedings, pages 87–98,
2017.
[26] T. M. Chan and E. Y. Chen. Multi-pass geometric algorithms. Discrete & Computational
Geometry, 37(1):79–102, 2007.
26
[27] T. M. Chan, J. I. Munro, and V. Raman. Faster, space-efficient selection algorithms in read-only
memory for integers. In Algorithms and Computation - 24th International Symposium, ISAAC
2013, Hong Kong, China, December 16-18, 2013, Proceedings, pages 405–412, 2013.
[28] T. M. Chan, J. I. Munro, and V. Raman. Selection and sorting in the ”restore” model. In
25th-SODA, pages 995–1004, 2014.
[29] S. A. Cook and C. Rackoff. Space lower bounds for maze threadability on restricted machines.
SIAM J. Comput., 9(3):636–652, 1980.
[30] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to Algorithms (3. ed.).
MIT Press, 2009.
[31] B. Courcelle and M. Mosbah. Monadic second-order evaluations on tree-decomposable graphs.
Theor. Comput. Sci., 109(1&2):49–82, 1993.
[32] O. Darwish and A. Elmasry. Optimal time-space tradeoff for the 2d convex-hull problem. In
22th ESA, pages 284–295, 2014.
[33] S. Datta and R. Kulkarni. Space complexity of optimization problems in planar graphs. In
11th Annual Conference, TAMC 2014, Chennai, India, April 11-13, 2014. Proceedings, pages
300–311, 2014.
[34] S. Datta, R. Kulkarni, and A. Mukherjee. Space-efficient approximation scheme for maximum
matching in sparse graphs. In 41st MFCS 2016, August 22-26, 2016 - Kraków, Poland, pages
28:1–28:12, 2016.
[35] S. Datta, N. Limaye, P. Nimbhorkar, T. Thierauf, and F. Wagner. Planar graph isomorphism
is in log-space. In 24th CCC, pages 203–214, 2009.
[36] E. D. Demaine and M. Taghi Hajiaghayi. Equivalence of local treewidth and linear local
treewidth and its algorithmic applications. In Fifteenth SODA 2004, New Orleans, Louisiana,
USA, January 11-14, 2004, pages 840–849, 2004.
[37] E. D. Demaine and M. Taghi Hajiaghayi. Approximation Schemes for Planar Graph Problems,
pages 1–99. Springer US, Boston, MA, 2008.
[38] Y. Dodis, M. Patrascu, and M. Thorup. Changing base without losing space. In Proceedings of
the 42nd ACM Symposium on Theory of Computing (STOC), pages 593–602, 2010.
[39] J. Edmonds, C. K. Poon, and D. Achlioptas. Tight lower bounds for st-connectivity on the
NNJAG model. SIAM J. Comput., 28(6):2257–2284, 1999.
[40] M. Elberfeld, A. Jakoby, and T. Tantau. Logspace versions of the theorems of bodlaender and
courcelle. In 51th FOCS, pages 143–152, 2010.
[41] M. Elberfeld and K. Kawarabayashi. Embedding and canonizing graphs of bounded genus in
logspace. In Symposium on Theory of Computing, STOC 2014, New York, NY, USA, May 31 June 03, 2014, pages 383–392, 2014.
[42] M. Elberfeld and P. Schweitzer. Canonizing graphs of bounded tree width in logspace. In 33rd
Symposium on Theoretical Aspects of Computer Science, STACS 2016, February 17-20, 2016,
Orléans, France, pages 32:1–32:14, 2016.
27
[43] A. Elmasry. Why depth-first search efficiently identifies two and three-connected graphs. In
21st ISAAC, pages 375–386, 2010.
[44] A. Elmasry, T. Hagerup, and F. Kammer. Space-efficient basic graph algorithms. In 32nd
STACS, pages 288–301, 2015.
[45] A. Elmasry, D. D. Juhl, J. Katajainen, and S. R. Satti. Selection from read-only memory with
limited workspace. Theor. Comput. Sci., 554:64–73, 2014.
[46] D. Eppstein. Diameter and treewidth in minor-closed graph families. Algorithmica, 27(3):275–
291, 2000.
[47] J. Feigenbaum, S. Kannan, A. McGregor, S. Suri, and J. Zhang. On graph problems in a
semi-streaming model. Theor. Comput. Sci., 348(2-3):207–216, 2005.
[48] F. V. Fomin, S. Gaspers, D. Lokshtanov, and S. Saurabh. Exact algorithms via monotone local
search. In Proceedings of the 48th Annual ACM SIGACT Symposium on Theory of Computing,
STOC 2016, Cambridge, MA, USA, June 18-21, 2016, pages 764–775, 2016.
[49] G. Franceschini and J. Ian Munro. Implicit dictionaries with O(1) modifications per update
and fast search. In Proceedings of the Seventeenth Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 404–413, 2006.
[50] G. Franceschini and S. Muthukrishnan. In-place suffix sorting. In Automata, Languages and
Programming, 34th International Colloquium, ICALP 2007, Wroclaw, Poland, July 9-13, 2007,
Proceedings, pages 533–545, 2007.
[51] G. Franceschini, S. Muthukrishnan, and M. Patrascu. Radix sorting with no extra space. In
Algorithms - ESA 2007, 15th Annual European Symposium, Eilat, Israel, October 8-10, 2007,
Proceedings, pages 194–205, 2007.
[52] G. N. Frederickson. Upper bounds for time-space trade-offs in sorting and selection. J. Comput.
Syst. Sci., 34(1):19–26, 1987.
[53] T. Hagerup and F. Kammer. Succinct choice dictionaries. CoRR, abs/1604.06058, 2016.
[54] F. Kammer, D. Kratsch, and M. Laudahn. Space-efficient biconnected components and
recognition of outerplanar graphs. In 41st MFCS, 2016.
[55] Donald E. Knuth. The Art of Computer Programming, Volume I: Fundamental Algorithms,
2nd Edition. Addison-Wesley, 1973.
[56] M. Koucký. Catalytic computation. Bulletin of the EATCS, 118, 2016.
[57] T. W. Lai and D. Wood. Implicit selection. In SWAT 88, 1st Scandinavian Workshop on
Algorithm Theory, Halmstad, Sweden, July 5-8, 1988, Proceedings, pages 14–23, 1988.
[58] R. J. Lipton and R. E. Tarjan. Applications of a planar separator theorem. SIAM J. Comput.,
9(3):615–627, 1980.
[59] A. McGregor. Graph stream algorithms: a survey. SIGMOD Record, 43(1):9–20, 2014.
[60] J. I. Munro. Tables. In FSTTCS, pages 37–42, 1996.
28
[61] J. I. Munro and M. Paterson. Selection and sorting with limited storage. Theor. Comput. Sci.,
12:315–323, 1980.
[62] J. I. Munro and V. Raman. Selection from read-only memory and sorting with minimum data
movement. Theor. Comput. Sci., 165(2):311–323, 1996.
[63] J. Ian Munro. An implicit data structure supporting insertion, deletion, and search in O(log2
n) time. J. Comput. Syst. Sci., 33(1):66–74, 1986.
[64] J. Pagter and T. Rauhe. Optimal time-space trade-offs for sorting. In 39th Annual Symposium
on Foundations of Computer Science, FOCS ’98, November 8-11, 1998, Palo Alto, California,
USA, pages 264–268, 1998.
[65] J. H. Reif. Symmetric complementation. J. ACM, 31(2):401–421, 1984.
[66] J. H. Reif. Depth-first search is inherently sequential. Inf. Process. Lett., 20(5):229–234, 1985.
[67] O. Reingold. Undirected connectivity in log-space. J. ACM, 55(4), 2008.
[68] T. Tantau. Logspace optimization problems and their approximability properties. Theory
Comput. Syst., 41(2):327–350, 2007.
[69] T. Thierauf and F. Wagner. The isomorphism problem for planar 3-connected graphs is in
unambiguous logspace. Theory Comput. Syst., 47(3):655–673, 2010.
[70] M. Tompa. Two familiar transitive closure algorithms which admit no polynomial time,
sublinear space implementations. SIAM J. Comput., 11(1):130–137, 1982.
29
| 8cs.DS
|
DUAL BRAID MONOIDS, MIKADO BRAIDS AND
POSITIVITY IN HECKE ALGEBRAS
arXiv:1508.06817v3 [math.GR] 31 Mar 2016
FRANÇOIS DIGNE AND THOMAS GOBET
Abstract. We study the rational permutation braids, that is the elements of an Artin-Tits group of spherical type which can be written
x−1 y where x and y are prefixes of the Garside element of the braid
monoid. We give a geometric characterization of these braids in type
An and Bn and then show that in spherical types different from Dn
the simple elements of the dual braid monoid (for arbitrary choice of
Coxeter element) embedded in the braid group are rational permutation braids (we conjecture this to hold also in type Dn ). This property
implies positivity properties of the polynomials arising in the linear expansion of their images in the Iwahori-Hecke algebra when expressed in
the Kazhdan-Lusztig basis. In type An , it implies positivity properties
of their images in the Temperley-Lieb algebra when expressed in the
diagram basis.
Contents
1. Introduction
2
2. Classical braid monoid
4
3. Dual braid monoids
4
3.1. Coxeter elements
4
3.2. Dual braid monoids
6
3.3. Hurwitz action on reduced decompositions
7
3.4. Groups of fractions of dual braid monoids are Artin-Tits groups 7
3.5. A formula for dual atoms
8
4. Rational permutation braids
9
4.1. Square-free braids
9
4.2. Rational permutation braids
9
5. Mikado braids of type An
11
5.1. Square-free braids
11
5.2. Mikado braids
11
5.3. Simple dual braids are Mikado braids
13
6. Mikado braids of type Bn
18
6.1. Coxeter and Artin-Tits groups of type Bn
18
6.2. Mikado braids of type Bn
18
6.3. Simple dual braids are Mikado braids
19
7. Dihedral and exceptional types
20
8. Positivity properties
21
8.1. Iwahori-Hecke algebra of a Coxeter system
21
8.2. Positivity properties of simple dual braids
21
8.3. Temperley-Lieb algebra and monomial basis
22
8.4. Projection of the canonical basis
23
1
2
FRANÇOIS DIGNE AND THOMAS GOBET
8.5. Zinno basis
8.6. Positivity consequences
References
23
24
25
1. Introduction
This paper is motivated by positivity properties arising when expanding
simple dual braids in the canonical basis of the Iwahori-Hecke algebra.
More precisely, a dual braid monoid associated to a finite Coxeter group
and a choice of Coxeter element is a Garside monoid ([16]) whose group
of fractions is isomorphic to the corresponding Artin-Tits group. The dual
braid monoids were introduced by Bessis [4], extending definitions by BirmanKo-Lee [6] and Bessis-Michel and the first author [5]. As Garside monoids,
they possess a finite set of simple elements which generates the whole monoid
and forms a lattice under the left-divisibility order and they embed into their
group of fractions. Similar constructions have been made for some infinite
Coxeter groups and for complex reflection groups, but in this paper, unless explicitely stated, “dual braid monoid” will mean a dual braid monoid
associated with a finite Coxeter group.
In the framework of Artin-Tits groups, the standard example of a Garside
monoid is the braid or Artin-Tits monoid, also known as positive or classical
braid monoid. Its set of simple elements is the canonical positive lift of
the Coxeter group in the Artin-Tits group. The embedding property in
this specific case was shown by Brieskorn-Saito ([8]) and by Deligne ([17]).
In the dual approach, the set of simple generators of the Coxeter group is
replaced in the spherical case by the whole set of reflections, that is, by the
conjugates of the simple generators. A key tool in this approach is an action
of the braid group on the set of reduced factorizations of a Coxeter element
as a product of reflections, called the Hurwitz action. This action is known
to be transitive ([4], [3]), providing an analogue in the dual approach of the
Tits-Matsumoto property.
The dual braid monoid is then an analogue to the classical braid monoid,
but having as set of generators a copy of the set of reflections of the Coxeter
group. In the classical setting, the lattice of simple elements is isomorphic to
the lattice obtained by endowing the Coxeter group with the left weak order
defined by the classical length function. By mimicking this approach in the
dual setting, that is, by replacing the classical length function by the length
function with respect to the whole set of reflections, one obtains an order
called reflection or absolute order on the Coxeter group. The absolute order
does not in general endow the Coxeter group with a lattice structure, but by
restricting such an order to the order ideal of a Coxeter element, one obtains
a lattice ([4], [11]). The dual braid monoid is built using this property. Its
lattice of simple elements is isomorphic to the order ideal of a given Coxeter
element ordered by the restriction of the absolute order ([4]).
There is a morphism from the Artin-Tits group to the group of invertible
elements of the Iwahori-Hecke algebra attached to the corresponding Coxeter
group. The images of the simple elements of the classical braid monoid
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
3
through this morphism yield a basis of the Iwahori-Hecke algebra, the socalled standard basis. It is therefore natural to ask about properties of the
images of the simple elements of the dual Garside structure of an Artin-Tits
group in the associated Iwahori-Hecke algebra.
In type An , it is known that their images in the Temperley-Lieb algebra
yield a basis ([40], [32], [39]) such that the base change to the diagram basis
is triangular ([40], [25]).
One of the main points of this paper consists of understanding how one
can express the simple dual braids in terms of the classical braid group
generators. Indeed, the isomorphism between the group of fractions of a
dual braid monoid and the Artin-Tits group is proven by a case-by-case
analysis, and there is no known general formula for the simple dual braids
in the braid group, even for the atoms. We recall the construction of the
classical and dual braid monoids in Sections 2 and 3 where we also prove
some new results on dual Coxeter systems. In particular we give a formula
for the reflections in terms of the simple reflections (Proposition 3.4) which
lifts well to the Artin-Tits group, allowing us to get a uniform formula for
the dual atoms in terms of the classical generators (Proposition 3.13) and
this with a case-free proof.
We then introduce in Section 4 a finite set of braids, which we call rational
permutation braids. These braids already appear in a paper by Dehornoy
([14]) in type An and in unpublished work of Dyer ([19]) for Artin-Tits
groups attached to arbitrary (not necessarily finite) Coxeter groups. They
may be defined in spherical types as the braids of the form xy −1 where x
and y are simple braids for the classical Garside structure. The images of
such braids in the Iwahori-Hecke algebra turn out to have positive KazhdanLusztig expansions; this is shown by Dyer-Lehrer ([20]) for finite Weyl groups
using perverse sheaves and by Dyer ([18]) for arbitrary finite Coxeter groups
using Soergel bimodules. These braids are closely related to the so-called
mixed braid relations introduced by Dyer ([19]).
We give uniform Garside-theoretic characterizations of the rational permutation braids in the spherical case in Proposition 4.3, and then turn to
a case-by-case analysis in types An (Section 5) and Bn (Section 6) to characterize them in terms of geometrical braids (Propositions 5.7 and 6.2). In
these cases, they turn out to be what we call the Mikado braids, that is, the
braids where one can inductively remove a strand which is above all the other
strands. With this geometric characterization, we can show that all the simple dual braids, for any choice of Coxeter element, are rational permutation
braids (Theorems 5.12 and 6.6)–the argument given here is therefore topological. We also prove the same result for dihedral groups and, by computer,
for exceptional types (Theorem 7.1); we conjecture the result to also hold in
type Dn (Conjecture 8.7).
It follows that the simple dual braids have a positive Kazhdan-Lusztig
expansion in all types except possibly type Dn (Theorem 8.6). Positivity
results in the Temperley-Lieb algebra (of type An ) can also be derived (Theorem 8.16).
4
FRANÇOIS DIGNE AND THOMAS GOBET
Acknowledgments. We thank Patrick Dehornoy, Matthew Dyer and
Jean Michel for useful discussions. We also thank the anonymous referee for
his careful reading of the manuscript and many interesting comments and
remarks.
2. Classical braid monoid
Starting with a finite Coxeter System (W, S), one can define the associated
Artin-Tits monoid or classical braid monoid B + (W ) and for each choice of
a Coxeter element c an associated dual braid monoid Bc∗ (W ) ([4]).
The classical braid monoid is defined by the following presentation:
B + (W ) = hS | s1 s2 . . . = s2 s1 . . . for s1 , s2 ∈ Si+
| {z } | {z }
ms1 ,s2
ms1 ,s2
where S is a set in bijection s 7→ s with S and ms1 ,s2 is the order of s1 s2 in
W.
The monoid B + (W ) is a Garside monoid (see for instance [15, I, 2.1]
for the definition), hence it satisfies the Ore conditions and embeds in its
group of fractions, the braid group B(W ) of W . This group has the same
presentation as B + (W ) but as a group:
B(W ) = hS | s1 s2 . . . = s2 s1 . . . for s1 , s2 ∈ Si.
| {z } | {z }
ms1 ,s2
ms1 ,s2
The group W is a quotient of B(W ) by the normal subgroup generated
by the squares of the elements of S. We denote by p : B(W ) → W this
surjective morphism. It has a set-wise section obtained by lifting S-shortest
expressions. We denote by W the image of this section. This set W is
the set of simples of the Garside monoid B + (W ). The Garside element of
B + (W ) is the lift ∆ = w0 of the longest element w0 of W .
By definition of Garside monoids the left- (resp. right-) divisibility gives
a lattice structure to B + (W ) which restricts to a lattice structure on W .
Moreover by construction of W , the map p provides an isomorphism of
lattices from W to W endowed with the Coxeter theoretic left- (resp. right-)
weak order.
3. Dual braid monoids
3.1. Coxeter elements. The dual braid monoids are defined using Coxeter
elements. Before defining these monoids we prove some properties of the
Coxeter elements that we will need.
Definition 3.1. Let (W, S) be a Coxeter system. An element c ∈ W is
a standard Coxeter element of (W, S) if it is a product of the elements of
S in someSorder. An element c ∈ W is a Coxeter element if there exists
S ′ ⊂ T = w∈W wSw−1 such that (W, S ′ ) is a Coxeter system and c is the
product of the elements of S ′ in some order, that is, c is a standard Coxeter
element of (W, S ′ ). The set T is the set of reflections of (W, S).
Remark 3.2. If S ′ is as above then by [9, Lemma 3.7] one has T =
S
′ −1
′
w∈W wS w . Moreover if W is finite and S is as above, then by [9,
Theorem 3.10], the Coxeter systems (W, S) and (W, S ′ ) are isomorphic. In
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
5
particular, the type of (W, S) depends only on T . If (W, S) is infinite irreducible with no infinite entry in its Coxeter matrix, a stronger property
holds: any S ′ such that (W, S ′ ) is a Coxeter system is conjugate to S ([24,
Theorem 1]). The example of the dihedral group I2 (5) with S = {s, t} and
S ′ = {s, ststs} shows that when W is finite S and S ′ need not be conjugate.
Definition 3.3. Let (W, S) be a Coxeter system.
(1) We write ℓS , resp. ℓT , for the length function on W with respect to
S, resp. T .
(2) We say that x ∈ W divides y ∈ W , written x 4T y, if ℓT (x−1 y) +
ℓT (x) = ℓT (y). We write Div(y) := {x ∈ W | x 4T y} and call the
elements of Div(y) the divisors of y.
Note that the definitions of ℓS and ℓT use the fact that each one of the
sets S and T positively generates W . Note also that, since T is invariant
under conjugation, the relation ℓT (x−1 y) + ℓT (x) = ℓT (y) is equivalent to
ℓT (yx−1 ) + ℓT (x) = ℓT (y): there is no need to distinguish between left- and
right-divisibility.
The following proposition shows how to recover all reflections from a Coxeter element when W is finite. Recall that if c is a Coxeter element in a
Coxeter system of rank n, one has ℓT (c) = n (see e.g., [3, Lemma 1.2]).
Proposition 3.4. Let (W, S) be a finite Coxeter system with S = {s1 , . . . , sn }.
Let c = s1 s2 · · · sn . Then the set of reflections of W is given by
{ck s1 s2 . . . si si−1 . . . s1 c−k | k ∈ N, 1 ≤ i ≤ n}.
Proof. Let Tc be the set of the statement and T be the set of reflections
of (W, S). We have Tc ⊂ T . If c is such that ℓS (ci ) = iℓS (c) for any
i ≤ |T |/h, where h = 2|T |/n is the Coxeter number (the order of c), then
Tc is equal to T (this is a consequence of e.g., the proof of [12, 3.9]; see also
[7, Chapter V, §6 exercise 2] in the case where c is bipartite), whence the
result holds in that case. We now remark that if c′ = si+1 . . . sn s1 . . . si , with
1 ≤ i ≤ n − 1 then Tc′ contains (si . . . s2 s1 )Tc (si . . . s2 s1 )−1 so has cardinality
at least |Tc |, whence the result holds: indeed, since W is finite, all standard
Coxeter elements of (W, S) are cyclically conjugate (that is by a sequence of
conjugations like the one deriving c′ from c above): see e.g., [23, Theorem
3.1.4].
We take from [3, Section 1] the second item of the following definition.
Definition 3.5.
(1) A parabolic subgroup W ′ of a Coxeter system (W, S)
is a subgroup generated by a conjugate S ′ of a subset of S. We shall
say that (W ′ , S ′ ) is a parabolic subsystem of (W, S).
(2) Let (W, S) be a Coxeter system with set of reflections T . Let x ∈ W .
Then x is a parabolic Coxeter element in W if there exists a subset
S ′ = {s′1 , . . . , s′n } ⊂ T such that x = s′1 · · · s′m for some m ≤ n and
(W, S ′ ) is a Coxeter system.
Corollary 3.6. Let (W, S) be a finite Coxeter system and let x ∈ W . The
following are equivalent:
(1) There exists a Coxeter element c ∈ W such that x 4T c.
6
FRANÇOIS DIGNE AND THOMAS GOBET
(2) The element x is a parabolic Coxeter element.
Proof. Assume that x4T c for some Coxeter element c. Let S ′ = {s1 , . . . , sn }
be a simple Coxeter system of (W, S) such that c = s1 . . . sn and let T be the
set of reflections of W . We argue by reverse induction on ℓT (x). First assume
that ℓT (x) = n−1, so that there exists a decomposition of c into a product of
n reflections c = t1 . . . tn such that x = t2 . . . tn . By the above proposition we
k
have t1 = c s1 ...si si+1 for some k ∈ N and some 1 ≤ i ≤ n − 1, where for two
elements a and b in a group, we write ab for aba−1 . Let y = si . . . s1 c−k ; we
have y c = si+1 . . . sn s1 . . . si and yt1 = si+1 . Then yx = si+2 . . . sn s1 . . . si is a
standard Coxeter element in the standard parabolic subgroup W1 generated
by S1 = {s1 , . . . , si , si+2 , . . . sn }) so that x is a standard Coxeter element in
−1
the parabolic subsystem y (W1 , S1 ). Since a conjugate of a simple system
is again a simple system, we deduce that x is a parabolic Coxeter element.
Now the reflections of a parabolic subgroup are precisely the reflections of W
which lie in this subgroup, so that the induction can go on: a divisor of x in
W is in the parabolic subsystem of which x is a standard Coxeter element.
The converse is immediate.
Remark 3.7. In his case, that is for finite Coxeter groups, Bessis gives [4,
Section 1.4] an alternative definition of parabolic Coxeter elements and shows
[4, Lemma 1.4.3] that an element x of W is a parabolic Coxeter element in
his sense if and only if there exists a bipartite Coxeter element c such that
x 4T c. Since in finite Coxeter groups any Coxeter element is conjugate to
a bipartite one, one can drop the word bipartite from this definition. This
last fact and the above lemma show that Bessis approach and the one of [3],
that is of Definition 3.5 (2), actually lead to equivalent definitions for finite
Coxeter groups; this equivalence is not obvious if we compare the definitions
from Bessis and [3].
3.2. Dual braid monoids. The dual braid monoids are defined as follows:
let (W, S) be a Coxeter system and let T denote the set of reflections of
(W, S). Let c be a standard Coxeter element. The dual braid monoid Bc∗
associated to c is defined by generators and relations as follows. Let D c be
a set in one-to-one correspondence with the set Div(c) of divisors of c. Then
Bc∗ (W ) is generated by Dc with only relations xy = z for x, y, z in D c such
that xy = z ∈ Div(c) and ℓT (z) = ℓT (x) + ℓT (y), where x, y, z ∈ Div(c) are
the elements corresponding to x, y, z respectively. The canonical bijection
D c → Div(c) extends to a (surjective) morphism of monoids from Bc∗ (W ) to
W . The above definition is valid for any Coxeter system but in the sequel
we study only dual braid monoids associated to finite Coxeter groups.
In the spherical case the monoid Bc∗ (W ) is a Garside monoid with D c as
its set of simples. In this case the monoid Bc∗ (W ) has a presentation with
a generating set smaller than D c : first it can be shown that any reflection
divides c (see [4, Lemma 1.3.3]); if we denote by T c the lift of T in D c ⊂
Bc∗ (W ) we have
Proposition 3.8 ([4], Theorem 2.1.4). In the spherical case the dual braid
monoid has the following presentation:
Bc∗ (W ) = hT c | t1 t2 = t2 t3 for t1 , t2 , t3 in T c with t1 t2 = t2 t3 4T ci+ .
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
7
The relations in the above presentation of Bc∗ (W ) are called the dual braid
relations.
Proof. We recall the proof. To this end, we need to recall a few facts on the
Hurwitz action on sets of T -reduced decompositions of elements.
3.3. Hurwitz action on reduced decompositions. We write Bn for the
Artin braid group on n strands, that is, the Artin-Tits group B(W ) where
W is of type An−1 .
Definition 3.9 (Hurwitz action). Let G be a group and g be in G; the braid
group Bn acts on the set of n-tuples of elements of G whose product is equal
to g as follows: if σ1 , . . . , σn−1 are the generators of the braid group Bn , the
Hurwitz action of σi maps a sequence (t1 , . . . , tn ) such that g = t1 t2 . . . tn to
the sequence (t1 , . . . , ti−1 , ti ti+1 t−1
i , ti , ti+2 , . . . , tn ).
When considering the Hurwitz action, we will always denote the generators
of Bn by σ1 , . . . , σn−1 . Otherwise we write the generators of B(W ) for W of
type An−1 as s1 , . . . , sn−1 . In the case of reduced factorizations of a Coxeter
element into reflections we have:
Theorem 3.10 ([3], Theorem 1.3). Let (W, S) be a (not necessarily finite)
Coxeter system with T its set of reflections. If x is a parabolic Coxeter element in W , then the Hurwitz action is transitive on the set of decompositions
of x into a product of ℓT (x) reflections.
Note that for finite Coxeter groups, the assumption that x is a parabolic
Coxeter element can be replaced by x 4T c for some Coxeter element c of
(W, S) thanks to Corollary 3.6.
Let c be a standard Coxeter element in a Coxeter system (W, S) of rank
n. By definition of Bc∗ , any decomposition c = t1 . . . tn into a product of n
reflections is lifted in Bc∗ (W ) to c = t1 . . . tn where c is the lift of c in Dc .
Hence if x divides c, writing x = t1 . . . tℓT (x) with ti ∈ T and c = t1 . . . tn ,
we see that the lift of x is a product of ℓT (x) elements of T c , so that T c
generates Bc∗ (W ). Moreover by Theorem 3.10 and Corollary 3.6 one can pass
from any decomposition of x ∈ D c to any other one by dual braid relations.
This gives Proposition 3.8.
3.4. Groups of fractions of dual braid monoids are Artin-Tits groups.
The following is proved in [4]:
Proposition 3.11. The group of fractions of Bc∗ (W ) is isomorphic to B(W)
and the restriction of the projection p : B(W ) → W is the canonical surjection of the dual braid monoid onto W .
We explain how this isomorphism is defined and sketch a proof that it is
an isomorphism.
Sketch of proof. Let us enumerate the elements s1 , . . . , sn of S in such a way
that c = s1 . . . sn . We have S ⊂ T and an easy computation using the dual
braid relations shows that the lift S ′ of S in T c ⊂ Bc∗ (W ) satisfies the braid
relations. This gives a morphism of monoids B + (W ) → Bc∗ (W ) mapping si
to s′i where s′i is the lift of si in S ′ . This extends to a group morphism from
B(W ) to the group of fractions of Bc∗ (W ).
8
FRANÇOIS DIGNE AND THOMAS GOBET
Since c = s1 . . . sn , whence c = s′1 . . . s′n , Theorem 3.10 implies that all
elements of T c can be obtained from S ′ using dual braid relations, so that
S ′ generates the group of fractions of Bc∗ (W ), hence the above morphism
from B(W ) to this group is surjective.
To show that it is injective one first considers the particular case of a
Coxeter element c which is bipartite, that is, such that either for m = ⌊n/2⌋
or for m = ⌊(n + 1)/2⌋ the elements s1 , s2 , . . . , sm pairwise commute and
so do sm+1 , . . . , sn . Bessis [4] proves that in this case the set ∪k∈Z ck Sc−k
is a lift in B(W ) of T and that its elements satisfy the dual braid relations
(this last fact by a case by case analysis, see [4, Fact 2.2.4]). Hence we get
a morphism in the other direction such that the composition from B(W ) to
itself is the identity (since it is the identity on S).
To get the result for an arbitrary standard Coxeter element, one can use
the fact that the lifts in B + (W ) of all Coxeter elements of a given Coxeter
system are conjugate in B(W ) (see [7, Chapter V, §6 Lemme 1]) and that
the Hurwitz action commutes with conjugation.
We will now identify Bc∗ (W ) with its image in B(W ) and S ′ with S. We
then see T c as a subset of B(W ). Using this identification, we note that,
by the above proof, the lift of the Coxeter element c = s1 . . . sn in Bc∗ (W ) is
the same as its lift to the classical braid monoid, that is c = s1 . . . sn ∈ W .
Theorem 3.10 implies:
Corollary 3.12. The set of decompositions of c into a product of rank(W )
elements of T c in B(W ) is a single orbit under the Hurwitz action and the
restriction of π : B(W ) → W to T c induces a bijection from this Hurwitz
orbit to the set of decompositions of c into rank(W ) reflections.
We recall the proof for sake of completeness.
Proof. Let Decn (c) (resp. Decn (c)) be the set of decomposition of c (resp.
c) into n reflections (resp. into n elements of T c ) where n = rank(W ). The
morphism π bijectively maps T c to T , hence maps Decn (c) to Decn (c). By
definition of Bc∗ (W ) a decomposition (t1 , . . . , tn ) ∈ Decn (c) is lifted in B(W )
to (t1 , . . . , tn ) ∈ Decn (c). Hence π induces a bijection from Decn (c) to
Decn (c).
By the dual braid relations, the Hurwitz action in B(W ) preserves Decn (c).
Since π is compatible with the Hurwitz action and by Theorem 3.10 Decn (c)
is a single orbit, we get the result.
A question is then to understand which elements of B(W ) correspond to
the subset D c of Bc∗ (W ) through the embedding of Bc∗ (W ) into B(W ). This
may be of interest for both computational reasons and for a possible case-free
proof of Proposition 3.11.
We will address this question in Sections 5 and 6 when W is of type An
or Bn . In the case of reflections the following subsection gives an answer.
3.5. A formula for dual atoms. The following proposition is the analog
for the braid monoid of Proposition 3.4 and the proof is parallel.
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
9
Proposition 3.13. Let c = s1 . . . sn be a standard Coxeter element, then,
through the embedding of Bc∗ (W ) into B(W ), we have
−1
−1
T c = {s1 s2 . . . si si+1 s−1
i si−1 . . . s1 , | 0 ≤ i < 2|T |},
where the index i in si is taken modulo n.
−1
−1
Proof. Let T ′c be the set of elements s1 s2 . . . si si+1 s−1
i si−1 . . . s1 with i ≥ 0.
′
The elements of T c are components of the elements of the Hurwitz orbit of
(s1 , . . . , sn ): this is clear if i ≤ n and the elements with i > n are conjugates
from the previous ones by a power of c. Hence by Corollary 3.12 we have
T ′c ⊂ T c . Since by definition |T c | = |T |, for proving T c = T ′c it is sufficient
to prove |T ′c | ≥ |T |. If c is a Coxeter element such that ℓS (ci ) = iℓS (c) for
any i ≤ |T |/h, where h = 2|T |/n is the Coxeter number (the order of c),
e.g., if c is bipartite, then the image in W of T ′c is the full set T of reflections
of W as seen in the proof of Proposition 3.4, so that the result holds in that
case. We now remark that if c = s1 . . . sn and c′ = si . . . sn s1 . . . si−1 then
T ′c′ contains (si . . . sn )T ′c (si . . . sn )−1 so has cardinality at least |T ′c |, whence
the result holds, using the fact that all standard Coxeter elements of a given
Coxeter system are cyclically conjugate. By [7, Chapter V, §6 exercise 2] if c
is bipartite, we have ch = w20 , where w0 is the lift in B + (W ) of the longest
element of W . Since w 20 is central in B(W ), one gets the same set T ′c of
−1
−1
elements s1 s2 . . . si si+1 s−1
i si−1 . . . s1 for 0 ≤ i < 2|T | as for i ∈ N.
4. Rational permutation braids
4.1. Square-free braids. Let (W, S) be a finite Coxeter system. Recall
that we denote by p : B(W ) → W the canonical surjection.
Definition 4.1. An element β ∈ B(W ) is a square-free braid if it can be
represented by a braid word
sεi11 · · · sεikk
with εj ∈ {−1, 1} such that k = ℓS (p(β)). We denote by B(W )per ⊂ B(W )
the set of square-free braids.
One has W = B + (W ) ∩ B(W )per which is also the set of prefixes of the
Garside element ∆ (see [16, Example 1]). It follows from the definition that
any square-free braid is obtained by replacing the various sij for j = 1, . . . , k
in a reduced S-decomposition si1 · · · sik of an element w ∈ W by s±1
ij .
4.2. Rational permutation braids. As any group of fractions of a cancellative monoid B(W ) is endowed with a left-invariant partial order that
we denote by 4 defined by u 4 v if u−1 v ∈ B + (W ) which extends the leftdivisibility relation on B + (W ).
Definition 4.2. An element β ∈ B(W ) is a rational permutation braid if it
lies in the interval [∆−1 , ∆] for the partial order 4.
The following proposition can be seen as a particular case of [14, Proposition 5.3]. It explains the terminology as the elements of W are usually
called permutation braids.
Proposition 4.3. Let β ∈ B(W ). The following are equivalent:
10
FRANÇOIS DIGNE AND THOMAS GOBET
(1) The braid β is a rational permutation braid,
(2) There exist x, y ∈ W such that β = x−1 y,
(3) There exist x, y ∈ W such that β = xy −1 .
Proof. We prove the equivalence of (1) and (2). The equivalence of (1) and
(3) is similar.
Since B + (W ) satisfies the Ore conditions, any element of B(W ) can be
uniquely written as x−1 y with x and y in B + (W ) having no common leftdivisor. We have ∆−1 4 x−1 y 4 ∆ if and only if x−1 y = ∆−1 a = ∆b−1 with
a, b in B + (W ). So (1) implies that a is a divisor of ∆2 which by the general
properties of Garside monoids (see [15, V, Proposition 3.24]) means that
−1
a = a1 a2 with a1 , a2 ∈ W , whence x−1 y = (a−1
1 ∆) a2 . The Ore conditions
imply then that x and y are right-divisors of respectively a−1
1 ∆ and a2 in
B + (W ), so that x and y are in W . Conversely, if x and y are in W , that is
divide ∆, then x−1 y 4 x−1 ∆ 4 ∆ and ∆−1 4 ∆−1 y 4 x−1 y.
We thank Matthew Dyer for having pointed to us the property stated in the
following lemma which is a particular case of [19, Section 9.4]. We reprove
it here in our particular case.
Lemma 4.4. If β is a rational permutation braid and if s1 s2 · · · sk is any
reduced expression of its image in W , then for i = 1, . . . , k there exists
εi = ±1 such that β = sε11 s2ε2 · · · sεkk . In particular, any rational permutation
braid is a square-free braid.
Proof. We prove by induction on k that if s1 s2 . . . sk is a reduced decomposition of an element of W , then for y ∈ W the element sε11 . . . sεkk y, where
εi = 1 if ℓS (si si+1 . . . sk y) = ℓS (si+1 . . . sk y) + 1 and εi = −1 otherwise is in
W . This concludes the proof, since if β = xy −1 is a rational permutation
braid with x and y in W and if s1 . . . sk is a reduced decomposition of the
image of β in W , then sε11 . . . sεkk y is in W and has same image as x, hence
is equal to x, so that β = xy −1 = sε11 . . . sεkk is a square-free braid.
By induction hypothesis x′ = sε21 . . . sεkk y is in W . If ε1 = 1 then s1 x′ is
in W . If ε1 = −1 then, by the exchange lemma, x′ has a reduced expression
of the form s1 s′2 . . . s′m so that sε11 x′ = s′2 . . . s′m is again in W .
Example 4.5. Let W be of type A2 , with simple generating reflections s1
and s2 ; the element w := s1 s2 s1 in W has exactly two reduced expressions
s1 s2 s1 = s2 s1 s2 . There are six square-free braids having w as image. For
example, β = s−1
1 s2 s1 . Here the reduced expression s1 s2 s1 has been lifted
to the braid word s−1
1 s2 s1 ; Lemma 4.4 says that we can lift any reduced
expression, in particular we can also lift the reduced expression s2 s1 s2 of w
in which case one has β = s2 s1 s−1
2 .
Remark 4.6. Note that the set of rational permutation braids is the set of
braids where "one can apply (mixed) braid relations in reduced words as in
the Coxeter group". Let us be more precise: if β is a rational permutation
braid, then one can pass from any shortest expression of β with respect to S∪
S −1 to any other one by mixed braid relations as defined in [19], that is, braid
relations possibly involving inverses. Moreover, to any reduced expression
of the image of β corresponds a shortest braid word for β. Precisely given
any two reduced expressions of the image of a braid β in W which differ by
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
11
a braid relation, then by Lemma 4.4 both reduced expressions can be lifted
to reduced braid words for β which differ by a mixed braid relation. For
example the relation s1−1 s2 s1 = s2 s1 s−1
2 in Example 4.5 is a mixed braid
relation.
5. Mikado braids of type An
In this section (W, S) is a Coxeter system of type An . The group W is
identified with the symmetric group Sn+1 on {1, 2 . . . , n + 1} and S with the
set of simple transpositions, writing si := (i, i + 1). The Artin-Tits group
B(W ) is identified with Bn+1 .
5.1. Square-free braids. We give some geometrical properties of squarefree braids of type An . Firstly, notice that β ∈ Bn+1 is a square-free braid
if and only if there is a braid diagram for β where any two strands cross at
most once. We call a braid diagram reduced if it has the minimal number of
crossings.
Lemma 5.1. Removing any strand in a diagram of a square-free braid β ∈
Bn+1 gives a diagram for a square-free braid β ′ ∈ Bn .
Proof. Let D be a diagram for β and D ′ be the diagram obtained by removing
a given strand. Since β is a square-free braid, there is an isotopy which allows
e where any two strands cross at most
one to pass from D to a diagram D
once. If we forget the strand we want to remove, our isotopy deforms D ′ in
e ′ obtained from D
e by removing the strand, and in D
e ′ any two
a diagram D
strands among the remaining n strands cross each other at most once since
e
they were crossing each other at most once in D.
5.2. Mikado braids.
Definition 5.2. A strand of a given reduced diagram D of a braid β ∈ Bn+1
which is over the other strands it crosses is called good.
per
Remark 5.3. A strand is good in a reduced diagram for a braid β ∈ Bn+1
if and only if it is good in any reduced diagram for β: indeed, we first
claim that the number of crossing between any two strands is constant on
the set of reduced diagrams of a square-free braid. To see this, notice that
this property holds for the reduced diagrams of the permutation obtained as
image of β in the symmetric group, and that the reduced diagrams for the
various square-free braids having this permutation as image are obtained by
replacing the crossings in the permutation diagrams by positive or negative
crossings. Hence the claim holds. Now since β is a square-free braid, any two
strands cross at most once, and obviously one cannot continuously deform a
positive crossing in a negative one. Hence the number and type of crossing
between any two strands is constant on the set of reduced diagrams for β. It
follows that a strand which is above all the other in some reduced diagram
must be above all the others in all the reduced diagrams.
Definition 5.4. We define Mikado braids recursively as
(1) The braid e is a Mikado braid in B1 .
12
FRANÇOIS DIGNE AND THOMAS GOBET
(2) A braid β ∈ Bn+1 is a Mikado braid if in any reduced diagram for β,
there exists at least one good strand and if removing any such strand
yields a Mikado braid in Bn .
Mik the set of Mikado braids. An example of a Mikado
We denote by Bn+1
braid in B7 is given in Figure 1.
Remark 5.5. Notice that Definition 5.4 is equivalent to the definition of an
f -realizable braid from [14, Section 2].
b
b
b
b
b
b
b
b
b
b
b
b
b
b
Figure 1. A mikado braid in B7 .
per
Remark 5.6. If β ∈ Bn+1
with good ith strand such that removing it yields
Mik (i.e., we can assume that the inductive
′
Mik
a braid β ∈ Bn , then β ∈ Bn+1
condition in point (2) of Definition 5.4 is true for one strand instead of any
strand). We prove it by induction on rank: indeed, assume that β has
another good j th strand. We must show that βe ∈ Bn obtained by removing
the j th strand is in BnMik . The ith and j th strand do not cross. The j th
strand of β may have become the (j − 1)th strand of β ′ but it is still good.
Mik . But β ′′ is also
In particular, we can remove it, yielding a braid β ′′ ∈ Bn−1
obtained from βe by removing the strand corresponding to the ith strand in
e hence by induction βe ∈ B Mik
β (which may be the (i − 1)th strand of β),
n
Mik is.
since β ′′ ∈ Bn−1
Proposition 5.7. Let β ∈ Bn+1 . Then β is a Mikado braid if and only if
β = x−1 y for some x, y ∈ W , hence by Proposition 4.3 if and only if β a
rational permutation braid.
Proof. First, assume that β is a rational permutation braid. We argue by
induction on n. The trivial braid e has the claimed form. Now assume
Mik and consider any reduced braid diagram for it. Then there exists
β ∈ Bn+1
at least one good strand. Removing the rightmost such strand we get a braid
β ′ ∈ BnMik which by induction can be written as x′−1 y ′ for x′ , y ′ ∈ W ′ , where
W ′ = hs1 , . . . , sn−1 i. Assuming that the strand removed joins the ith point
of the sequence above to the j th point of the sequence below, one then has
−1
−1 ′−1 y ′ s s
′
that β = s−1
n n−1 . . . sj (see Figure 2), where x and
i si+1 . . . sn x
y ′ are seen in Bn+1 under the embedding ιn : Bn ֒→ Bn+1 . But for any
k, the parabolic Coxeter element sn sn−1 . . . sk is left-reduced with respect
to W ′ . In particular for k = i, j, it implies that both y ′ sn sn−1 . . . sj and
x′ sn sn−1 . . . si lie in W .
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
13
b
−1
−1
s−1
i si+1 · · · sn
β′
b
b
b
b
b
b
b
b
b
b
b
b
b
b
ι(β ′ )
b
b
b
b
sn sn−1 · · · sj
Figure 2. Illustration of the inductive process used in the
proof of Proposition 5.7.
Conversely, assume that β = x−1 y with x, y ∈ W . Note that in a reduced
braid diagram for a (positive) permutation braid, i.e., an element of W , for
all i the ith strand is above the stands numbered 1 to i − 1. Hence in x−1 the
strand which ends at position n + 1 is above all the others, the one ending
at position n is just below, and so on. It implies that when concatenating
reduced diagrams for x−1 and y, in the resulting diagram, the (n + 1)th
strand of y is above all the others, the nth strand of y is just below, and so
on. Hence the Mikado property is satisfied, so that β is a Mikado braid.
This gives in particular a non-inductive algebraic characterization of rational permutation braids. To summarize, putting Propositions 4.3, 5.7 and
Remark 5.5 together we get:
Theorem 5.8. Let β ∈ Bn+1 . The following are equivalent:
(1)
(2)
(3)
(4)
(5)
The braid β is a Mikado braid.
The braid β is a rational permutation braid.
There exist x, y ∈ W such that β = x−1 y.
There exist x, y ∈ W such that β = xy −1 .
The braid β is f -realizable in the sense of [14].
Remark 5.9. If one is interested in enumerative combinatorics, it is natural
to ask for the number Mik(n) of Mikado braids in Bn . Using Theorem
5.8, we see that counting Mikado braids is equivalent to counting pairs of
permutations (x, y) ∈ Sn × Sn with no common left descents. Indeed, there
may be distinct pairs (x, y) and (x′ , y ′ ) ∈ W ×W such that x−1 y = x′−1 y ′ ,
but taking x and y with no common left divisor in B(W )+ gives unicity.
These pairs have been counted in [13] and correspond to the coefficients of
a power series given by a Bessel function. The first values are Mik(1) = 1,
Mik(2) = 3, Mik(3) = 19, Mik(4) = 211, . . . . There does not seem to exist
a simple closed formula for Mik(n).
5.3. Simple dual braids are Mikado braids. The aim of this Section
is to show that the images in Bn+1 of the simple elements of a dual braid
monoid Bc∗ , where c is a standard Coxeter element, are Mikado braids. To
this end we first describe a model for the group of fractions of Bc∗ using braids
in a cylinder and then explain how the embedding of Bc∗ into the Artin braid
group Bn+1 can be viewed using this model.
14
FRANÇOIS DIGNE AND THOMAS GOBET
5.3.1. Noncrossing partitions. Given a standard Coxeter element c, we graphically describe the elements of Div(c) as follows. The Coxeter elements in W
are exactly the (n+1)-cycles; a Coxeter element c is standard if and only if it
is an (n+1)-cycle c = (i1 , i2 , . . . , ik , . . . in+1 ) with i1 = 1, ik = n+1, with the
property that the sequence i1 i2 · · · ik increasing and the sequence ik · · · in in+1
decreasing (see [27, Lemma 8.2]). To represent the elements of Div(c) we use
n + 1 points labeled by i1 , i2 , . . . , ik , . . . in+1 in clockwise order on a circle.
Like in the case where c = s1 s2 · · · sn , the elements of Div(c) can be represented as a union of polygons having vertices the marked points as follows:
to each cycle occurring in the decomposition of x ∈ Div(c) into a product of
disjoint cycles, one associates the polygon obtained as convex hull of the set
of points on the circle labeled by the elements in the support of the cycle (in
particular a polygon can be reduced to an edge or even a single point). Such
a polygon with vertices labelled (i1 , i2 , . . . , ik ) in clockwise order corresponds
to the cycle (i1 , i2 , . . . , ik ) which is the unique cycle dividing c with set of
vertices {i1 , . . . , ik }. The polygons obtained are pairwise disjoint; equivalently the partition defined by the cycle decomposition of x is noncrossing
for this choice of labeling of the circle depending on c. To emphasize that the
property to be a noncrossing partition depends on the labeling we will speak
of a noncrossing partition of the sequence (i1 , i2 , . . . , ik , . . . in+1 ). So we have
defined a bijection between Div(c) and the set of noncrossing partitions of
the sequence (i1 , i2 , . . . , ik , . . . in+1 ).
It will be convenient for the proofs to draw the point with label 1 at the
top of the circle, the point with label n + 1 at the bottom, the points with
label in ik · · · in in+1 on the left and the points with label in i1 i2 · · · ik on the
right, each point having a specific height depending on its label: that is, if
P , Q are two points with respective labels i, j ∈ {1, . . . , n + 1}, i < j, then
P is higher than Q. An example is given in Figure 3. In case we represent
a noncrossing partition we may use curvilinear polygons instead or regular
polygons for a more comfortable reading (see Figure 6 on the left).
1
b
1
b
3
b
2
b
2
b
b
5
3
b
4
5
b
b
b
4
bb
6
b
6
Figure 3. Example of a labeling of the vertices given by the
standard Coxeter element c = s2 s1 s3 s5 s4 = (1, 3, 4, 6, 5, 2).
5.3.2. Dual braids and graphical representation. We follow in this part [5], in
particular loc. cit. Section 1 and Corollary 3.5, and refer to it for the results.
Definition 5.10. A dual braid is an element of the group of fractions of
Bc∗ . A simple dual braid is a dual braid which is a simple element of Bc∗ ⊂
Frac(Bc∗ ) as described in Section 3.2.
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
15
To describe the simple dual braids we consider braids where the starting
points and the ending points of the strands are on two parallel circles. The
points of the two circles are labeled as in Section 5.3.1. More precisely,
we consider one unit circle in a plane with one complex coordinate at level
t = 0 and a second unit circle in another such plane at level t = 1. So the
starting points of the strands have coordinates (zk , 0)k=1,...,n+1 with |zk | =
1, and the ending points have coordinates (zk , 1). The Coxeter element is
c = (i1 , . . . , in+1 ) as in Section 5.3.1, when along the circle in clockwise
order starting from (zi1 , 0) the startpoints of the strands appear in the order
(zi1 , 0), (zi2 , 0), . . . , (zin+1 , 0), and the same for the endpoints, replacing the
second coordinate by 1.
For each pair {i, j} one lifts the reflection (i, j) to the braid δi,j where the
strands starting with (zk , 0) for k 6= i, j has fixed first coordinate and the
strands starting with zi (resp. zj ) is represented by the path t 7→ (t, (zi +
zj )/2 + (zi − zj )/2(cos(πt) + iε sin(πt)) (resp. same with i and j exchanged)
where ε is small enough (see Figure 4). We draw the figures with the level
t = 0 above and the level t = 1 below.
i
j
i
j
Figure 4. Braid diagram for the dual braid δi,j .
These braids statisfy the relations in the dual presentation of Proposition
3.8 for the braid group Bn+1 , that is δi,j δj,k = δj,k δk,i when zi , zj , zk are in
clockwise order and δi,j δk,l = δk,l δi,j when i, j, k, l are 4 distinct points with
[i, j] and [k, l] noncrossing (these are exactely the cases where the product
of two reflections divide the Coxeter element). The simple dual braids are in
one-to-one correspondence with the noncrossing partitions of the sequence
(i1 , . . . , in+1 ) as described in Section 5.3.1. The braid diagram associated
to a simple dual braid represented by a polygon (zj1 , . . . , zjk ) where the
points are in clockwise ordering is the product δj1 ,j2 δj2 ,j3 . . . δjk−1 ,jk . In the
resulting braid diagram the strand starting with (zjm , 0) ends with (zjm−1 , 1)
for 1 < m ≤ k while the strand starting with (zj1 , 0) ends with (zjk , 1). (see
Figure 5).
If the Coxeter element is standard the dual braids δi,i+1 satisfy the usual
braid relations and the identification of the group of fractions of the dual
braid monoid with the braid group Bn+1 maps δi,i+1 to σi . Graphically to
16
FRANÇOIS DIGNE AND THOMAS GOBET
Figure 5. The braid associated to a quadrangle.
recover the usual braid diagram of the image in Bn+1 of a simple dual braid
β, one has to put the points zi1 , . . . zin+1 on the circle in such a way that
the imaginary parts are in decreasing order and “look from the right” to the
braid. This gives a diagram for an Artin braid, which is a horizontal mirror
diagram of a diagram for the image in Bn+1 of β. Equivalently starting from
the noncrossing partition representation of β, one orders the polygons in
counterclockwise order and then projects everything to the right; this gives
the braid seen from the bottom (see Figure 6).
The point in this description of the embedding which will be crucial later
and follows from our graphical descriptions of the embedding above is that
if a polygon in the noncrossing partition representation of β has no polygon
at its right (we assume that an edge and a single point are polygons), then
the strands from this polygon are above all the strands coming from other
polygons in the classical Artin braid representation of the image of β, as one
can see in Figure 6 with the polygon reduced to the edge labeled by 4: on
the very right, the strand labeled by 4 appears above all the other strands.
Remark 5.11. Note that the point where the Coxeter element needs to be
standard is the identification of δi,i+1 with σi . If the Coxeter element is not
standard the braids δi,i+1 do not satisfy the braid relations.
1
b
2
b
b
3
b
b
4
b
5
b
6
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
1 2 3 4 5 6
Figure 6. Passing from a dual to a classical braid.
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
17
5.3.3. Removing the strands of simple dual braids. The aim here is to use
the graphical representation of simple elements from the previous section to
show that simple elements of dual braid monoids are rational permutation
braids. The argumentation is therefore topological.
Any u ∈ D c always has at least one polygon. The polygons of the graphical representation correspond to the cycles (or blocks) of the corresponding
permutation (or noncrossing partition). Consider any block represented by
either a polygon or a point which we denote by P and which is right to any
other polygon or point, that is, with no other block at its right in the picture
(with our conventions described above). It always exists, since the polygons
are disjoint and with vertices on the circle. One can then inductively remove
all the strands corresponding to edges of P one after the other such that
the removed strand at each step is good, starting with any strand having
no other strand from P at its right (it always exists and one can keep going
inductively since P is a polygon): such a strand is necessarily good. After
having removed it one can find a strand which is good at this step, and so
on.
After having removed all the strands from P , we go on with another
polygon Q which is right to all the remaining polygons, and so on, until all
the strands have been removed. Hence we showed:
Theorem 5.12. Let c be a Coxeter element in a Coxeter group (W, S) of
type An . Any element of Dc ⊂ Bc∗ ⊂ B(W ) is a Mikado braid, and thus a
rational permutation braid.
5.3.4. Additional properties in case the Coxeter element is linear. We denote
by < the Bruhat order on W . In the next proposition, we show that in case
c = s1 s2 · · · sn , one can always remove a strand in a braid representing
u ∈ D c such that the resulting braid v lies in Dc′ for c′ = s1 s2 · · · sn−1 ; this
allows us to say a bit more in that case.
Proposition 5.13. Let c = s1 s2 · · · sn , u ∈ D c . Then u 6= e can be written
in the form x−1 y with x, y ∈ W and p(x) < p(y).
Proof. We argue by induction on the rank. We prove the statement with
the inequality < replaced by ≤. It is then clear that the inequality is an
equality if and only if u = e. The result is trivially true for n = 1. Assume
n > 1. With our choice of Coxeter element, all indices lie on the right
part of the circle. If there is a block of u reduced to a single index i, then
the ith strand of u must be unbraided, and moreover since all the points
lie on the right of the circle, the ith strand lies over all the other strands.
We then argue as in the proof of Proposition 5.7; that is, one has u =
−1
′
−1 ′
s−1
i si+1 . . . sn u sn sn−1 . . . si where u ∈ Bn is the braid obtained from u
by removing the ith strand. But u′ is then in D c′ , where c′ = s1 · · · sn−1 :
indeed, it is graphically obtained from the noncrossing representation of u by
removing the point with label i and subtracting 1 from each label larger than
i, still yielding a diagram of a simple element u′ for the smaller linear Coxeter
element c′ . By induction u′ = x′−1 y ′ with p(x′ ) < p(y ′ ). Since sn · · · si is
left-reduced with respect to the parabolic subgroup hs1 , . . . , sn−1 i we get
that p(x′ sn · · · si ) < p(y ′ sn · · · si ), hence we have the claimed property with
x = x′ sn · · · si , y = y ′ sn · · · si .
18
FRANÇOIS DIGNE AND THOMAS GOBET
Hence we can assume that there is no unbraided strand. But in that case,
since the Coxeter element is linear, there must be a polygon in the graphical
representation of u having an edge (i − 1, i), that is an ith strand ending at
i − 1 which is over all the other strands that it crosses. We argue as above,
−1
′
−1 ′
writing this time u = s−1
i si+1 . . . sn u sn sn−1 . . . si−1 . The element u is
the simple element for c′ obtained by contracting the edge (i − 1, i), hence
also identifying the points with labels i and i − 1 and subtracting 1 from any
point with label bigger than i. Arguing as above we get the claim.
6. Mikado braids of type Bn
6.1. Coxeter and Artin-Tits groups of type Bn . The Coxeter group
W of type Bn is the group of fixed points in the Coxeter group W ′ of type
A2n−1 under the diagram automorphism τ which exchanges si and s2n−i .
This fact lifts to the corresponding Artin-Tits groups (see e.g., [33, Corollary 4.4]). So we shall see the Artin-Tits group B(W ) of type Bn as the
group of fixed points in the Artin-Tits group of type A2n−1 under the diagram automorphism which we still denote by τ , that is, B(W ′ )τ = B(W ).
Topologically this means that the Artin-Tits group B(W ) can be seen as the
group of symmetric braids in B2n (the symmetry exchanges the ith and the
(2n + 1 − i)th strands and exchanges under-/over-crossings).
B for the braids of type B identified with B τ . For conveWe write B2n
n
2n
B consists of symmetric braids, we label the strands of B
nience, since B2n
2n by
−n, −n+1, . . . , −1, 1, . . . , n rather than by 1, 2, . . . , 2n. We write t0 , . . . , tn−1
for the Coxeter generators of W with relations
t0 t1 t0 t1 = t1 t0 t1 t0 ,
ti ti+1 ti = ti+1 ti ti+1 if 1 ≤ i ≤ n − 2,
ti tj
= tj ti
if |i − j| > 1.
′
Inside W one has t0 = sn and ti = si s2n−i for i = 1, . . . , n − 1, where
s1 = (−n, −n + 1), s2 = (−n + 1, −n + 2), . . . , sn = (−1, 1), sn+1 =
(1, 2), . . . , s2n−1 = (n − 1, n) in S2n seen as the group of permutations of
{−n, . . . , −1, 1, . . . , n}.
This lifts to the braid groups: the generators of B(W ) seen in B2n are
t0 = sn and ti = si s2n−i for i = 1, . . . , n − 1.
6.2. Mikado braids of type Bn .
B is a Mikado braid (of type B ) if when
Definition 6.1. A braid β ∈ B2n
n
viewed in B2n , starting from any diagram D for β one can inductively remove
pairs of symmetric strands, one being above all the other strands (so that
its symmetric is under all the other strands).
B,Mik
We write B2n
for the set of Mikado braids of type Bn . It follows from
B,Mik
B ∩ B Mik . Hence we can use the geometric
the definition that B2n
= B2n
2n
properties of Mikado braids of type An given in Section 5.2.
B ; then β is a Mikado braid of type B if and
Proposition 6.2. Let β ∈ B2n
n
−1
only if β = x y for some x, y ∈ W , hence by Proposition 4.3 if and only
if β a rational permutation braid.
Note that here W refers to the lift of the Coxeter group W of type Bn in
the braid group of type Bn .
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
19
B,Mik
Proof. A braid β ∈ B2n is in B2n
if and only if it is a Mikado braid in B2n
and it is fixed by τ . By Proposition 5.7, this is equivalent to β being of the
form x−1 y with x and y in W ′ and β fixed by τ . Since there are gcds in
Artin-Tits monoids of spherical type, the expression x−1 y is unique under
the condition that x and y have no common left-divisor (in the Artin-Tits
monoid of type A2n−1 ). Since τ induces an automorphism of the monoid,
if x and y satisfy the unicity condition, then x−1 y is τ -fixed if and only if
each one of x and y is τ -fixed, that is, is in W ′τ = W .
This gives in particular a non-inductive algebraic characterization of rational permutation braids of type Bn . To summarize, putting Propositions
4.3 and 6.2 together we get:
B . The following are equivalent:
Theorem 6.3. Let β ∈ B(W ) ∼
= B2n
(1)
(2)
(3)
(4)
The braid β is a Mikado braid.
The braid β is a rational permutation braid.
There exist x, y ∈ W such that β = x−1 y.
There exist x, y ∈ W such that β = xy −1 .
Problem 6.4. What is the number of Mikado braids of type Bn ? Unlike
in type An (see Remark 5.9), we have no hint towards an answer to this
question.
6.3. Simple dual braids are Mikado braids.
6.3.1. Graphical representation of simple elements. Through the embbeding
W = W ′τ ⊂ W ′ , the standard Coxeter elements of W identify with the
τ -fixed standard Coxeter elements of W ′ . Moreover if T ′ denotes the set
of reflections of W ′ and T the set of reflections of W , and if c is a τ -fixed
standard Coxeter element of W ′ , then for x and y in W one has x 4T y 4T c
if and only if x 4T ′ y 4T ′ c (see [10, Lemma 4.8]).
Remark 6.5. Note that a reflection of W is not a reflection of W ′ in general.
Since the dual braid monoid of type A2n−1 is defined by a presentation
with generators the left-divisors of the chosen Coxeter element c and with
relations given by ℓT ′ -shortest decompositions of c, the above observations
imply that the dual braid monoid of type Bn identifies with the fixed points
of τ in the dual braid monoid of type A2n−1 for the same Coxeter element.
The τ -fixed standard Coxeter elements of W ′ are the 2n-cycles of the form
c = (i1 , i2 , . . . , in , −i1 , −i2 , . . . , −in ) with {i1 , i2 , . . . , in } = {1, 2, . . . , n} and
the sequence i1 i2 · · · in first increasing, then decreasing.
We also get that the elements of Div(c) for a Coxeter element c ∈ W are
in one-to-one correspondence with the τ -fixed noncrossing partitions of 2n
points labelled −n, . . . , −1, 1, 2, . . . , n on a circle with the clockwise order on
the labels in the order given by c (see Figure 7).
These symmetric partitions are called noncrossing partitions of type Bn
associated to c (see [35]).
Such a partition corresponds to a τ -fixed braid in B2n by the same process
as explained in Subsection 5.3.2.
20
FRANÇOIS DIGNE AND THOMAS GOBET
−5
b
−3
−4
b
b
−2
b
b
−1
1
b
b
2
4
5
b
b
3
b
Figure 7. Example of a labeling of the vertices given by the
Coxeter element c = t1 t2 t0 t4 t3 and of a noncrossing partition of type B5 for this Coxeter element. The corresponding
element of Div(c) is (−5, −4, −1)(5, 4, 1)(2, 3, −2, −3).
6.3.2. Removing pairs of strands of simple dual braids. The argumentation
to show inductively that any simple dual braid x ∈ D c of type Bn is a
Mikado braid can be led as in type An , that is, by looking at the graphical
representation of x, equivalently the number of blocks of the noncrossing
partition x ∈ Div(c), and removing pairs of strands. We start from a polygon
P which is to the right of any other polygon. Either P has a symmetric
polygon P̄ such that j indexes a vertex of P if and only if −j indexes a
vertex of P̄ ; or P is its own symmetric, in particular P has vertices on both
sides of the circle and an index j indexes a vertex of P if and only if −j
indexes a vertex of P .
In the first case, we remove inductively pairs of strands of P , one corresponding to an edge of P and its symmetric which is an edge of P̄ as we did
in type An , in such an order that at each step, the removed strand from P is
above all the others (which implies that the corresponding symmetric strand
which is also removed is below all the others), that is, the corresponding edge
should be right to any other edge of P . In the second case, we do exactly
the same, except that pairs of strands of P are removed simultaneously.
After having removed all the strands from P , we go on with another
polygon Q which is right to all the remaining polygons, and so on, until all
the strands have been removed. Hence we showed:
Theorem 6.6. Let c be a Coxeter element in a Coxeter group (W, S) of type
Bn . Any element of D c ⊂ Bc∗ ⊂ B(W ) is a Mikado braid of type Bn , and
thus a rational permutation braid.
7. Dihedral and exceptional types
Theorem 7.1. Let c be a Coxeter element in a Coxeter system (W, S) of
type H3 , H4 , I2 (m) (m ≥ 3), En (n = 6, 7, 8) or F4 ; any element of D c ⊂
Bc∗ ⊂ B(W ) is a rational permutation braid.
Proof. For exceptional types we have checked the result using the program
CHEVIE ([34]). Let us prove it for dihedral types. By Proposition 4.3., it
suffices to show that any element in D c has the form xy −1 for x, y ∈ W .
Since S = {s, t} there are only two possible choices of Coxeter element. Let
c = st. Then the elements of D c are the identity element, c and the elements
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
21
of T c . By considering the Hurwitz action on (s, t) we see that the elements
−1
of T c are exactly the (sts
| {z· ·}·)(sts
| {z· ·}·) with 1 ≤ k ≤ m, which concludes
k
k−1
the proof.
8. Positivity properties
The whole paper is motivated by a positivity conjecture on the expansion
of the images of simple dual braids in a canonical basis of the Iwahori-Hecke
algebra of the Coxeter group. Using work from the previous sections we prove
these properties here for finite irreducible Coxeter groups of type other than
type Dn .
8.1. Iwahori-Hecke algebra of a Coxeter system.
Definition 8.1. Let (W, S) be a Coxeter system. For s, t ∈ S, recall that
ms,t denotes the order of st in W . The Iwahori-Hecke algebra H(W ) =
H(W, S) of (W, S) is the associative, unital Z[v, v −1 ]-algebra generated by a
copy {Ts | s ∈ S} of S with relations
Ts2 = (v −2 − 1)Ts + v −2 ∀s ∈ S,
∀s, t ∈ S.
T Tt · · · = Tt Ts · · · ,
| {z }
}
| s {z
ms,t copies
ms,t copies
The algebra H(W ) has a standard basis {Tw }w∈W where Tw is the image
of w ∈ W under the unique group morphism a : B(W ) → H(W )× such that
s 7→ Ts for all s ∈ S. It also has two canonical bases {Cw }w∈W and {Cw′ }w∈W
defined in [31]. The algebra H(W ) has a unique semilinear involution jH
such that jH (Ts ) = −v 2 Ts for all s ∈ S and jH (v) = v −1 . The two canonical
bases are then related by the equalities
(8.2)
Cw = (−1)ℓS (w) jH (Cw′ ), ∀w ∈ W.
A reference for this material is [29, Section 7.9].
8.2. Positivity properties of simple dual braids. There are many positivity statements involving the canonical bases of H(W ). In this section we
prove positivity results on the expansion of images of simple dual braids in
H(W ) when they are expressed in the basis {Cw }. This comes as a corollary
of the results of the previous sections and of the following:
Theorem 8.3 ([20, Corollary 2.9]). Let (W, S) be a finite Weyl group, then
for all x, y ∈ W , one has
X
Tx−1 Ty ∈
N[v, v −1 ]Cw .
w∈W
Remark 8.4. Dyer and Lehrer showed that in case (W, S) is a finite Coxeter
group, the statement of the above theorem (for fixed x, y) is equivalent to the
statement that Cx′ −1 Ty has a positive expansion in the standard basis; they
then show that this last statement holds for Weyl groups ([20, Proposition
2.7 and Theorem 2.8]). Dyer later showed in [18, Conjecture 7(b)]1 (using
partially unpublished results) that the Kazhdan-Lusztig positivity conjecture
1The authors thank Matthew Dyer for pointing out this fact to the second author.
22
FRANÇOIS DIGNE AND THOMAS GOBET
(now proven in [21] as a corollary of Soergel conjecture [36]) implies the
positivity of the expansion of Cx′ −1 Ty in the standard basis for all finite
Coxeter groups. Hence we can assume that Theorem 8.3 holds for all finite
Coxeter groups.
Theorem 8.3 together with Proposition 4.3 yields:
Proposition 8.5. Let β ∈ B(W ) be a rational permutation braid. Then
X
a(β) ∈
N[v, v −1 ]Cw .
w∈W
Proof. This is an immediate consequence of Theorem 8.3, Proposition 4.3
and the fact that a(w) = Tw for any w ∈ W .
Theorem 8.6. Let (W, S) be a finite irreducible Coxeter system of type other
than Dn and let c be any Coxeter element in W ; then for any u ∈ Dc , one
has that
X
a(u) ∈
N[v, v −1 ]Cw .
w∈W
Proof. This follows from Theorems 5.12, 6.6, 7.1 and Proposition 8.5.
Notice that the proof of this theorem required to prove that simple dual
braids can be written in the form x−1 y, which was done in types An and Bn
by using the geometry of Artin braids.
We are not able to prove this property for type Dn , but both computations
and the fact that it holds in any other spherical type leads us to conjecture
that it again holds:
Conjecture 8.7. Let (W, S) be a Coxeter system of type Dn . Let c be any
Coxeter element in W ; then any element of D c ⊂ Bc∗ ⊂ B(W ) is a rational
permutation braid.
Note that if this conjecture is true we will have the same positivity property for type Dn as for all other spherical types.
Problem 8.8. Does there exist a model for the braid group of type Dn by
Artin-like braids which could be used to prove that simple elements of dual
braid monoids are rational permutation braids? A model for type Dn can
be found in [1], but the analogues of the Mikado braids in this model are not
the expected ones.
Problem 8.9. Is there a uniform approach to show that simple elements of
dual braid monoids are rational permutation braids?
8.3. Temperley-Lieb algebra and monomial basis. Since elements of
W , that is, images of the simple elements for the classical Garside structure
provide a basis of the Iwahori-Hecke algebra, one might investigate the linear
independence of images of elements of Dc in H(W ). The set Dc is too small
to give a basis of H(W ): indeed, the algebra H(W ) has rank |W | while
Div(c) is in general a strict subset of W counted by the generalized Catalan
number of type W (see for instance [2]). Nevertheless, in type An , the set Dc
gives a basis of a remarkable quotient of H(W ), the Temperley-Lieb algebra.
From now on, (W, S) is a Coxeter system of type An .
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
23
Definition 8.10. The Temperley-Lieb algebra TLn is the associative, unital Z[v, v −1 ]-algebra obtained as quotient of H(W ) by the two-sided ideal
generated by the elements
X
Tw ,
w∈hsi ,si+1 i
for i = 1, . . . , n − 1, where si = (i, i + 1). We write θ : H(W ) ։ TLn for the
quotient map. Alternatively, one gets an isomorphic algebra by taking the
quotient by the two-sided ideal generated by the elements
X
(−1)ℓS (w) v 2ℓS (w) Tw ,
w∈hsi ,si+1 i
for i = 1, . . . , n − 1 (see e.g., [28, Section 2.3 and Remark 2.4]). We write
θ ′ : H(W ) ։ TLn for this alternative quotient map.
The algebra TLn
tions
bsi bsi±1 bsi =
bs i bs j =
b2si =
has a presentation by generators bs1 , . . . , bsn and relabsi
bsj bsi
(v + v −1 )bsi
∀i, i ± 1 ∈ {1, . . . , n},
∀i, j ∈ {1, . . . , n} with |i − j| > 1,
∀i ∈ {1, . . . , n}.
We have (see [28, Section 2.3 and Remark 2.4])
(8.11)
θ(Tsi ) = v −1 bsi − 1, θ ′ (Tsi ) = v −2 − v −1 bsi .
Definition 8.12. An element w ∈ W is fully commutative if one can pass
from any reduced S-decomposition of w to any other one by applying only
relations of the form si sj = sj si for |i − j| > 1. We denote by Wf the set of
fully commutative elements. For details on fully commutative elements we
refer to [37].
The algebra TLn has a basis indexed by the elements of Wf . It is built
as follows; given any w ∈ Wf with reduced expression st · · · u, the above
presentation shows that the product bs bt · · · bu does not depend on the choice
of the reduced expression and therefore we can denote it by bw . The set
{bw }w∈Wf turns out to yield a basis of TLn (see [30, Corollary 5.32]).
8.4. Projection of the canonical basis. The basis {bw }w∈Wf turns out
to be related to the canonical basis {Cw′ }w∈W of H(W ) as follows:
Theorem 8.13 ([22], Section 3). The basis {bw }w∈Wf is the projection of
{Cw′ }w∈W by θ. In symbols, for w ∈ Wf , one has θ(Cw′ ) = bw while θ(Cw′ ) =
0 for w ∈
/ Wf .
8.5. Zinno basis. The Temperley-Lieb algebra comes equipped with a semilinear involution jTLn such that jTLn (v) = v −1 and jTLn (bs ) = bs for any
s ∈ S. To realize the algebra TLn as a quotient of Z[v, v −1 ][Bn ], we use the
composition of a′ : B(W ) → H(W ), s 7→ vTs with θ ′ , that is, by (8.11) any
generator si ∈ Bn is mapped to v −1 − bsi . We set ω := θ ′ ◦ a′ .
Theorem 8.14 ([40], Theorem 2, [32], Theorem 1). Let c = sn sn−1 · · · s1 ;
then ω(D c ) yields a basis of TLn .
24
FRANÇOIS DIGNE AND THOMAS GOBET
The above Theorem is stated differently in [40] since dual braid monoids
had not been introduced when that paper appeared. Notice that the conventions used in [40] correspond to the quotient map θ ◦ a, which sends si
to v −1 bsi − 1. Since one passes from one to the other by composing with a
semilinear automorphism of Z[v, v −1 ]-modules and we are here interested in
bases, this does not affect the result.
This result is also generalized to arbitrary standard Coxeter elements in
[39, Corollary 5.2.9] and [25, Theorem 3.8.28 and Remark 3.8.29]:
Theorem 8.15. For any standard Coxeter element c, the set ω(D c ) yields
a basis of TLn .
Hence the images of the simple elements of the Garside monoids Bc∗ yield
a basis of TLn , in much the same way that the images of the simple elements
of the Garside monoid B + (W ) yield the standard basis of H(W ).
Note that one may have x ∈ Div(c) ∩ Div(c′ ) for some c′ 6= c, in which
case the corresponding elements of D c and Dc′ , hence also the corresponding
basis elements of TLn are not equal in general. For example, in type A2
with c = s1 s2 , c′ = s2 s1 , the reflection s1 s2 s1 lies in Div(c) ∩ Div(c′ ) (as any
reflection). The corresponding element of D c , resp. Dc′ is s1 s2 s−1
1 , resp.
s−1
s
s
.
The
two
elements
are
not
equal.
2
1
1
8.6. Positivity consequences. The two quotient maps from Section 8.3
are related by θ = jTLn ◦ θ ′ ◦ jH . Using this fact together with Theorems
8.6, 8.13 and Equality (8.2) we get:
Theorem 8.16. Let c be any standard Coxeter element, let {Zx }x∈Div(c) be
the corresponding basis of TLn , i.e., Zx := ω(x) for x ∈ Div(c). Then for
any x ∈ Div(c), one has
X
N[v, v −1 ](−1)ℓS (w) bw .
Zx ∈
w∈Wf
One can show that there exist total orders on the sets Div(c) and Wf
such that the base change from {bw }w∈Wf to {Zx }x∈Div(c) is upper triangular. Zinno shows it for c = s1 s2 · · · sn in [40] and the result is extended
in the second author’s thesis [25] to arbitrary c. As noticed in [26], for
c = s1 s2 · · · sn , the order given by Zinno can be refined into the Bruhat
order on Div(c), which is studied extensively and combinatorially in [27],
where the orders to consider for other choices of Coxeter elements are also
described. Closed formulas for some of the coefficients (in the inverse matrix) are given in [26] in case the Coxeter element is c = s1 s2 · · · sn (which
can be obviously adapted for c = sn sn−1 · · · s1 ), but we do not have a closed
formula for them in full generality.
Remark 8.17. In type Bn , Theorem 8.15 also holds (see [38]). In that case,
the diagram basis of the Temperley-Lieb quotient is not the projection of the
canonical basis of the corresponding Iwahori-Hecke algebra (see [22]), hence
there is no analogue of Theorem 8.16 in that case. In type Dn , the cardinality
of |D c | is bigger than the dimension of the Temperley-Lieb algebra, but the
images of the elements of D c in it still generate the whole algebra, as shown
in [39].
DUAL BRAID MONOIDS, MIKADO BRAIDS AND POSITIVITY
25
References
[1] D. Allcock, Braid pictures for Artin groups, Trans. Amer. Math. Soc. 354 (2002),
345–3474.
[2] D. Armstrong, Generalized noncrossing partitions and combinatorics of Coxeter
groups, Mem. Amer. Math. Soc. 202 (2009), no. 949.
[3] B. Baumeister, M. Dyer, C. Stump, and P. Wegener, A note on the transitive Hurwitz
action on decompositions of parabolic Coxeter elements, Proc. Amer. Math. Soc., Ser.
B, 1 (2014), 149–154.
[4] D. Bessis, The dual braid monoid, Ann. Sci. École Normale Supérieure 36 (2003),
647–683.
[5] D. Bessis, F. Digne, and J. Michel, Springer theory in braid groups and the BirmanKo-Lee monoid, Pacific J. Math. 205 (2002), 287–309.
[6] J. Birman, K.H. Ko, and S.J. Lee, A New Approach to the Word and Conjugacy
Problems in the Braid Groups, Adv. in Math. 139 (1998), 322–353.
[7] N. Bourbaki, Groupes et algèbres de Lie, chapitres 4,5 et 6, Masson (1981).
[8] E. Brieskorn and K. Saito, Artin-Gruppen und Coxeter-Gruppen, Invent. Math. 17
(1972), 245–271.
[9] N. Brady, J. McCammond, B. Mühlherr, and W. Neumann, Rigidity of Coxeter groups
and Artin groups, Geom. Dedicata 94 (2002), 91–109.
[10] T. Brady and C. Watt, K(π, 1)’s for Artin groups of finite type, Geom. Dedicata 94
(2002), 225–230.
[11] T. Brady and C. Watt, Noncrossing partitions lattices in finite real reflection groups,
Trans. Amer. Math. Soc. 360 (2008), 1983–2005.
[12] M. Broué and J. Michel, Sur certains éléments réguliers des groupes de Weyl et les
variétés de Deligne-Lusztig associées, Prog. Math. 141 (1997), 73–139.
[13] L. Carlitz, R. Scoville, and T. Vaughan, Enumeration of pairs of permutations and
sequences, Bull. Amer. Math. Soc. 80 (1974), 881–884.
[14] P. Dehornoy, Three-dimensional realizations of braids, J. London Math. Soc. 60
(1999), 108–132.
[15] P. Dehornoy, F. Digne, D. Krammer, E. Godelle, and J. Michel. Foundations of
Garside theory, Tracts in Mathematics 22, Europ. Math. Soc. (2015).
[16] P. Dehornoy and L. Paris, Gaussian groups and Garside groups, two generalizations
of Artin groups, Proc. London Math. Soc. 79 (1999), 569–604.
[17] P. Deligne, Les immeubles des groupes de tresses généralisés, Invent. Math. 17 (1972),
273–302.
[18] M.J. Dyer, Representation theories from Coxeter groups, Representations of groups
(Banff, AB, 1994), CMS Conf. Proc. 16, (1995), 105–139 Amer. Math. Soc., Providence, RI.
[19] M.J.
Dyer,
Modules
for
the
dual
nil
Hecke
ring,
http://www3.nd.edu/~dyer/papers/nilhecke.pdf.
[20] M.J. Dyer and G.I. Lehrer, On positivity in Hecke algebras, Geom. Dedicata 35
(1990), 115–125.
[21] B. Elias and G. Williamson, The Hodge theory of Soergel bimodules, Ann. of Math.
180 (2014), 1089–1136.
[22] C.K. Fan and R.M. Green, Monomials and Temperley-Lieb algebras, J. Algebra 190
(1997), 498–517.
[23] M. Geck and G. Pfeiffer, Characters of finite Coxeter groups and Iwahori-Hecke algebras, London Math. Soc. Monographs, New Series 21, Oxford University Press (2000).
[24] W.N. Franzsen, R.B. Howlett, and B. Mühlherr Reflections in abstracts Coxeter
groups Comment. Math. Helv. 81 (2006) 665–697
[25] T. Gobet, Bases of Temperley-Lieb algebras, PhD thesis, Université de Picardie,
(2014), http://www.mathematik.uni-kl.de/fileadmin/AGs/agag/gobet/files/these.pdf.
[26] T. Gobet, Noncrossing partitions, fully commutative elements and bases of the
Temperley-Lieb algebra, accepted in Journal of Knot Theory and its Ramifications,
http://arxiv.org/abs/1409.6500.
26
FRANÇOIS DIGNE AND THOMAS GOBET
[27] T. Gobet and N. Williams, Noncrossing partitions and Bruhat order, European Journal of Combinatorics 53 (2016), 8-34.
[28] T. Halverson, M. Mazzocco and A. Ram, Commuting families in Hecke and
Temperley-Lieb algebras, Nagoya Math. J. 195 (2009), 125–152.
[29] J. Humphreys, Reflection groups and Coxeter groups, Cambridge Studies in Advanced
Mathematics 29, Cambridge University Press (1990).
[30] C. Kassel, V. Turaev, Braid groups, Graduate Texts in Mathematics 247 (2008),
Springer, New York.
[31] D. Kazhdan and G. Lusztig, Representations of Coxeter Groups and Hecke Algebras,
Invent. Math. 53 (1979), 165–184.
[32] E.K. Lee and S.J. Lee, Dual presentation and linear basis of the Temperley-Lieb
algebras, J. Korean Math. Soc. 47 (2010), 445–454.
[33] J. Michel, A note on words in braid monoids, J. Algebra 215 (1999), 366–377.
[34] J. Michel, The development version of the CHEVIE package of GAP3, J. Algebra
435 (2015), 308–336.
[35] V. Reiner, Non-crossing partitions for classical reflection groups, Discrete Math. 177
(1997), 195–222.
[36] W. Soergel, Kazhdan-Lusztig polynomials and indecomposable bimodules over polynomial rings, J. Inst. Math. Jussieu 6 (2007), 501–525.
[37] J.R. Stembridge, On the fully commutative elements of Coxeter groups J. of Algebraic
Comb. 5 (1996), 353–385.
[38] C. Vincenti, Algèbre de Temperley-Lieb de type B, C. R. Math. Acad. Sci. Paris 342
(2006), 233–236.
[39] C. Vincenti, Monoïde dual, antichaînes de racines et algèbres de Temperley-Lieb,
PhD thesis, Université de Picardie, (2007).
[40] M.G. Zinno, A Temperley-Lieb basis coming from the braid group, J. Knot Theory
Ramifications 11 (2002), 575–599.
LAMFA, Université de Picardie Jules Verne, 33 Rue Saint-Leu, 80039
Amiens Cedex 1, France.
E-mail address: [email protected]
TU Kaiserslautern, Fachbereich Mathematik, Postfach 3049, 67653 Kaiserslautern, Germany.
E-mail address: [email protected]
| 4math.GR
|
1
Analysis of Feedback Error in Automatic
Repeat reQuest
Saeed R. Khosravirad and Harish Viswanathan
arXiv:1710.00649v1 [cs.IT] 2 Oct 2017
Nokia - Bell Labs
Abstract
The future wireless networks envision ultra-reliable communication with efficient use of the limited
wireless channel resources. Closed-loop repetition protocols where retransmission of a packet is enabled
using a feedback channel has been adopted since early days of wireless telecommunication. Protocols
such as automatic repeat request (ARQ) are used in today’s wireless technologies as a mean to provide
the link with reduced rate of packet outage and increased average throughput. The performance of such
protocols is strongly dependent to the feedback channel reliability. This paper studies the problem of
feedback error and proposes a new method of acknowledging packet delivery for retransmission protocols
in unreliable feedback channel conditions. The proposed method is based on backwards composite
acknowledgment from multiple packets in a retransmission protocol and provides the scheduler of
the wireless channel with additional parameters to configure ultra-reliable communication for a user
depending on channel quality. Numerical analysis are presented which show orders of magnitude increase
in reliability of the proposed method as compared to ARQ at the cost of a small increase in average
experienced delay.
I. I NTRODUCTION
Repetition of a packet over non-deterministic channel conditions is a prominent approach to
reliable packet delivery. Wireless telecommunications technologies such as high speed packet
access (HSPA), worldwide interoperability for microwave access (WiMax) and long term evolution (LTE), to mention a few, have relied on the performance boost provided by retransmission
techniques such as ARQ and hybrid automatic repeat request (HARQ) [1]. Such retransmission
protocols add to the robustness of transmission and increase link throughput. In LTE, and as
expected for the 5th generation mobile networks (5G) [2], ARQ is used in the radio link control
(RLC) layer while HARQ in the lower Media Access Control (MAC) and upper physical layer
(PHY) layer. Performing together, these retransmission protocols provide the system with high
October 3, 2017
DRAFT
2
reliability where failure in the MAC layer HARQ operation is compensated for by the RLC layer
ARQ in acknowledged mode at the expense of extra experienced latency for the packet [3].
The role of feedback channel is to limit repetitions to only when the initial attempt is failed
thus, increasing data channel efficiency. However, inevitable feedback channel impairments may
cause unreliability in packet delivery. A decoding failure report, i.e. negative acknowledgement
(NAK), falsely received as positive acknowledgement (ACK) results in undesirable packet outage.
Attempts to increase feedback reliability, e.g., by means of repetition coding, is costful to
the receiver node while erronoeus feedback detection may cause an increased packet delivery
latency and diminish throughput and reliability key performance indicators (KPIs). E.g., in LTE
a single-bit ACK/NAK spans over multiple resource element (RE) up to a physical resource
block (PRB) in up-link (UL) and down-link (DL) HARQ respectively to reduce false feedback
detection [4], making feedback bits too costly to the network. In newer releases of LTE, blind
HARQ retransmissions of a packet is considered as a solution to avoid feedback complexity
of broadcast HARQ and increase reliability [5]. Such approach, despite the offered simplicity,
can severely decrease resource utilization efficiency of the system considering that typically a
high percentage of transmissions are successfully decoded in the initial attempt in typical link
adaptation configurations.
The core question this paper tries to answer is how to reliably design a feedback-based
retransmission protocol in unreliabile feedback conditions. We first study the effect of erroneous
feedback on performance of retransmission protocols. We assume a simple stop-and-wait (SAW)
mode of operation in a narrow-band wireless link where the receiver node is a low-cost and
low-energy device with limited power for feedback channel acknowledgement reports. Such
model portrays well the unreliable feedback channel problem where the straight-forward solution
to acquire reliable packet delivery is by either adding diversity gain to the feedback link or
relaxing the dependency to feedback channel and performing blind or conservative retransmission
of the packet. Specifically, for low-cost narrow-band communication such diversity gain can
be achieved by increasing time diversity order of the feedback channel. We study different
approaches of increasing feedback channel time diversity and establish achievable reliability
regions with respect to feedback channel error rate. We show that in reasonably reliable feedback
channel conditions where the product of packet error rate and feedback error rate is comparable
to packet outage rate, conservative asymmetric feedback detection can provide the required
October 3, 2017
DRAFT
3
reliability level by slightly increasing false NAK rate while reducing false ACK rate. Further, in
extremely un-reliable feedback channel conditions we see that blind retransmission of packet is
the viable solution in terms of reliability while it zeros the receiver node’s energy consumption
over feedback channel.
Next, we propose a new method of backwards composite acknowledgment that helps improves
reliability of repetition process without the need to increase time diversity order of the feedback
channel. The proposed scheme relies on collaboration between transmitter and receiver nodes
to provide ultra-reliable communication of packets even in poor feedback channel conditions.
Furthermore, thanks to the additional design parameters provided by the proposed method, the
scheduler of wireless network is able to configure each communication node with desirable ultrareliability only using one layer of retransmission protocol. This enables the wireless technologies
such as LTE to adopt one layer of retransmission protocol with configurable reliability level as
opposed to stacked two-layer ARQ/HARQ operation that is currently deployed.
The rest of this paper is organized as follows: in Sec. II the unreliability problem of retransmission protocols caused by feedback channel unreliability feedback error problem is studied;
Sec. III introduces the backwards composite feedback solution for reliable packet delivery; in
Sec. IV numerical results are presented; to evaluate the performance of the proposed solution;
finally, Sec. V covers the concluding remarks.
II. P ROBLEM
DESCRIPTION
In this study we adopt the SAW mode of operation for retransmission protocols which works
as follows. First, at time i the jth packet arrived from a higher layer application denoted by Pji
is transmitted by transmitter node for the first time. Next, receiver node attempts decoding on
the observed packet denoted by P̃ji . Using a feedback channel, receiver node sends the decoding
success report Ai at corresponding feedback instance i, where Ai = 1 in case of ACK and Ai = 0
in case of NAK (respectively, decoding success and decoding failure). Feedback transmission,
similar to the data transmission, is assumed to be subject to channel impairments. We use Ãi
to denote the feedback observed by the transmitter node at feedback time instance i. In case of
observing a NAK the same data packet 1 is retransmitted at the next transmit time instance i + 1
1
In practice the same message can be conveyed in different set of coded bits called redundancy versions.
October 3, 2017
DRAFT
4
(i.e., Pji+1 ), otherwise, transmission of a new data packet is initiated (i.e., Pj+1
i+1 ). Retransmission
of a NAKed packet continues until ACK is observed over the feedback channel or maximum M
transmission attempts for the packet is reached. Therefore, at transmitter node a packet is only
regarded as delivered if ACK is observed and otherwise it is regarded as failed. The transmitter
node sends a single-bit new data indicator (NDI) message per transmit data packet P. The singlebit NDI is toggled every time a packet is transmitted for the first time. We assume that receiver
is able to detect NDI error-free. The duration between transmit occasions i and i + 1 is denoted
by round trip time (RTT) where only one packet transmit occasion and one feedback occasion
are considered in each RTT.
Reliability of packet delivery in SAW operation with feedback channel is limited by both
packet transmission block error probability (BLEP) and feedback detection error rate. We use
pe = p1e and pm
e for integer m to denote BLEP of a packet after one and m transmission attempts
respectively where, by definition p0e = 1. We assume independent and identically distributed
m
(i.i.d.) block fading channel model for packet transmission. Therefore, we have pm
e ≤ (pe )
where equality holds only if the decoder utilizes no combining gain (e.g., in case of ARQ
operation). Feedback channel is assumed to follow the binary asymmetric channel (BAC) model
where error probability varies depending on the input symbol to the channel. Error probabilities
for such channel model are described as follows.
n
o
p0 = Pr Ãi = 1|Ai = 0
n
o
p1 = Pr Ãi = 0|Ai = 1
(1)
(2)
We assume that instances of the feedback channel are independent from each other and from data
channel. Such model for the feedback channel is simplified as compared to real-life feedback
channel where an extra message (e.g., discontinued transmission (DTX)) may also be considered
as input to the feedback channel. E.g., in case of LTE technology, DTX may indicate failure in
detection of the scheduling grant for data transmission [3]. Throughout this paper we reserve
the notation ā to denote ā = 1 − a for any real valued a where a ∈ [0, 1].
In a retransmission protocol where retransmissions are triggered by NAK feedback, in case of
NAK→ACK error the transmitter node will mistakenly drop the packet assuming it is successfully
decoded at the receiver. Therefore, it is crucial to reduce the effective chances of a packet
October 3, 2017
DRAFT
5
being discarded as a result of false ACK. The straightforward solution to reduce chances of
NAK→ACK is to increase reliability of the feedback channel (i.e., reducing p0 ) e.g., by increasing
repetition order of feedback transmission by factor of L > 1. However, such solution stretches
feedback message in time, frequency or power domains, requiring extra resources. Specifically,
in scenarios where receiver node has limited power and bandwidth for feedback transmission
(e.g., narrow-band and low-cost massive machine type of communication (mMTC) receivers)
the cost of increasing feedback reliability is additional time diversity for feedback which in turn
increases the experienced delay and receiver node power consumption. We use T to denote the
number of feedback occasions utilized for a packet before it is dropped at the transmitter (either
considered delivered or failed). The average number of feedback occasions utilized per packet
is then denoted by T̄ = E{T } which, in time diversity scenario for feedback, is equivalent to
the average delay experienced by the higher-layer application per packet. Further, we denote
the events of decoding failure and success for Pj in maximum m transmission attempts by Fjm
and Sjm respectively. Outage probability, Pout , is defined as the probability of decoding failure
after maximum M attempts, i.e., Pout = Pr FjM . Data channel utilization is measured by the
average number of transmission attempts per packet and is denoted by N̄ . We use τ to denote the
packet delivery latency defined as the time duration it takes from the first transmission attempt
of a packet until it is correctly decoded at the receiver. Assuming zero processing time at the
receiver node we have τ = TTI + k × RTT where k > 0.
In the following we first study the effects of feedback error on Pout performance of retransmission protocols. Later in the next section we proposes a new feedback reporting scheme that
suits low-cost and low-energy narrow-band wireless devices in ultra-reliable packet delivery. We
start by investigating different feedback reporting approaches and analyze the trade-off between
reliability and feedback time diversity order L.
1) Regular SAW (Reg-SAW): We assume the binary symmetric channel (BSC) model with
p0 = p1 = p for feedback channel. As shown in Fig. 1, in a SAW process, packet Pji is initially
transmitted over ith (blue color) transmit occasion which has duration of a transmission time
interval (TTI). The numbers inside the transmit occasion blocks indicates data packet index j
from Pji where j is a positive integer. Followed by transmission of the packet, after a given
propagation and receiver processing time [6], acknowledgement for the packet arrives. Next,
transmitter node transmits the next packet Pj+1
i+1 in case of ACK (green color blocks) or retransmits
October 3, 2017
DRAFT
6
a version of the same packet Pji+1 (grey color blocks) in case of NAK (red color blocks). NDI is
transmitted with each packet transmit occasion to notify the receiver of whether a new packet is
transmitted (toggled NDI) or the same packet is retransmitted (un-toggled NDI). In principle ACK
observance can be result of either a successful packet decoding followed by correct feedback
detection or decoding failure followed by false feedback. In order to make it simple to follow
the illustrations the packet index corresponding to each feedback occasion is shown inside the
feedback occasion blocks.
The duration between transmit occasions i and i + 1 is denoted by RTT. Without loss of
generality, the propagation and processing time duration will be skipped in the illustrations
after Fig. 1. Therefore, acknowledgement of each packet will be shown below it while the next
transmit occasion starts immediately after.
RTT
transmit occasion
feedback occasion
j+1
j
j
j+1
j+1
j+1
j +1
j+2
j+1
j +2
Figure 1: Stop-and-wait operation.
2) Increased feedback repetition order (L-Rep-ACK): To increase feedback reliability for a receiver node with narrow-band low-energy feedback transmission a simple solutions is to increase
time diversity of feedback transmission by L > 1. In this model each feedback transmission is
stretched over L feedback occasions where packet is declared as delivered at the transmitter
node only if all L observances of feedback are ACK. Otherwise, the packet is retransmitted by
the L-Rep-ACK process. Therefore, probability of false ACK reduces to pL0 compared to that
of p0 in case of Reg-SAW. Further, due to feedback repetition, RTT of L-Rep-ACK is L times
that of Reg-SAW. NDI is used in similar way as in Reg-SAW to notify receiver node about
retransmissions.
3) L required ACK per packet (L-ACK-SAW): In this approach, the acknowledgment generated
for a packet is repeated over feedback occasions by the receiver node until L > 1 number of
ACK observances are made at the transmitter node which in turn will trigger initiating the
transmission of a new packet. A retransmission of the same packet is followed immediately if
October 3, 2017
DRAFT
7
transmit occasion
1
feedback observation
1
1
1
1
1
2
1
1
2
Figure 2: L-Rep-ACK operation for L = 3.
NAK is observed while using NDI receiver is notified about the retransmission. Note that in this
approach transmitter node keeps counting the number of ACK observances for a packet and L
required ACK observances may be received in non-consecutive occasions unlike L-Rep-ACK
approach where L observances of ACK must be counted consecutively.
transmit occasion
1
feedback observation
1
1
1
1
2
1
2
Figure 3: L-ACK-SAW operation for L = 3.
4) Retransmission until L ACKs are observed (ReTx-L-ACK): In this approach, similar to
L-ACK-SAW, transmitter node requires L observance of ACK before considering a packet
as delivered. However, transmitter continues retransmission of the packet while observing the
feedback channel. Therefore, ReTx-L-ACK transmits each packet at least L times and stops
retranmission when L times ACK observances are made or maximum M transmission attempts
is reached. Fig. 4 depicts ReTx-L-ACK process for L = 3 where retransmission of packet P1 is
continued until 3 non-consecutive ACKs are detected.
transmit occasion
1
1
1
1
2
feedback observation
1
1
1
1
2
Figure 4: ReTx-L-ACK operation for L = 3.
5) Asymmetric feedback detection for SAW (Asym-SAW): A different approach to decrease the
false ACK rate is by using asymmetric detection of the binary feedback channel. For instance,
let’s assume an additive white Gaussian noise (AWGN) feedback channel and a binary phase shift
October 3, 2017
DRAFT
8
keying (BPSK) symbol that conveys the single-bit feedback acknowledgement, where the binary
0 and 1 inputs of the feedback channel are represented in the signal constellation for BPSK in
√
√
terms of energy per bit Eb respectively by − Eb and Eb . Assuming coherent detection and
perfect recovery of the carrier frequency and phase, from signal modulation and detection theory
we know that the bit error probability (BEP) with symmetric decision regions [7] is as follows,
r
Eb
1
,
(3)
p = erfc
2
N0
where p denotes detection error probability, N0 denotes the additive noise power spectral density
R∞
and the complementary error function erfc (·) is defined as, erfc (x) = π2 x exp(−t2 ) dt. Asymmetric decision regions, e.g., by moving the detection threshold in the BPSK constellation from
√
√
√
origin to the point α× Eb (closer to Eb than − Eb for α > 0), decreases the modified chances
of false ACK detection q0 , while the modified false NAK rate q1 increases accordingly. This
reduces the chances of discarding unsuccessful packets at the transmitter while in turn increases
chances of unnecessary retransmissions. The modified error probabilities for such asymmetric
detection is then as follows.
r
Eb
1
q0 = erfc (1 + α)
2
N0
r
Eb
1
q1 = erfc (1 − α)
2
N0
(4)
(5)
Asym-SAW follows similar algorithm as in Reg-SAW where NDI signal is utilized to notify
receiver node of retransmissions. For performance evaluation of this approach we assume BPSK
modulation is used for the feedback channel where Eb is chosen based on a given p in (3). Then,
the detection threshold is adjusted using parameter α in (4) and (5) to provide the required q0 .
6) Blind retransmission (Blind-ReTx): We further investigate the performance of blind retransmission without feedback. In such approach each packet is transmitted M times by the
transmitter node without requiring a feedback message from the receiver node.
Closed-form formulation for Pout , N̄, T̄ and cumulative density function (cdf) of packet delivery
latency are shown in Table I, Table II, Table III and Table IV respectively for the approaches
described in this section.
For an infinite allowed number of transmission attempts (M → ∞) and assuming zero
m
combining gain at the receiver (i.e., pm
e = (pe ) ), in Fig. 5 the outage probability Pout of
October 3, 2017
DRAFT
9
Table I: Outage probability, Pout for different feedback approaches.
M →∞
finite M
Reg-SAW
M
−1
P
m−1
M −1
pm
+ pM
e p0 p¯0
e p¯0
m=1
L-Rep-ACK
M
−1
P
m=1
L-ACK-SAW
L
L m−1
L M −1
pm
+ pM
e p0 (1 − p0 )
e (1 − p0 )
PM −1
m=1
L m−1
pm
e p0 p¯0
L+m−2
m−1
PL−1
l+M −2
M −2
M −1
pM
e p¯0
ReTx-L-ACK
PM −1
m=L
M
−1
P
PL−1
l=0
m−1
L−1
( pp¯00 )l
+
+
M −1
l
m−1
M −1
pm
+ pM
e q0 q¯0
e q¯0
m=1
Blind Retx
pl0
L m−L
pm
e p0 p¯0
M −1
pM
e p¯0
Asym-SAW
l=0
pM
e
≤ pe p0 1−p1 p¯0
e
≤ pe pL0 1−p
1
L
e (1−p0 )
≈ pe pL0 1 +
pe p¯0
(L−1)! (1−pe p¯0 )L
≈ pLe pL0 1 +
pe p¯0
(L−1)! (1−pe p¯0 )L
≤ pe q0 1−p1 q¯0
e
→0
the above-listed feedback approaches is illustrated. Reliability of Reg-SAW is proportional to
feedback reliability metric p even in unbounded M scenario. Specifically, the operation regions
on Fig. 5 that are labeled by R1, ..., R6 for Pout below 10−4 and 10−5 are not achievable using
Reg-SAW. Therefore, for ultra-reliable communication it is required to either increase feedback
diversity order L or to perform blind retransmission without reliance on unreliable feedback
channel. By increasing feedback time diversity order by L = 2, reliability regions R1 and
R4 are achievable with proper choice of M. However, achieving regions R2 and R5 requires
L > 2. Interestingly, Blind-ReTx with M = 5 and 6 can achieve reliability in regions R2 and
R5 respectively by refusing the dependency on feedback channel. However, as we see later
in this paper such approach can be harmfully inefficient in resource utilization. Nevertheless,
for highly unreliable feedback channel conditions such as in region R6, very large feedback
diversity order L >> 4 can have reverse effect on the performance efficiency parameters such
as channel utilization and average delay. This makes Blind-ReTx a viable solution for when
the feedback channel is unable to offer a reasonable level of reliability. Further, Asym-SAW
feedback operation requires stringent q0 adjustment of, e.g., q0 < 10−4 for Pout < 10−5 . In highly
October 3, 2017
DRAFT
10
Table II: Average number of transmission attempt per packet, N̄
M →∞
finite M
M
P
pm−1
p¯0 m−1 +
e
m=1
M
−1
P
Reg-SAW
m=1
L-Rep-ACK
M
P
m=1
M
−1
P
m=1
m−1
(pm−1
− pm
p
e
e )p¯0
<
−m
1−pM
1
1
p¯1
(1 − pL0 )m−1 +
pm−1
e
<
L M −m
L m−1
(pm−1
− pm
(1 − p¯1 L ) 1−(1−pp¯¯11L )
e
e )(1 − p0 )
p¯1 +p¯e p1
p¯1 (1−pe p¯0 )
p¯1 L +p¯e (1−p¯1 L )
p¯1 L (1−pe (1−pL
0 ))
1 + gM XLM −1 + g¯M YLM −1
where, ∀l ∈ {1, ..., L} and ∀m ∈ {1, ..., M − 1},
L-ACK-SAW
Xlm = p¯0
l
P
ĺ
pl−
1 + gm Xĺm−1 + g¯m Yĺm−1 ,
0
ĺ=1
Ylm = p1
l
P
< 1+
pe (1−p¯0 L )+p¯e (1−p¯1 L )
p¯1 L
p¯1 l−ĺ (1 + Yĺm )
ĺ=1
and, Xl0 , Yl0 = 0, X0m , Y0m = 0, and gm =
M
P
peM −m+1
−m
pM
e
mηm , where,
m=L
L m−L
ηm = pm
e p0 p¯0
ReTx-L-ACK
̺m,l =
m−1
L−1
min{l−1,L−1}
P
k=max{L−m+l−1,0}
+
m
P
(pl−1
− ple )̺m,l , with
e
l=1
m−l
pk0 p¯0 l p¯1 L pm+k+1
l−1
1
L+l
k+1
k
k
L−k−1
p¯0
p¯1 p1
...
and,
̺M,l =
min{l−1,n}
P
L−1
P
n=0 k=max{n−M +l,0}
Asym-SAW
M
P
Blind Retx
October 3, 2017
pm−1
q¯0 m−1 +
e
m=1
M
−1
P
m=1
+k
pk0 p¯0 l p¯1 n pM
l−1 M −l
1
l+n
k
n−k
p¯0 k+1 p¯1 k p1
m−1
(pm−1
− pm
q1
e
e )q¯0
M
1−q1M −m
q¯1
<
q¯1 +p¯e q1
q¯1 (1−pe q¯0 )
∞
DRAFT
11
Table III: Average experienced delay, T̄ , in number of RTT.
assuming one feedback occasion per RTT
N̄
Reg-SAW
N̄ ∗ L
L-Rep-ACK
1 + gM XLM −1 + g¯M YLM −1
where, ∀l ∈ {1, ..., L} and ∀m ∈ {1, ..., M − 1},
Xlm = p¯0
L-ACK-SAW
l
P
ĺ=1
Y l m = p1
l
P
´
l=1
ĺ
pl−
l − ´l + 1 + gm Xĺm−1 + g¯m Yĺm−1 ,
0
p¯1 l−ĺ (1 + Yĺm )
and, Xl0 , Yl0 = 0, X0m , Y0m = 0, and gm =
ReTx-L-ACK
N̄
Asym-SAW
N̄
Blind Retx
M
peM −m+1
−m
pM
e
Table IV: cdf of packet delivery latency τ .
Pr {τ ≤ TTI + k ∗ RTT}
Reg-SAW
L-Rep-ACK
L-ACK-SAW
p¯0 k (pke − pk+1
e ),
(1 − pL0 )n (pne − pn+1
),
e
min{k,M
P −1}
m=max{1,k−L+1}
ReTx-L-ACK
n = k ∗ L, ∀k ∈ {0, ..., M − 1}
p¯0 k (pke − pk+1
)p0k−m
e
l−1
m−1
,
∀k ∈ {0, ..., M + L − 2}
(pke − pk+1
e ) for k ∈ {0, ..., L − 1}, and
k
P
(pke − pk+1
)p0k−m p¯0 m
e
max{1,k−L+1}
Asym-SAW
q¯0 k (pke − pk+1
e ),
Blind Retx
pke − pk+1
,
e
October 3, 2017
∀k ∈ {0, ..., M − 1}
k
m
for k ∈ {L, ..., M − 1}
∀k ∈ {0, ..., M − 1}
∀k ∈ {0, ..., M − 1}
DRAFT
12
10-1
Reg-SAW
2-Rep-ACK
4-Rep-ACK
2-ACK-SAW
4-ACK-SAW
Asym-SAW, q0 = min{10−3 , p}
Asym-SAW, q0 = min{10−4 , p}
ReTx-2-ACK
ReTx-4-ACK
10-2
PSfrag replacements
Pout
10-3
10-4
R2
R3
R1
10-5
10-6
R4
R5
R6
10-1
10-2
10-3
10-4
p
Figure 5: Outage probability for M → ∞ against a range of p = p0 = p1 values with pe = 0.1.
unreliable feedback channel conditions, with q0 → 0, false NAK rate increases drastically (i.e.,
q1 → 1) which increases number of transmission attempts resulting in similar performance as
Blind-ReTx approach.
Moreover, from Fig. 5 it is observed that reliability performance of L-Rep-ACK and LACK-SAW approaches are tightly similar in practical range of feedback diversity order L. The
downside of increasing feedback diversity order is the increased energy consumption at the
receiver to report more than one feedback per packet transmission (i.e., T̄ ). In particular, for use
cases where battery life-time is of critical importance, less energy consumption over feedback
reports is desirable. Thus, it is required to reduce feedback energy consumption while configuring
high reliability for the retransmission protocol. This motivates next section of this paper where
we propose a variant of the L-ACK-SAW approach which similarly requires L observances of
ACK for a data packet to be considered as delivered. However, the new solution reduces number
October 3, 2017
DRAFT
13
of average feedback occasions used per packet thanks to the proposed backwards composite
feedback operation. This way, T̄ improves compared to L-ACK-SAW while thanks to the required
multiple ACK observance per packet, a higher reliability of operation is expected compared to
Reg-SAW.
III. BACKWARDS
COMPOSITE FEEDBACK
In this section we propose a composite feedback solution to provide highly-reliable SAW
operation in unreliable feedback channel condition, which can be applied to retransmission
protocols such as ARQ/HARQ. In the proposed backwards composite feedback (BCF) solution,
the aim is to observe a given L > 1 times ACK for a packet before the packet is labeled reliably
as delivered. In order to avoid drastic increase in T̄ , in the proposed BCF solution we suggest to
repeat the feedback for each packet in a composite manner. We assume that a BCF process has
at most L active packets in its buffers at each time instance. An active packet is identified as a
packet that has been transmitted m times, where 1 ≤ m < M, and ACK feedback is observed
for it less than L times. We define composite feedback at the receiver node at time i as follows,
where l denotes index of the active packets set.
l
1 A = 1, ∀l
Ci = &Al =
l
0 otherwise
(6)
Therefore, an observed composite feedback C̃i = 1 at the transmitter is counted as ACK for
all active packets. In the case C̃i = 0 is observed, a retransmission phase cycle starts which
attempts on retransmitting active packets one after another following a given order of packets.
The retransmission phase then continues until C̃ = 1 is observed or the maximum transmission
attempt is reached for all active packets. The propose BCF-SAW (BCF-SAW) operation at
transmitter and receiver nodes is as follows.
A. Operation at the transmitter
Alg. 1 presents the BCF algorithm at the transmitter side. We assume that all the active packets
are stored in separate buffers at the transmitter for in case a retransmission is needed. An active
packet is then represented by Buffer(l) for l ∈ [0, ..., L − 1]. The variable NDItoggle stores the
October 3, 2017
DRAFT
14
Algorithm 1: Operation at the transmitter
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
Input : observed composite feedback C̃i
Output: transmit packet Pi ; new data indicator NDIi
if C̃i == 1 then
ACKcounter(l) + + ∀l;
NDIi+1 ← toggle(NDIi );
NDItoggle + + mod L;
Buffer(NDItoggle) ←get new packet ;
ACKcounter(NDItoggle) = 0;
NAKcounter(NDItoggle) = 0;
Pi+1 ← Buffer(NDItoggle);
TXcountre(NDItoggle) = 1;
clear Indx;
else
if TXcounter(l) == 0 or M ∀l, then
go to 3
else
NDIi+1 ← NDIi ;
NAKcounter(l) + + ∀l;
Indx ← look up reTx index TXcounter, ACKcounter, NAKcounter ;
Pi+1 ← Buffer(Indx);
TXcountre(Indx) + +;
end
end
i + +;
return Pi , NDIi ;
go to 1;
index l of the last active packet which was transmitted for the first time (i.e., NDI was toggled
for it). When transmitter node observes ACK over the feedback channel it increments counters
ACKcounter(l) by one for all l. Then, NDI is toggled and NDItoggle index is incremented by
one
mod L. The updated NDItoggle points either to an empty buffer or to an active packet
where ACKcounter(NDItoggle) = L. Buffer(NDItoggle) is therefore reset and substituted by a
new packet taken from the higher layer application (this process is denoted by function get new
packet). Next transmit occasion is then utilized to transmit the newly initiated active packet. A
toggle in NDI bit informs the receiver node about transmission of a new packet.
In the case where NAK is observed, the retransmission phase of the operation starts where
active packets are retransmitted one after another following a given order until ACK is observed
or the maximum transmission attempt is reached for all active packets. The order in which
active packets are retransmitted in this phase is given in a look up table that is pre-shared
between receiver and transmitter nodes. The look-up table identifies the next active packet index
denoted by variable Indx that is to be retransmitted. This process is denoted by function look up
October 3, 2017
DRAFT
15
reTx index which inputs counters TXcounter(l), ACKcounter(l) and NAKcounter(l) respectively
denoting number of transmission attempts, observed NAKs and observed ACKs for active packet
l. NDI signal remains untoggled during retransmission phase.
Algorithm 2: Operation at the receiver
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Input : observed NDI; received packet P̃i
Output: composite feedback Ci
if NDI is toggled then
ACKcounter(l) + + ∀l ∈ [0, ..., L − 1];
NDItoggle + + mod L;
Buffer(NDItoggle) ← P̃i ;
RXcountre(NDItoggle) = 1;
ANDItoggle ← decode success(P̃i )
else
NAKcounter(l) + + ∀l;
Indx ← look up reTx index RXcounter, ACKcounter, NAKcounter ;
RXcountre(Indx) + +;
AIndx ← decode success(P̃i , Buffer(Indx));
Buffer(Indx) ← combine(P̃i , Buffer(Indx));
end
i + +;
Ci ← &Al where l = NDItoggle or, RXcounter(l) 6= 0, M ;
l
16
17
return Ci ;
go to 1;
B. Operation at the receiver
Receiver node follows Alg. 2 where firstly it detects whether the received packet is a new
transmission or it is retransmission of one of the earlier received active packets. This is done at
each time i by observing NDIi signal and comparing it with NDIi−1 . The variable NDItoggle ∈
[0, ..., L − 1] store the index of the active packet which is most recently received for the first
time. If NDI is detected to be toggled, NDItoggle index at the receiver is incremented by one
mod L. Then, Buffer(NDItoggle) is substituted with the newly received packet. Function decode
success outputs the feedback generated from decoding lth active packet, denoted using upper
case index by Al . We assume zero chance of error detection failure where the acknowledgement
for a packet is generated based on error detection for the packet, e.g., using cyclic redundancy
check (CRC).
In case NDI signal is detected as untoggled, the algorithm follows retransmission phase
operation where the pre-shared look-up table is used to find the index of the active packet that
is being retransmitted (denoted by Indx). Decoding in such case may be based on the received
October 3, 2017
DRAFT
16
retransmission and the stored version of the active packet from previous (re)transmission attempts
to provide the decoder with combing gain. Similarly, the buffer content may be updated (e.g.,
in case of AIndx = 0) after combining the two versions of the packet. Composite feedback Ci is
generated using Al for all the active packet indexes l according to (6).
The retransmission phase order of packets follows a pre-shared look-up table. As discussed
above we assume that both nodes keep track of the number of transmission attempts and the
number of observed ACK and NAK for each active packet. Given the value of these counters
the next packet to be retransmitted in case of NAK is found.
C. Example case for L = 2
The proposed BCF operation is explained below for the case of L = 2, i.e., twice ACK
observances required for a packet to be considered as delivered). The cases of larger L will
follow a similar approach.
Operation in case of observed composite ACK, C̃i = 1: As depicted in Fig. 6, two active
packets are assumed at the transmitter each having a separate TXcounter and ACKcounter. The
process starts by transmitting packet P11 thus, the first feedback occasion only acknowledges
decoding status for packet 1, C1 = A1 . Transmitter then initiates the second active packet by
transmitting packet P22 in the following transmit occasion. We assume that receiver is notified
of the new packet using a single-bit NDI signal.
In the second feedback occasion, receiver composites acknowledgement the two active packets
resulting in composite NAK due to decoding failure for P22 , C2 = A1 &A2 = 0. The observed
composite feedback at time i = 2 is however erroneously detected as ACK, i.e., C̃2 = 0.
The observed ACK feedback counts as one ACK for both active packets resulting. As a result,
ACKcounter for packet 1 reaches L = 2 and the packet is regarded as delivered and thus
discarded from BCF-SAW process. Note that for ease of following the illustration, TXcounter
blocks in Fig. 6 show the counter value and the corresponding packet number in brackets.
Next, P33 is transmitted and decoded successfully. However, since A2 = 0, the composite
feedback at time i = 3 is C3 = 0. A second feedback error at time i = 3 results in C̃3 = 1 and
as a result P2 is discarded from the transmitter buffer even though it failed in decoding. Such
outage case requires L = 2 times NAK→ACK errors during the time a packet is in an active
packet buffer of the transmitter node.
October 3, 2017
DRAFT
17
transmit occasion
1
2
3
4
5
composite feedback
1
1&2
2&3
3&4
4&5
feedback observation
1
1&2
2&3
3&4
4&5
1(1)
1(1)
1(3)
1(3)
1(5)
transmitter ACKcounter(1)
1
2
1
2
1
transmitter TXcounter(2)
0
1(2)
1(2)
1(4)
1(4)
transmitter ACKcounter(2)
0
1
2
1
2
transmitter TXcounter(1)
Figure 6: Backwards feedback bundling operation in case of composite ACK.
Operation in case of observed composite NAK, C̃i = 0: An observed composite NAK initiates
retransmission phase where active packets are retransmitted one after another according to a preshared order of packets. NDI remains un-toggled during retransmission phase. The retransmission
phase continues until ACK is observed or the transmit counter for all active packets reaches M.
In Fig. 7 we assume similar events have encountered as in Fig. 6 up to feedback occasion i = 3.
Let’s assume that at i = 3 in Fig. 7 composite feedback is correctly detected, C̃3 = 0. From
transmitter point of view, observed composite NAK may be caused by several events including
decoding failure of one of the active packets or feedback channel error. For instance, the observed
composite NAK C̃3 in Fig. 7 may be the result of any of the following events.
•
E1: F31 &S21 followed by successful feedback detection at i = 3
•
E2: F21 &S31 with successful feedback detection at i = 3 and feedback error at i = 2
•
E3: F21 &F31 with successful feedback detection at i = 3 and feedback error at i = 2
•
E4: S21 &S31 followed by feedback error at i = 3
The likelihood of such events can be evaluated as it was explained in Sec. ??. E.g., for packet
transmission with target pe = 0.1 and feedback channel reliability of p = 0.001, E1 is the
more likely event making P3 the best candidate for retransmission in next transmit occasion.
Nevertheless, the retransmission packet order as a function of TXcounter, ACKcounter and
NAKcounter can be established prior to start of the communication process and shared between
communicating nodes.
October 3, 2017
DRAFT
18
In Fig. 7, retransmission of P34 is performed at i = 4. Due to failure of P22 , NAK composite
feedback C4 = 0 is generated at the receiver node and correctly detected at the transmitter.
Therefore, retransmission phase continues and transmitter uses the look-up table to pull the
next index of active packet that must be retransmitted. Thus, at time i = 5, P25 is retransmitted
resulting in S22 .
transmit occasion
1
2
3
3
2
composite feedback
1
1&2
2&3
2&3
2&3
feedback occasion
1
1&2
2&3
2&3
2&3
1(1)
1(1)
1(3)
2(3)
2(3)
transmitter ACKcounter(1)
1
2
0
0
1
transmitter TXcounter(2)
0
1(2)
1(2)
1(2)
2(2)
transmitter ACKcounter(2)
0
1
1
1
2
transmitter TXcounter(1)
Figure 7: Backwards composite feedback operation in case of observed NAK. Feedback error
has occurred in the second feedback occasion.
The proposed BCF-SAW uses the same number of feedback occasions per packet transmit
occasions as in Reg-SAW in Fig. 1 without the need to increase time diversity of feedback
channel. However, each packet is ACKed L times over the feedback channel which in turn
increases the reliability of packet delivery.
IV. N UMERICAL
RESULTS
In this section we evaluate the packet outage probability of the proposed BCF-SAW against
the range of feedback channel reliability metrics p and compare it with benchmark approaches
introduced in Sec. II. All the results are presented for the case of M = 4 for repeated Monte
m−1 g
Carlo analysis where pe = 0.1 and the combining gain is modeled by g in pm
) for
e = (pe
m > 1. For the case of ARQ operation, combining gain is set to g = 1 while for the case
of HARQ operation we assume g = 1.2. The retransmission protocols are allowed to transmit
only one packet (either initial transmission or retransmission) per transmit occasion. We further
October 3, 2017
DRAFT
10-1
Asym-HARQ, q0 = min{10−5 , p}
Asym-HARQ, q0 = min{10−4 , p}
BCF-ARQ, L = 2
BCF-ARQ, L = 4
BCF-HARQ, L = 2
BCF-HARQ, L = 4
2-ACK-ARQ
4-ACK-ARQ
10-2
outage probability, Pout
g replacements
19
10
2-ACK-HARQ
4-ACK-HARQ
-3
10-4
R2
R3
R1
10-5
10-6
R4
R5
R6
10-1
10-2
10-3
10-4
p
Figure 8: Outage probability for p0 = p1 = p. The four black-colored markers from top to
bottom represent Reg-SAW for ARQ and HARQ and Blind-ReTx for ARQ and HARQ.
adopt the assumption of an error-free NDI detection at the receiver node to solely focus on the
effects of unreliable feedback channel. Performance of the proposed BCF-SAW is evaluated for
different number of required ACK observances, L. We assume a simple retransmission phase
packet order where upon observing a composite NAK the last active packet is retransmitted until
an ACK is observed. Otherwise, when transmit counter reaches M for the packet retransmission
phase order switches to the next last active packet and repeats the same process until all active
packets reach M transmission attempts or an ACK is observed.
Best case outage probability for HARQ and ARQ operations reaches
Q
m
pm
e resulting in
4.285e − 6 and 1e − 4 limits respectively as shown in Fig. 8. The proposed BCF-SAW reduces
October 3, 2017
DRAFT
20
7
Reg-SAW
BCF, L = 2
BCF, L = 4
6
L-ACK-SAW, L = 2
L-ACK-SAW, L = 4
L-Rep-ACK, L = 2
T̄ [× RTT]
5
L-Rep-ACK, L = 4
Asym, q0 = min{10−5 , p}
Asym, q0 = min{10−4 , p}
4
Blind-ReTx
ag replacements
3
2
1
10-1
10-2
10-3
10-4
p
Figure 9: Average number of feedback occasions utilized per packet packets T̄ for p0 = p1 = p
and combining gain g = 1.2.
outage probability by orders of magnitude e.g., for L = 2 and L = 4 as compared to Reg-SAW
even in highly unreliable feedback cases while its outage performance is bounded by L-ACKSAW. The latter performs more reliably because the L observances of ACK are separately
received for a given packet and a NAK feedback will trigger retransmission of the same packet.
On the other hand, NAK observance in BCF-SAW may be followed by retransmission of a
packet other than the failed packet incurring additional feedback occasions which may increase
the false ACK rate. The better outage performance of L-ACK-SAW and L-Rep-ACK is thanks to
the increased number of channel uses per packet N̄ as shown in Fig. 10. However, the penalty paid
for improved reliability by the two latter approaches, as shown in Fig. 9, is an increased average
October 3, 2017
DRAFT
21
number of feedback occasions per packet T̄ which is equivalent to the average experienced delay
by higher layer application. T̄ increases almost linearly by increasing L for those two methods
resulting in a significantly higher penalty as compared to the the proposed BCF-SAW.
In Fig. 11, complementary cumulative density function (ccdf) of packet delivery latency is
shown for all the approaches achieving Pout ≤ 10−5 in R5 from Fig. 8. The best case latency
performance is achieved using Blind-ReTx and Asym-SAW with q0 = 10−5 . While L-ACK-
SAW with L = 4 provides a better latency statistic than BCF-SAW, it fails in the average delay
experienced by the higher layer application. This increases number of feedback reporting per
packet transmit occasion roughly by L. On the other hand, BCF-SAW uses roughly the same
average number of feedback reporting per packet as compared to Reg-SAW while providing
higher reliability. By comparison of the presented numerical results in different ultra-reliability
operation regions the following observations can be made.
•
In fairly reliable feedback channel conditions, e.g., lower p regime in region R4, AsymSAW provides high reliability with low T̄ which makes it a viable choice for only when p
is ideally low.
•
In unreliable feedback conditions, e.g., region R5 and higher p regime in R4, BCF-SAW
is the most viable solution for ultra-reliable communication with low energy and low cost
receiver type where low T̄ is required. For use cases with low latency requirement, LACK-SAW approach performs better if L < M however, it requires relaxed limitations on
receiver node energy consumption and assuming low traffic channel (i.e., where high T̄ is
tolerated). Otherwise, for L ≥ M, Blind-ReTx performs more efficiently than L-ACK-SAW
with less energy consumption for feedback and guaranteed ultra reliability.
•
In extremely unreliable feedback channel conditions, e.g., region R6, Blind-ReTx is the
better choice over Asym-SAW, providing similar performance without the need for feedback
channel.
V. C ONCLUSIONS
We proposed a new method of acknowledging packet delivery for unreliable feedback channel
conditions. The proposed method, dubbed BCF-SAW, relies on backwards composite acknowledgement and provides the retransmission protocols with configurable ultra-reliability. It further
October 3, 2017
DRAFT
22
ag replacements
average number of channel use per packet, N̄
4
Reg-SAW
BCF, L = 2
BCF, L = 4
L-ACK-SAW, L = 2
L-ACK-SAW, L = 4
L-Rep-ACK, L = 2
L-Rep-ACK, L = 4
Asym, q0 = min{10−5 , p}
Asym, q0 = min{10−4 , p}
3.5
3
2.5
Blind-ReTx
2
1.5
1
10-1
10-2
10-3
p
Figure 10: Average channel use per packet in number of transmit occasions for M = 4, p0 =
p1 = p and pe = 0.1.
provides the scheduler of the wireless system with new degrees of freedom to configure the communication link in order to meet the desirable reliability requirement even in highly-unreliable
feedback channel conditions. The presented numerical analysis show orders of magnitude increase in reliability of the retransmission protocols over the practical range of target block error
rate only at the cost of a negligible increase in average experienced packet delay. System-level
performance analysis of the proposed method in more realistic multi-user communication systems
with time-varying channel conditions will be studied as future work.
October 3, 2017
DRAFT
23
100
BCF, L = 4
L-ACK-SAW, L = 4
L-Rep-ACK, L = 4
ag replacements
10-1
Asym-SAW, q0 = 10−5
Blind-ReTx
ccdf
10-2
10-3
10-4
10-5
t0
t1
t2
t3
t4
t5
t6
t7
t8
t9
t10
t11
t12
t13
t14
t15
packet delivery latency
Figure 11: ccdf of the packet delivery latency at p = 2e − 2 with combining gain g = 1.2, where
tk = TTI + k ∗ RTT.
R EFERENCES
[1] S. Lin et al., “Automatic-repeat-request error-control schemes,” Comm. Mag., vol. 22, no. 12, pp. 5–17, Dec. 1984.
[2] “3GPP TR 38.802; study on new radio access technology physical layer aspects,” 3GPP, Tech. Rep., Mar. 2017.
[3] “3GPP TS 36.212; evolved universal terrestrial radio access (E-UTRA); multiplexing and channel coding,” 3GPP, Tech.
Spec., Dec. 2016, release 10.
[4] “3GPP TS 36.213; evolved universal terrestrial radio access (E-UTRA); physical layer procedures,” 3GPP, Tech. Spec., Jan.
2015, release 10.
[5] “3GPP TR 36.877; LTE device to device (D2D) proximity services (ProSe),” 3GPP, Tech. Rep., Mar. 2015, release 12.
[6] S. R. Khosravirad et al., “Enhanced HARQ design for 5G wide area technology,” in IEEE Vehicular Technology Conference
(VTC’16), May 2016.
[7] A. Goldsmith, Wireless Communications.
October 3, 2017
Cambridge University Press, Aug. 2007.
DRAFT
| 7cs.IT
|
Under review as a conference paper at ICLR 2017
A RECURRENT NEURAL NETWORK WITHOUT CHAOS
James von Brecht
Department of Mathematics
California State University, Long Beach
Long Beach, CA 90840, USA
[email protected]
arXiv:1612.06212v1 [cs.NE] 19 Dec 2016
Thomas Laurent
Department of Mathematics
Loyola Marymount University
Los Angeles, CA 90045, USA
[email protected]
A BSTRACT
We introduce an exceptionally simple gated recurrent neural network (RNN) that
achieves performance comparable to well-known gated architectures, such as
LSTMs and GRUs, on the word-level language modeling task. We prove that our
model has simple, predicable and non-chaotic dynamics. This stands in stark contrast to more standard gated architectures, whose underlying dynamical systems
exhibit chaotic behavior.
1
I NTRODUCTION
Gated recurrent neural networks, such as the Long Short Term Memory network (LSTM) introduced
by Hochreiter & Schmidhuber (1997) and the Gated Recurrent Unit (GRU) proposed by Cho et al.
(2014), prove highly effective for machine learning tasks that involve sequential data. We propose
an exceptionally simple variant of these gated architectures. The basic model takes the form
ht = θt
tanh(ht−1 ) + ηt
tanh(W xt ),
(1)
where stands for the Hadamard product. The horizontal/forget gate (i.e. θt ) and the vertical/input
gate (i.e. ηt ) take the usual form used in most gated RNN architectures. Specifically
θt := σ (Uθ ht−1 + Vθ xt + bθ )
and
ηt := σ (Uη ht−1 + Vη xt + bη )
(2)
where σ(x) := (1 + e−x )−1 denotes the logistic sigmoid function. The network (1)–(2) has quite
intuitive dynamics. Suppose the data xt present the model with a sequence
10 if t = T
(W xt )(i) =
(3)
0
otherwise,
where (W xt )(i) stands for the ith component of the vector W xt . In other words we consider an
input sequence xt for which the learned ith feature (W xt )(i) remains off except at time T . When
initialized from h0 = 0, the corresponding response of the network to this “impulse” in the ith
feature is
if t < T
0
ht (i) ≈ ηT if t = T
(4)
αt if t > T
with αt a sequence that relaxes toward zero. The forget gate θt control the rate of this relaxation.
Thus ht (i) activates when presented with a strong ith feature, and then relaxes toward zero until
the data present the network once again with strong ith feature. Overall this leads to a dynamically
simple model, in which the activation patterns in the hidden states of the network have a clear cause
and predictable subsequent behavior.
Dynamics of this sort do not occur in other RNN models. Instead, the three most popular recurrent
neural network architectures, namely the vanilla RNN, the LSTM and the GRU, have complex,
irregular, and unpredictable dynamics. Even in the absence of input data, these networks can give
rise to chaotic dynamical systems. In other words, when presented with null input data the activation
patterns in their hidden states do not necessarily follow a predictable path. The proposed network
(1)–(2) has rather dull and minimalist dynamics in comparison; its only attractor is the zero state,
1
Under review as a conference paper at ICLR 2017
and so it stands at the polar-opposite end of the spectrum from chaotic systems. Perhaps surprisingly,
at least in the light of this comparison, the proposed network (1) performs as well as LSTMs and
GRUs on the word level language modeling task. We therefore conclude that the ability of an RNN
to form chaotic temporal dynamics, in the sense we describe in Section 2, cannot explain its success
on word-level language modeling tasks.
In the next section, we review the phenomenon of chaos in RNNs via both synthetic examples and
trained models. We also prove a precise, quantified description of the dynamical picture (3)–(4) for
the proposed network. In particular, we show that the dynamical system induced by the proposed
network is never chaotic, and for this reason we refer to it as a Chaos-Free Network (CFN). The
final section provides a series of experiments that demonstrate that CFN achieve results comparable
to LSTM on the word-level language modeling task. All together, these observations show that
an architecture as simple as (1)–(2) can achieve performance comparable to the more dynamically
complex LSTM.
2
C HAOS IN R ECURRENT N EURAL N ETWORKS
The study of RNNs from a dynamical systems point-of-view has brought fruitful insights into
generic features of RNNs (Sussillo & Barak, 2013; Pascanu et al., 2013). We shall pursue a brief
investigation of CFN, LSTM and GRU networks using this formalism, as it allows us to identify
key distinctions between them. Recall that for a given mapping Φ : Rd 7→ Rd , a given initial time
t0 ∈ N and a given initial state u0 ∈ Rd , a simple repeated iteration of the mapping Φ
ut+1 = Φ(ut ) t > t0 ,
ut0 = u0
t = t0 ,
defines a discrete-time dynamical system. The index t ∈ N represents the current time, while the
point ut ∈ Rd represents the current state of the system. The set of all visited states O+ (u0 ) :=
{ut0 , ut0 +1 , . . . , ut0 +n , . . .} defines the forward trajectory or forward orbit through u0 . An attractor
for the dynamical system is a set that is invariant (any trajectory that starts in the set remains in
the set) and that attracts all trajectories that start sufficiently close to it. The attractors of chaotic
dynamical systems are often fractal sets, and for this reason they are referred to as strange attractors.
Most RNNs generically take the functional form
ut = Ψ(ut−1 , W1 xt , W2 xt , . . . , Wk xt ),
(5)
th
where xt denotes the t input data point. For example, in the case of the CFN (1)–(2), we have
W1 = W , W2 = Vθ and W3 = Vη . To gain insight into the underlying design of the architecture of
an RNN, it proves usefull to consider how trajectories behave when they are not influenced by any
external input. This lead us to consider the dynamical system
ut = Φ(ut−1 )
Φ(u) := Ψ(u, 0, 0, . . . , 0),
(6)
which we refer to as the dynamical system induced by the recurrent neural network. The timeinvariant system (6) is much more tractable than (5), and it offers a mean to investigate the inner
working of a given architecture; it separates the influence of input data xt , which can produce
essentially any possible response, from the model itself. Studying trajectories that are not influenced
by external data will give us an indication on the ability of a given RNN to generate complex and
sophisticated trajectories by its own. As we shall see shortly, the dynamical system induced by a
CFN has excessively simple and predictable trajectories: all of them converge to the zero state. In
other words, its only attractor is the zero state. This is in sharp contrast with the dynamical systems
induced by LSTM or GRU, who can exhibit chaotic behaviors and have strange attractors.
The learned parameters Wj in (5) describe how data influence the evolution of hidden states at each
time step. From a modeling perspective, (6) would occur in the scenario where a trained RNN has
learned a weak coupling between a specific data point xt0 and the hidden state at that time, in the
sense that the data influence is small and so all Wj xt0 ≈ 0 nearly vanish. The hidden state then
transitions according to ut0 ≈ Ψ(ut0 −1 , 0, 0, . . . , 0) = Φ(ut0 −1 ).
We refer to Bertschinger & Natschläger (2004) for a study of the chaotic behavior of a simplified
vanilla RNN with a specific statistical model, namely an i.i.d. Bernoulli process, for the input data as
well as a specific statistical model, namely i.i.d. Gaussian, for the weights of the recurrence matrix.
2
Under review as a conference paper at ICLR 2017
Figure 1: Strange attractor of a 2-unit LSTM. Successive zooms (from left to right) reveal the selfrepeating, fractal nature of the attractor. Colored boxes depict zooming regions.
2.1
C HAOTIC BEHAVIOR OF LSTM AND GRU IN THE ABSENCE OF INPUT DATA
In this subsection we briefly show that LSTM and GRU, in the absence of input data, can lead to
dynamical systems ut = Φ(ut−1 ) that are chaotic in the classical sense of the term (Strogatz, 2014).
Figure 1 depicts the strange attractor of the dynamical system:
h
o tanh (f c + i g)
ut = t
u 7→ Φ(u) =
(7)
ct
f c+i g
i := σ(Wi h + bi ) f := σ(Wf h + bf ) o := σ(Wo h + bo ) g := tanh(Wg h + bg )
induced by a two-unit LSTM with weight matrices
−1 −4
4
1
−2
Wi =
Wo =
Wf =
−3 −2
−9 −7
0
6
−6
Wg =
−1
6
−6
−9
(8)
(9)
and zero bias for the model parameters. These weights were randomly generated from a normal
distribution with standard deviation 5 and then rounded to the nearest integer. Figure 1(a) was
obtained by choosing an initial state u0 = (h0 , c0 ) uniformly at random in [0, 1]2 × [0, 1]2 and
plotting the h-component of the iterates ut = (ht , ct ) for t between 103 and 105 (so this figure
should be regarded as a two dimensional projection of a four dimensional attractor, which explain
its tangled appearance). Most trajectories starting in [0, 1]2 × [0, 1]2 converge toward the depicted
attractor. The resemblance between this attractor and classical strange attractors such as the Hénon
attractor is striking (see Figure 5 in the appendix for a depiction of the Hénon attractor). Successive
zooms on the branch of the LSTM attractor from Figure 1(a) reveal its fractal nature. Figure 1(b) is
an enlargement of the red box in Figure 1(a), and Figure 1(c) is an enlargement of the magenta box
in Figure 1(b). We see that the structure repeats itself as we zoom in.
The most practical consequence of chaos is that the long-term behavior of their forward orbits can
exhibit a high degree of sensitivity to the initial states u0 . Figure 2 provides an example of such
behavior for the dynamical system (7)–(9). An initial condition u0 was drawn uniformly at random
in [0, 1]2 × [0, 1]2 . We then computed 100, 000 small amplitude perturbations û0 of u0 by adding a
small random number drawn uniformly from [−10−7 , 10−7 ] to each component. We then iterated
(7)–(9) for 200 steps and plotted the h-component of the final state û200 for each of the 100, 000
trials on Figure 2(a). The collection of these 100, 000 final states essentially fills out the entire
attractor, despite the fact that their initial conditions are highly localized (i.e. at distance of no more
than 10−7 ) around a fixed point. In other words, the time t = 200 map of the dynamical system
will map a small neighborhood around a fixed initial condition u0 to the entire attractor. Figure 2(b)
additionally illustrates this sensitivity to initial conditions for points on the attractor itself. We take
an initial condition u0 on the attractor and perturb it by 10−7 to a nearby initial condition û0 . We
then plot the distance kût − ut k between the two corresponding trajectories for the first 200 time
steps. After an initial phase of agreement, the trajectories strongly diverge.
The synthetic example (7)–(9) illustrates the potentially chaotic nature of the LSTM architecture.
We now show that chaotic behavior occurs for trained models as well, and not just for synthetically
generated instances. We take the parameter values of an LSTM with 228 hidden units trained on the
3
Under review as a conference paper at ICLR 2017
(a) Final state û200 for 105 trials
(b) Distance kût − ut k between 2 trajectories
Figure 2: (a): A small neighborhood around a fixed initial condition u0 , after 200 iterations, is
mapped to the entire attractor. (b): Two trajectories starting starting within 10−7 of one another
strongly diverge after 50 steps.
(a) No input data
(b) With input data
Figure 3: 228-unit LSTM trained on Penn Treebank. (a): In the absence of input data, the system
is chaotic and nearby trajectories diverge. (b): In the presence of input data, the system is mostly
driven by the external input. Trajectories starting far apart converge.
Penn Treebank corpus without dropout (c.f. the experimental section for the precise procedure). We
then set all data inputs xt to zero and run the corresponding induced dynamical system. Two trajectories starting from nearby initial conditions u0 and û0 were computed (as before û0 was obtained
by adding to each components of u0 a small random number drawn uniformly from [−10−7 , 10−7 ]).
Figure 3(a) plots the first component h(1) of the hidden state for both trajectories over the first
1600 time steps. After an initial phase of agreement, the forward trajectories O+ (u0 ) and O+ (û0 )
strongly diverge. We also see that both trajectories exhibit the typical aperiodic behavior that characterizes chaotic systems. If the inputs xt do not vanish, but come from actual word-level data, then
the behavior is very different. The LSTM is now no longer an autonomous system whose dynamics
are driven by its hidden states, but a time dependent system whose dynamics are mostly driven by
the external inputs. Figure 3(b) shows the first component h(1) of the hidden states of two trajectories that start with initial conditions u0 and û0 that are far apart. The sensitivity to initial condition
disappears, and instead the trajectories converge toward each other after about 70 steps. The memory of this initial difference is lost. Overall these experiments indicate that a trained LSTM, when it
is not driven by external inputs, can be chaotic. In the presence of input data, the LSTM becomes a
forced system whose dynamics are dominated by external forcing.
Like LSTM networks, GRU can also lead to dynamical systems that are chaotic and they can also
have strange attractors. The depiction of such an attractor, in the case of a two-unit GRU, is provided
in Figure 6 of the appendix.
2.2
C HAOS - FREE BEHAVIOR OF THE CFN
The dynamical behavior of the CFN is dramatically different from that of the LSTM. In this subsection we start by showing that the hidden states of the CFN activate and relax toward zero in a
predictable fashion in response to input data. On one hand, this shows that the CFN cannot produce
non-trivial dynamics without some influence from data. On the other, this leads to an interpretable
model; any non-trivial activations in the hidden states of a CFN have a clear cause emanating from
4
Under review as a conference paper at ICLR 2017
data-driven activation. This follows from a precise, quantified description of the intuitive picture
(3)–(4) sketched in the introduction.
We begin with the following simple estimate that sheds light on how the hidden states of the CFN
activate and then relax toward the origin.
Lemma 1. For any T, k > 0 we have
H
|hT +k (i)| ≤ Θk |hT (i)| +
max |(W xt )(i)|
1 − Θ T ≤t≤T +k
where Θ and H are the maximum values of the ith components of the θ and η gate in the time
interval [T, T + k], that is:
Θ=
max
T ≤t≤T +k
θt (i) and
H=
max
T ≤t≤T +k
ηt (i).
This estimate shows that if during a time interval [T1 , T2 ] one of
(i) the embedded inputs W xt have weak ith feature (i.e. maxT ≤t≤T +k |(W xt )(i)| is small),
(ii) or the input gates ηt have their ith component close to zero (i.e. H is small),
occurs then the ith component of the hidden state ht will relaxes toward zero at a rate that depends
on the value of the ith component the the forget gate. Overall this leads to the following simple
picture: ht (i) activates when presented with an embedded input W xt with strong ith feature, and
then relaxes toward zero until the data present the network once again with strong ith feature. The
strength of the activation and the decay rate are controlled by the ith component of the input and
forget gates. The proof of Lemma 1 is elementary —
Proof of Lemma 1. Using the non-expansivity of the hyperbolic tangent, i.e. | tanh(x)| ≤ |x|, and
the triangle inequality, we obtain from (1)
|ht (i)| ≤ Θ |ht−1 (i)| + H
max
T ≤t≤T +k
|(W xt )(i)|
whenever t is in the interval [T, T + k]. Iterating this inequality and summing the geometric series
then gives
1 − Θk
|hT +k (i)| ≤ Θk |hT (i)| +
H max |(W xt )(i)|
T ≤t≤T +k
1−Θ
from which we easily conclude.
We now turn toward the analysis of the long-term behavior of the the dynamical system
ut = ht ,
u 7→ Φ(u) := σ (Uθ u + bθ )
tanh(u).
(10)
induced by a CFN. The following lemma shows that the only attractor of this dynamical system is
the zero state.
Lemma 2. Starting from any initial state u0 , the trajectory O+ (u0 ) will eventually converge to the
zero state. That is, limt→+∞ ut = 0 regardless of the the initial state u0 .
Proof. From the definition of Φ we clearly have that the sequence defined by ut+1 = Φ(ut ) satisfies
−1 < ut (i) < 1 for all t and all i. Since the sequence ut is bounded, so is the sequence vt :=
Uθ ut + bθ . That is there exists a finite C > 0 such that (Uθ ut )(i) + bθ (i) < C for all t and i. Using
the non-expansivity of the hyperbolic tangent, we then obtain that |ut (i)| ≤ σ(C)|ut−1 (i)|, for all t
and all i. We conclude by noting that 0 < σ(C) < 1.
Lemma 2 remains true for a multi-layer CFN, that is, a CFN in which the first layer is defined by (1)
and the subsequent layers 2 ≤ ` ≤ L are defined by:
(`)
(`)
ht = θt
(`)
(`)
tanh(ht−1 ) + ηt
(`−1)
tanh(W (`) ht
).
Assume that W xt = 0 for all t > T , then an extension of the arguments contained in the proof of
the two previous lemmas shows that
(`)
|hT +k | ≤ C(1 + k)(`−1) Θk
5
(11)
Under review as a conference paper at ICLR 2017
where 0 < Θ < 1 is the maximal values for the input gates involved in layer 1 to ` of the network,
and C > 0 is some constant depending only on the norms kW (j) k∞ of the matrices and the sizes
(j)
|hT | of the initial conditions at all previous 1 ≤ j ≤ ` levels. Estimate (11) shows that Lemma 2
remains true for multi-layer architectures.
(a) First layer
(b) Second layer
Figure 4: A 2-layer, 224-unit CFN trained on Penn Treebank. All inputs xt are zero after t = 1000,
i.e. the time-point indicated by the dashed line. At left: plot of the 10 “slowest” units of the first
layer. At right: plot of the 10 slowest units of the second layer. The second layer retains information
much longer than the first layer.
Inequality (11) shows that higher levels (i.e. larger `) decay more slowly, and remain non-trivial,
while earlier levels (i.e. smaller `) decay more quickly. We illustrate this behavior computationally
with a simple experiment. We take a 2-layer, 224-unit CFN network trained on Penn Treebank and
feed it the following input data: The first 1000 inputs xt are the first 1000 words of the test set of
Penn Treebank; All subsequent inputs are zero. In other words, xt = 0 if t > 1000. For each of the
two layers we then select the 10 units that decay the slowest after t > 1000 and plot them on Figure
4. The first layer retains information for about 10 to 20 time steps, whereas the second layer retains
information for about 100 steps. This experiment conforms with the analysis (11), and indicates
that adding a third or fourth layer would potentially allow a multi-layer CFN architecture to retain
information for even longer periods.
3
E XPERIMENTS
In this section we show that despite its simplicity, the CFN network achieves performance comparable to the much more complex LSTM network on the word level language modeling task. We
use two datasets for these experiments, namely the Penn Treebank corpus (Marcus et al., 1993)
and the Text8 corpus (Mikolov et al., 2014). We consider both one-layer and two-layer CFNs and
LSTMs for our experiments. We train both CFN and LSTM networks in a similar fashion and always compare models that use the same number of parameters. We compare their performance with
and without dropout, and show that in both cases they obtain similar results. We also provide results
published in Mikolov et al. (2014), Jozefowicz et al. (2015) and Sukhbaatar et al. (2015) for the sake
of comparison.
For concreteness, the exact implementation for the two-layer architecture of our model is
(0)
= W (0) xt
(0)
= Drop(ht , p)
(1)
= θt
(1)
= Drop(ht , p)
(2)
= θt
(2)
= Drop(ht , p)
ht
ĥt
ht
ĥt
ht
ĥt
(0)
(1)
(1)
(1)
tanh(W (1) ĥt )
(2)
(2)
tanh(W (2) ĥt )
tanh(ht−1 ) + ηt
(0)
(1)
(2)
tanh(ht−1 ) + ηt
(1)
(2)
(2)
yt = LogSoftmax(W (3) ĥt + b)
6
(12)
Under review as a conference paper at ICLR 2017
Table 1: Experiments on Penn Treebank without dropout.
Model
Vanilla RNN
GRU
LSTM
LSTM (1 layer)
CFN (2 layers)
Size
5M parameters
5M parameters
5M parameters
5M parameters
5M parameters
Training
Jozefowicz et al. (2015)
Jozefowicz et al. (2015)
Jozefowicz et al. (2015)
Trained by us
Trained by us
Val. perp.
108.4
109.3
Test perp.
122.9
108.2
109.7
105.1
106.3
Table 2: Experiments on Text8 without dropout
Model
Vanilla RNN
SCRN
LSTM
MemN2N
LSTM (2 layers)
CFN (2 layers)
Size
500 hidden units
500 hidden units
500 hidden units
500 hidden units
46.4M parameters
46.4M parameters
Training
Mikolov et al. (2014)
Mikolov et al. (2014)
Mikolov et al. (2014)
Sukhbaatar et al. (2015)
Trained by us
Trained by us
Perp. on development set
184
161
156
147
139.9
142.0
where Drop(z, p) denotes the dropout operator with a probability p of setting components in z to
zero. We compute the gates according to
(`)
(`) (`)
(`) (`−1)
(`)
(`)
(`−1)
θt := σ Uθ h̃t−1 + Vθ h̃t
+ bθ and ηt := σ Uη(`) h̃t−1 + Vη(`) h̃t
+ bη
(13)
where
(`)
(`)
h̃t−1 = Drop(ht−1 , q) and
(`−1)
h̃t
(`−1)
= Drop(ht
, q),
(14)
and thus the model has two dropout hyperparameters. The parameter p controls the amount of
dropout between layers; the parameter q controls the amount of dropout inside each gate. We use a
similar dropout strategy for the LSTM, in that all sigmoid gates f, o and i receive the same amount
q of dropout.
To train the CFN and LSTM networks, we use a simple online steepest descent algorithm. We update
the weights w via
∇w L
,
(15)
w(k+1) = w(k) − lr · p~ where p~ =
k∇w Lk2
and ∇w L denotes the approximate gradient of the loss with respect to the weights as estimated
from a certain number of presented examples. We use the usual backpropagation through time
approximation when estimating the gradient: we unroll the net T steps in the past and neglect longer
dependencies. In all experiments, the CFN and LSTM networks are unrolled for T = 35 steps and
we take minibatches of size 20. In the case of an exact gradient, the update (15) simply corresponds
to making a step of length lr in the direction of steepest descent. As all search directions p~ have
Euclidean norm k~
pk2 = 1, we perform no gradient clipping during training.
We initialize all the weights in the CFN, except for the bias of the gates, uniformly at random in
[−0.07, 0.07]. We initialize the bias bθ and bη of the gates to 1 and −1, respectively, so that at the
beginning of the training
θt ≈ σ(1) ≈ 0.73
and
ηt ≈ σ(−1) ≈ 0.23.
We initialize the weights of the LSTM in exactly the same way; the bias for the forget and input
gate are initialized to 1 and −1, and all the other weights are initialized uniformly in [−0.07, 0.07].
This initialization scheme favors the flow of information in the horizontal direction. The importance
of a careful initialization of the forget gate was first pointed out in Gers et al. (2000) and further
emphasized in Jozefowicz et al. (2015). Finally, we initialize all hidden states to zero for both
models.
Dataset Construction. The Penn Treebank Corpus has 1 million words and a vocabulary size of
10,000. We used the code from Zaremba et al. (2014) to construct and split the dataset into a training
set (929K words), a validation set (73K words) and a test set (82K words). The Text8 corpus has 100
million characters and a vocabulary size of 44,000. We used the script from Mikolov et al. (2014) to
7
Under review as a conference paper at ICLR 2017
Table 3: Experiments on Penn Treebank with dropout.
Model
Vanilla RNN
GRU
LSTM
LSTM (2 layers)
CFN (2 layers)
LSTM (2 layers)
CFN (2 layers)
Size
20M parameters
20M parameters
20M parameters
20M parameters
20M parameters
50M parameters
50M parameters
Training
Jozefowicz et al. (2015)
Jozefowicz et al. (2015)
Jozefowicz et al. (2015)
Trained by us
Trained by us
Trained by us
Trained by us
Val. perp.
103.0
95.5
83.3
78.4
79.7
75.9
77.0
Test perp.
97.7
91.7
78.8
74.3
74.9
71.8
72.2
construct and split the dataset into a training set (first 99M characters) and a development set (last
1M characters).
Experiments without Dropout. Tables 1 and 2 provide a comparison of various recurrent network
architectures without dropout evaluated on the Penn Treebank corpus and the Text8 corpus. The last
two rows of each table provide results for LSTM and CFN networks trained and initialized in the
manner described above. We have tried both one and two layer architectures, and reported only the
best result. The learning rate schedules used for each network are described in the appendix.
We also report results published in Jozefowicz et al. (2015) were a vanilla RNN, a GRU and an
LSTM network were trained on Penn Treebank, each of them having 5 million parameters (only
the test perplexity was reported). Finally we report results published in Mikolov et al. (2014) and
Sukhbaatar et al. (2015) where various networks are trained on Text8. Of these four networks, only
the LSTM network from Mikolov et al. (2014) has the same number of parameters than the CFN
and LSTM networks we trained (46.4M parameters). The vanilla RNN, Structurally Constrained
Recurrent Network (SCRN) and End-To-End Memory Network (MemN2N) all have 500 units, but
less than 46.4M parameters. We nonetheless indicate their performance in Table 2 to provide some
context.
Experiments with Dropout. Table 3 provides a comparison of various recurrent network architectures with dropout evaluated on the Penn Treebank corpus. The first three rows report results
published in (Jozefowicz et al., 2015) and the last four rows provide results for LSTM and CFN
networks trained and initialized with the strategy previously described. The dropout rate p and q are
chosen as follows: For the experiments with 20M parameters, we set p = 55% and q = 45% for the
CFN and p = 60% and q = 40% for the LSTM; For the experiments with 50M parameters, we set
p = 65% and q = 55% for the CFN and p = 70% and q = 50% for the LSTM.
4
C ONCLUSION
Despite its simple dynamics, the CFN obtains results that compare well against LSTM networks
and GRUs on word-level language modeling. This indicates that it might be possible, in general, to
build RNNs that perform well while avoiding the intricate, uninterpretable and potentially chaotic
dynamics that can occur in LSTMs and GRUs. Of course, it remains to be seen if dynamically
simple RNNs such as the proposed CFN can perform well on a wide variety of tasks, potentially
requiring longer term dependencies than the one needed for word level language modeling. The
experiments presented in Section 2 indicate a plausible path forward — activations in the higher
layers of a multi-layer CFN decay at a slower rate than the activations in the lower layers. In theory,
complexity and long-term dependencies can therefore be captured using a more “feed-forward”
approach (i.e. stacking layers) rather than relying on the intricate and hard to interpret dynamics of
an LSTM or a GRU.
Overall, the CFN is a simple model and it therefore has the potential of being mathematically wellunderstood. In particular, Section 2 reveals that the dynamics of its hidden states are inherently more
interpretable than those of an LSTM. The mathematical analysis here provides a few key insights
into the network, in both the presence and absence of input data, but obviously more work is needed
before a complete picture can emerge. We hope that this investigation opens up new avenues of
inquiry, and that such an understanding will drive subsequent improvements.
8
Under review as a conference paper at ICLR 2017
R EFERENCES
Nils Bertschinger and Thomas Natschläger. Real-time computation at the edge of chaos in recurrent
neural networks. Neural computation, 16(7):1413–1436, 2004.
Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder
for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.
Felix A Gers, Jürgen Schmidhuber, and Fred Cummins. Learning to forget: Continual prediction
with lstm. Neural computation, 12(10):2451–2471, 2000.
Michel Hénon. A two-dimensional mapping with a strange attractor. Communications in Mathematical Physics, 50(1):69–77, 1976.
Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):
1735–1780, 1997.
Rafal Jozefowicz, Wojciech Zaremba, and Ilya Sutskever. An empirical exploration of recurrent network architectures. In Proceedings of the 32nd International Conference on Machine Learning,
2015.
Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated
corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993.
Tomas Mikolov, Armand Joulin, Sumit Chopra, Michael Mathieu, and Marc’Aurelio Ranzato.
Learning longer memory in recurrent neural networks. arXiv preprint arXiv:1412.7753, 2014.
Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural
networks. ICML (3), 28:1310–1318, 2013.
Steven H Strogatz. Nonlinear dynamics and chaos: with applications to physics, biology, chemistry,
and engineering. Westview press, 2014.
Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. End-to-end memory networks. In Advances
in neural information processing systems, pp. 2440–2448, 2015.
David Sussillo and Omri Barak. Opening the black box: low-dimensional dynamics in highdimensional recurrent neural networks. Neural computation, 25(3):626–649, 2013.
Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. Recurrent neural network regularization.
arXiv preprint arXiv:1409.2329, 2014.
9
Under review as a conference paper at ICLR 2017
A PPENDIX
Strange attractor of the Hénon map. For the sake of comparison, we provide in Figure 5 a depiction of a well-known strange attractor (the Hénon attractor) arising from a discrete-time dynamical
system. We generate these pictures by reproducing the numerical experiments from Hénon (1976).
The discrete dynamical system considered here is the two dimensional map
xt+1 = yt + 1 − ax2t ,
yt+1 = bxt ,
with parameters set to a = 1.4 and b = 0.3. We obtain Figure 5(a) by choosing the initial state
(x0 , y0 ) = (0, 0) and plotting the iterates (xt , yt ) for t between 103 and 105 . All trajectories starting
close to the origin at time t = 0 converge toward the depicted attractor. Successive zooms on the
branch of the attractor reveal its fractal nature. The structure repeats in a fashion remarkably similar
to the 2-unit LSTM in Section 2.
Figure 5: Strange attractor of the Hénon map. From left to right: The Hénon attractor, enlargement
of the red box, enlargement of the magenta box.
Strange attractor of a 2-unit GRU. As with LSTMs, the GRU gated architecture can induce a
chaotic dynamical system. Figure 6 depicts the strange attractor of the dynamical system
ut = ht ,
u 7→ Φ(u) := (1 − z) u + z tanh (U (r
z := σ (Wz u + bz ) r := σ (Wr u + br ) ,
u))
induced by a two-dimensional GRU, with weight matrices
0 1
0 1
−5 −8
Wz =
Wr =
U=
1 1
1 0
8
5
and zero bias for the model parameters. Here also successive zooms on the branch of the attractor
reveal its fractal nature. As in the LSTM, the forward trajectories of this dynamical system exhibit
a high degree of sensitivity to initial states.
Figure 6: Strange attractor of a two-unit GRU. Successive zooms reveal the fractal nature of the
attractor.
10
Under review as a conference paper at ICLR 2017
Network sizes and learning rate schedules used in the experiments. In the Penn Treebank experiment without dropout (Table 1), the CFN network has two hidden layers of 224 units each for
a total of 5 million parameters. The LSTM has one hidden layer with 228 units for a total of 5
million parameters as well. We also tried a two-layer LSTM with 5 million parameters but the result
was worse (test perplexity of 110.6) and we did not report it in the table. For the Text8 experiments
(Table 2), the LSTM has two hidden layers with 481 hidden units for a total 46.4 million parameters.
We also tried a one-layer LSTM with 46.4 million parameters but the result was worse (perplexity
of 140.8). The CFN has two hidden layers with 495 units each, for a total of 46.4 million parameters
as well.
For both experiments without dropout (Table 1 and 2), we used a simple and aggressive learning
rate schedule: at each epoch, lr is divided by 3. For the CFN the initial learning rate was chosen to
be lr0 = 5.5 for PTB and lr0 = 5 for Text8. For the LSTM we chose lr0 = 7 for PTB and lr0 = 5
for Text8.
In the Penn Treebank experiment with dropout (Table 3), the CFN with 20M parameters has two
hidden layers of 731 units each and the LSTM with 20M parameters trained by us has two hidden
layers of 655 units each. We also tried a one-layer LSTM with 20M parameters and it led to similar
but slightly worse results than the two-layer architecture. For both network, the learning rate was
divided by 1.1 each time the validation perplexity did not decrease by at least 1%. The initial
learning rate were chosen to be lr0 = 7 for the CFN and lr0 = 5 for the LSTM.
11
| 9cs.NE
|
Hall polynomials for the torsion free nilpotent groups
of Hirsch length at most 5
arXiv:1605.01591v2 [math.GR] 1 Sep 2016
Bettina Eick and Ann-Kristin Engel
November 30, 2017
Abstract
A theorem by Hall asserts that the multiplication in torsion free nilpotent groups of
finite Hirsch length can be facilitated by polynomials. In this note we exhibit explicit
Hall polynomials for the torsion free nilpotent groups of Hirsch length at most 5.
1
Introduction
Let T denote the class of torsion free finitely generated nilpotent groups. By [7, p. 137],
each T -group has a central series with infinite cyclic quotients. The length of such a
central series is the Hirsch length of the group; see [7, p. 152] for background.
Let G be a group. We say that a sequence of elements (g1 , . . . , gn ) is a T -basis for G if
the subgroups Gi = hgi , . . . , gn i for 1 ≤ i ≤ n + 1 define a central series G = G1 > G2 >
. . . > Gn > Gn+1 = {1} with infinite quotients Gi /Gi+1 for 1 ≤ i ≤ n. The construction
implies that each quotient Gi /Gi+1 is cyclic and generated by gi Gi+1 . It is not difficult
to observe that a group G has a T -basis of length n if and only if G is T -group of Hirsch
length n.
We note that a T -basis (g1 , . . . , gn ) of a T -group G is a polycyclic sequence for G in the
sense of [5, Sec. 8.1]. Thus as in [5, Lemma 8.3], it follows that each element g of G can be
written uniquely as g = g1a1 · · · gnan for certain a1 , . . . , an ∈ Z. Hence the T -basis induces
a bijection
β : Zn → G : (a1 , . . . , an ) → g1a1 · · · gnan .
The bijection β is not a group homomorphism in general. More precisely, β translates the
multiplication in G to certain functions pi : Zn × Zn → Z for 1 ≤ i ≤ n with
p (a,b)
(g1a1 · · · gnan ) · (g1b1 · · · gnbn ) = g1 1
· · · gnpn (a,b) .
A fundamental theorem by Hall [3, 4] asserts that each function pi (a, b) can be described
by a polynomial. These polynomials are nowadays called Hall polynomials.
Our aim here is to determine Hall polynomials for all T -groups of Hirsch length at most 5.
Methods to compute Hall polynomials for a fixed T -group given by a consistent polycyclic
presentation have been exhibited by Sims [8, page 441ff] and by Leedham-Green & Soicher
[6]. Our method is based on the approach by Sims and uses this in a generic form so that
it applies to the infinitely many T -groups of Hirsch length at most 5 simultaneously.
An application of the Hall polynomials determined here is the classification of the torsion
free nilpotent groups of Hirsch length at most 5 as described in [2].
1
2
Nilpotent presentations and consistency
As a first step towards our aim we exhibit consistent nilpotent presentations for the T groups of Hirsch length at most 5. We refer to [5, Sec. 8.1] for details on this type of
n
presentation. For n ∈ N and t = (tijk | 1 ≤ i < j < k ≤ n) ∈ Z( 3 ) we denote
t
t
i,j,j+1
G(t) = hg1 , . . . , gn | [gj , gi ] = gj+1
· · · gni,j,n for 1 ≤ i < j ≤ ni.
Then the presentation defining G(t) is a polycyclic presentation and the relation imply
that G(t) is a nilpotent group of Hirsch length at most n. This presentation is consistent
if and only if G(t) has Hirsch length precisely n, or, equivalently, if the generators of the
presentation form a T -basis for G(t).
1 Lemma: Let G be a T -group of Hirsch length 5 and let (h1 , . . . , h5 ) be a T -basis for
G. Then there exist a unique t ∈ Z10 so that the map G(t) → G : gi 7→ hi extends to an
isomorphism. This t satisfies the relations t123 t345 = 0 and t124 t345 + t145 t234 = t134 t245 .
Proof: We first construct a suitable t ∈ Z10 . The definition of T -basis implies directly that
[hj , hi ] ∈ Gj+1 for each 1 ≤ i < j ≤ 5. As each element of Gj+1 can be written (uniquely)
aj+1
as hj+1
· · · ha55 , this yields that there exist (unique) tijk with
t
t
i,j,j+1
[hj , hi ] = hj+1
· · · h5i,j,5
for 1 ≤ i < j ≤ 5. Using this constructed t ∈ Z10 then induces that the map G(t) → G :
gi 7→ hi is an epimorphism. It is an isomorphism, as G has Hirsch length 5 and G(t) has
at most Hirsch length 5. We now observe that the restrictions on t are satisfied. For this
purpose we evaluate the equation (h4 h2 )h1 = h4 (h2 h1 ) in G and obtain that
(h4 h2 )h1 = (h2 h4 ht5245 )h1
= h1 hh2 1 hh4 1 ht5245
= h1 h2 ht3123 ht4124 ht5125 h4 ht5145 ht5245
= h1 h2 ht3123 ht4124 +1 ht5125 +t145 +t245 and
h4 (h2 h1 ) = h4 (h1 h2 ht3123 ht4124 ht5125 )
= h1 h4 h2 ht3123 ht4124 ht5125 +t145
= h1 h2 h4 ht3123 ht4124 ht5125 +t145 +t245
= h1 h2 ht3123 ht4124 +1 h5t125 +t145 +t245 +t123 t345 .
Comparison of the two results yields that t123 t345 = 0. Evaluating the equation (h3 h2 )h1 =
h3 (h2 h1 ) in a similar form yields that t124 t345 + t145 t234 = t134 t245 .
•
As a next step, we show that the restrictions on t in Lemma 1 yield that the presentation
defining G(t) is consistent.
n
2 Lemma: Let n ∈ {1, . . . , 5} and t ∈ Z( 3 ) .
(a) Let n ≤ 4. Then the presentation defining G(t) is consistent.
2
(b) Let n = 5. Then the presentation defining G(t) is consistent if and only if t123 t345 = 0
and t124 t345 + t145 t234 = t134 t245 .
Proof: We prove (b) only and note that (a) follows by a similar (and easier) calculation.
⇒: Suppose that G(t) is consistent. Then the generators of this presentation form a T basis for G(t). Thus Lemma 1 yields the result.
⇐: Suppose that t ∈ Z10 with t123 t345 = 0 and t124 t345 + t145 t234 = t134 t245 is given. We
show that G(t) is consistent in this case. For this purpose we evaluate the finitely many
consistency relations, see [8, page 424] (or [1, Lemma 2.10]). In the case considered here,
these consistency relations have the form
(gk gj )gi = gk (gj gi ) for 1 ≤ i < j < k ≤ 5, and
gk = (gk gi−1 )gi for 1 ≤ i < k ≤ 5.
It is not difficult to observe that these relations always hold if k = 5, since g5 is central in
G. Hence it remains to consider the relations
(R1)
(g4 g3 )g1 = g4 (g3 g1 ),
(R2)
(g4 g2 )g1 = g4 (g2 g1 ),
(R3)
(g3 g2 )g1 = g3 (g2 g1 ),
(R4)
(g4 g3 )g2 = g4 (g3 g2 ),
(R5)
g2 = (g2 g1−1 )g1 ,
(R6)
g3 = (g3 g1−1 )g1 ,
(R7)
g4 = (g4 g1−1 )g1 ,
(R8)
g3 = (g3 g2−1 )g2 ,
(R9)
g4 = (g4 g2−1 )g2 ,
(R10)
g4 = (g4 g3−1 )g3 .
The relations (R1) - (R4) can be evaluated similar to the proof of Lemma 1 and are
satisfied due to the equations imposed on the exponents tijk . For the relations (R5) (R10) we note that
g4 g1−1 = g1−1 g4 g5−t145 ,
g3 g1−1 = g1−1 g3 g4−t134 g5t134 t145 −t135 ,
g2 g1−1 = g1−1 g2 g3−t123 g4t123 t134 −t124 g5−t123 t134 t145 +t123 t135 +t124 t145 −t125 ,
g4 g2−1 = g2−1 g4 g5−t245 ,
g3 g2−1 = g2−1 g3 g4−t234 g5t234 t245 −t235 ,
g3 g2−1 = g2−1 g3 g5−t235 .
These conjugates allow to evaluate the relations (R5)-(R10) and thus to determine that
the presentation of G(t) is consistent.
•
3
3
Hall polynomials
Next, we consider Hall polynomials for the T -groups of Hirsch length at most 5. We
assume that the considered groups are given by consistent presentations of the form G(t)
as exhibited in Section 2. We denote s2 (x) = x(x − 1)/2 and s3 (x) = x(x − 1)(x − 2)/6.
3 Theorem: Let t ∈ Z10 so that G(t) is consistent. Then the multiplication in G(t) can
be described by Hall polynomials pi (a, b) with 1 ≤ i ≤ 5 and such polynomials are given by
p 1 = a 1 + b1 ,
p 2 = a 2 + b2 ,
p3 = a3 + b3 + t123 a2 b1 ,
p4 = a4 + b4 + t124 a2 b1 + t134 a3 b1 + t234 a3 b2
+t123 t134 a2 s2 (b1 ) + t123 t234 s2 (a2 )b1 + t123 t234 a2 b1 b2 ,
p5 = a5 + b5 + t345 a4 b3 + t245 a4 b2 + t235 a3 b2 + t145 a4 b1 + t135 a3 b1 + t125 a2 b1
+t234 t345 s2 (a3 )b2 + t234 t245 a3 s2 (b2 ) + t134 t345 s2 (a3 )b1 + t134 t145 a3 s2 (b1 )
+t234 t345 a3 b2 b3 + t134 t345 a3 b1 b3 + t134 t245 a3 b1 b2 + t124 t345 a2 b1 b3
+t124 t345 a2 a3 b1 + (t123 t235 + t124 t245 )a2 b1 b2
+(t123 t235 + t124 t245 )s2 (a2 )b1 + (t123 t135 + t124 t145 )a2 s2 (b1 )
+t123 t234 t245 a2 b1 s2 (b2 ) + t123 t134 t245 s2 (a2 )s2 (b1 ) + t123 t234 t245 s3 (a2 )b1
+t123 t134 t145 a2 s3 (b1 ) + t123 t234 t245 s2 (a2 )b1 b2 + t123 t134 t245 a2 s2 (b1 )b2 .
Proof: We use the approach of Sims [8, page 441ff] to determine the desired polynomials.
As a first step, we determine the polynomials ri,j,k with
r
i,j,i+1
gix gjy = gjy gix gi+1
(x,y)
r
· · · gni,j,n
(x,y)
for 1 ≤ j < i ≤ 5 and all x, y ∈ Z. Using [7, 5.1.5], we note that the following conditions
hold for arbitrary group elements g, h and x ∈ Z:
(1) [[g, h], g] = 1 implies that [g x , h] = [g, h]x ,
(2) [[g, h], h] = 1 implies that [g, hx ] = [g, h]x ,
(3) [[g, h], g] = [[g, h], h] = 1 implies that (gh)x = gx hx [g, h]s2 (x) .
As [g4 , gi ] is central in G(t), we thus obtain that
g4x g1y = g1y g4x [g4x , g1y ]
= g1y g4x [g4 , g1y ]x
= g1y g4x [g4 , g1 ]xy
= g1y g4x g5xyt145 ,
g4x g2y = g2y g4x g5xyt245 ,
g4x g3y = g3y g4x g5xyt345 .
4
Next, we compute the equations for g3x g1y and g3x g2y in two steps. First, we show that
induction on y yields that
gy
yt135 +s2 (y)t134 t145
(∗) g31 = g3 g4yt134 g5
.
This equation is clearly valid for y = 0 and y = 1. Assume that it is valid for y − 1. Then
gy
g31
g y−1 g1
= (g31
)
(y−1)t134 (y−1)t235 +s2 (y−1)t134 t145 g1
g5
)
g1 g1 (y−1)t134 g1 (y−1)t135 +s2 (y−1)t134 t145
g3 (g4 )
(g5 )
t134 t135
t145 (y−1)t134 (y−1)t134 +s2 (y−1)t134 t145
g3 g4 g5 (g4 g5 )
g5
(y−1)t134 (y−1)t134 t145 (y−1)t135 +s2 (y−1)t134 t145
g3 g4t134 g5t135 g4
g5
g5
yt134 yt135 +s2 (y−1)t134 t145 +(y−1)t134 t145
g3 g4 g5
yt +((y−2)(y−1)/2)t134 t145 +(y−1)t134 t145
g3 g4yt134 g5 135
yt +(y(y−1)/2)t134 t135
g3 g4yt134 g5 135
yt +s (y)t t
g3 g4yt134 g5 135 2 134 135 .
= (g3 g4
=
=
=
=
=
=
=
gy
This proves the desired formula for y ≥ 0. An analogue computation leads to g31 =
yt +s (y)t t
g3 g4yt134 g5 135 2 134 145 for y ≤ 0.
gy
y
Next note that (g3x )g1 = (g31 )x holds. Hence (3) and (∗) yield
y
(g3x )g1
gy
= (g31 )x
yt135 +s2 (y)t134 t145 x
= (g3 g4yt134 g5
=
=
=
)
xyt +xs (y)t t
(g3 g4yt134 )x g5 135 2 134 145
xyt +xs (y)t t
g3x g4xyt134 [g3 , g4 ]s2 (x)yt134 g5 135 2 134 145
s (x)yt134 t345 +xyt135 +xs2 (y)t134 t145
g3x g4xyt134 g52
.
Using this and the similar calculation for g3x g2y we obtain that
s (x)yt134 t345 +xyt135 +xs2 (y)t134 t145
g3x g1y = g1y g3x g4xyt134 g52
g3x g2y =
,
s (x)yt234 t345 +xyt235 +xs2 (y)t234 t245
g2y g3x g4xyt234 g52
.
We have now determined all polynomials rijk in the subgroup G2 = hg2 , . . . , g5 i. Based
on this, we can use collection to determine the Hall polynomials of G2 with this:
ga gb = g2a2 g3a3 g4a4 g5a5 g2b2 g3b3 g4b4 g5b5
= g2a2 g3a3 g2b2 g4a4 g5a4 b2 t245 g5a5 g3b3 g4b4 g5b5
s (a3 )b2 t234 t345 +a3 b2 t235 +a3 s2 (b2 )t134 t245 a4 b2 t245 +a5 b3 b4 b5
g5
g3 g4 g5
a2 +b2 a3 +b3 a3 b2 t234 +a4 b3 a3 b2 t234 t345 +b3 a4 t345
g2
g3
g4
g5
s2 (a3 )b2 t234 t345 +a3 b2 t235 +a3 s2 (b2 )t134 t245 +a4 b2 t245 +a5 b4 b5
g5
g4 g5
a2 +b2 a3 +b3 a3 b2 t234 +a4 +b4
g2
g3
g4
b3 a3 b2 t234 t345 +b3 a4 t345 +s2 (a3 )b2 t234 t345 +a3 b2 t235 +a3 s2 (b2 )t134 t245 +a4 b2 t245 +a5 +b5
g5
.
= g2a2 +b2 g3a3 g4a3 b2 t234 +a4 g52
=
=
5
Thus the the multiplication in G2 is given by
f (a,b)
g2a2 · · · g5a5 · g2b2 · · · g5b5 = g22
f (a,b)
· · · g55
with
f2 (a, b) = a2 + b2 ,
f3 (a, b) = a3 + b3 ,
f4 (a, b) = a4 + b4 + a3 b2 t234 ,
f5 (a, b) = a5 + b5 + a3 b2 t235 + a4 b2 t245 + a4 b3 t345
+a3 s2 (b2 )t234 t245 + s2 (a3 )b2 t234 t345 + a3 b2 b3 t234 t345 .
Further, these multiplication polynomials allow to determine powering polynomials for
G2 . The resulting polynomials with
k (a,x)
(g2a2 · · · g5a5 )x = g2 2
k (a,x)
· · · g5 5
are given by
k2 (a, x) = f2 (k2 (a, x − 1), a)
= k2 (a, x − 1) + a2
..
.
= k2 (a, 1) + (x − 1)a2
= xa2 ,
k3 (a, x) = f3 (k3 (a, x − 1), a)
= k3 (a, x − 1) + a3
..
.
= k3 (a, 1) + (x − 1)a3
= xa3 ,
k4 (a, x) = f4 (k4 (a, x − 1), a)
= k4 (a, x − 1) + a4 + k3 (a, x − 1)a2 t234
= k4 (a, x − 1) + a4 + (x − 1)a3 a2 t234
= k4 (a, x − 2) + 2a4 + (x − 1 + x − 2)a3 a2 t234
..
.
x−1
X
i)a3 a2 t234
= k4 (a, 1) + (x − 1)a4 + ((x − 1)x −
i=1
= xa4 + s2 (x)a2 a3 t234 ,
k5 (a, x) = f5 (k5 (a, x − 1))
6
= k5 (a, x − 1) + a5 + k3 (a, x − 1)a2 t235 + k4 (a, x − 1)a2 t245
+k4 (a, x − 1)a3 t345 + k3 (a, x − 1)s2 (a2 )t234 t245
+s2 (k3 (a, x − 1))a2 t234 t345 + k3 (a, x − 1)a2 a3 t234 t345
= k5 (a, x − 1) + a5 + (x − 1)a3 a2 t235 + ((x − 1)a4 + s2 (x − 1)a2 a3 t234 )a2 t245
+((x − 1)a4 + s2 (x − 1)a2 a3 t234 )a3 t345 + (x − 1)a3 s2 (a2 )t234 t245
+(a23 s2 (x − 1) + (x − 1)s2 (a3 ))a2 t234 t345 + (x − 1)a3 a2 a3 t234 t345
= k5 (a, x − 2) + 2a5 + (x − 1 + x + 1)(a2 a3 t235 + a2 a4 t245 + a3 a4 t345 )
+(s2 (x − 2) + s2 (x − 2))(a22 a3 t234 t245 + a2 a23 t234 t345 + a23 a2 t234 t345 )
+(x − 1 + x − 2)(s2 (a3 )a2 t234 t345 + a2 a23 t234 t345 + s2 (a2 )t234 t345 )
= ···
..
.
= k5 (a, 1) + (x − 1)a5 s2 (x)(a2 a3 t235 + a2 a4 t245 + a3 a4 t345 )
+s3 (x)(a22 a3 t234 t245 + a2 a23 t234 t345 + a23 a2 t234 t345 )
+s2 (x)(s2 (a3 )a2 t234 t345 + a2 a23 t234 t345 + s2 (a2 )t234 t345 )
= xa5 + s2 (x)(a2 a3 t235 + a2 a4 t245 + a3 a4 t345 )
+s2 (x)a2 a3 t234 (t245 ((2x − 1)a2 − 3) + t345 ((4x + 1)a3 − 3))/6.
We now consider the computation for g2x g1y . This is done in two steps. At first we invesgy
r(y) s(y) t(y)
tigate the conjugate g21 and write this as g2 g3 g4 g5 for polynomials r(y), s(y), t(y).
We determine these recursively using the polynomials f2 , . . . , f5 :
y
(g2 )g1
y−1
= ((g2 )g1
)g1
r(y−1) s(y−1) t(y−1) g1
g4
g5
)
g1 g1 r(y−1) g1 s(y−1) g1 t(y−1)
g2 (g3 )
(g4 )
(g5 )
t123 t124 t125 r(y−1) r(y−1)t134 s2 (r(y−1))t134 t345 +r(y−1)t135 s(y−1) s(y−1)t145
g2 g3 g4 g5 g3
g4
g5
g4
g5
t(y−1)
g5
r(y−1) r(y−1)t134 +s(y−1)
(g2 g3t123 g4t124 g5t125 )(g3
g4
s2 (r(y−1))t134 t345 +r(y−1)t135 +s(y−1)t145 +t(y−1)
g5
)
t +r(y−1) t124 +r(y−1)t134 +s(y−1)
g21+0 g3123
g4
t125 +t(y−1)+s2 (r(y−1))t134 t345 +r(y−1)t135 +s(y−1)t145 +t124 r(y−1)t345
g5
.
= (g2 g3
=
=
=
=
We obtain the following conditions:
r(y) = r(y − 1) + t123 ,
s(y) = s(y − 1) + t124 + r(y − 1)t134 ,
t(y) = t(y − 1) + t125 + s2 (r(y − 1))t134 t345 + r(y − 1)(t124 t345 + t135 ) + s(y − 1)t145 .
Solving these recursions and using that t123 t345 = 0 yields
r(y) = r(y − 1) + t123
7
= r(y − 2) + t123 + t123
..
.
= r(1) + (y − 1)t123
= yt123 ,
s(y) = s(y − 1) + t124 + (y − 1)t123 t134
= s(y − 2) + 2t124 + (y − 1 + y − 2)t123 t134
..
.
y−1
X
i)t123 t134
= s(1) + (y − 1)t124 + ((y − 1)y −
i=1
= yt124 + (y(y − 1)/2)t123 t134
= yt124 + s2 (y)t123 t134 ,
t(y) = t(y − 1) + t125 + (t2123 s2 (y − 1) + (y − 1)s2 (t123 ))t134 t345
+(y − 1)t123 (t124 t345 + t135 ) + (y − 1)t124 t145 + s2 (y − 1)t123 t134 t145
= t(y − 2) + 2t125 + (s2 (y − 1) + s2 (y − 2))t2123 t134 t345
+(y − 1 + y − 2)s2 (t123 )t134 t345 + (y − 1 + y − 2)t123 (t124 t345 + t135 )
+(y − 1 + y − 2)t124 t145 + (s2 (y − 1) + s2 (y − 2))t123 t134 t145
= ···
..
.
= t(1) + (y − 1)t125
+(y(y − 1)/2)(s2 (t123 )t134 t345 + t123 (t124 t345 + t135 ) + t124 t145 )
y−1
X
s2 (i))t123 t134 (t123 t345 + t145 )
+(
i=1
= yt125 + s2 (y)(s2 (t123 )t345 t134 + t123 (t124 t345 + t135 ) + t124 t145 )
+s3 (y)t123 t134 (t123 t345 + t145 )
= yt125 + s2 (y)(t123 t135 + t124 t145 ) + s3 (y)t123 t134 t145 .
Using the polynomials k2 , . . . , k5 we now determine
gy
r(y) s(y) t(y) x
g4 g5 )
g2x g1y = g1y (g21 )x = g1y (g2 g3
R(x,y) S(x,y) T (x,y)
g4
g5
= g1y g2x g3
with
R(x, y) = xr(y) = xyt123 ,
S(x, y) = xs(y) + s2 (x)r(y)t234
= xyt124 + xs2 (y)t123 t134 + s2 (x)yt123 t234 ,
T (x, y) = xt(y) + s2 (x)(r(y)t235 + s(y)t245 + r(y)s(y)t345 )
+s2 (x)r(y)t234 (t245 ((2x − 1) − 3) + t345 ((4x + 1)r(y) − 3))/6.
8
Thus we have determined all polynomials ri,j,k . Based on these, one can use the following
GAP program to perform a symbolic collection and thus obtain p1 , . . . , p5 . This program
assumes that r and t are global variables so that r[i][j][k] contains the polynomial rijk and
t[i][j][k] is the indeterminate tijk .
CollectSymbolic := function( )
local a, b, v, stack, c, i, j, k, m;
# initiate variables
a := List([1..5],
x -> Indeterminate(Rationals, Concatenation("a",String(x))));
b := List([1..5],
x -> Indeterminate(Rationals, Concatenation("b",String(x))));
v := StructuralCopy(a);
v[5] := v[5] + b[5];
stack := Reversed(List([1..4], x -> [x, b[x]]));
# perform collection until stack is empty
while Length(stack) > 0 do
c := stack[Length(stack)];
Unbind(stack[Length(stack)]);
if c[2] <> 0*c[2] then
i := c[1];
v[i] := v[i] + c[2];
for j in Reversed([i+1..4]) do
for k in Reversed([1..Length(r[i][j])]) do
m := StructuralCopy(r[i][j][k]);
m[2] := Value(m[2], [x,y], [v[j], c[2]]);
if m[1] = 5 then
v[5] := v[5] + m[2];
elif m[2] <> 0*m[2] then
Add(stack, m);
fi;
od;
v[j] := 0;
od;
fi;
od;
# divide by consistency relations to simplify polynomial
v[5] := PolynomialReduction(v[5], [t[1][2][3]*t[3][4][5]],
MonomialLexOrdering())[1];
return v;
end;
•
9
If G is a T -group of Hirsch length n < 5, then Hall polynomials for G can also be read off
from Theorem 3: These are given by p1 , . . . , pn .
References
[1] B. Eick. Algorithms for polycyclic groups. Habilitationsschrift, Universität Kassel,
2001.
[2] B. Eick and A.-K. Engel. The torsion free nilpotent groups of hirsch length at most 5.
Submitted, 2016.
[3] P. Hall. The Edmonton notes on nilpotent groups. Queen Mary College Mathematics
Notes. Mathematics Department, Queen Mary College, London, 1969.
[4] P. Hall. The collected works of Philip Hall. Oxford Science Publications. The Clarendon
Press Oxford University Press, New York, 1988. Compiled and with a preface by K.
W. Gruenberg and J. E. Roseblade, With an obituary by Roseblade.
[5] D. F. Holt, B. Eick, and E. A. O’Brien. Handbook of computational group theory.
Discrete Mathematics and its Applications (Boca Raton). Chapman & Hall/CRC,
Boca Raton, FL, 2005.
[6] C. R. Leedham-Green and L. H. Soicher. Symbolic collection using Deep Thought.
LMS J. Comput. Math., 1:9 – 24, 1998.
[7] D. J. S. Robinson. A course in the theory of groups. Springer, Graduate texts in
mathematics 80, 1982.
[8] C. C. Sims. Computation with finitely presented groups. Cambridge University Press,
Cambridge, 1994.
10
| 4math.GR
|
Logical Methods in Computer Science
Vol. 1 (1:4) 2005, pp. 1–22
www.lmcs-online.org
Submitted
Published
Sep. 17, 2004
Apr. 21, 2005
CONTEXTUAL EQUIVALENCE FOR HIGHER-ORDER π-CALCULUS REVISITED
ALAN JEFFREY a AND JULIAN RATHKE b
a
Bell Labs, Lucent Technologies, and CTI, DePaul University
e-mail address: [email protected]
b
School of Informatics, University of Sussex
e-mail address: [email protected]
A BSTRACT. The higher-order π-calculus is an extension of the π-calculus to allow communication
of abstractions of processes rather than names alone. It has been studied intensively by Sangiorgi in
his thesis where a characterisation of a contextual equivalence for higher-order π-calculus is provided
using labelled transition systems and normal bisimulations. Unfortunately the proof technique used
there requires a restriction of the language to only allow finite types.
We revisit this calculus and offer an alternative presentation of the labelled transition system and
a novel proof technique which allows us to provide a fully abstract characterisation of contextual
equivalence using labelled transitions and bisimulations for higher-order π-calculus with recursive
types also.
1. I NTRODUCTION
It is evident that there is growing interest in the study of mobile code in process languages
[3, 1, 9, 15]. It is also clear that there is some relationship between the use of higher-order features
and mobility. Indeed, code mobility can be expressed as communication of process abstractions.
For this reason then it is important for us to develop a clear understanding of the use of higher-order
features in process languages.
Work towards this began several years ago with various proposals for higher-order versions of
known calculi [14, 4], including the higher-order π-calculus or HOπ [10]. This calculus was studied
intensively by Sangiorgi and one of his achievements was to provide a translation of the higherorder language which supports code mobility, to a first-order π-calculus which supports only name
mobility. This translation is proved to be fully abstract with respect to barbed congruence, but with
the restriction to a language of finite types.
While the translation is of interest in its own right, it also turned out to be very useful for providing a powerful fully abstract characterisation of barbed congruence in terms of labelled transition
systems and normal bisimulations. Providing direct proof techniques for contextual equivalences
in higher-order process languages is often considered to be hard [13]. In this paper, the difficulty
2000 ACM Subject Classification: D.3.1.
Key words and phrases: Higher-order languages, concurrency, full abstraction.
a This material is based upon work supported by the National Science Foundation under Grant No. 0430175.
b Research partially funded by the Nuffield Foundation.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-1 (1:4) 2005
CC
A. Jeffrey and J. Rathke
Creative Commons
2
A. JEFFREY AND J. RATHKE
arises in establishing soundness of the proof technique, which is tantamount to establishing some
sort of contextuality property. It has been seen that the use of a translation of higher- to first-order
communication can alleviate this problem and such translations have been employed to this effect
[11, 7].
However, due to the restriction to finite types for the correctness of these translations, the
soundness of the proof technique is only guaranteed for finite types. Given that recursive types are
used extensively in π-calculus, for encodings of datatypes and functions, this poses a significant
restriction. Sangiorgi has shown that by studying various subcalculi, such as the asynchronous πcalculus, he is able to remove the restriction to finite types [13]. To date, there has been no proof of
full abstraction for full HOπ in the presence of recursive types.
In this paper we present an alternative description of labelled transition systems and normal
bisimulations for HOπ, which is informed by Sangiorgi’s translation of higher-order to first-order
communication. Our alternative presentation allows a direct proof of soundness for contextual
equivalence which makes no use of the translation to first-order π-calculus and, more importantly,
makes no restriction on types.
The innovation here lies in the introduction of operators τk and hk ⇐ vi which simulate the
triggers Trk and meta-notation {k := v} of Sangiorgi [11] where k is a unique identifier for the
trigger and v is a process abstraction. The crucial difference is that where Sangiorgi gives definitions
as HOπ terms for these devices:
Trk = (x)khxi
and
{k := v} = ∗k(x)v · x
where khxi represents an output on name k and ∗k(x)P represents a replicated input on name k, we
leave the operators uninterpreted. There are no interactions between the operators τk and hk ⇐ vi.
Rather, we just mimic the behaviour of triggers in the labelled transition systems. The benefit of
doing this is that it allows us to obtain a direct soundness proof that (normal) bisimilarity implies
contextual equivalence without recourse to any translation in its correctness proof.
A challenge of approaching the problem in this way is that it is not immediately clear that
bisimilarity will be complete for contextual equivalence in HOπ. That is to say, it is not obvious
whether each transition has a genuine HOπ context which validates it. At this point however we can
interpret the operators τk and hk ⇐ vi as HOπ terms exactly as Sangiorgi does. It is then a simple
matter to demonstrate completeness following familiar techniques [3, 7, 5]. The real payoff is that
not only do we obtain a direct soundness proof but the postponement of interpreting the triggers
allows us to finesse any restrictions to finite types.
The remainder of the paper is organised as follows: in Section 2 we recall the syntax and
semantics of HOπ along with the definition of contextual equivalence which we will be using. This
is followed in Section 3 by a presentation of the novel labelled transition system using the operators
τk and hk ⇐ vi. We prove that bisimilarity over this labelled transition system is sound for contextual
equivalence in Section 4 and conversely, that it is complete for contextual equivalence in Section 5.
We conclude in Section 6 with some closing remarks.
2. H IGHER - ORDER π CALCULUS
Except for small changes in notation the language is as can be found in [13] with three main
differences:
(1) We assume two distinct countably infinite sets of identifiers, V and N , for variables and
channel names respectively. In general we will use x, y, z to range over variables and a, b, c
to range over channel names. This variable/name distinction makes the algebraic properties
CONTEXTUAL EQUIVALENCE FOR HIGHER-ORDER π-CALCULUS REVISITED
T, U
::=
·
ch[T ]
T →⋄
Z
rec Z.T
Value Types
Unit type
Channel type
Abstraction type
Type variable
Recursive type
v·w
v(x : T )P
vhwiP
if v = w then P else Q
ν(a : T ) . (P)
PkQ
∗P
0
Terms
Application
Input
Output
Matching
Name creation
Concurrency
Repetition
Termination
·
a
x
(x : T )P
Values
Unit value
Channel name
Variable
Abstractions
P, Q ::=
v, w ::=
3
Figure 1: The Syntax
of the language a little cleaner and we are confident that the techniques proposed here would
also be applicable if we identified these sets.
(2) Since we have adopted a variable/name distinction, we have used Honda and Yoshida’s
definition of observational equivalence [6] in Section 2.4 rather than Sangiorgi’s. See [2]
for a discussion of this issue.
(3) We allow communication of channel names as well as process abstractions so that there is
a core π-calculus as a direct subcalculus of HOπ.
2.1. Syntax. We present the syntax of HOπ in Figure 1. The grammar of types for values includes:
• (·): a singleton type just containing the value (·).
• ch[T ]: the type of channels which can be used for communicating data of type T . Note that
in this paper we are not considering input-only or output-only channels.
• T → ⋄: the type of an abstraction (x : T )P. Such an abstraction can be applied to a value v
of type T to return a well-typed process P[v/x].
• Z and rec Z.T : these allow recursive types, such as the type for monomorphic π-calculus
channels rec Z.ch[Z]. We require Z to be guarded: any free occurrence of Z lies within a
subexpression of T of the form ch[U ] or U → ⋄.
The grammar of process terms includes:
4
A. JEFFREY AND J. RATHKE
• v · w: the application of abstraction v to argument w. During execution, v will be instantiated
by an abstraction of the form (x : T )P, and β-reduction will give the process P[w/x].
• v(x : T )P and vhwiP, which are the standard synchronous input and output of the π-calculus,
except that since abstractions are first-class values, we can communicate higher-order data
as well as first-order data.
• if v = w then P else Q: an equality test on values, where the type system will ensure that v
and w are channels, and so we will never compare abstractions for syntactic identity.
• ν(a : T ) . (P), P k Q, ∗P and 0: the standard π-calculus processes for channel generation,
concurrency, replication and termination.
The grammar of values includes:
• (·): the only value of type (·).
• a and x: channel names and variables respectively.
• (x : T )P: an abstraction, which can be applied to a value v to return a process P[v/x]. Since
abstractions are considered first-class values, they can be communicated on channels, or
passed as arguments to other abstractions. This feature gives HOπ its higher-order power.
2.2. Reduction semantics. The reduction semantics for the language is defined in a standard manner: we first introduce the evaluation contexts
E ::= [ · ] | E k P | νa . E
Structural equivalence, ≡ is defined to be the least congruence with respect to E contexts such that
it makes (k, 0) into a commutative monoid and moreover satisfies
νa . (P k Q) ≡ νa . P k Q
∗P ≡ ∗P k P
if a 6∈ fn(P)
We will now consider processes up to structural equivalence throughout the remainder. We define
the reduction relation → as the least precongruence with respect to E contexts such that the
following axioms hold
(comm)
ahviP k a(x)Q
→
P k (x)Q · v
(β − redn)
(x)P · v
→
P[v/x]
(cond—tt)
if a = a then P else Q
→
P
(cond—ff)
if a = b then P else Q
→
Q
(a 6= b)
In a standard notation we write ==⇒ to denote the reflexive, transitive closure of → .
2.3. Type system. We introduce a simple type system for the language which comprises types for
channels and abstractions, together with recursive types. To allow us to infer recursive types for
terms we make use of type isomorphism. We define this by letting ∼iso be the least congruence on
types which includes
rec Z.T ∼iso T [rec Z.T /Z]
A type environment Γ is a finite set of mappings from identifiers (channel names or variables) to
types with the restriction that channel names a must be mapped to channel types of the form ch[T ].
We write Γ, n : T to represent the environment made up of the disjoint union of Γ and the mapping
n to T . We will call an environment closed if it contains mappings of channel names only and will
CONTEXTUAL EQUIVALENCE FOR HIGHER-ORDER π-CALCULUS REVISITED
Γ⊢·:·
Γ(v) = T
Γ⊢v:T
Γ, x : T ⊢ P
Γ ⊢ (x : T )P : T → ⋄
Γ ⊢ v : ch[T ], w : ch[T ]
Γ⊢P
Γ⊢Q
Γ ⊢ if v = w then P else Q
Γ⊢v:T →⋄ Γ⊢w:T
Γ ⊢ v·w
Γ, a : T ⊢ P
Γ ⊢ ν(a : T ) . (P)
Γ, x : T ⊢ P Γ ⊢ v : ch[T ]
Γ ⊢ v(x : T )P
5
Γ ⊢ v : T T ∼iso U
Γ⊢ v:U
Γ ⊢ P, Q
Γ ⊢ P k Q, ∗P, 0
Γ ⊢ P Γ ⊢ w : T Γ ⊢ v : ch[T ]
Γ ⊢ vhwiP
Figure 2: The Typing Rules
write ∆ to indicate this. Type inference rules for the calculus are given in Figure 2. We will call a
well-typed process, P, closed if it can be typed as ∆ ⊢ P for some closed ∆. It is easily shown that
subject reduction holds for closed terms for the reduction relation and type inference system given.
2.4. Contextual equivalence. We will now define an appropriate notion of behavioural equivalence based on contexts and barbs.
Contexts are defined by extending the syntax of processes by allowing typed holes [ ·Γ ] in
terms. The type inference system is extended to contexts by using the rule
Γ, Γ′ ⊢ [ ·Γ ]
We write C[] to denote contexts with at most one hole and C[P] for the term which results from
substituting P into the hole.
For any given channel name a such that ∆ ⊢ a : ch[·] we write ∆ |= P ⇓ a if there exists some
P′ , P′′ such that P ==⇒ ν∆′ . (ah·iP′′ k P′ ) with a 6∈ ∆′ .
We use type-indexed families of relations {R ∆ } between closed process terms to describe equivalence. We will write R to refer to the whole family of relations and
∆ |= P R Q
to indicate that P and Q are well-typed with respect to ∆ and related by R ∆ . For general process
terms we define the open extension R o of a typed relation R as
∆, x1 : T1 , . . . , xn : Tn |= P R o Q
holds if for every ∆′ disjoint from ∆ and every vi such that ∆, ∆′ ⊢ vi : Ti (for 1 ≤ i ≤ n) we have
∆, ∆′ |= P[v1 , . . . , vn /x1 , . . . , xn ] R Q[v1 , . . . , vn /x1 , . . . , xn ]
Note that, in general, for closed terms ∆ |= P R Q is not equivalent to ∆ |= P R o Q as R o enjoys the
weakening property that ∆, ∆′ |= P R o Q whenever ∆ |= P R o Q, even when R does not. However,
the contextual equivalence which we study in this paper is defined as an open extension and therefore
will satisfy this weakening.
There are a number of properties of type-indexed relations that we must define:
Symmetry:: A type-indexed relation R is symmetric whenever ∆ |= P R Q implies ∆ |= Q R P.
6
A. JEFFREY AND J. RATHKE
Reduction closure:: A type-indexed relation R is reduction-closed whenever ∆ |= P R Q and
P → P′ implies there exists some Q′ such that Q ==⇒ Q′ and ∆ |= P′ R Q′ .
Contextuality:: A type-indexed relation R is contextual whenever Γ′ |= P R o Q and Γ ⊢ C[·Γ′ ]
implies Γ |= C[P] R o C[Q].
Barb preservation:: A type-indexed relation R is barb-preserving if ∆ |= P R Q and ∆ |= P ⇓ a
implies ∆ |= Q ⇓ a.
Definition 2.1 (Contextual equivalence). Let ∼
= be the open extension of the largest type-indexed
relation which is symmetric, reduction-closed, contextual and barb-preserving.
2
For technical convenience it will be useful to work with a lighter definition of contextuality.
We say that a relation R is k-contextual if it is preserved by all contexts of the form [ ·Γ ] k R and we
let ∼
= p denote the open extension of the largest typed relation over processes which is symmetric,
k-contextual, reduction-closed and barb-preserving. The following lemma demonstrates that this
lighter definition is sufficient.
Lemma 2.2 (Context lemma). Γ |= P ∼
=Q
if and only if
Γ |= P ∼
=p Q
Proof. In Appendix A.
3. F ULL
ABSTRACTION
In this section, we will present a bisimulation equivalence for HOπ, and show that this equivalence is fully abstract for contextual equivalence.
3.1. Labelled transitions. We will use a labelled transition system to characterize ∼
= over higherorder π-calculus terms. The style of the labelled transition system differs a little from previous
transition systems offered for HOπ. Most notably, the nodes of the transition system are described
using an augmented syntax rather than process terms alone. Specifically, for each k drawn from a
countable set of names disjoint from N and V , we introduce two new operators:
τk
hk ⇐ vi
and
with the intuitive reading that τk is an indirect reference to an abstraction and hk ⇐ vi stores the abstraction to which k refers so that access to v is provided through interaction with k. The augmented
syntax for nodes is given the grammar of configurations C obtained by extending Figure 1 with:
v ::= . . . (as Figure 1) . . . | τk
C ::= P | hk ⇐ vi | νa : T . (C) | C k C
We impose a syntactic restriction on the augmented syntax so that in any configuration C for any
given k then hk ⇐ vi appears at most once in C. Structural equivalence and reduction lift to C in
the obvious manner — note that there are no reduction rules given for τk and hk ⇐ vi though. We
augment the type rules by considering judgements of the form
Γ; Θ⊢v:T
and
Γ ; Θ ⊢C
where Θ represents a set of mappings from reference names to types T . The rules in Figure 2 are
easily decorated with the extra Θ environment. The further rules required are given by
Θ(k) = T
Γ ; Θ ⊢ τk : T → ⋄
Θ(k) = T Γ ; Θ ⊢ v : T → ⋄
Γ ; Θ ⊢ hk ⇐ vi
CONTEXTUAL EQUIVALENCE FOR HIGHER-ORDER π-CALCULUS REVISITED
7
Nodes of our labelled transition system then are well-typed closed terms of the augmented language
of the form
(∆ ; Θ ⊢ C)
α
τ
The transitions are of the form (∆ ; Θ ⊢ C) → (∆ ; Θ ⊢ C) or (∆ ; Θ ⊢ C) → (∆ ; Θ ⊢ C) where
visible labels α are given by the grammar:
α ::= νa . α | νk . dhτk i! | νk . dhτk i? | dhvi? | dhvi!
where write d to mean either a channel name a or an indirect reference name k. The transitions are
presented in Figures 3,4,5. The intuition for these transitions is (eliding types for readability):
ahvi?
• P → P′ : indicates that P is prepared to input a value v on channel a and then perform as
P′ . The type system enforces that v is a first-order value, and not an abstraction. Moreover,
in this case both a and v are pre-existing values, and were not generated fresh for this
transition.
khvi?
• P → P′ : indicates that P has provided a named abstraction reference k to the environment,
and that the environment is calling the abstraction with pre-existing argument v.
νb.ahbi?
• P
→ P′ : indicates that P is prepared to input a fresh channel b on channel a and then
ahbi?
perform as P′ . This is the same as P → P′ , except that b is now a fresh channel generated
by the environment, and has not been seen before by the process.
νb.khbi?
• P
→ P′ : indicates that P has provided a named abstraction reference k to the environment, and that the environment is calling the abstraction with fresh argument b.
νl.ahτl i?
→ P′ : indicates that P is prepared to input an abstraction l on channel a and then
• P
perform as P′ . In this case, we do not record the abstraction itself in the label, but instead
we just generate a fresh reference l to the abstraction.
νl.khτl i?
→ P′ : indicates that P has provided a named abstraction reference k to the environ• P
ment, and that the environment is calling that abstraction with argument l. In this case, k
must be a higher-order abstraction, so is expecting an abstraction as an argument. Rather
than recording the abstraction itself in the label, we instead generate a fresh reference l to
the abstraction.
• Each of the above input transitions has a dual output transition, where the role of the process
and environment are exchanged.
We write ᾱ to denote the complement of an action α, which is defined to be the action α with the
input/output annotation inversed. We will often write ==⇒ to mean the reflexive transitive closure
τ
α
α
of → and ==⇒ to mean ==⇒ → ==⇒ . The following proposition states that the labelled
transition system is well-defined in the sense that the transition relation only relates well-typed
terms.
α
Proposition 3.1. If ∆ ; Θ ⊢ C and (∆ ; Θ ⊢ C) → (∆, ∆′ ; Θ, Θ′ ⊢ C′ ) then ∆, ∆′ ; Θ, Θ′ ⊢ C′ is a
valid typing judgement.
Proof. Straightforward induction.
8
A. JEFFREY AND J. RATHKE
α
(∆ ; Θ ⊢ C) → (∆′ ; Θ′ ⊢ C′ )
C → C′
τ
α
(∆ ; Θ ⊢ C) → (∆ ; Θ ⊢ C′ )
(∆ ; Θ ⊢ C k D) → (∆′ ; Θ′ ⊢ C′ k D)
α
(∆, a : T ; Θ ⊢ C) → (∆, a : T, ∆′ ; Θ, Θ′ ⊢ C′ ) (a 6∈ fn(α))
α
(∆ ; Θ ⊢ νa : T .C) → (∆, ∆′ ; Θ, Θ′ ⊢ νa : T .C′ )
(∆, b : T ; Θ ⊢ C)
dhbi!
→ (∆, b : T ; Θ ⊢ C′ ) (d 6= b)
(∆ ; Θ ⊢ νb : T .C)
(∆, b : T ; Θ ⊢ C)
(∆ ; Θ ⊢ C)
νb.dhbi!
→ (∆, b : T ; Θ ⊢ C′ )
dhbi?
→ (∆, b : T ; Θ ⊢ C′ )
(d 6= b)
νb.dhbi?
→ (∆, b : T ; Θ ⊢ C′ )
Figure 3: Structural labelled transition rules
3.2. Bisimilarity. We use a standard definition of (weak) bisimilarity to provide our characterisation of ∼
= for HOπ:
Definition 3.2. We call a symmetric relation, R , between nodes of the labelled transition system a
bisimulation if whenever (n, m) ∈R we have
τ
• n → n′ implies there exists some m′ such that m ==⇒ m′ and (n′ , m′ ) ∈R
α
α
• n → n′ implies there exists some m′ such that m ==⇒ m′ and (n′ , m′ ) ∈R
Let bisimulation equivalence, or bisimilarity, ≈ be the largest bisimulation relation.
2
We will write
∆ ; Θ |= C ≈ D
to mean that ∆ ; Θ ⊢ C and ∆ ; Θ ⊢ D are valid typing judgements and moreover, they are related
by ≈ as nodes of the lts. In order to provide a bisimulation characterisation of ∼
= over HOπ we will
consider a subrelation of ≈ by restricting our attention to nodes of the form
(∆ ; ⊢ P)
whose terms are clearly definable in HOπ. We will simply write (when Θ is empty)
∆ |= P ≈ Q
to indicate bisimilarity between such terms of HOπ considered as nodes of the labelled transition
system.
3.3. Soundness of bisimilarity for contextual equivalence. We need to demonstrate that bisimilarity implies contextual equivalence for all HOπ processes. In particular, because of Lemma 2.2,
we need only show that bisimilarity is contained in some symmetric, reduction-closed, barb preserving and k-contextual relation. The key to achieving this is to study the k-context closure of
bisimilarity. If we can demonstrate that this is reduction-closed then we have our result. To do this
we must establish a decomposition theorem for interactions. For instance, if P and Q are bisimilar
and we compose each of them with a process R then suppose
PkR →S
CONTEXTUAL EQUIVALENCE FOR HIGHER-ORDER π-CALCULUS REVISITED
T ∼iso U
νk.ahτk i?
(∆ ; Θ ⊢ a(x : T )P)
9
→⋄
→ (∆ ; Θ, k : U ⊢ (x : T )P · τk )
Θ(k) ∼iso T → ⋄
(∆ ; Θ ⊢ hk ⇐ vi)
νl.khτl i?
→ (∆ ; Θ, l : T ⊢ v · τl k hk ⇐ vi)
∆; Θ⊢v:T →⋄
(∆ ; Θ ⊢ ahviP)
νk.ahτk i!
→ (∆ ; Θ, k : T ⊢ hk ⇐ vi k P)
Θ(k) ∼iso T → ⋄
(∆ ; Θ ⊢ τk · v)
νl.khτl i!
→ (∆ ; Θ, l : T ⊢ hl ⇐ vi)
Figure 4: Basic higher-order labelled transition rules
represents an interaction between P and R. We decompose this into complementary actions
α
P → P′
and
ᾱ
R → R′
respectively. Note however that S is not necessarily obtained by a parallel composition of the targets
of the transitions: P′ k R′ . Instead, P′ and R′ may contain indirect references and their corresponding
resources. These need to be matched up correctly to obtain S. We achieve this by introducing the
merge (partial) operator hh·ii which will match up these terms and replace every indirect reference
to an abstraction with the abstraction itself. We write
C[v/τk ]
to denote the substitution of the value v for every instance of the indirect reference τk . We define
hhCii then as the operator on terms of the augmented syntax (up to ≡) such that
hhCii = C
if C doesn’t contain hk ⇐ vi for any k, v
hhν(~a : ~T ) . (hk ⇐ vi k C)ii = hhν(~a : ~T ) . (C[v/τk ])ii if τk 6∈ v
Intuitively, this says that we substitute any values stored at a hk ⇐ vi through for the corresponding
τk . Note that this need not substitute for all the indirect reference identifiers in C. It is clear that the
above definitions are only partial. For example, if C contains an occurrence of hk ⇐ vi for which
τk occurs in v, then hhCii is undefined. In order to identify for which terms the merge is defined we
make use of the notion of reference graph: For a term C we define the graph rg(C) to be the graph
which has nodes as the indirect reference identifiers k in C and edges
k 7→ l
if τl ∈ v for
hk ⇐ vi in C
Proposition 3.3. hh·ii is a well-defined partial function such that hhCii is defined if and only if rg(C)
is acyclic.
Proof. Given in Appendix B.
10
A. JEFFREY AND J. RATHKE
∆ ⊢ v : T a base type
(∆ ; Θ ⊢ a(x : T )P)
Θ(k) = T
(∆ ; Θ ⊢ hk ⇐ vi)
ahvi?
→ (∆ ; Θ ⊢ (x : T )P · v)
∆ ⊢ w : T a base type
khwi?
→ (∆ ; Θ ⊢ v · w k hk ⇐ vi)
∆ ⊢ v : T a base type
(∆ ; Θ ⊢ ahviP)
Θ(k) = T
ahvi!
→ (∆ ; Θ ⊢ P)
T a base type
(∆ ; Θ ⊢ τk · v)
khvi!
→ (∆ ; Θ ⊢ 0)
Figure 5: Basic first-order labelled transition rules
Lemma 3.4 (Composition/Decomposition). For ∆ ; Θ ⊢ C, D
(i) If hhC k Dii ≡ E and
α
(∆ ; Θ ⊢ C) → (∆, ∆′ ; Θ, Θ′ ⊢ C′ )
and
ᾱ
(∆ ; Θ ⊢ D) → (∆, ∆′ ; Θ, Θ′ ⊢ D′ )
then there exists a E ′ such that E ==⇒ E ′ and hhν∆′ . (C′ k D′ )ii = E ′
(ii) If hhCii ≡ E and C → C′ then there exists a E ′ such that E → E ′ and hhC′ ii ≡ E ′
(iii) If hhC k Dii ≡ E and E → E ′ then one of the following hold
C → C′ with hhC′ k Dii ≡ E ′
or D → D′ with hhC k D′ ii ≡ E ′
α
ᾱ
or (∆ ; Θ ⊢ C) ==⇒ (∆, ∆′ ; Θ, Θ′ ⊢ C′ ) and (∆ ; Θ ⊢ D) ==⇒ (∆, ∆′ ; Θ, Θ′ ⊢ D′ ) with
hhν∆′ . (C′ k D′ )ii ≡ E ′ .
Proof. Part (ii) is straightforward as the merge operator hh ii simply removes subterm of the form
hk ⇐ vi, which can’t be involved in reductions, and substitutes higher-order values through for
variables of higher-order type. Reductions are based on structure alone except for the conditionals
which can be affected by first-order substitutions of channel names only.
To show (i) we must consider all the possible cases for α. By symmetry there are four distinct
pairs of complementary actions. We only consider the cases where α is νk . ahτk i? and νl . khτl i? as
the first-order actions can be treated similarly.
νk.ahτk i?
Case: ∆ ; Θ ⊢ C
we see that
–
–
–
–
→ ∆ ; Θ, k : U ⊢ C′ and ∆ ; Θ ⊢ D
νk.ahτk i!
C ≡ ν∆′ . (a(x : T )P k C′′ )
with T ∼iso U → ⋄
C′ ≡ ν∆′ . ((x : T )P · τk k C′′ )
D ≡ ν∆′′ . (ahviQ k D′′ )
D′ ≡ ν∆′′ . (hk ⇐ vi k Q k D′′ )
→ ∆ ; Θ, k : U ⊢ D′ . By inspection
CONTEXTUAL EQUIVALENCE FOR HIGHER-ORDER π-CALCULUS REVISITED
11
It is easy to see that hhC k Dii → hhν∆′ , ∆′′ . ((x : T )P · v k C′′ k Q k D′′ )ii let us call the target
of this reduction E ′ . We simply need to check
E ′ ≡ hhν∆′ , ∆′′ . ((x : T )P · v k C′′ k Q k D′′ )ii
(τk 6∈ v)
≡ hhν∆′ . ((x : T )P · τk k C′′ ) k ν∆′′ . (hk ⇐ vi k Q k D′′ )ii
≡ hhC′ k D′ ii
νl.khτl i?
→ ∆ ; Θ, l : T ⊢ C′ and ∆ ; Θ ⊢ D
Case: ∆ ; Θ ⊢ C
inspection we see that
–
–
–
–
νl.khτl i!
→ ∆ ; Θ, l : T ⊢ D′ . Again, by
C ≡ ν∆′ . (hk ⇐ vi k C′′ )
C′ ≡ ν∆′ . (v · τl k hk ⇐ vi k C′′ )
D ≡ ν∆′′ . (τk · w k D′′ )
D′ ≡ ν∆′′ . (hl ⇐ wi k D′′ )
Note that the previous proposition tells us that rg(C k D) must be acyclic — in particular,
τk 6∈ v. Here we see that
hhC k Dii ≡
(τk 6∈ v)
≡
(τl 6∈ v, w,C′′ , D′′ )
≡
≡
hhν∆′ , ∆′′ . (hk ⇐ vi k C′′ k τk · w k D′′ )ii
hhν∆′ , ∆′′ . (hk ⇐ vi k C′′ k v · w k D′′ )ii
hhν∆′ , ∆′′ . (hk ⇐ vi k C′′ k v · τl k hl ⇐ wi k D′′ )ii
hhC′ k D′ ii
So by letting E ′ be hhC′ k D′ ii we note that hhC k Dii ==⇒ E ′ as required.
To show (iii) we suppose hhC k Dii ≡ E and that E → E ′ . We must consider all possible ways
in which this reduction can occur. If the reduction arises from a conditional then it is clear that we
must have C → C′ or D → D′ for some C′ or D′ . Moreover it is easy to check that hhC′ k Dii (resp
hhC k D′ ii) ≡ E ′ . There are two more possibilities to consider:
Case: the reduction arises from a β-reduction. In this case either C → C′ or D → D′ as above and
the result follows easily, or v is (x : U )P and
– C ≡ ν∆′ . (τk · w k C′′ )
with all names in ∆′ appearing in w
– D ≡ ν∆′′ . (hk ⇐ vi k D′′ )
with
τk 6∈ v
′
′
′′
′′
– E ≡ hhν∆ , ∆ . (P[w/x] k C k hk ⇐ vi k D′′ )ii
or a symmetric version of these with the roles of C and D reversed. So we notice that if
U ∼iso T → ⋄, we have
∆ ; Θ ⊢C
νl.khτl i!
→ ∆ ; Θ, l : T ⊢ C′
and
νl.khτl i?
∆ ; Θ ⊢ D ====⇒ ∆ ; Θ, l : T ⊢ D′
where C′ ≡ ν∆′ . (hl ⇐ wi k C′′ ) and D′ ≡ ν∆′′ . (P[τl /x] k hk ⇐ vi k D′′ ). We check:
(τl 6∈
v, w,C′′ , D′′ )
hhC′ k D′ ii ≡ hhν∆′ . (hl ⇐ wi k C′′ ) k ν∆′′ . (P[τk /x]) k hk ⇐ vi k D′′ ii
≡ hhν∆′ , ∆′′ . (C′′ k P[w/x] k hk ⇐ vi k D′′ )ii
≡ E′
as required. Alternatively, it could be that U is a base type, in which case
∆ ; Θ⊢C
ν∆′ .khwi!
→ ∆, ∆′ ; Θ ⊢ C′
and
ν∆′ .khwi?
∆ ; Θ ⊢ D ====⇒ ∆, ∆′ ; Θ ⊢ D′
where C′ ≡ C′′ and D′ ≡ ν∆′′ .(P[w/x] k hk ⇐ vi k D′′ ). It is easy to check that hhC′ k D′ ii ≡ E ′
as required.
12
A. JEFFREY AND J. RATHKE
Case: the reduction arises from communication. Again we see that either C → C′ or D → D′ , in
which case we easily obtain the result, or
– C ≡ ν∆′ . (ahviP k C′′ )
– D ≡ ν∆′′ . (a(x : T )Q k D′′ )
– E ′ ≡ hhν∆′ . (P k C′′ ) k ν∆′′ . ((x : T )Q · v k D′′ )ii
or a symmetric version of this with the roles of C and D reversed. Again we must consider
whether the type T is a base type or higher-order. We omit the details of the former case.
Suppose then that ∆ ; Θ ⊢ v : T ∼iso U → ⋄ we know
∆; Θ⊢C
νk.ahτk i!
→ ∆ ; Θ, k : U ⊢ C′
and
∆; Θ⊢D
νk.ahτk i?
→ ∆ ; Θ, k : U ⊢ D′
where C′ ≡ ν∆′ . (hk ⇐ vi k P k C′′ ) and D′ ≡ ν∆′′ . ((x : T )Q · τk k D′′ ). We check:
(τk
6 v, P,C′′ , D′′ )
∈
hhC′ k D′ ii ≡ hhν∆′ . (hk ⇐ vi k P k C′′ ) k ν∆′′ . ((x : T )Q · τk k D′′ )ii
≡ hhν∆′ , ∆′′ . (P k C′′ k (x : T )Q · v k D′′ )ii
≡ E′
as required.
Definition 3.5. Let ≈m be defined to be
∆ ; Θ |= hhC1 k Dii ≈m hhC2 k Dii
if and only if
∆ ; Θ |= C1 ≈ C2
and
∆; Θ⊢D
whenever hhC1 k Dii and hhC2 k Dii are defined.
2
Note that in the case where Θ is empty we have that hhCi k Dii = Ci k D, and hence ≈m and ∼
=p
coincide.
Lemma 3.6. ≈m is reduction-closed.
Proof. Follows easily from the previous lemma. Take ∆ ; Θ |= hhC1 k Dii ≈m hhC2 k Dii and suppose
hhC1 k Dii → E. We must show that hhC2 k Dii → E ′ for some E ′ such that ∆ ; Θ |= E ≈m E ′ . We
know from Part (iii) of the previous lemma that one of three cases must hold. Either, C1 → C1′ ,
D → D′ or there are complementary actions from both C1 and D. We only deal with the last case
as the others follow easily from the hypothesis that ∆ ; Θ |= C1 ≈ C2 and Part (ii) of the previous
lemma.
α
ᾱ
We have then that ∆ ; Θ ⊢ C1 =⇒ ∆, ∆′ ; Θ, Θ′ ⊢ C1′ and ∆ ; Θ ⊢ D =⇒ ∆, ∆′ ; Θ, Θ′ ⊢ D′ such
that E ≡ hhC1′ k D′ ii. We know by hypothesis that there must exist some
α
∆ ; Θ ⊢ C2 ==⇒ ∆, ∆′ ; Θ, Θ′ ⊢ C2′
such that
∆, ∆′ ; Θ, Θ′ |= C1′ ≈ C2′ .
(†)
We can now use Parts (i) and (ii) of the previous lemma to see that hhC2 k Dii ==⇒ E ′ such that
E ′ ≡ hhC2′ k D′ ii. Note that (†) guarantees ∆ ; Θ |= E ≈m E ′ to finish.
CONTEXTUAL EQUIVALENCE FOR HIGHER-ORDER π-CALCULUS REVISITED
13
Theorem 3.7. For all closed terms P, Q of HOπ:
∆ |= P ≈ Q
implies
∆ |= P ∼
=p Q
Proof. We let ≈ p denote the relation
∆, ∆′ |= (P k R) ≈ p (Q k R) iff ∆ |= P ≈ Q and ∆, ∆′ ⊢ R
It is easy to see that ≈ p is a k-contextual relation over terms of HOπ. It is also easy to see that ≈ p
is symmetric and barb preserving and coincides with ≈m for closed terms of HOπ, thus Lemma 3.6
can be instantiated to demonstrate that ≈ p is reduction-closed and, given that ∼
= p is defined to be the
largest symmetric, k-contextual, reduction-closed, and barb-preserving relation over terms of HOπ,
then we have our result.
Corollary 3.8 (Soundness). For all terms P, Q of HOπ:
Γ |= P ≈o Q
implies
Γ |= P ∼
=Q
Proof. Follows from the previous theorem and Lemma 2.2.
3.4. Completeness of bisimilarity for contextual equivalence. The interactions described by the
labelled transition system are not obviously derived by genuine contextual observations in HOπ because of the use of the extra syntax for indirect references. In order to show completeness of our
bisimilarity for contextual equivalence we must demonstrate that the indirect references are in fact
definable as terms of the language proper. Following Sangiorgi [13], we implement the implicit protocol outlined by the indirect references by using the following translation of the augmented terms
into HOπ:
[[k1 : T1 , . . . , kn : Tn ]] = k1 : ch[T1 ], . . . , kn : ch[Tn ]
[[Γ ; Θ ⊢ C]] = Γ, [[Θ]] ⊢ [[C]]Θ
[[τk ]]Θ = (x : T )khxi0
if Θ(k) = T
[[hk ⇐ vi]]Θ = ∗k[[v]]Θ
The translation acts homomorphically on all other terms. We abuse notation here by using identifiers
k as channel names in the translation. It is evident that this translation is well-defined in the sense
that the translation of well-typed augmented terms are indeed well-typed terms of HOπ.
We would now like to prove a correspondence between reductions from the terms of the augmented syntax and reductions between their translations. However, we note that in translating a term
containing both hk ⇐ vi and τk we provide matching input and output prefixes, which, in HOπ may
create a communication which was not possible in the source term. This turns out not to be of particular concern to us though as we see that if we starting with terms of HOπ, then terms reachable
by transitions are balanced in the following sense: we call a term C of the augmented language
balanced if for each k then C contains at most one of τk (possible multiple times) or hk ⇐ vi. Unfortunately the translation may introduce extra reductions which aren’t present in the source term.
These arise through the translation of terms of the form τk · v. Note that
τ
[[τk · v]] = (x : T )khxi0 · [[v]] → kh[[v]]i0
but τk · v has no corresponding reduction. We will identify these rogue reductions as housekeeping
h
reductions and indicate them with → defined as any reduction which can be derived using the
axiom
(h − redn)
(x : T )khxi0 · v → khvi0
Lemma 3.9. If ∆ ; Θ ⊢ C is balanced then
14
A. JEFFREY AND J. RATHKE
(1) If C ==⇒ C′ then [[C]]Θ ==⇒ [[C′ ]]Θ
h
(2) If [[C]]Θ ==⇒ P then [[C]]Θ ==⇒ [[D]]Θ →∗ P for some ∆ ; Θ ⊢ D such that C ==⇒ D.
Proof. We will omit mention of the environment Θ in the proof as it plays no role. Part 1 is straightforward. For Part 2 we use induction on the length of the reductions. If there are no reductions
then we are done. We examine the base case in which [[C]] → P. If this reduction happens to be a
h
housekeeping move, that is, [[C]] → P then there is nothing to prove. Suppose otherwise, then it is
not too difficult to check that P ≡ [[D]] for some D such that C → D. For the inductive case suppose
that
[[C]] → ==⇒ P
(†)
By inspecting the translation [[·]] and using the fact that C is balanced we see that
h
[[C]] → → Q
h
[[C]] → → Q
implies
h
thus we may assume that the first reduction in (†) above is not of the form → . This means that
[[C]] → [[C′ ]] ==⇒ P for some C′ such that C → C′ . It is clear that C′ is also balanced so we may
apply the inductive hypothesis to
[[C′ ]] ==⇒ P
h
to obtain a D such that C′ ==⇒ D′ and [[C′ ]] ==⇒ [[D]] →∗ P. Putting these together we obtain
C → C′ ==⇒ D
h
[[C]] → [[C′ ]] ==⇒ [[D]] →∗ P
and
as required.
When ∆′ is of length at most one, we shall write δh∆′ i as shorthand, defined:
/ = δh·i δha : T i = δhai
δh0i
α
Moreover, note that whenever (∆ ; Θ ⊢ D) ==⇒ (∆, ∆′ ; Θ, Θ′ ⊢ D′ ), we have that ∆′ has at length
most one, and so δh∆′ i is well-defined.
Proposition 3.10. For each α, ∆ and fresh channels δ, δ′ of appropriate type given by α and ∆, there
exists a process T α∆ (defined in Figure 6) in HOπ such that if
α
∆ ; Θ ⊢ C → ∆, ∆′ ; Θ, Θ′ ⊢ C′
then
∆,[[Θ]]
∆, [[Θ, Θ′ ]], δ : ch[T0 ], δ′ : ch[·] ⊢ T α
and moreover, for balanced D
α
(∆ ; Θ ⊢ D) ==⇒ (∆, ∆′ ; Θ, Θ′ ⊢ D′ )
if and only if ∆ ; Θ ⊢ D and
T α∆,[[Θ]] k [[D]]Θ ==⇒ ν∆′ . (δh∆′ i k P)
with
h
[[D′ ]]Θ,Θ′ →∗ P.
CONTEXTUAL EQUIVALENCE FOR HIGHER-ORDER π-CALCULUS REVISITED
15
Proof. It is straightforward to check that ∆, [[Θ, Θ′ ]], δ : ch[T0 ], δ′ : ch[·] ⊢ T α∆ whenever
α
∆ ; Θ ⊢ C → ∆, ∆′ ; Θ, Θ′ ⊢ C′ .
For the remainder, to show the ‘only if’ direction we use Lemma 3.9 Part 1 to reduce our obligation
α
to the case of a single transition → , and we must consider each label α. By way of example we
show the case for α = νl . khτl i! (the other cases can be treated similarly). Suppose:
α
(∆ ; Θ ⊢ D) → (∆ ; Θ, l : U ⊢ D′ ).
then we know that
D ≡ ν∆′′ . (τk · v k D′′ )
and
D′ ≡ ν∆′′ . (hl ⇐ vi k D′′ ).
We see that for T ∼iso U → ⋄
T α∆,[[Θ]] k [[D]]Θ
≡
k(x : T )(∗l(y : U )x · y k (δhi ⊕ δ′ hi)) k ν∆′′ . (((z : T )khzi0) · [[v]]Θ k [[D′′ ]]Θ )
==⇒
(δhi ⊕ δ′ hi) k ν∆′′ . (∗l(y : U )[[v]]Θ · y k [[D′′ ]]Θ )
==⇒
δhi k [[D′ ]]Θ,l:U
as required.
For the converse direction we suppose that
T α∆,[[Θ]] k [[D]]Θ ==⇒ ν∆′ . (δh∆′ i k P)
Again, we must perform a case analysis on α. We show the case in which α is νl . khτl i? (the other
∆,[[Θ]]
cases can be treated similarly). We know ∆′ is empty so T α
k [[D]]Θ ==⇒ δhi k P. Note that
∆,[[Θ]]
Tα
has no reductions of its own and can only interact with [[D]]Θ so we can detail the assumed
reductions as
T α∆,[[Θ]] k [[D]]Θ ==⇒ T α∆,[[Θ]] k P0 → (δhi ⊕ δ′ hi) k P1 ==⇒ δhi k P
where [[D]] ==⇒ P0 and P1 ==⇒ P. We assumed that D is balanced so Lemma 3.9 Part 2 applied to
h
[[D]] ==⇒ P0 tells us that [[D]] ==⇒ [[D0 ]]Θ →∗ P0 for some D0 such that D ==⇒ D0 . We know that
P0 is obtained from [[D0 ]]Θ by housekeeping reductions and that it interacts with T α∆ . This tells us
that we must have the forms
P0 ≡ ν∆′′ . (∗k[[v]]Θ k P0′ )
and
P1 ≡ ν∆′′ . ([[v]]Θ · [[τl ]]Θ,l:U k ∗k[[v]]Θ k P0′ )
This in turn tells us that
D0 ≡ ν∆′′ . (hk ⇐ vi k D′0 )
h
such that [[D′0 ]]Θ →∗ P0′ . Now it is clear that
(∆ ; Θ ⊢ D0 )
νl.khτl i?
→ (∆ ; Θ, l : U ) ⊢ D1 )
16
∆
T dhvi?
∆
T dhvi!
∆
T νb.dhbi?
∆
T νb.dhbi!
∆
T νk.dhτ
k i?
∆
T νk.dhτ
k i!
A. JEFFREY AND J. RATHKE
=
=
=
=
=
=
dhvi(δhi ⊕ δ′ hi)
d(x : T )if x = v then (δhi ⊕ δ′ hi) else 0
νb : T . (dhbi(δhbi ⊕ δ′ hi))
d(x : T )if x 6∈ ∆ then (δhxi ⊕ δ′ hi) else 0
dh(x : U )khxi0i(δhi ⊕ δ′ hi)
d(x : T )(∗l(y : U )x · y k (δhi ⊕ δ′ hi))
where ∆(d) = ch[T ]
where ∆(d) = ch[T ]
where ∆(d) = ch[T ]
where ∆(d) = ch[T ] and T ∼iso U → ⋄
where ∆(d) = ch[T ] and T ∼iso U → ⋄
⊕ represents an encoding of internal choice in HOπ
if x 6∈ 0/ then P else Q = P
if x 6∈ (a : T, ∆) then P else Q = if x = a then Q else if x 6∈ ∆ then P else Q
Figure 6: Testing processes for labelled transitions
where D1 ≡ ν∆′′ . (v · τl k hk ⇐ vi k D′0 ). We check
[[D1 ]]Θ,l:U
≡
h
→∗
≡
==⇒
ν∆′′ . ([[v]]Θ · [[τl ]]Θ,l:U k ∗k[[v]] k [[D′0 ]]Θ )
ν∆′′ . ([[v]]Θ · [[τl ]]Θ,l:U k ∗k[[v]] k P0′ )
P1
P
h
Therefore [[D1 ]] =⇒ P and we can apply Lemma 3.9 Part 2 to this to see that [[D1 ]] =⇒ [[D′ ]] →∗ P
for some D′ such that D1 ==⇒ D′ . By collecting the above together we obtain
α
(∆ ; Θ ⊢ D) ==⇒ (∆ ; Θ ⊢ D0 ) → (∆ ; Θ, l : U ⊢ D1 ) ==⇒ (∆ ; Θ, l : U ⊢ D′ )
h
with [[D′ ]]Θ,l:U →∗ P as required.
Lemma 3.11 (Extrusion). If ∆ |= ν∆′ . (δh∆′ i k P) ∼
= p ν∆′ . (δh∆′ i k Q) then ∆, ∆′ |= P ∼
= p Q.
Proof. Follows a similar argument found in [7]: define a relation R such that
∆, ∆′ |= P R Q
iff
∆ |= ν∆′ . (δh∆′ i k P) ∼
= p ν∆′ . (δh∆′ i k Q)
and show that R is barb-preserving, reduction-closed and k-contextual. These properties follow
from the corresponding property for ∼
= p and an extra piece of context to interact with δh∆′ i.
Theorem 3.12 (Completeness). For all closed terms P, Q of HOπ:
∆ |= P ∼
implies
∆ |= P ≈ Q
=p Q
Proof. We define R over terms of the augmented language to be
∆ ; Θ |= C R D
iff
∆, [[Θ]] |= [[C]]Θ ∼
= p [[D]]Θ
and show that R is a bisimulation. Take ∆ ; Θ |= C R D and suppose that
α
(∆ ; Θ ⊢ C) → (∆, ∆′ ; Θ, Θ′ ⊢ C′ ).
We know from Proposition 3.10 that
∆,[[Θ]]
∆, [[Θ, Θ′ ]], δ : ch[T0 ], δ′ : ch[·] ⊢ T α
and that
T α∆,[[Θ]] k [[C]]Θ ==⇒ ν∆′ . (δh∆′ i k P)
CONTEXTUAL EQUIVALENCE FOR HIGHER-ORDER π-CALCULUS REVISITED
17
h
with [[C′ ]]Θ,Θ′ →∗ P. We know that
∆, [[Θ]] |= [[C]]Θ ∼
= p [[D]]Θ
by the definition of R , and hence, by contextuality we also have
∆,[[Θ]]
∆, [[Θ, Θ′ ]], δ : ch[T0 ], δ′ : ch[·] |= T α
This tells us that
such that
∆,[[Θ]]
k [[C]]Θ ∼
k [[D]]Θ
=p T α
T α∆,[[Θ]] k [[D]]Θ ==⇒ Q′
∆, [[Θ, Θ′ ]] |= ν∆′ . (δh∆′ i k P) ∼
= p Q′ .
(†)
∆,[[Θ]]
of T α
But by the construction
we notice that ν∆′ .(δh∆′ i k P) barbs on δ but not on δ′ . Therefore,
by the preservation of barbs property of ∼
= p , we know that Q′ must also barb on δ but not on δ′ . This
∆,[[Θ]]
constrains Q′ so that Q′ ≡ ν∆′ . (δh∆′ i k Q). We apply Lemma 3.9 Part 2 to T α
∆,[[Θ]]
to see that there is some D′′ such that T α
k [[D]]Θ ==⇒ Q′
h
k [[D]]Θ =⇒ [[D′′ ]]Θ,Θ′ →∗ ν∆′ . (δh∆′ i k Q) from which
h
it clearly follows that D′′ ≡ ν∆′ . (δh∆′ i k D′ ) and [[D′ ]]Θ,Θ′ →∗ Q. We use Proposition 3.10 again to
see that
α
(∆ ; Θ ⊢ D) ==⇒ (∆, ∆′ ; Θ, Θ′ ⊢ D′ )
and we now must show that ∆, ∆′ ; Θ, Θ′ |= C′ R D′ . To do this we use Lemma 3.11 on (†) (note
that Q′ ≡ ν∆′ . (δh∆′ i k Q)) to see that ∆, ∆′ , [[Θ, Θ′ ]] |= P ∼
= p Q. It is also easy to check that hreductions are confluent with respect to all other reductions and hence preserve contextual equivah
h
lence, that is →∗ ⊆∼
= p , so we also have ∆, ∆′ , [[Θ, Θ′ ]] |= [[C′ ]]Θ,Θ′ ∼
= p [[D′ ]]Θ,Θ′ because [[C′ ]]Θ,Θ′ →∗ P
h
and [[D′ ]]Θ,Θ′ →∗ Q. This allows us to conclude ∆, ∆′ ; Θ, Θ′ |= C′ R D′ as required.
We must also consider transitions of the form
τ
(∆ ; Θ ⊢ C) → (∆, ∆′ ; Θ, Θ′ ⊢ C′ ).
These can be dealt with as above but in this case no T α∆ is needed.
Corollary 3.13 (Full abstraction). For all terms P, Q of HOπ:
Γ |= P ≈o Q
if and only if
Γ |= P ∼
=Q
Proof. Follows from Corollary 3.8, Lemma 2.2, and the previous theorem.
4. C ONCLUDING
REMARKS
We have re-examined the use of labelled transitions to characterise contextual equivalence in
the higher-order π calculus. The technique of augmenting the core syntax with extra operators to
assist in the definition of the labelled transitions allows use to give a direct proof of soundness of
bisimilarity for contextual equivalence. This advances Sangiorgi’s analagous result by allowing
recursive types also.
We believe that the technique of using extra operators to describe the points of interaction with
the environment in the lts is fairly robust and should be applicable to many higher-order languages.
Indeed, this was the approach that the authors developed for their work on concurrent objects [8].
We have only concerned ourselves with the characterisation of contextual equivalence in HOπ and
so far have not studied Sangiorgi’s translation of higher-order to first-order mobility. Thus, the restriction to finite types for his translation is still necessary. It would be interesting to investigate
whether the current work could be of use in removing this type restriction for his translation also.
18
A. JEFFREY AND J. RATHKE
A PPENDIX A. P ROOF OF T HE C ONTEXT L EMMA
We recall the statement of Lemma 2.2 and detail its proof here.
Γ |= P ∼
= Q if and only if Γ |= P ∼
= p Q.
The force of this lemma is to show that the simplified form of observational testing allowed by ∼
=p
is sufficient to capture the power of full contextual testing. In order to prove this we essentially need
to show that ∼
= p is preserved by the operators of HOπ. For the most part, this can be done directly
and is stated in Lemma A.1 below.
Lemma A.1.
∼ p (x : T )Q · v.
(1) If ∆, x : T |= P ∼
= p Q and ∆ ⊢ v : T then ∆ |= (x : T )P · v =
(2) If ∆, x : T |= P ∼
= p a(x : T )Q.
= p Q and ∆ ⊢ a : ch[T ] then ∆ |= a(x : T )P ∼
(3) If ∆ |= P ∼
= p ahwiQ.
= p Q, ∆ ⊢ w : T and ∆ ⊢ a : ch[T ] then ∆ |= ahwiP ∼
∼
∼
(4) If ∆ |= P1 = p Q1 and ∆ |= P2 = p Q2 then ∆ |= if v = w then P1 else P2 ∼
= p if v = w then Q1 else Q2 .
(5) If ∆, a : T |= P ∼
= p ν(a : T ) . (Q).
= p Q then ∆ |= ν(a : T ) . (P) ∼
(6) If ∆ |= P1 ∼
= p Q1 and ∆ |= P2 ∼
= p Q2 then ∆ |= P1 k P2 ∼
= p Q1 k Q2 .
(7) If ∆ |= P ∼
= p Q then ∆ |= ∗P ∼
= p ∗Q.
Proof. The majority of these are straightforward by exhibiting appropriate symmetric, reductionclosed, k-contextual, barb-preserving relations. As an example of this we show the case for input
prefixing (Case 2). We define R so that ∼
= p ⊆ R and moreover
∆ |= a(x : T )P k R R a(x : T )Q k R for any ∆ ⊢ R
(†)
It is clear that R is symmetric, barb-preserving and k-contextual so if we can show that it is
reduction-closed then we may conclude that R coincides with ∼
= p and we have our result.
Suppose that (†) holds and
a(x : T )P k R → P′ .
We know then that either R → R′ and P′ ≡ a(x : T )P k R′ or the reduction came about by interaction, that is R ≡ ν∆′ . (ahviR′′ k R′′′ ) with a 6∈ ∆′ and by writing R′ for R′′ k R′′′ we have
P′ ≡ ν∆′ . (P[v/x] k R′ ) for some ∆, ∆′ ⊢ v and ∆, ∆′ ⊢ R′ . If the former is true then we see immediately that
a(x : T )Q k R → a(x : T )Q k R′
where
∆ |= a(x : T )P k R′ R a(x : T )Q k R′ .
If instead the latter is true then we use the fact that
∆, x : T |= P ∼
=p Q
to see that ∆, ∆′ |= P[v/x] ∼
= p Q[v/x] and note that
a(x : T )Q k R → ν∆′ . (Q[v/x] k R′ )
where (using k-contextuality and Case 5)
∆ |= ν∆′ . (P[v/x] k R′ ) ∼
= p ν∆′ . (Q[v/x] k R′ )
as required.
CONTEXTUAL EQUIVALENCE FOR HIGHER-ORDER π-CALCULUS REVISITED
19
Notice that there are two particular cases which are not covered by this lemma: application of
a function to, and output of higher-order ∼
= p -related values (c.f. Corollary A.11). Establishing that
∼
is
preserved
in
these
cases
can
be
done
directly but is a little more involved. We notice that the
=p
property we require in both cases follows immediately from Substitutivity (cf. Corollary A.10) , that
is (ignoring types):
if P ∼
= p Q then R[(x)P/y] ∼
= p R[(x)Q/y].
The remainder of the appendix is devoted to achieving this. The proof follows a very similar scheme
to the proof of Proposition 4.2.6 in [10] but simplified to avoid any use of induction on type as
appeared there.
Lemma A.2. If ∆ ⊢ (x : T )P · w then ∆ |= (x : T )P · w ∼
= p P[w/x].
In the following we will make use of a “bisimulation up to” argument [12].
∼ p ) whenever ∆ |= P R Q
Definition A.3. A type-indexed relation R is reduction-closed up to (=, =
2
and P → P′ implies there exists some Q′ such that Q ==⇒ Q′ and ∆ |= P′ R ∼
= p Q′ .
Lemma A.4. For any type-indexed relation R which is symmetric, reduction-closed up to (=, ∼
= p ),
k-contextual and barb-preserving, R ⊆ ∼
.
=p
Definition A.5. We say that x is (un)guarded in P whenever:
(1) if x 6∈ P then x is (un)guarded in P,
(2) if x 6∈ w then x is unguarded in x · w,
(3) if v 6= x then x is guarded in v · w,
(4) x is guarded in v(y : T )P, vhwiP, and if v = w then P else Q, and
(5) if x is (un)guarded in P and Q then x is (un)guarded in ν(a : T ) . (P), P k Q and ∗P.
2
Lemma A.6. For any ∆, y : T → ⋄ ⊢ R with y guarded in R and for any ∆ ⊢ v : T → ⋄ and ∆ ⊢ w : T → ⋄,
if R[v/y] → R′ then R′ = R′′ [v/y] for some R′′ and moreover, R[w/y] → R′′ [w/y].
Proof. We first observe that as ∆ ⊢ v : T → ⋄ it must be the case that v is an abstraction and not
a channel name. From this it is routine to check that the required property holds for the reduction
axioms. Furthermore, if y is guarded in E [P] then y is guarded in P and so the required property is
preserved by reduction in evaluation contexts.
Lemma A.7. For any P and x we can find Q and y such that x is guarded in Q, y is unguarded in Q
and P = Q[x/y].
Proof. A routine induction on P.
Lemma A.8 (Unguarded Substitutivity). If ∆, x : T |= P ∼
= p Q and ∆, y : T → ⋄ ⊢ R and y is un∼
guarded in R then ∆ |= R[(x : T )P/y] = p R[(x : T )Q/y].
Proof. We proceed by induction on the structure of R. If y 6∈ R then the result is immediate. If R is
not of the form v·w, the result follows easily by induction by making use of Lemma A.1. Otherwise,
since y is unguarded in R we must have that R is of the form y · w with y 6∈ w. Hence:
∆ |= R[(x : T )P/y]
as required.
=
∼
=p
∼
=p
∼
=p
=
(x : T )P · w
P[w/x]
Q[w/x]
(x : T )Q · w
R[(x : T )P/y]
(as R = y · w and y 6∈ w)
(by Lemma A.2)
(by hypothesis)
(by Lemma A.2)
(as R = y · w and y 6∈ w).
20
A. JEFFREY AND J. RATHKE
Lemma A.9 (Guarded Substitutivity). If ∆, x : T |= P ∼
= p Q and ∆, y : T → ⋄ ⊢ R and y is guarded
in R then ∆ |= R[(x : T )P/y] ∼
= p R[(x : T )Q/y].
Proof. Let R be defined as
∆ |= R′ [(x : T )P/y] R R′ [(x : T )Q/y] whenever ∆, y : T → ⋄ ⊢ R′ and y is guarded in R′
We show that R is symmetric, reduction-closed up to (=, ∼
= p ), k-contextual, and barb-preserving
and so the result follows by Lemma A.4. Symmetry, k-contextuality, and barb-preservation are
direct. For reduction-closure up to (=, ∼
= p ) we suppose:
R′ [(x : T )P/y] → R′′
By Lemma A.6 we have that R′′ = R′′′ [(x : T )P/y] and moreover:
R′ [(x : T )Q/y] → R′′′ [(x : T )Q/y]
We use Lemma A.7 to find a R′′′′ and z such that y is guarded in R′′′′ , z is unguarded in R′′′′ and
R′′′ = R′′′′ [z/y]. Hence:
R′′
=
=
R
∼
=p
=
R′′′ [(x : T )P/y]
R′′′′ [(x : T )P/y, (x : T )P/z]
R′′′′ [(x : T )Q/y, (x : T )P/z]
R′′′′ [(x : T )Q/y, (x : T )Q/z]
R′′′ [(x : T )Q/y]
(from above)
(from above)
(from definition of R and y guarded in R′′′′ [(x : T )P/z])
(from Lemma A.8 and z unguarded in R′′′′ [(x : T )Q/y])
(from above)
as required.
Corollary A.10. If ∆, x : T |= P ∼
= p Q and ∆, y : T → ⋄ ⊢ R then ∆ |= R[(x : T )P/y] ∼
= p R[(x : T )Q/y].
Proof. Follows from Lemmas A.7, A.8 and A.9.
Corollary A.11.
∼ p v · (x : T )Q.
(1) If ∆, x : T |= P ∼
= p Q and ∆ ⊢ v : T → ⋄ then ∆ |= v · (x : T )P =
(2) If ∆, x : T |= P ∼
= p Q, ∆ ⊢ a : ch[T → ⋄] and ∆ ⊢ R then ∆ |= ah(x : T )PiR ∼
= p ah(x : T )QiR.
Proof. Follows from Corollary A.10.
Proof of Lemma 2.2: The ‘only if’ direction is immediate. For the converse it is sufficient to
show that ∼
= p is preserved by each process operator of HOπ as demonstrated by Lemma A.1 and
Corollary A.11.
A PPENDIX B. M ERGE
IS A PARTIAL FUNCTION
Proof of Proposition 3.3: We consider the rewriting relation ։ which we will define as the one-step
rewriting used to define the merge operation:
C ։ X
if C doesn’t contain hk ⇐ vi for any k, v
ν(~a : ~T ) . (hk ⇐ vi k C) ։ ν(~a : ~T ) . (C[v/τk ]) if τk 6∈ v
It is easy to see that ։ is a terminating rewriting relation. Moreover, the rewriting will terminate
with a X from C (so that hhCii is defined) exactly when rg(C) is acyclic. To see this we consider the
effect of ։ on reference graphs: for
hk ⇐ vi k C
։
C[v/τk ]
CONTEXTUAL EQUIVALENCE FOR HIGHER-ORDER π-CALCULUS REVISITED
21
the reference graph of hk ⇐ vi k C has the node k removed and any edges such that
l ′ 7→ k 7→ l
for l ′ , l 6= k, are replaced with an edge
l ′ 7→ l
all other edges involving k are removed. So if node k is involved in a cycle before rewriting occurs,
that is
l 7→∗ k 7→∗ l
for some l, then either it is a tight loop, that is l = k and k 7→ k, or l 6= k and the cycle still exist after
rewriting as l 7→∗ l. The side-condition on the rewrite rule forbids tight loops hence we see that ։
preserves cyclicity. That is:
if C ։ C′ then rg(C) is acyclic if and only if rg(C′ ) is acyclic.
Now, suppose that hhCii is defined. We know that there exists a finite sequence
C ։ C1 ։ · · · ։ Cn ։ X
with hhCii = Cn . We know that rg(Cn ) is acyclic as it contains no edges. Thus, rg(C) is acyclic also.
Conversely, suppose that rg(C) is acyclic. Then as ։ is terminating there must be a finite sequence
C ։ C1 ։ · · · ։ Cn
such that Cn cannot be rewritten. There are two possibilities for this: either rg(Cn ) contains a tight
loop, or Cn is X. We see that rg(C) is acyclic, so Cn is acyclic too and therefore cannot contain a
tight loop. Thus Cn is X and hhCii is defined.
To show that hh·ii is a well-defined partial function it suffices to show that it is strongly confluent
for acyclic terms. Note that if νa : T . (C) ։ C′ then either C′ is X or C′ ≡ νa : T . (C′′ ) such that
C ։ C′′ . So without loss of generality suppose that
C ։ C1
and
C ։ C2
C ≡ C1′ k hk1 ⇐ v1 i
and
C ≡ C2′ k hk2 ⇐ v2 i
for
so that
C1 ≡ C1′ [v1 /τk1 ]
and
C2 ≡ C2′ [v2 /τk2 ].
So either, k1 = k2 in which case C1 ≡ C2 or k1 6= l2 and
C1′ ≡ C3′ k hk2 ⇐ v2 i
and
C2′ ≡ C3′ k hk1 ⇐ v1 i
We notice that
C1
(acyclicity implies τk2 6∈ v2 [v1 /τk1 ])
(acyclicity)
(def)
≡
≡
≡
։
≡
≡
≡
C1′ [v1 /τk1 ]
(C3′ k hk2 ⇐ v2 i)[v1 /τk1 ]
C3′ [v1 /τk1 ] k hk2 ⇐ v2 [v1 /τk1 ]i
C3′ [v1 /τk1 ][v2 [v1 /τk1 ]/τk2 ]
C3′ [v1 [v2 [v1 /τk1 ]/τk2 ]/τk1 , v2 [v1 /τk1 ]/τk2 ]
C3′ [v1 [v2 /τk2 ]/τk1 , v2 [v1 /τk1 ]/τk2 ]
C3
By a symmetric argument we see that C2 ։ C3′ [v2 [v1 /τk1 ]/τk2 , v1 [v2 /τk2 ]/τk1 ] and, by definition, this
is just C3 so we have C2 ։ C3 . Thus ։ is strongly confluent for acyclic terms and hence hh·ii is
well-defined.
22
A. JEFFREY AND J. RATHKE
R EFERENCES
[1] L. Cardelli and A. Gordon. Mobile ambients. In Proc. Foundations of Software Science and Computation Structures
(FoSSaCS), Lecture Notes in Computer Science. Springer-Verlag, 1998.
[2] C. Fournet and G. Gonthier. A hierarchy of equivalences for asynchronous calculi. In Proc. Int. Conf. Automata,
Languages and Programming (ICALP), volume 1443 of Lecture Notes in Computer Science. Springer-Verlag, 1998.
[3] C. Fournet, G. Gonthier, J-J. Levy, L. Maranget, and D. Remy. A calculus of mobile agents. In Proc. CONCUR,
volume 1119 of Lecture Notes in Computer Science. Springer-Verlag, 1996.
[4] A. Giacalone, P. Mishra, and S. Prasad. Facile: A symmetric integration of concurrent and functional programming.
In Proc. TAPSOFT, volume 352 of Lecture Notes in Computer Science, pages 184–209. Springer-Verlag, 1989.
[5] M. Hennessy and J. Rathke. Typed behavioural equivalences for processes in the presence of subtyping. In Proc.
Computing: the Australasian Theory Symposium (CATS), Electronic Notes in Theoretical Computer Science. Elsevier, 2002.
[6] K. Honda and N. Yoshida. On reduction-based process semantics. Theoretical Computer Science, 152(2):437–486,
1995.
[7] A.S.A Jeffrey and J. Rathke. A theory of bisimulation for a fragment of Concurrent ML with local names. In Proc.
IEEE Symp. Logic in Computer Science (LICS), pages 311–321. Computer Society Press, 2000.
[8] A.S.A Jeffrey and J. Rathke. A fully abstract may testing semantics for concurrent objects. In Proc. IEEE Symp.
Logic in Computer Science (LICS), pages 101–112. Computer Society Press, 2002.
[9] J. Riely and M. Hennessy. A typed language for distributed mobile processes. In Proc. ACM Conf. Principles of
Programming Languages (POPL). ACM Press, 1998.
[10] D. Sangiorgi. Expressing Mobility in Process Algebras: First-Order and Higher-Order Paradigms. PhD thesis,
University of Edinburgh, 1993.
[11] D. Sangiorgi. Bisimulation for higher-order process calculi. Information and Computation, 131(2):141–178, 1996.
[12] D. Sangiorgi and R. Milner. On the problem of ‘weak bisimulation up to’. In Proc. CONCUR, volume 630 of Lecture
Notes in Computer Science, pages 32–46. Springer-Verlag, 1992.
[13] D. Sangiorgi and D. Walker. The pi-calculus: A Theory of mobile processes. Cambridge University Press, 2001.
[14] B. Thomsen. Calculi for Higher-Order Communicating Systems. PhD thesis, University of London, 1990.
[15] J. Vitek and G. Castagna. Seal: A framework for secure mobile computations. In Internet Programming Languages,
volume 1686 of Lecture Notes in Computer Science. Springer-Verlag, 1999.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http://creativecommons.org/licenses/by-nd/2.0/ or send a letter to
Creative Commons, 559 Nathan Abbott Way, Stanford, California 94305, USA.
| 6cs.PL
|
Discrepancy Analysis of a New Randomized Diffusion Algorithm
for Weighted Round Matrices
Takeharu Shiraga∗
arXiv:1802.06532v2 [cs.DS] 21 Feb 2018
February 22, 2018
Abstract
For an arbitrary initial configuration of indivisible (discrete) loads over vertices of a distributed network (graph), we consider the problem of minimizing the discrepancy between the maximum and minimum load among all vertices. For this problem, diffusion-based algorithms are well studied because
of its simplicity. In diffusion-based algorithms, each vertex distributes its loads as evenly as possible
among its neighbors in each synchronous round.
This paper presents a new randomized diffusion-based algorithm inspired by multiple random walks.
In multiple random walks, at each vertex, each token (load) generates a random variable in [0, 1), and
move to a vertex corresponding to the given probability distribution (transition matrix). In our algorithm,
at each vertex, each token k (k ∈ {0, 1, . . . , X −1}) generate a random number in [k/X, (k+1)/X), and
moves to a vertex corresponding to the given probability distribution. Our algorithm is adaptive to any
transition transition probabilities while almost all previous works are concerned with uniform transition
probabilities.
For this algorithm, we analyze the discrepancy between the token configuration and its expected
√
value, and give an upper bound depending on the local 2-divergence of the transition matrix and log n,
where n is the number of vertices. The local 2-divergence is a measure which often appeared in previous
works. We also give an upper bound of the local-2 divergence for any reversible and lazy transition
matrix.
√
These yield the following specific results. First, our algorithm achieves O( d log n) discrepancy
for any d regular graph, which matches the best result on previous works of diffusion model. Note
that our algorithm does not need any assumption of the number of tokens such as negative loads which
are often assumed in previous
works. Second, for general graphs with maximum degree dmax , our
√
algorithm achieves O( dmax log n) discrepancy using the transition matrix based on the metropolis
hasting algorithm. Note that this algorithm does not need information of dmax while almost all previous
works use it.
Key words: Load balancing, Diffusion, Markov chain
1
Introduction
This paper is concerned with the load balancing algorithms on distributed networks. Let G = (V, E) be
an undirected and connected network (graph), and X (0) ∈ Zn≥0 be an initial configuration of loads (tokens)
over V , where n = |V |. Then, we consider a distributed and iterative algorithm to balance the tokens over
vertices, i.e., each vertex moves its tokens to its neighbors iteratively at each discrete and synchronous time
step to minimize the discrepancy between the maximum and minimum tokens among all vertices.
∗
Department of Information and System Engineering, Faculty of Science and Engineering, Chuo University, Tokyo, Japan
[email protected]
1
1.1
Previous works
Load balancing algorithms studied in previous works are classified as diffusion-based algorithms and matchingbased algorithms. Roughly speaking, each vertex sends its tokens to neighbors as evenly as possible (according to round matrix) at each round in diffusion-based algorithms. Matching-based algorithms generate
a (different) matching of the graph in a distributed way at each round, and the endpoints of each matching
balances tokens as evenly as possible.
Rabani et al. [10] studied deterministic diffusion-based algorithms and matching-based algorithms, and
gave a framework of the analysis of these models. They introduced the local 1-divergence and show that
the discrepancy is upper bounded by the local 1-divergence within O(log(Kn)/(1 − λ)) step, where K =
(0)
(0)
maxx,y |Xx − Xy | is the initial discrepancy and λ is the second largest eigenvalue of the round matrix.
They also showed that the local 1-divergence is upper bounded by O(d log n/(1 − λ)) for any d regular
graph.
Randomization is a natural approach to get smaller discrepancy. Friedrich and Sauerwald [6] studied
randomized version√of matching models. They showed that the discrepancy is upper bounded by the local
2-divergence
times log n with in O(log(Kn)/(1 − λ)) step and the local 2-divergence is upper bounded
p
by O( d/(1 − λ)).
Berenbrink et al. [3] studied a√randomized diffusion-based algorithm. They showed that the discrepancy
is upper bounded by O(d max{ log n, log log n/(1 − λ)}) within O(log(Kn)/(1 − λ)) step for any d
regular graph.
Akbari and Berenbrink [1] studied randomized and deterministic diffusion algorithm based on rotorrouter mechanism. They gave same bound of [3] for their randomized diffusion algorithm using fewer
random bits compared to [3]. Furthermore, they gave upper bounds of deterministic diffusion algorithm on
specific graphs, O(d1.5 ) for hypercube and O(1) for constant-dimensional tori.
The results of Sauerwald and Sun [11] is the best result of the discrepancy so far. They showed that
constant discrepancy within O(log(Kn)/(1 − λ)) step for a randomized matching model.
Recently, Berenbrink et al. [4] studied deterministic diffusion based algorithms. They showed that O(d)
discrepancy within O((log(Kn) + d log2 n)/(1 − λ)) step. This result improves [10, 1].
There are some other results concerned with the randomized diffusion-based algorithms which use negative loads [5, 2, 11]. In [11], authors showed that O(dmax log n) discrepancy within O(log(Kn)/(1 − λ))
step for any graphs.
1.2
This work
In this paper, we consider the following diffusion-based algorithm. Let P be an arbitrary transition matrix
over V , which is related to the round matrix in previous works. For example, Px,y = 1/dx if (x, y) ∈ E,
(t)
where dx is the degree size of x. Now, let Xv denote the number of tokens on v at time t in our algorithm.
(t)
Algorithm. At each time step t and at each vertex v ∈ V , each token k (k ∈ {0, 1, . . . , Xv − 1})
(t)
(t)
generates a random number in [k/Xv , (k + 1)/Xv ) uniformly at random. Then, each token moves to its
corresponding neighbor.
See Figure 1 for an example. There are 5 tokens (k = 0, 1, 2, 3, 4) on v and (Pv,v0 , Pv,v1 , Pv,v2 , Pv,v3 ) =
(1/10, 3/5, 1/5, 1/10). In this example, token 0 moves to v0 (v1 ) with probability 1/2, token 1 (2) moves
to v1 with probability 1, token 3 moves to v1 (v2 ) with probability 1/2, and token 4 moves to v2 (v3 ) with
probability 1/2.
More precisely, we assume that for each v ∈ V , each u is mapped into a interval in [0, 1) whose length
(t)
(t)
is Pv,u . Then each token k moves to u if the generated random number in [k/Xv , (k + 1)/Xv ) is in the
interval of u. (See equation (1) in Section 2.2 for the concrete definition).
2
𝑘:
0
1
2
3
4
𝑣𝑖 :
𝑣0
𝑣1
𝑣2
𝑣3
𝑃𝑣,𝑣𝑖 :
1Τ10
3Τ5
1Τ5
1Τ10
Figure 1:
This algorithm is quite simple. At each round, each vertex only sends its tokens using the number of
tokens on it and the probability distribution around it (e.g., its degree size), does not need to communicate
to neighbors or use the number of tokens of its neighbor.
Furthermore, this algorithm allows to use arbitrary transition matrix. Almost all previous works are
corresponding to “uniform” transition matrices, i.e., Pv,x = Pv,y holds for any x, y in the neighboring set of
v (x, y 6= v).
The main result of this paper is the following general upper bound.
Main results in general form. First, we introduce the local p-divergence of P .
Definition 1.1 (local p-divergence, [10, 6]). For any p ∈ Z≥0 , the local-p divergence of P is defined by
1/p
∞
X
Ψp (P ) := max
w∈V
X
t=0 (v,u)∈V ×V
:Pv,u >0
t
t
|Pv,w
− Pu,w
|p
.
Then, we show the following main theorem.
Theorem 1.2. For any initial configuration X (0) and transition matrix P ,
√
2
(T )
(0) T
Pr max Xw − (X P )w ≤ 4Ψ2 (P ) ln n ≥ 1 −
w∈V
n
holds for each time T ∈ Z≥0 .
Note that X (0) P T is the expected value of X (T ) (See Section 3.2 for details). X (0) P T is converge to
1
1 π, where π is the stationary distribution of P (See Appendix A for details).
Second, we obtain the following upper bound of the local 2-divergence for weighted P .
kX (0) k
Theorem 1.3. Suppose that P is reversible 2 and lazy 3 . Then,
s
2 maxw∈V πw
Ψ2 (P ) ≤
min(v,u)∈E◦P πv Pv,u
holds, where E◦P = {(v, u) ∈ V × V | Pv,u > 0 and v 6= u}.
1
Probability distribution such that πP = π holds.
P is reversible if the detailed balance equation πv Pv,u = πu Pu,v holds for any u, v ∈ V .
3
P is lazy if Pv,v ≥ 1/2 holds for any v ∈ V .
2
3
This is a generalized version of the previous result [11], which is corresponding to uniform P .
These results in general form give the following specific result.
Contribution for regular graphs. Now, we consider the lazy transition matrix PL on d-regular graph:
(PL )x,x = 1/2, (PL )x,y = 1/2d if (x, y) ∈ E and (PL )x,y = 0 otherwise. For this transition matrix, we
obtain the following.
Corollary 1.4. Suppose that G = (V, E) is an arbitrary connected d-regular graph and the transition
matrix is PL . Then, for each T ≥ log(2Kn)
1−λ , it holds that
Pr max
x,y∈V
|Xx(T )
−
Xy(T ) |
√
2
≤ 16 d ln n ≥ 1 − .
n
This upper bound improves the previous result of [3]. Corollary 1.4 achieves the same upper bound of
the diffusion model in [11] without negative loads.
Contribution for general graphs. Since our algorithm allows to use any transition matrix, we can use
the Metropolis chain PM for general graphs, where (PM )x,y = (1/2) min{1/dx , 1/dy } if (x, y) ∈ E,
(PM )x,y = 0 if (x, y) ∈
/ E and the self loop is equal to the remaining probability [9]. For this transition
matrix, we obtain the following.
Corollary 1.5. Suppose that G = (V, E) is an arbitrary connected graph and the transition matrix is PM .
Then, for each T ≥ log(2Kn)
1−λ , it holds that
Pr max
x,y∈V
|Xx(T )
−
Xy(T ) |
p
2
≤ 16 dmax ln n ≥ 1 − .
n
This also matches the previous result of the diffusion model on general graph [11] without negative loads.
Furthermore, each vertex only needs the degree of its neighbors to calculate (PM )x,y = (1/2) max{dx , dy },
while previous work [11] needs dmax , which is the maximum value of the degree among all vertices.
2
Notations and Model description
In this section we describe our model precisely.
2.1
Notations
P
Let V be a vertex set, and let n = |V |. Let P ∈ [0, 1]n×n be a transition matrix on V , i.e., u∈V Pv,u = 1
t denotes (v, u) entry of P t . In this paper we
holds for any v ∈ V , where Pv,u denotes (v, u) entry of P . Pv,u
assume that P 0 is the identity matrix.
Let NvP be the set of neighbors of v ∈ V , i.e., NvP := {u ∈ V | Pv,u > 0}. In this paper, we assume an
arbitrary ordering on NvP , i.e., we denote NvP = {v0 , v1 , . . . , vdPv −1 }, where dPv = |NvP |. Let E P be the set
of edges of transition diagram of P , i.e., E P := {(v, u) ∈ V × V | Pv,u > 0}. Note that E P may contain
self-loop edges.
2.2
Model description
Let X (0) ∈ Zn≥0 be a initial configuration of M tokens over V , and let X (t) ∈ Zn≥0 denote the configuration
of M tokens over V at time t ∈ Z≥0 in our algorithm. In an update from X (t) to X (t+1) of our algorithm,
(t)
(t)
(t)
(t)
(t)
at each vertex v ∈ V , we generate Xv random numbers rv (0), rv (1), . . . , rv (Xv − 1), where each
4
(t)
rv (k) is picked from
operators
(t)
Dv (k)
h
k
k+1
(t) ,
(t)
Xv
Xv
uniformly at random. From these random numbers, we define destination
such that
i−1
i
X
X
Dv(t) (k) = vi if rv(t) (k) ∈
Pv,vj ,
Pv,vj .
j=0
(1)
j=0
(t)
Dv (k) denotes the destination of (k + 1)-th token on v ∈ V at time t. Then, let
(t)
Xv(t+1)
u −1
X XX
:=
1 Du(t) (k) = v .
u∈V
(2)
k=0
(t)
According to the above definition, destination operators Dv (k) satisfy the following properties.
(t)
Observation 2.1. Suppose that X (t) is fixed. Then, for any v, x ∈ V , k ∈ {0, 1, . . . Xv − 1} and ` ∈
(t)
(t)
(t)
{0, 1, . . . Xx − 1}, Dv (k) and Dx (`) are independent if v 6= x or k 6= `.
Observation 2.2.
h
Pr Dv(t) (k) = vi | Xv(t)
i
=
"
!
i−1
i
X
X
k
+
1
k
, (t)
· Xv(t)
Pv,vj ,
Pv,vj ∩
(t)
X
X
v
v
j=0
j=0
(t)
This paper is concerned with the behavior of X (t) defined by destination operators Dv (k) satisfying
Observation 2.1 and 2.2.
3
Proof of Theorem 1.2
This section gives the proof of Theorem 1.2.
Theorem 3.1 (Theorem 1.2). For any initial configuration X (0) and transition matrix P ,
√
2
(T )
(0) T
Pr max Xw − (X P )w ≤ 4Ψ2 (P ) ln n ≥ 1 −
w∈V
n
holds for each time T ∈ Z≥0 .
3.1
Framework of the proof
First, we introduce some notations for the proof. Let V = {0, 1, . . . , n − 1}. For any t ∈ {0, 1, . . . , T − 1},
v ∈ V , and k ∈ {0, 1, . . . , M − 1}, let
lv(t) (k) := (M n)t + (M )v + k,
(3)
and let
Dl(t) (k)
v
(
(t)
Dv (k)
:=
−1
(if k ∈ {0, 1, . . . , M − 1})
.
(otherwise)
5
(4)
(t)
(t)
This definition means that Pr[Dl(t) (k) = u | Xv ] = 0 for any k ≥ Xv
v
convenience, let
and u ∈ V . For the notational
D` := (Dl )l<` = D0 , D1 , . . . , D`−1 ,
(5)
which is a sequence of random variables. Then, we observe the following from the definition of the configuration of tokens (2).
Observation 3.2. For any t ∈ Z≥0 , X (t) is determined by X (0) and DM nt = D0 , D1 , . . . , Dl(t−1) (M −1) .
n−1
Now, let
h
i
Y` = Y` (w, T ) = E Xw(T ) − (X (0) P T )w | D` .
(6)
(T )
We define Y0 := E[Xw − (X (0) P T )w ]. Note that Y0 , Y1 , . . . is a martingale with respect to D0 , D1 , . . .,
thus
"
#
MX
nT −1
2
2
(c` )
(7)
Pr |YM nT − Y0 | ≥ η ≤ 2 exp −η /2
`=0
holds from Azuma-Hoeffding inequality (See Appendix A), where c` is a value satisfies |Y`+1 − Y` | ≤ c` .
(T )
(T )
Since YM nT = Xw − (X (0) P T )w from Observation 3.2 and Y0 = E[Xw − (X (0) P T )w ] = 0 (See
Lemma 3.4 and (17) in Section 3.2 for the detail),
v
uM nT −1
u X
2
Pr |Xw(T ) − (X (0) P T )w | ≥ 2t
(8)
(c` )2 ln n ≤ 2
n
`=0
q P
nT −1
(c` )2 ln n2 . Thus, using the union bound to (8) and apply the following
holds by taking η = 2 M
`=0
lemma, we obtain Theorem 1.2.
Lemma 3.3. For any T ∈ Z≥0 and w ∈ V , it holds that
MX
nT −1
2
2
Y`+1 (w, T ) − Y` (w, T )
≤ 4 Ψ2 (P ) .
`=0
To complete the proof, we prove Lemma 3.3 in the following subsection.
3.2
Proof of Lemma 3.3
For the notational convenience, let
(
∅
e
k
I(X,
k) :=
k+1
X, X
(if X = 0)
(otherwise)
(9)
and
Pev,vi
i
i−1
X
X
:=
Pv,vj ,
Pv,vj .
j=0
(10)
j=0
First, we introduce the following lemma inspired by Lemma 4.1 in [12]. We use the definition of our
algorithm (2) to prove Lemma 3.4.
6
Lemma 3.4. For any P , X (0) , w ∈ V and T ∈ Z≥0 , it holds that
Xw(T ) − (X (0) P T )w
=
T
−1 X
X
−1
X M
X
T −t−1
T −t−1
e (t) , k)
1 Dl(t) (k) = u − Xv(t) Pev,u ∩ I(X
Pu,w
− Pv,w
.
v
v
t=0 v∈V u∈NvP k=0
Proof. It is not difficult to see that
T
−1
X
T
−1
X
X (t+1) − X (t) P P T −t−1 =
X (t+1) P T −t−1 − X (t) P T −t
t=0
t=0
(T )
= X
P 0 − X (0) P T = X (T ) − X (0) P T .
(11)
From (11) we obtain
Xw(T )
− (X
(0)
T
−1
X
T
P )w =
X
T −t−1
Xu(t+1) − (X (t) P )u Pu,w
.
(12)
t=0 u∈V
Then, from the definitions (2) and (4) we have
(t)
Xu(t+1)
v −1
X XX
=
1 Dv(t) (k) = u
v∈V
k=0
(X (t) P )u =
X
=
−1
XM
X
1 Dl(t) (k) = u .
v
(13)
v∈V k=0
We also have
X
Xv(t) Pv,u =
v∈V
Xv(t)
v∈V
M
−1
X
e (t) , k)
Pev,u ∩ I(X
v
(14)
k=0
from the definitions (9) and (10). Finally, we have
X
(t)
(t) e
e
1 Dl(t) (k) = u − Xv Pv,u ∩ I(Xv , k) Pv,w = 0
v
(15)
u∈NvP
since
X
1 Dl(t) (k) = u
v
u∈NvP
=
X
e (t) , k) = 1.
Xv(t) Pev,u ∩ I(X
v
(16)
u∈NvP
Combining (12), (13) and (14) and subtracting (15), we obtain the claim.
Note that from Observation 2.2 and the chain rule of the conditional expectation,
h
h h
i
ii
E 1 Dl(t) (k) = u
= E E 1 Dl(t) (k) = u | Xv(t)
v
v
h h
ii
= E Pr Dl(t) (k) = u | Xv(t)
v
h
i
e (t) , k)
= E Xv(t) Pev,u ∩ I(X
v
(T )
holds, thus we have Y0 = E[Xw − (X (0) P T )w ] = 0.
Next, we show the following lemma using Lemma 3.4.
7
(17)
(τ )
Lemma 3.5. For any τ ∈ {0, 1, . . . , T − 1}, ν ∈ V and κ ∈ {0, 1, . . . , Xν − 1}, it holds that
X
T −τ −1
T −τ −1
e (τ ) , κ)
1 Dl(τ ) (κ) = u − Xν(τ ) Peν,u ∩ I(X
Pu,w
− Pν,w
.
Yl(τ ) (κ)+1 − Yl(τ ) (κ) =
ν
ν
ν
ν
u∈NνP
Proof. From the definition of Y` (6) and Lemma 3.4, we have
Yl(τ ) (κ)+1 − Yl(τ ) (κ) =
T
−1 n−1
−1
X
XM
X
X
ν
ν
T −t−1
T −t−1
(∗) Pu,w
− Pv,w
,
(18)
t=0 v=0 k=0 u∈NvP
where
h
i
∗ = E 1 Dl(t) (k) = u | Dl(τ ) (κ)+1
ν
h v
i
(t) e
(t)
e
− E Xv Pv,u ∩ I(X
,
k)
|
D
(τ )
v
l (κ)+1
h
iν
− E 1 Dl(t) (k) = u | Dl(τ ) (κ)
v
ν
i
h
(t) e
(t)
e v , k) | D (τ )
.
+ E Xv Pv,u ∩ I(X
l (κ)
(19)
ν
To obtain the claim, let we start with showing
h
i
E 1 Dl(t) (k) = u | Dl(τ ) (κ)
ν
v h
i
(t)
(τ )
E Xv(t) Pev,u ∩ I(X
e v(t) , k) | D (τ )
if
l
(k)
≥
l
(κ)
v
ν
lν (κ)
=
.
1 D (t) = u
(otherwise)
l (k)
(20)
v
h
i
(t)
(τ )
Obviously, E 1 Dl(t) (k) = u | Dl(τ ) (κ) = 1 Dl(t) (k) = u holds if lv (k) < lν (κ). Hence we
v
consider if
(t)
lv (k)
ν
v
(τ )
lν (κ)
≥
holds. In this case, from the chain rule of the conditional expectation,
h
i
h h
i
i
E 1 Dl(t) (k) = u | Dl(τ ) (κ) = E E 1 Dl(t) (k) = u | Dl(t) (k) | Dl(τ ) (κ)
v
ν
v
v
ν
(21)
holds. From Observation 2.1, we have
h h
i
i
(21) = E E 1 Dl(t) (k) = u | Dl(t) (0) | Dl(τ ) (κ)
v
ν
0
i
h h
i
= E Pr Dl(t) (k) = u | DM nt | Dl(τ ) (κ) .
v
ν
(22)
(t)
Observation 3.2 says that DM nt determines Xv . Using Observation 2.2, we have
h
i
e v(t) , k) .
Pr Dl(t) (k) = u | DM nt = Xv(t) Pev,u ∩ I(X
(23)
Thus combining equations (21)-(23), we obtain (20).
Note that Observation 3.2 also says that
h
i
e v(t) , k) | D (τ )
e v(t) , k) (if t ≤ τ )
E Xv(t) Pev,u ∩ I(X
= Xv(t) Pev,u ∩ I(X
l (κ)
(24)
v
ν
holds.
Now we prove the lemma using above discussion. We consider the following 3 cases.
8
case 1. [t ≤ τ − 1] or [t = τ and v ≤ ν − 1] or [t = τ , v = ν and k ≤ κ − 1 ]:
(t)
(τ )
In this case, lv (k) ≥ lν (κ) + 1 holds. Using (20), we have
(∗) = 0
(25)
in case 1.
case 2. [t = τ , v = ν and k ≥ κ + 1] or [t = τ and v ≥ ν + 1] or [t ≥ τ + 1]:
(t)
(τ )
In this case, we have lv (k) ≤ lν (κ) − 1. From (20), we have
h
i
h
i
E 1 Dl(t) (k) = u | Dl(τ ) (κ)+1 = E 1 Dl(t) (k) = u | Dl(τ ) (κ) .
v
ν
v
(26)
ν
Furthermore,
i
i
h
h
(t) e
(t)
e
e v(t) , k) | D (τ )
|
D
=
E
X
P
∩
I(X
,
k)
E Xv(t) Pev,u ∩ I(X
(τ
)
v,u
v
v
l (κ)
l (κ)+1
ν
ν
(27)
holds from (24). Thus
(∗) = 0
(28)
in case 2.
case 3. [t = τ , v = ν and k = κ]: First, we have
h
i
h
i
(t) e
(t)
e v(t) , k) | D (τ )
e
=
E
X
P
∩
I(X
,
k)
|
D
E Xv(t) Pev,u ∩ I(X
(τ
)
v,u
v
v
l (κ)+1
l (κ)
ν
ν
(29)
from (24). We also have
h
i
E 1 Dl(τ ) (κ) = u | Dl(τ ) (κ)+1 = 1 Dl(τ ) (κ) = u
ν
ν
(30)
ν
and
h
i
h
i
e ν(τ ) , κ) | D (τ )
E 1 Dl(τ ) (κ) = u | Dl(τ ) (κ) = E Xν(τ ) Peν,u ∩ I(X
l (κ)
ν
ν
ν
=
Xν(τ )
Peν,u ∩
e (τ ) , κ)
I(X
ν
(31)
from (20). Thus
e ν(τ ) , κ)
(∗) = 1 Dl(τ ) (κ) = u − Xν(τ ) Peν,u ∩ I(X
ν
(32)
in case 3.
Combining (18), (25), (28) and (32), we obtain the claim.
Proof of Lemma 3.3. First, we observe that
1 Dl(τ ) (κ) = u −
ν
(τ )
since Xν
Xν(τ )
Peν,u ∩
e (τ ) , κ)
I(X
ν
e ν(τ ) , κ) = Pr D (τ )
Peν,u ∩ I(X
l (κ)
ν
X
=
(
1−1=0
(τ )
(if Xν
e ν(τ ) , κ) = 1)
Peν,u ∩ I(X
e ν(τ ) , κ) = 0)
Peν,u ∩ I(X
(33)
(τ )
0 − 0 = 0 (if Xν
(τ )
= u | Xν
from Observation 2.2. Furthermore, we have
e ν(τ ) , κ)
1 Dl(τ ) (κ) = u − Xν(τ ) Peν,u ∩ I(X
ν
u∈NνP
≤
X
1 Dl(τ ) (κ) = u
ν
+
u∈NνP
X
e ν(τ ) , κ)
Xν(τ ) Peν,u ∩ I(X
u∈NνP
= 1 + 1 = 2.
(34)
9
Now, let
o
n
e (τ ) , κ) < 1 .
Sν(τ ) (κ) = u ∈ NvP | 0 < Xν(τ ) Peν,u ∩ I(X
ν
(35)
(τ )
(τ )
e ν(τ ) , κ) = 0 since (33) holds. Combining
Note that if u 6= Sν (κ), then 1 Dl(τ ) (κ) = u − Xν Peν,u ∩ I(X
ν
Lemma 3.5, Cauchy–Schwarz and (34), we have
2
Yl(τ ) (κ)+1 − Yl(τ ) (κ)
ν
ν
2
X
T −τ −1
T −τ −1
e (τ ) , κ)
=
1 Dl(τ ) (κ) = u − Xν(τ ) Peν,u ∩ I(X
Pu,w
− Pν,w
ν
ν
u∈NνP
2
X
T −τ −1
T −τ −1
e (τ ) , κ)
=
1 Dl(τ ) (κ) = u − Xν(τ ) Peν,u ∩ I(X
Pu,w
− Pν,w
1 u ∈ Sν(τ ) (κ)
ν
ν
u∈NνP
≤
2 X
2
X
T −τ −1
T −τ −1
e (τ ) , κ)
Pu,w
− Pν,w
1 u ∈ Sν(τ ) (κ)
1 Dl(τ ) (κ) = u − Xν(τ ) Peν,u ∩ I(X
ν
ν
u∈NνP
u∈NνP
≤
X
e (τ ) , κ)
1 Dl(τ ) (κ) = u − Xν(τ ) Peν,u ∩ I(X
ν
X
ν
u∈NνP
≤ 2
X
T −τ −1
T −τ −1
Pu,w
− Pν,w
2
1 u ∈ Sν(τ ) (κ)
u∈NνP
T −τ −1
T −τ −1 2
Pu,w
− Pν,w
1 u ∈ Sν(τ ) (κ) .
(36)
u∈NνP
(τ )
e ν(τ ) , κ) ≤ 1. Thus
The second inequality holds since 1 Dl(τ ) (κ) = u − Xν Peν,u ∩ I(X
ν
MX
nT −1
(Y`+1 − Y` )
2
=
T
−1 n−1
−1
X
XM
X
Yl(τ ) (κ)+1 − Yl(τ ) (κ)
ν
2
ν
τ =0 ν=0 κ=0
`=0
≤
T
−1 n−1
−1
X
XM
X
τ =0 ν=0 κ=0
= 2
T
−1 n−1
X
X
X
2
T −τ −1
T −τ −1
Pu,w
− Pν,w
u∈NνP
X
T −τ −1
Pu,w
−
T −τ −1 2
Pν,w
τ =0 ν=0 u∈NνP
≤ 4
T
−1 n−1
X
X
2
1 u ∈ Sν(τ ) (κ)
X
M
−1
X
1 u ∈ Sν(τ ) (κ)
κ=0
T −τ −1
T −τ −1
Pu,w
− Pν,w
2
τ =0 ν=0 u∈NνP
= 4
T
−1
X
X
τ
τ
Pu,w
− Pν,w
2
τ =0 (u,ν)∈E P
≤ 4(Ψ2 (P ))2
holds, and we obtained the claim. Note that the second inequality holds since for each u ∈ NνP , the number
(τ )
(τ )
e ν(τ ) , κ) < 1) holds is at most two. (For example, see
of κ such that Sν (κ) 3 u (0 < Xν Peν,u ∩ I(X
Figure 1. If u = v1 , such κ is 0 and 3.)
10
4
Upper bound of the local 2-divergence and specific results
This section shows Theorem 1.3 and other specific results.
Theorem 4.1 (Theorem 1.3). Suppose that P is reversible and lazy. Then,
s
2 maxw∈V πw
Ψ2 (P ) ≤
min(v,u)∈E◦P πv Pv,u
holds, where E◦P = {(v, u) ∈ V × V | Pv,u > 0 and v 6= u}.
4.1
Proof of Theorem 1.3
To prove Theorem 1.3, we introduce the following lemma. We use the reversibility of P to prove Lemma 4.2.
Lemma 4.2. For any reversible P , it holds that
X X
t
t
2t
2t+1
πv Pv,u (Pu,w
− Pv,w
)2 = 2πw (Pw,w
− Pw,w
).
v∈V u∈NvP
Proof. From the definition of NvP , we have
X X
XX
t
t
t
t
πv Pv,u (Pu,w
− Pv,w
)2 =
πv Pv,u (Pu,w
− Pv,w
)2
v∈V u∈NvP
v∈V u∈V
=
XX
t
t
t
t
πv Pv,u (Pu,w
)2 + (Pv,w
)2 − 2Pu,w
Pv,w
v∈V u∈V
=
XX
v∈V u∈V
−2
XX
t
πv Pv,u (Pu,w
)2 +
t
πv Pv,u (Pv,w
)2
v∈V u∈V
XX
t
t
πv Pv,u Pu,w
Pv,w
.
(37)
v∈V u∈V
Then, from the reversibility of P ,
XX
XX
XX
XX
t
t
t
t
πv Pv,u (Pu,w
)2 +
πv Pv,u (Pv,w
)2 =
πu Pu,v (Pu,w
)2 +
πv Pv,u (Pv,w
)2
v∈V u∈V
v∈V u∈V
v∈V u∈V
=
X
u∈V
=
X
+
X
t
πv (Pv,w
)2
v∈V
t
t
πw Pw,u
Pu,w
+
u∈V
=
v∈V u∈V
t
πu (Pu,w
)2
X
t
t
πw Pw,v
Pv,w
v∈V
2t
2πw Pw,w
,
(38)
and
XX
t
t
πv Pv,u Pu,w
Pv,w
= πw
v∈V u∈V
=
holds. Thus we obtain the claim.
11
XX
v∈V u∈V
2t+1
πw Pw,w
t
t
Pw,v
Pv,u Pu,w
(39)
Proof of Theorem 1.3. From Lemma 4.2, we have
∞ X X
X
t
πv Pv,u (Pu,w
−
t
Pv,w
)2
t=0 v∈V u∈NvP
∞
X
2t
2t+1
)
= 2πw
(Pw,w
− Pw,w
(40)
t=0
∞
X
2t
2t+2
(Pw,w
− Pw,w
)
≤ 2πw
(41)
t=0
≤ 2πw .
(42)
t
t+1 holds for lazy P . Thus we obtain the claim from (42).
Note that Pw,w
≥ Pw,w
4.2
Specific results
4.2.1
Lazy chain on regular graphs
Let G = (V, E) be an undirected and connected graph. Additionally, we assume G is d-regular graph. Then,
we consider the following transition matrix
1
2d (if (v, u) ∈ E)
.
(43)
(PL )v,u = 12
(if v = u)
0
(othrwise)
Then, the following corollary is obtained from Theorem 1.2 and Theorem 1.3.
Corollary 4.3. Suppose that G = (V, E) is an arbitrary d-regular graph and the transition matrix is PL .
Then, for each T ∈ Z≥0 , it holds that
√
2
(T )
(0) T
Pr max |Xw − (X PL )w | ≤ 8 d ln n ≥ 1 − .
w∈V
n
Combining Corollary 4.3 and Proposition A.2, we obtain Corollary 1.4.
4.2.2
Metropolis chain on general graphs
Now, we consider an arbitrary undirected and connected graph G = (V, E). We do not assume the regularity
of G. Let dv be the degree of v ∈ V , i.e., dv = |{u ∈ V | (v, u) ∈ E}|. Then, we consider the following
transition matrix on V
n
o
1
1
1
min
,
(if (v, u) ∈ E)
dv du
2
P
(PM )v,u = 1 − u:(v,u)∈E (PM )v,u (if v = u)
.
(44)
0
(othrwise)
(PM ) is known as the Metropolis chain [9]. Let dmax = maxv∈V dv . Then, the following corollary is
obtained from Theorem 1.2 and Theorem 1.3.
Corollary 4.4. Suppose that G = (V, E) is an arbitrary graph and the transition matrix is PM . Then, for
each T ∈ Z≥0 , it holds that
p
2
(T )
(0) T
Pr max |Xw − (X PM )w | ≤ 8 dmax ln n ≥ 1 − .
w∈V
n
Combining Corollary 4.4 and Proposition A.2, we obtain Corollary 1.5.
12
Acknowledgements
This work is supported by JSPS KAKENHI Grant Number 17H07116.
References
[1] H. Akbari and P. Berenbrink, Parallel rotor walks on finite graphs and applications in discrete load
balancing, Proc. SPAA 2013, 186–195.
[2] H. Akbari, P. Berenbrink and T. Sauerwald, A simple approach for adapting continuous load balancing
processes to discrete settings, Distributed Computing, 29(2) (2016), 143–161.
[3] P. Berenbrink, C. Cooper, T. Friedetzky, T. Friedrich, T. Sauerwald, Randomized diffusion for indivisible loads, Journal of Computer and System Sciences 81(1) (2015), 159–185.
[4] P. Berenbrink, R. Klasing, A. Kosowski, F. Mallmann-Trenn, and P. Uznanski, Improved analysis of
deterministic load-balancing schemes, Proc. PODC 2015, 301–310.
[5] T. Friedrich, M. Gairing, and T. Sauerwald, Quasirandom load balancing, SIAM Journal on Computing, 41 (2012), 747–771.
[6] T. Friedrich and T. Sauerwald, Near-perfect load balancing by randomized rounding, Proc. STOC
2009, 121–130.
[7] D. A. Levin and Y. Peres, Markov Chain and Mixing Times: Second Edition, The American Mathematical Society, 2017.
[8] M. Mitzenmacher and E. Upfal, Probability and Computing Randomization and Probabilistic Techniques in Algorithms and Data Analysis 2nd edition, Cambridge University Press, 2017.
[9] Y. Nonaka, H. Ono, K. Sadakane, M. Yamashita, The hitting and cover times of Metropolis walks,
Theoretical Computer Science, 411 (2010), 1889–1894.
[10] Y. Rabani, A. Sinclair, and R. Wanka, Local divergence of Markov chains and analysis of iterative load
balancing schemes, Proc. FOCS 1998, 694–705.
[11] T. Sauerwald and H. Sun, Tight bounds for randomized load balancing on arbitrary network topologies,
Proc. FOCS 2012, 341–350.
[12] T. Shiraga, Y. Yamauchi, S. Kijima, and M. Yamashita, Deterministic random walks for rapidly mixing
chains, arXiv:1311.3749.
A
A.1
APPENDIX
Preliminaries of Markov chains
Proposition A.1. Suppose that P is reversible. Then, for any X (0) , w ∈ V and
M
1
log
,
T ≥
1−λ
πmin ε
it holds that
(X (0) P T )w − M πw ≤ ε.
13
Proof. We have
(X (0) P T )w − M πw
=
X
T
T
Xv(0) (Pv,w
− πw ) ≤ M max |Pv,w
− πw |.
v∈V
v∈V
Then, from Theorem 12.4 in [7], we obtain the claim.
Proposition A.2. Suppose that P is symmetric. Then, for any X (0) , w ∈ V and
1
2Kn
T ≥
,
log
1−λ
ε
it holds that
(X (0) P T )w −
M
≤ ε,
n
(0)
(0)
where K := maxx,y∈V |Xx − Xy |.
Proof. We have
(X (0) P T )w − M/n
=
X
T
Xv(0) (Pv,w
− 1/n)
v∈V
=
X
T
Xv(0) (Pw,v
− 1/n)
v∈V
=
X
T
(Xv(0) − Xx(0) )(Pw,v
− 1/n)
v∈V
≤ 2K·
1X T
|Pw,v − 1/n|.
2
v∈V
Then, from Theorem 12.4 in [7], we obtain the claim.
A.2
Concentration inequality
Theorem A.3 (Asuma-Hoeffding Inequality, [8]). Let X0 , . . . , Xn be a martingale such that
|Xk − Xk−1 | ≤ ck .
Then, for all t ≥ 1 and any λ > 0,
"
λ2
#
Pr [|Xt − X0 | ≥ λ] ≤ 2 exp − Pt
.
2 k=1 (ck )2
14
| 8cs.DS
|
Real Options for Project Schedules (ROPS)
Lester Ingber
Lester Ingber Research
Ashland Oregon
[email protected], [email protected]
http://www.ingber.com/
Abstract
Real Options for Project Schedules (ROPS) has three recursive sampling/optimization shells. An outer
Adaptive Simulated Annealing (ASA) optimization shell optimizes parameters of strategic Plans
containing multiple Projects containing ordered Tasks. A middle shell samples probability distributions
of durations of Tasks. An inner shell samples probability distributions of costs of Tasks. PATHTREE is
used to develop options on schedules. Algorithms used for Trading in Risk Dimensions (TRD) are
applied to develop a relative risk analysis among projects.
KEYWORDS: options; simulated annealing; risk management; copula; nonlinear; statistical
† L. Ingber, “Real Options for Project Schedules (ROPS)”, Report 2007:ROPS, Lester Ingber Research,
Ashland, OR, 2007. URL http://www.ingber.com/markets07_rops.pdf.
$Id: markets07_rops,v 1.22 2007/04/01 14:32:24 ingber Exp ingber $
Lester Ingber
-2-
Real Options for Project Schedules (ROPS)
1. Introduction
This paper is a brief description of a methodology of developing options (in the sense of financial options,
e.g., with all Greeks), to be applied in collaboration with Michael Bowman, as a first example to
scheduling a massive US Army project, Future Combat Systems (FCS) [1].
The major focus is to develop Real Options for non-financial projects, as discussed in other earlier
papers [3,4,12]. Data and some guidance on its use has been reported in a previous study of FCS [2,5].
The need for tools for fairly scheduling and pricing such a complex project has been emphasized in
Recommendations for Executive Action in a report by the U.S. General Accounting Office (GAO) on
FCS [14], and they also emphasize the need for management of FCS business plans [13].
2. Goals
A given Plan results in S(t), money allocated by the client/government is defined in terms of Projects
S i(t),
S(t) = Σ S i(t)
i
where ai(t) may be some scheduled constraints. PATHTREE processes a probability tree developed over
the life of the plan T , divided into N nodes at times {t n }, each with mean epoch length dt [11]. Options,
including all Greeks, familiar to financial markets, are calculated for quite arbitrary nonlinear means and
variances of multiplicative noise [6,9]. This ability to process nonlinear functions in probability
distributions is essential for real-world applications.
Each Task has a range of durations, with nonzero Ai , with a disbursement of funds used, defining S i(t n ).
Any Task dependent on a Task completion is slaved to its precursor(s).
We develop the Plan conditional probability density (CPD) in terms of differenced costs, dS,
P(S ± dS; t n + dt|S; t n)
P is modeled/cast/fit into the functional form
−
P(S ± dS; t n + dt|S; t n ) = (2π g2 dt)
L=
1
2
exp(−Ldt)
(dS − fdt)2
(2g2 dt 2)
where f and g are nonlinear function of cost S and time t. The g2 variance function absorbs the multiple
Task cost and schedule statistical spreads, to determine P(dS, t), giving rise to the stochastic nature of
dollars spent on the Plan.
A given Project i with Task k has a mean duration i ik , with a a mean cost S ik . The spread in dS has two
components arising from: (1) a stochastic duration around the mean duration, and (2) a stochastic spread
of mean dollars around a deterministic disbursement at a given time. Different finite-width asymmetric
distributions are used for durations and costs. For example, the distribution created for Adaptive
Simulated Annealing (ASA) [8], originally called Very Fast Simulated Re-annealing [7], is a finite-ranged
distribution with shape determined by a parameter “temperature” q. For each state (whether duration or
cost): (a) A random binary choice can be made to be higher or lower than the mean, using any ratio of
probabilities selected by the client. (b) Then, an ASA distribution is used on the chosen side. Each side
has a different q, each falling off from the mean. This is illustrated and further described in Fig. 1.
At the end of the tree at a time T (T also can be a parameter), there is a total cost at each node S(T ),
called a final “strike” in financial language. (A final strike might also appear at any node before T due to
cancellation of the Project using a particular kind of schedule alternative.) Working backwards, options
are calculated at time t0 . Greeks (functional derivatives of the option) assess sensitivity to various
variables, e.g., like those discussed in previous papers [12], but here we deliver precise numbers based on
as much real-world information available.
Lester Ingber
-3-
2
Real Options for Project Schedules (ROPS)
1/(2 * (abs(y) + q) * log(1 + 1/q))
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
ASA (q = 0.1)
0
-1
-0.5
0
0.5
1
Fig. 1. The ASA distribution can be used to develop finite-range asymmetric distributions
from which a value can be chosen for a given state of duration or cost. (a) A random binary
distribution is selected for a lower-than or higher-than mean, using any ratio of probabilities
selected by the client. Each side of the mean has its own temperature q. Here an ASA distribution is given for q = 0.1. The range can be scaled to any finite interval and the mean
placed within this range. (b) A uniform random distribution selects a value from [-1,1], and
a normalized ASA value is read off for the given state.
3. Data
The following data are used to develop Plan CPD. Each Task i has
(a) a Projected allocated cost, C i
(b) a Projected time schedule, T i
(c) a CPD with a statistical width of funds spent, SW S i
(d) a distribution with a statistical width of duration, SW T i
(e) a range of durations, RT i
(f) a range of costs, R S i
Expert guesses need to be provided for (c)-(f) for the prototype study.
A given Plan must be constructed among all Tasks, specified the ordering of Tasks, e.g., obeying any
sequential constraints among Tasks.
4. Three Recursive Shell
4.1. Outer Shell
There may be several parameters in the Project, e.g., as coefficients of variables in means and variances of
different CPD. These are optimized in an outer shell using ASA [8]. This end product, including
MULTI_MIN states returned by ASA, gives the client flexibility to apply during a full Project [12]. We
may wish to minimize Cost/T , or (CostOverrun - CostInitial)/T , etc.
Lester Ingber
-4-
Real Options for Project Schedules (ROPS)
4.2. Middle Shell
To obtain the Plan CPD, an middle shell of Monte Carlo (MC) states are generated from recursive
calculations. A Weibull or some other asymmetric finite distribution might be used for Task durations.
For a given state in the outer middle, a MC state has durations and mean cost disbursements defined for
each Task.
4.3. Inner Shell
At each time, for each Task, the differenced cost ( S ik (t + dt) − S ik (t))) is subjected to a inner shell
stochastic variation, e.g., some asymmetric finite distribution. The net costs dS ik (t) for each Project i and
Task k are added to define dS(t) for the Plan. The inner shell cost CPD is re-applied many times to get a
set of {dS} at each time.
5. Real Options
5.1. Plan Options
After the Outer MC sampling is completed, there are histograms generated of the Plan’s dS(t) and
dS(t)/S(t − dt) at each time t. The histograms are normalized at each time to give P(dS, t). At each time
t, the data representing P is “curve-fit” to the form of Eq. (0), where f and g are functions needed to get
good fits, e.g., fitting coefficients of parameters {x}
f = x f 0 + x f 1 S + x f 2 S 2 + ...
g = x g0 + x g1 S + x g2 S 2 + ...
At each time t, the functions f and g are fit to the function ln(( P(dS, t)), which includes the prefactor
containing g and the function L which may be viewed as a Padé approximate of these polynomials.
Complex constraints as functions of S ik (t) can be easily incorporated in this approach, e.g., due to regular
reviews by funding agencies or executives. These P’s are input into PATHTREE to calculate options for a
given strategy or Plan.
5.2. Risk Management of Project Options
If some measure of risk among Projects is desired, then during the MC calculations developed for the toplevel Plan, sets of differenced costs for each Project, dS i(t) and dS i(t)/S i(t − dt), stored from each of the
Project’s Tasks. Then, histograms and Project CPDs are developed, similar to the development of the
Plan CPD. A copula analysis, coded in TRD for risk management of financial markets, are applied to
develop a relative risk analysis among these projects [10]. In such an analysis, the Project marginal CPDs
are all transformed to Gaussian spaces, where it makes sense to calculate covariances and correlations.
An audit trail back the original Project spaces permits analysis of risk dependent on the tails of the Project
CPDs.
6. Generic Applications
ROPS can be applied to any complex scheduling of tasks similar to the FCS project. The need for
government agencies to plan and monitor such large projects is becoming increasingly difficult and
necessary [15]. Many large businesses have similar projects and similar requirements to manage their
complex projects.
Lester Ingber
-5-
Real Options for Project Schedules (ROPS)
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
M. Bowman and L. Ingber, ‘‘Real Options for US Army Future Combat Systems,’’ Report
2007:ROFCS, Lester Ingber Research, Ashland, OR, 2007.
G.G. Brown, R.T. Grose, and R.A. Koyak, ‘‘Estimating total program cost of a long-term, hightechnology, high-risk project with task durations and costs that may increase over time,’’ Military
Operations
Research
11,
41-62
(2006).
[URL
http://www.nps.navy.mil/orfacpag/resumePages/papers/Brownpa/Estimating_total_
program_cost.pdf]
T.E. Copeland and P.T. Keenan, ‘‘Making real options real,’’ McKinsey Quarterly 128-141 (1998).
[URL http://faculty.fuqua.duke.edu/˜charvey/Teaching/BA456_2006/McK98_3.pdf]
G. Glaros, ‘‘Real options for defense,’’ Transformation Trends June, 1-11 (2003). [URL
http://www.oft.osd.mil/library/library_files/trends_205_transformation_trends_9_june%202003_issue.pdf]
R. Grose, ‘‘Cost-constrained project scheduling with task durations and costs that may increase
over time: Demonstrated with the U.S. Army future combat systems,’’ Thesis, Naval Postgraduate
School, Monterey, CA, 2004. [URL http://www.stormingmedia.us/75/7594/A759424.html]
J.C. Hull, Options, Futures, and Other Derivatives, 4th Edition (Prentice Hall, Upper Saddle River,
NJ, 2000).
L. Ingber, ‘‘Very fast simulated re-annealing,’’ Mathl. Comput. Modelling 12, 967-973 (1989).
[URL http://www.ingber.com/asa89_vfsr.pdf]
L. Ingber, ‘‘Adaptive Simulated Annealing (ASA),’’ Global optimization C-code, Caltech Alumni
Association, Pasadena, CA, 1993. [URL http://www.ingber.com/#ASA-CODE]
L. Ingber, ‘‘Statistical mechanics of portfolios of options,’’ Report 2002:SMPO, Lester Ingber
Research, Chicago, IL, 2002. [URL http://www.ingber.com/markets02_portfolio.pdf]
L. Ingber, ‘‘Trading in Risk Dimensions (TRD),’’ Report 2005:TRD, Lester Ingber Research,
Ashland, OR, 2005.
L. Ingber, C. Chen, R.P. Mondescu, D. Muzzall, and M. Renedo, ‘‘Probability tree algorithm for
general diffusion processes,’’ Phys. Rev. E 64, 056702-056707 (2001).
[URL
http://www.ingber.com/path01_pathtree.pdf]
K.J. Leslie and M.P. Michaels, ‘‘The real power of real options,’’ McKinsey Quarterly 4-22 (1997).
[http://faculty.fuqua.duke.edu/˜charvey/Teaching/BA456_2006/McK97_3.pdf]
General Accounting Office, ‘‘Future Combat System Risks Underscore the Importance of
Oversight,’’ Report GAO-07-672T, GAO, Washington DC, 2007. [URL http://www.gao.gov/cgibin/getrpt?GAO-07-672T]
General Accounting Office, ‘‘Key Decisions to Be Made on Future Combat System,’’ Report
GAO-07-376,
GAO,
Washington
DC,
2007.
[URL
http://www.gao.gov/cgibin/getrpt?GAO-07-376]
B. Wysocki, Jr, ‘‘Is U.S. Government ’Outsourcing Its Brain’?,’’ Wall Street Journal March 30, 1
(2007).
| 5cs.CE
|
arXiv:0806.2735v2 [quant-ph] 21 Jul 2008
QPL/DCM 2008 Preliminary version
An overview of QML with a concrete
implementation in Haskell
Jonathan Grattage
Laboratoire d’Informatique de Grenoble, CNRS, France
[email protected]
Abstract
This paper gives an introduction to and overview of the functional quantum programming language QML.
The syntax of this language is defined and explained, along with a new QML definition of the quantum
teleport algorithm. The categorical operational semantics of QML is also briefly introduced, in the form
of annotated quantum circuits. This definition leads to a denotational semantics, given in terms of superoperators. Finally, an implementation in Haskell of the semantics for QML is presented as a compiler.
The compiler takes QML programs as input, which are parsed into a Haskell datatype. The output from
the compiler is either a quantum circuit (operational), an isometry (pure denotational) or a superoperator
(impure denotational). Orthogonality judgements and problems with coproducts in QML are also discussed.
Keywords: QML, language, functional, teleport, denotational semantics, operational semantics, Haskell.
1
Introduction and motivation
Language development for quantum computation is a rapidly developing research
area [4], motivated by the application of established formal reasoning and verification techniques within a quantum framework, understanding the behaviour of quantum computation, aiding the development of new algorithms and gaining a deeper
understanding of how they work. This paper discusses the syntax and features of,
and gives a compiler for, a language allowing both classical and quantum control:
QML [1,5]. The syntax and semantics for QML is a complete redevelopment of that
presented previously [1], as the language has been changed to remove a problematic
interpretation of coproducts (section 2.2); the interpretation of orthogonality has
also been updated (section 2.1). In addition, in this work the operational semantics
is made concrete by a compiler for QML programs implemented in Haskell.
QML has a syntax similar to other functional languages [4], and is based on
strict linear logic: i.e. linear logic with contraction, but without implicit weakening
(in contrast, Selinger’s QPL uses affine linear logic [8], which allows weakening but
1 The author gratefully acknowledges Thorsten Altenkirch in the development of QML and the operational
semantics. Ellie D’Hondt is also thanked for her constructive suggestions.
Grattage
not contraction). QML also integrates reversible and irreversible computations in
a single language, where weakenings (which can give rise to the collapse of superpositions and entanglement) must be explicit. Contractions are allowed and are
modelled as a form of sharing, analogous to the behaviour of classical functional
languages. Differences between QML and other languages include the use of a quantum control operation, provided by the quantum-if construct if ◦ . To use the if ◦
operation, the branches of the computation must be orthogonal (distinguishable),
the proof of which is supplied automatically by the type checker at compilation,
and hence the programmer need not supply the condition, nor need it appear in the
syntax of terms. Classical control is provided by a second classical-if construct if ,
which measures the term being branched over.
QML has both an operational and denotational semantics [5], supporting formal
reasoning principles with an algebra for equational reasoning and a normal form [2].
The operational semantics of QML is presented using a categorical formalisation of
the quantum circuit model, which is realised by a compiler that translates QML
programs into ‘typed’ quantum circuits. The denotational semantics is also implemented, which translates QML terms, via the operational semantics category FQC,
into either isometries Q◦ or superoperators Q.
Recent developments in QML include a fully operational compiler, with a type
inference algorithm for QML terms, and the automatic derivation and extension of
the orthogonality judgements and circuits required for quantum control, which have
all been implemented. The implementation of the orthogonality judgements is such
that it can be easily expanded as new rules for proving the orthogonality of terms are
added to QML. This short paper outlines the new syntax of QML and its semantics.
A new and faithful interpretation of the quantum teleportation algorithm in QML
is also presented as an example of using QML. Details for using the compiler, and
the output it generates, are then provided. Finally, the future development of the
language, the semantics and compiler, and the uses of the compiler are described.
2
The syntax of QML
This section introduces the syntax for QML (see also ref. [5]). The symbols σ, τ
are used to vary over QML types, given by σ = Q1 | Q2 | σ ⊗ τ , where the type
constructor is the tensor product ⊗ corresponding to a product type and Q2 is a
qubit type. x, y, z are used to vary over names. Typing contexts (Γ, ∆) are given
by Γ = • | Γ, x : σ where • is the empty context. Contexts correspond to functions
from a finite set of variables to types.
Constants κ, ι ∈ C are also used to define the syntax of expressions. Function
variables are used to refer to previously defined QML programs. The terms of QML
consist of those of a first-order functional language, extended with quantum data,
a quantum control structure, and a measurement operator. The vector notation y
is used for sequence variables to be measured (weakened). The grammar of QML
terms is defined thus:
(Variables)
x , y, ... ∈ Vars
(Prob amplitudes) κ, ι, ... ∈ C
(Patterns)
p, q
::= x | (x , y)
2
Grattage
(Terms)
::= x | x y | () | (t, u) | let p = t in u
| if t then u else u ′ | if ◦ t then u else u ′
| qfalsey | qtruey | κ × t | t + u
Quantum data is modelled using the constructs κ × t and t + u. The term κ × t,
associates the probability amplitude κ with the term t. The term t + u describes a
quantum superposition of t and u. Quantum superpositions are first class values,
and can be used in conditional expressions to provide quantum control. For example:
if ◦ (qtrue + qfalse) then t else u evaluates both t and u and combines the results
in a quantum superposition. Note that the term λ0 × t + λ1 × u, where t, u are not
qubits, is syntactic-sugar for if ◦ (λ0 × qtrue + λ1 × qfalse) then t else u. The
type-checker and orthogonality judgements ensure that this is a valid operation, by
providing a proof that t and u are orthogonal (distinguishable in some way), that
their types match, and that they are strict terms (they produce no garbage).
In a quantum control operation, the two branches must be orthogonal, otherwise the type system would accept terms that implicitly perform measurements.
Without this restriction “valid” programs could be written in QML that are not
physically realisable by a quantum computer. Orthogonality judgements are inferred automatically by static analysis of terms (see section 2.1).
As an example of superposition formation, the term ( √12 )×qfalse+( √12 )×qtrue
is an equal superposition of qfalse and qtrue. Normalisation factors that are equal
may be omitted.
Finally, a QML program is a sequence of function definitions, where a function
definition is given by f Γ = t : τ . A Haskell-style syntax is used to present program
examples. For example, the QML function below (left) is equivalent to the following
Haskell-like code (right):
t, u
f (x1 : σ1 , x2 : σ2 , . . , xn : σn )
=t :τ
2.1
f : σ1 ⊸ σ2 . . ⊸ σn ⊸ τ
f x1 x2 . . xn = t
QML orthogonality judgements
QML has a basic type system that tracks the use of variables, preventing them from
being weakened inappropriately. However, the type system still accepts terms which
implicitly perform measurements. As a consequence QML would accept programs
which are not realisable as quantum computations.
Consider the expression if ◦ x then qtrue else qtrue. This expression returns
qtrue without using any information about x . In order to maintain the invariant
that all measurements are explicit, the type system should reject the above expression. More precisely, the expression if ◦ x then t else u should only be accepted
if t ⊥ u. This notion intuitively ensures that the conditional operator does not
implicitly discard any information about x during the evaluation. The branches of
a superposition should also be orthogonal for similar reasons.
Mathematically, two terms, t, u, are orthogonal if their inner-product is equal to
zero, ht|ui = 0. If this is the case then the judgement t ⊥ u is true, but if the innerproduct yields any other value then t is not orthogonal to u. In the presentation of
an equational theory for QML [2] the orthogonality judgements are replaced by an
inner-product judgement on terms, to much the same effect. However, the inner3
Grattage
product approach is more informative and flexible, and gives a method of reasoning
about orthogonality, hence in future this method may be adopted for all terms.
The following rules give the current QML orthogonality judgements:
qtrue ⊥ qfalse
t ⊥u
(t, v ) ⊥ (u, w )
qfalse ⊥ qtrue
t ⊥u
⊥ pair0
(v , t) ⊥ (w , u)
λ∗0 κ0 = −λ∗1 κ1
t⊥u
λ0 × t + λ1 × u ⊥ κ0 × t + κ1 × u
⊥ pair1
⊥ sup
λ∗0 κ0 = −λ∗1 κ1
t ⊥u
◦
◦
if (λ0 × qtrue + λ1 × qfalse) then t else u ⊥ if (κ0 × qtrue + κ1 × qfalse) then t else u
t ⊥u
t ⊥ u′
t ⊥ if ◦ c then u else u ′
⊥
t ⊥u
if ◦0
t ⊥ u′
if ◦ c then u else u ′ ⊥ t
⊥ supif ◦
⊥ if ◦1
The first two axioms state that the basic states of qtrue and qfalse are orthogonal.
The third and fourth rules state that pairs of terms can be orthogonal, provided that
one component of a pair is orthogonal to the other pair’s corresponding component.
The two sup rules state when superpositions of terms can be orthogonal; the second
is a restatement of the first, translating superpositions using if ◦ . The final two rules
state that if ◦ statements can be orthogonal if all the component terms are.
The use of if ◦ in QML programs is valid only if the two branches are orthogonal.
Hence, for the Hadamard operation (section 4), it is required that −qtrue+qfalse ⊥
qtrue + qfalse, with the appropriate renormalisation. In this case the ⊥ sup rule
verifies orthogonality.
The rules for orthogonality given so far are incomplete, and may be extended.
Orthogonality judgements must also be interpreted by the operational semantics,
discussed in ref [5].
2.2
Removing coproducts from QML
In a previous version of QML [1], here referred to as QML⊕ , the language included
the notion of a tensorial coproduct, denoted by ⊕. This coproduct has now been
removed. The types of QML⊕ were generated by Q1 , σ ⊗ τ , and σ ⊕ τ . Qubits were
not primitive, but defined as Q2 = Q1 ⊕ Q1 . The coproduct allowed any finite type
to be directly represented in this way; not just limited to Q2 . The introduction
rules used for ⊕ were the usual coproduct rules of a left and a right injection:
Γ ⊢a s : σ
Γ ⊢a inl s : σ ⊕ τ
Γ ⊢a t : τ
+introl
+intror
Γ ⊢a inr t : σ ⊕ τ
The coproduct type was interpreted as σ ⊕ τ = Q2 ⊗ |σ ⊔ τ |, where |σ ⊔ τ | could
store a value of either |σ| or |τ |, by padding the smaller type. Using the coproduct
and injection rules, qfalse and qtrue were defined in QML⊕ as inl() : Q2 and
qfalse = inr() : Q2 , respectively, omitting the weakening property of QML⊕ .
Instead of if and if ◦ rules, QML⊕ implemented two ⊕-elimination rules: case,
providing classical-control (a generalisation of if ), and a quantum-control operation
4
Grattage
case◦ (generalising if ◦ ). The quantum (non-measuring) ⊕-elimination rule is similar
to the standard coproduct elimination rule, and is given as:
Γ ⊢a c : σ ⊕ τ
∆, x : σ ⊢◦ t : ρ
∆, y : τ ⊢◦ u : ρ
a
t⊥u
◦
Γ ⊗ ∆ ⊢ case c of {inl x ⇒ t | inr y ⇒ u} : ρ
⊕elim◦
The non-strict (measuring) case removes the orthogonality requirement, and
does not require sub-terms to be strict. The if a rules (a ∈ {◦, − }; if a = ◦ then
the rules are strict, i.e. measurement-free) would then be derived as:
if a b then t else u = casea b of {inl ⇒ t | inr ⇒ u }
The branches of a case◦ operation can be of different sizes, and this was dealt
with in the semantics of QML⊕ by padding the type of the smaller branch. The
padding of one type in this way can lead to the garbage becoming entangled with the
useful output in some way. This happens, for example, if branching over Q1 ⊗ Q2 .
The garbage created by padding may indirectly measure the qubit which is being
branched over. Consequently, this approach is not compositional, and has therefore
been rejected.
This version of QML resolves this issue by removing the coproduct. Qubits
are now primitive, as are if and if ◦ . Additionally, the strict if ◦ does not allow
any garbage to be produced. Coproducts may be reintroduced, possibly limited to
classical types.
3
An operational semantics for QML
The new operational semantics for QML is briefly discussed here, which is an updated version of that presented in [1]. This semantics is defined by giving a translation from QML terms into morphisms in the category FQC. FQC morphisms
consist of a reversible quantum circuit φ, the input context Γ, the output type σ,
and the size of the auxiliary heap h and any garbage g. Any heap qubits are initially
set to 0 (false) in the computational basis, and garbage qubits can be removed by
the partial trace operation at the end of the computation. A full development of
FQC is given in references [5,6].
As an example, the classical if construct is defined by the following typing rule
and operational semantic:
Γ ⊢ c : Q2
∆ ⊢ t, u : σ
c ∈ FQC Γ Q2
t ∈ FQC ∆ σ
u ∈ FQC ∆ σ
if
ifOp c t u ∈ FQC (Γ ⊗ ∆) σ
ifOp c t u = (hC + hc + ht|u , gc + 1, φ)
Γ ⊗ ∆ ⊢ if c then t else u : σ
where φ in the operational semantics is the following circuit:
Γ⊗∆
::
::
∆
Γ
φC
φc
Q2
33
33
>>
>>
φt |φu
33
33
σ
Q2
with φC as a “context-splitting” operation that copies any variables used by both
subterms. φt |φu denotes a conditional circuit which performs φt if the result of the
conditional circuit c is true (|1i) and φu if it is false (|0i).
5
Grattage
In this way, the semantics of QML is defined recursively over the syntax of terms,
and results in a valid FQC morphism for a valid QML term. Given a QML term,
the QML compiler follows the operational semantics and outputs an FQC morphism
represented as a typed circuit, which can then either be displayed, exported for use
with other programs, or used to directly implement the program on a quantumcircuit based quantum computer. The FQC morphism can be further evaluated
via the denotational semantics, producing a unitary matrix Q≃ (for strict-QML
terms without heap), an isometry Q◦ (for strict-QML terms), or a superoperator Q
(for QML terms that produce garbage). In addition, orthogonality judgements and
circuits for quantum control and superpositions of terms are automatically inferred
by the compiler at compile-time, so there can be no orthogonality errors during runtime. The relationships between QML without measurement (QML◦ ), full QML
(QML), typed quantum circuits (FQC), and isometries (Q◦ ) and superoperators
(Q), is shown in the following diagram:
QML◦ Γ σ
rr
rrr
r
r
xrrr
FQC◦ ΓLσ
LLL
LLL
J·Kd LL
L&
◦
J·K◦
Op
Q Γσ
b
.
/
/ QML Γ σ
JJ
JJJ·KOp
JJ
JJ
J$
/ FQC Γ σ
t
tt
tt
t
t J·Kd
ztt
/QΓσ
where J·KOp denotes the operational semantics, and J·KD denotes the additional
translation from quantum circuits into the denotational semantics. The full semantics can be found in reference [5].
4
Example: Teleportation in QML
The quantum teleportation describes how to transport a quantum state using a small
amount of classical communication. A QML interpretation of the teleportation
circuit with deferred measurement has previously been presented, along with a full
description of the algorithm [5]. However, this circuit relies on the existence of a
quantum channel. In order to demonstrate and explain the syntax of QML and the
compiler, a new, faithful, implementation of the quantum teleport algorithm, which
explicitly makes use of measurement, is presented. 2 This algorithm is similar to
that developed by Selinger and Valiron [9], and also in reference [3], which includes
a relevant discussion of teleportation, both with and without deferred measurement.
These examples show the elegance of allowing quantum control via the quantum if ◦
construct (CNot) and term superpositions (Epr ), in writing functions. For simple
one qubit functions (Had ), the branches of the quantum control are the columns
from the unitary matrices that describe the operation. A simple measurement
operation, using classical control, is also defined in the following example which
implements the teleportation algorithm (Tele):
2
Had, Qnot, Meas ∈ Q2 ⊸ Q2
Had b = if ◦ b then qfalse + − qtrue
else qfalse + qtrue
-- The Hadamard operation
Qnot b = if ◦ b then qfalse else qtrue
Meas b = if b then qtrue else qfalse
CNot
∈ Q2 ⊸ Q2 ⊸ Q2 ⊗ Q2
-- The Not operation (Pauli X)
-- A measurement operator, using if
-- A quantum cnot operation
The code is included with the compiler as teleport.qml, see section 5.
6
Grattage
CNot s t = if ◦ s then (qtrue, Qnot t)
else (qfalse, t)
Epr ∈ Q2 ⊗ Q2
-- The EPR pair, |00i + |11i
Epr = (qtrue, qtrue) + (qfalse, qfalse)
Bmeas
∈ Q2 ⊸ Q2 ⊸ Q2 ⊗ Q2
-- The Bell-measurement operation
Bmeas x y = let (x ′ , y ′ ) = CNot x y
in (Meas (Had x ′ ), Meas y ′ )
U
∈ Q2 ⊸ Q2 ⊗ Q2 ⊸ Q2
-- The unitary correction operations
U q xy = let (x , y) = xy in if x then (if y then U11 q else U10 q)
else
(if y then U01 q else q)
U01 , U10 , U11 ∈ Q2 ⊸ Q2
◦
U01 x = if x then qfalse else qtrue
U10 x = if ◦ x then − qtrue else qfalse
U11 x = if ◦ x then − qfalse else qtrue
-- The quantum teleportation algorithm
Tele
∈ Q2 ⊸ Q2
Tele q = let (a, b)
= Epr
-- a is given to Alice, b to Bob
f
= Bmeas q a -- Result of Alice’s Bell-measurement is classical data
in U b f
-- Bob applies U to his qubit, using classical data f
5
The QML compiler
This section briefly describes the design and operation of the Haskell QML compiler,
which can be found on the project website. 3 The objective is to take a file containing
a QML program, consisting of QML function definitions, and output an annotated
(typed) quantum circuit which realises the QML program as an FQC object. This
is achieved by implementing the operational semantics in Haskell. Additionally,
the circuit produced by the compiler can be further processed to produce either a
unitary matrix (Q≃ ), isometry (Q◦ ), or superoperator (Q), as appropriate, giving
an implementation of the denotational semantics of QML, as factored through the
operational semantics.
The compiler has a modular design, giving a clear logical structure. For example,
the QOrth module contains all the code for generating the orthogonality judgements
and circuits, while QCirc contains the definition of the circuit datatypes and associated functions. The compiler exploits advanced Haskell features, such as monads
and pattern matching, and making use of the ideas put forward by Vizzotto et
al [10]. The operational semantics is realised in the QComp (Q ML Compilation)
module.
To compile QML functions into the operational semantics (FQC morphisms,
represented as typed circuits), the “run typed circuit” command is used:
runTC "#filename" "#function". For example, to evaluate the typed circuit of
the function Epr (from section 4) the command is runTC "teleport.qml" "Epr".
This outputs the following typed circuit (as a Haskell datatype):
H
•
Q2
X
Q2
where there is no input context, the heap is two qubits (marked ⊢), and a pair
3
Instructions for downloading the QML compiler can be found on the QML compiler website
http://fop.cs.nott.ac.uk/qml/compiler. The Haskell compiler GHC is also required. To use the compiler, the QML system must be loaded into GHC via the command “ghci qml”. This will initialise the
system with all the modules required to interpret and evaluate QML programs.
7
Grattage
of qubits are the output. H and X are the Hadamard and Pauli-X operations.
As heap is initialised to qfalse, this circuit describes the function |00i → |00i +
|11i, producing an EPR pair. The compilation of a QML term into a circuit is
efficient; each QML term is recursively translated directly into an FQC morphism
(an annotated circuit), as described in section 3. No quantum computation is
simulated, so this does not effect the efficiency of the compiler - it is a simple
recursive translation into circuits. The output is further optimised after compilation,
by collapsing identities and permutations and other simple circuit manipulations
which do not effect the action of the circuit (see QCirc).
There are three main options for further evaluation of a QML term:
•
The QML term could be evaluated to a unitary matrix (Q≃ ), interpreting only the
reversible circuit φ from the FQC morphism. As this option classically computes
the full mathematical interpretation of the program it is inherently inefficient.
The output is actually a triple (h, g, φ) ∈ (Int, Int ,Unit), where h, g gives the size
of any heap required or garbage produced. The command for producing a unitary
matrix is runM ;
•
The function could be evaluated to an isometry (Q◦ ), which initialises any required heap, and is the full description for terms that produce no garbage. This
option is no less efficient than the runM option, and is the preferred evaluation
option. The output is actually a pair (g, φ) ∈ (Int , Q◦ ), where g gives the size of
any garbage (if the QML function is impure). The command for evaluating to an
isometry is runI ;
•
Thirdly, the QML term can be interpreted as a superoperator (Q), which initialises heap and traces out any garbage, using the command runS . This option
is substantially less efficient than the previous two options, as the state space is
doubled and the partial trace is a computationally expensive operation.
Together, the options runI and runS give an interpretation of the denotational
semantics of QML factored through the category FQC, as shown in the diagram
in section 3. A direct implementation of the denotational semantics, without using
the operational semantics, is an extension currently being developed. Please refer
to the project website for full details.
6
Conclusions and further work
This paper introduces the language QML and presents its semantics with a compiler
and example programs. The semantics and compiler give a realisable interpretation
of QML programs as quantum circuits in a formal, categorical, setting. The semantics can be extended in many ways, such as expanding the current orthogonality
judgements, or by the addition of non-linear, classical, data. The algebra for the
pure fragment of QML [2] is currently being extended to include measurement, following work on van Tonder’s quantum lambda calculus [3]. It would be instructive
to implement this algebra, especially the normal form, as part of the QML compiler.
Future possibilities for the development of the language also including developing a
notion of higher-order functions for QML, and adding iteration to the language.
The development of QML and the compiler is an ongoing project which has
8
Grattage
already reached a functional state. As the language and semantics evolve, extensions and new features can be incorporated into the compiler; which also provides
a useful testbed for the development of new language features and capabilities. For
example, an extension of the orthogonality circuits given in [5] was developed using
the compiler in this way. The compiler also facilitates the testing and development
of new QML algorithms, such as the described teleportation algorithm. It has also
been useful in allowing others to experiment with quantum programming and get
immediate feedback on the behaviour of their functions, in a style that is familiar to computer scientists, logicians, and physicists with functional programming
experience.
Further extensions to the compiler include adding the ability to export typed
circuits as images, or in notation compatible with tools such as MatLab and Mathematica. Possible relationships with the measurement calculus, the Haskell QIO
monad [7], and other formalisms are being studied, and may provide new insights.
This will lead to new features being developed, such as basis independence, and
further useful abstractions.
References
[1] Altenkirch, T. and J. Grattage, A functional quantum programming language, in: LICS 2005 proceedings
(2005), pp. 249–258, also arXiv:quant-ph/0409065.
URL http://fop.cs.nott.ac.uk/qml
[2] Altenkirch, T., J. Grattage, J. K. Vizzotto and A. Sabry, An algebra of pure quantum programming,
ENTCS 170 (2007), pp. 23–47, also arXiv:quant-ph/0506012v1.
URL http://fop.cs.nott.ac.uk/qml
[3] Dı́az-Caro, A., P. Arrighi, M. Gadella and J. Grattage, Measurements and confluence in quantum
lambda calculi with explicit qubits (2008), accepted by DCM/QPL (ICALP 2008).
URL http://equipes-lig.imag.fr/capp/qcg/people/jgrattage/papers/adc-lambdameas-qpl.pdf
[4] Gay, S. J., Quantum programming languages: Survey and bibliography, Mathematical Structures in
Computer Science 16 (2006).
URL http://www.dcs.gla.ac.uk/~ simon/publications/QPLsurvey.pdf
[5] Grattage, J., “QML: A functional quantum programming language,” Ph.D. thesis, School of Computer
Science and School of Mathematical Physics, The University of Nottingham (2006).
URL http://etheses.nottingham.ac.uk/archive/00000250/
[6] Green, A. and T. Altenkirch, From reversible to irreversible computations, in: P. Selinger, editor, QPL
2006 proceedings, ENTCS (2006).
URL http://fop.cs.nott.ac.uk/qml
[7] Green, A. and T. Altenkirch, Shor in Haskell: The quantum IO monad (2008), accepted by Trends in
Functional Programming 2008.
URL http://www.cs.nott.ac.uk/~ asg/pdfs/tfp08.pdf
[8] Selinger, P., Towards a Quantum Programming Language, Mathematical Structures in Computer
Science 14 (2004), pp. 527–586.
URL http://www.mathstat.dal.ca/~ selinger/papers.html#qpl
[9] Selinger, P. and B. Valiron, A lambda calculus for quantum computation with classical control, in:
TLCA 2005 proceedings, LNCS 3461 (2005).
URL http://www.mathstat.dal.ca/~ selinger/papers.html#qlambda
[10] Vizzotto, J. K., T. Altenkirch and A. Sabry, Structuring Quantum Effects: Superoperators as Arrows
(2005), also arXiv:quant-ph/0501151.
9
| 6cs.PL
|
arXiv:1211.0734v1 [math.AC] 4 Nov 2012
ON TEST MODULES FOR FLAT DIMENSION
JAVIER MAJADAS
Abstract. A ring with a test module of finite upper complete intersection
dimension is complete intersection.
1. Introduction
Let (A, m, k) be a noetherian local ring. We say that an A-module of finite type
M is a test module if for any local homomorphism of noetherian local rings A → B,
any B-module of finite type N such that T oriA (M, N ) = 0 for all i ≫ 0 has finite
flat dimension over A. A well known example of a test module is the residue field
k.
When A is regular, then any A-module of finite type is a test module. Over a
hypersurface, Huneke and Wiegand [7, Theorem 1.9] prove that test modules are
precisely the ones of infinite projective dimension (alternative proofs of this fact
were obtained also by Miller [9, Theorem 1.1] and Takahashi [11, Corollary 7.2]).
More generally, if A is a complete intersection, Celikbas, Dao and Takahashi [5,
Proposition 2.7] prove that test modules are those of maximal complexity. To be
precise, all these characterizations were obtained with a slightly different definition
of test modules, but we will show in Proposition 4 that they also hold in our case.
Some characterizations of these properties of rings in terms of test modules also
exist. First, the Auslander-Buchsbaum-Serre theorem can be stated in terms of
test modules as follows:
Theorem R. If (and only if) there exists a test module of finite projective dimension then A is regular.
A similar criterion for Gorensteinness was obtained in [5, Corollary 3.4] in terms
of the Gorenstein dimension of Auslander and Bridger [1] (related criteria also
appear in [12]):
Theorem G. Assume that A has a dualizing complex. If (and only if) there exists
a test module of finite Gorenstein dimension, then A is Gorenstein.
The same authors ask for a similar criterion for complete intersection rings [5,
Question 3.5] in terms of the complete intersection dimension of [4]:
Question CI. If (and only if) there exists a test module of finite complete intersection dimension, must A be a complete intersection?
Avramov [2] introduces the virtual projective dimension as follows (with a slight
modification when the residue field of A is finite). Let A → Â be the completion
homomorphism (for the topology of the maximal ideal). An A-module of finite type
Key words and phrases. Test modules, complete intersection, regular local ring, Gorenstein
ring, flat dimension.
2010 Mathematics Subject Classification. 13D05, 13D07, 13H05, 13H10.
1
2
JAVIER MAJADAS
M is said of finite virtual projective dimension if there exists a deformation Q → Â
(that is, a surjective homomorphism of noetherian local rings with kernel generated
by a regular sequence) such that the projective dimension pdQ (Â ⊗A M ) is finite.
A noetherian local ring A is complete intersection if and only if its residue field is
of finite virtual projective dimension if and only if any module of finite type is of
finite virtual projective dimension.
Subsequently a modification of this concept, the complete intersection dimension,
was introduced in [4]. The definition is similar, but instead of the completion
homomorphism A → Â, any flat local homomorphism of noetherian local rings
A → A′ is allowed. Complete intersection dimension shares many of the good
properties with virtual projective dimension. In particular, a noetherian local ring A
is complete intersection if and only if its residue field has finite complete intersection
dimension if and only if any module of finite type has finite complete intersection
dimension. Moreover it behaves well with respect to localization. An intermediate
definition, the upper complete intersection dimension, was given in [10]. It allows
only flat local homomorphism of noetherian local rings (A, m, k) → (A′ , m′ , k ′ ) with
regular closed fiber A′ ⊗A k.
The difficulty of the above Question CI, is that the property of being a test module does not ascend by local flat extensions. For instance, if (A, m, k) → (A′ , m′ , k ′ )
is a flat local homomorphism of noetherian local rings with A regular and A′ singular, and M is a test module for A, then A′ ⊗A M is never a test module for A′ ,
′
since T oriA (A′ ⊗A M, k ′ ) = T oriA (M, k ′ ) = 0 for all i ≫ 0. However, we circumvent
this difficulty by switching to upper complete intersection dimension, giving in this
context an affirmative answer to Question CI (Theorem 7). There are no surprises
in the proof. On the contrary, the crucial step is in the definition of test module.
Since the involved functor is T or, we modify a little the definition of test module
used in [5] allowing modules to test not only for projectivity (of finite type modules) but also for flatness (of modules which are finite over a local homomorphism).
Nevertheless, in Proposition 4 we prove that our definition is equivalent to the one
used in [5] at least over complete intersection rings.
In the last section, we see how this modification in the definition of test module,
allows us also to remove easily the supplementary hypothesis that the ring has a
dualizing complex in Theorem G.
2. Complete intersection
Definition 1. A local homomorphism of noetherian local rings (A, m, k) → (A′ , m′ , k ′ )
is weakly regular if it is flat and the closed fiber A′ ⊗A k is a regular local ring. Let
f : A → B be a local homomorphism of noetherian local rings. A Cohen factorization of f is a factorization A → A′ → B such that A → A′ is weakly regular and
A′ → B is surjective. If B is complete, a Cohen factorization of f always exists [3].
A surjective homomorphism of noetherian local rings with kernel generated by
a regular sequence is called a deformation.
Definition 2. Let (A, m, k) be a noetherian local ring. We say that an A-module of
finite type M is a test module (for A) if for any local homomorphism of noetherian
local rings A → B and any B-module of finite type N , T oriA (M, N ) = 0 for all
i ≫ 0 implies that the flat dimension of N over A, fdA (N ), is finite.
TEST MODULES FOR FLAT DIMENSION
3
Examples 3. (i) For any noetherian local ring (A, m, k), the residue field k is a
test module.
(ii) If A → B is a finite local homomorphism of local rings and B is regular, then
the change of rings spectral sequence
2
A
Epq
= T orpB (T orqA (B, N ), l) ⇒ T orp+q
(N, l)
(where l is the residue field of B) shows that B is a test module for A.
Note that our definition of test module is, a priori, more restrictive than the one
used in [5]: in that paper, an A-module of finite type M is called a Test module
if for any A-module of finite type N , T oriA (M, N ) = 0 for all i ≫ 0 implies that
fdA (N ) < ∞. Over a complete intersection ring, we will see that Test modules and
test modules are the same thing.
The complexity of an A-module of finite type M , cxA (M ), is defined as the least
integer c ≥ 0 such that there exists some real number α verifying dimk (T ornA (M, k)) ≤
αnc−1 for all n ≫ 0 [4, 5.1]. Over a complete intersection ring A, the possible complexities of A-modules of finite type are those values between 0 and codim(A) :=
edim(A) - dim(A) [4, 5.6, 5.7].
If A is a complete intersection ring, a finite type A-module M is a Test module
if and only if cxA (M ) = codim(A). The “if” part is a direct consequence of a
result of C. Miller [9, Proposition 2.2] and was also obtained in [8, Corollary 1.2]: if
cxA (M ) = codim(A), and N is a finite type A-module such that T oriA (M, N ) = 0
for all i ≫ 0, then T oriA (M, Ωt N ) = 0 for all i > 0 for some syzygy Ωt N of N .
By [9, loc.cit.], cxA (M ⊗A Ωt N ) = cxA (M ) + cxA (Ωt N ) which means, since the
complexity of M is maximal, that cxA (Ωt N ) = 0, that is, pdA (Ωt N ) < ∞ where
pd denotes projective dimension. Thus pdA (N ) < ∞.
The “only if” part was proved in [5, Proposition 2.7].
Proposition 4. Let A be a complete intersection ring, and M an A-module of
finite type. Then M is a test module if and only if it is a Test module.
Proof. Since test ⇒ Test ⇒ maximal complexity, we only have to prove that if
cxA (M ) = codim(A) then M is a test module. So let A → B be a local homomorphism of noetherian local rings and N a B-module of finite type such that
T oriA (M, N ) = 0 for all i ≫ 0. Let A → R → B̂ be a Cohen factorization where B̂
is the completion of B. By flat base change, T oriR (R ⊗A M, N̂ ) = T oriA (M, N̂ ) =
T oriA (M, N ) ⊗B B̂ = 0 for all i ≫ 0. Since A → R is weakly regular, codim(A)
= codim(R). Also, cxA (M ) = cxR (R ⊗A M ) by [4, Proposition 5.2]. That is,
cxR (R ⊗A M ) = codim(R), so as we have seen above, R ⊗A M is a Test module
for R, and then T oriR (R ⊗A M, N̂ ) = 0 for all i ≫ 0 implies pdR (N̂ ) < ∞. Since
A → R is flat, we have fdA (N̂ ) < ∞ and then fdA (N ) < ∞.
Lemma 5. Let (A, m, k) → (A′ , m′ , k ′ ) be a weakly regular homomorphism. If M
is a test module for A, then A′ ⊗A M is a test module for A′ .
Proof. Let A′ → B be a local homomorphism of noetherian local rings and N a
′
finite B-module such that T oriA (A′ ⊗A M, N ) = 0 for all i ≫ 0. Since A′ is A-flat,
we have T oriA (M, N ) = 0 for all i ≫ 0, and then by hypothesis fdA (N ) < ∞.
Consider the change of rings spectral sequence
′
′
′
2
A
Epq
= T orpA ⊗A k (T orqA (A′ ⊗A k, N ), k ′ ) ⇒ T orp+q
(N, k ′ ).
4
JAVIER MAJADAS
′
The ring A′ ⊗A k is regular and so T orpA ⊗A k (−, −) = 0 for all p ≫ 0. Also,
′
2
T orqA (A′ ⊗A k, N ) = T orqA (k, N ) = 0 for all q ≫ 0. That is Epq
= 0 for all
A′
′
p + q ≫ 0, and so T orn (N, k ) = 0 for all n ≫ 0. Therefore fdA′ (N ) < ∞.
Definition 6. [10] We say that a finite module M 6= 0 over a noetherian local
ring A has finite upper complete intersection dimension and denote it by CI*dimA (M ) < ∞ if there exist a weakly regular homomorphism A → A′ and a
deformation Q → A′ such that pdQ (A′ ⊗A M ) < ∞.
Theorem 7. Let M be a test A-module. If (and only if ) CI*-dimA (M ) < ∞ then
A is a complete intersection ring.
Proof. Consider a weakly regular homomorphism A → A′ and a deformation Q →
A′ such that pdQ (A′ ⊗A M ) < ∞. Since A′ ⊗A M is a test module for A′ by Lemma
5, then A′ is a Test module for Q [5, Corollary 2.6]. Since n := pdQ (A′ ) < ∞,
T oriQ (A′ , N ) = 0 for all i > n, and then pdQ (N ) ≤ n for any Q-module N of finite
type. Therefore Q is a regular local ring. Then A′ is a complete intersection and
by flat descent so is A.
3. Gorenstein
Definition 8. [1] Let A be a noetherian local ring and M an A-module of finite
type. We say that G-dimA (M ) = 0 if the following three conditions hold:
(i) The canonical homomorphism M → HomA (HomA (M ,A),A) is an isomorphism.
(ii) ExtiA (M, A) = 0 for all i > 0.
(iii) ExtiA (HomA (M, A), A) = 0 for all i > 0.
We say that G-dimA (M ) < ∞ if there exists an exact sequence
0 → Tn → ... → T0 → M → 0
where G-dimA (Ti ) = 0 for i = 0, ..., n.
Theorem 9. Let M be a test A-module. If (and only if ) G-dimA (M ) < ∞ then A
is a Gorenstein ring.
Proof. Using that for a homomorphism A → A′ and an A-module T , HomA′ (A′ ⊗A
T, A′ ) = HomA (T, A′ ), and that if T is of finite type and A′ is A-flat we have also
HomA (T, A′ ) = A′ ⊗A HomA (T, A), it is easy to prove that if  is the completion of
A, then G-dim (M̂ ) ≤ G-dimA (M ) < ∞ (a more general result can be seen in [6,
Corollary 5.11]).
By Lemma 5, M̂ is a test Â-module. So from [5, Corollary 3.4] we deduce that
 is Gorenstein, and then so is A.
References
[1] M. Auslander, M. Bridger. Stable module theory, Memoirs of the American Mathematical
Society, No. 94, 1969.
[2] L. L. Avramov. Modules of finite virtual projective dimension, Invent. Math. 96 (1989), no.
1, 71-101.
[3] L. L. Avramov, H.-B. Foxby, B. Herzog, Structure of local homomorphisms, J. Algebra 164
(1994), no. 1, 124-145.
TEST MODULES FOR FLAT DIMENSION
5
[4] L. L. Avramov, V. N. Gasharov, I. V. Peeva. Complete intersection dimension, Inst. Hautes
Études Sci. Publ. Math. No. 86 (1997), 67-114.
[5] O. Celikbas, H. Dao, R. Takahashi. Modules that detect finite homological dimensions,
arXiv:1207.5869v1.
[6] L. W. Christensen. Semi-dualizing complexes and their Auslander categories, Trans. Amer.
Math. Soc. 353 (2001), no. 5, 1839-1883.
[7] C. Huneke, R. Wiegand. Tensor products of modules, rigidity and local cohomology, Math.
Scand. 81 (1997), no. 2, 161-183.
[8] D. A. Jorgensen. Tor and torsion on a complete intersection, J. Algebra 195 (1997), no. 2,
526-537.
[9] C. Miller. Complexity of tensor products of modules and a theorem of Huneke-Wiegand, Proc.
Amer. Math. Soc. 126 (1998), no. 1, 53-60.
[10] R. Takahashi. Upper complete intersection dimension relative to a local homomorphism,
Tokyo J. Math. 27 (2004), no. 1, 209-219.
[11] R. Takahashi. Classifying thick subcategories of the stable category of Cohen-Macaulay modules, Adv. Math. 225 (2010), no. 4, 2076-2116.
[12] A. Tehranian, M. Tousi, S. Yassemi. Characterizing local rings via test modules, Comm.
Algebra 35 (2007), no. 8, 2524-2532.
Departamento de Álgebra, Facultad de Matemáticas, Universidad de Santiago de
Compostela, E15782 Santiago de Compostela, Spain
E-mail address: [email protected]
| 0math.AC
|
arXiv:1607.00062v1 [math.AG] 30 Jun 2016
LOCAL COHOMOLOGY AND BASE CHANGE
KAREN E. SMITH
f
Abstract. Let X −→ S be a morphism of Noetherian schemes, with S reduced.
For any closed subscheme Z of X finite over S, let j denote the open immersion
X \ Z ֒→ X. Then for any coherent sheaf F on X \ Z and any index r ≥ 1, the
sheaf f∗ (Rr j∗ F ) is generically free on S and commutes with base change. We prove
this by proving a related statement about local cohomology: Let R be Noetherian
algebra over a Noetherian domain A, and let I ⊂ R be an ideal such that R/I is
finitely generated as an A-module. Let M be a finitely generated R-module. Then
there exists a non-zero g ∈ A such that the local cohomology modules HIr (M )⊗A Ag
are free over Ag and for any ring map A → L factoring through Ag , we have
r
HIr (M ) ⊗A L ∼
(M ⊗A L) for all r.
= HI⊗
AL
1. Introduction
In his work on maps between local Picard groups, Kollár was led to investigate the
behavior of certain cohomological functors under base change [Kol]. The following
theorem directly answers a question he had posed:
f
Theorem 1.1. Let X → S be a morphism of Noetherian schemes, with S reduced.
Suppose that Z ⊂ X is closed subscheme finite over S, and let j denote the open
embedding of its complement U. Then for any coherent sheaf F on U, the sheaves
f∗ (Rr j∗ F ) are generically free and commute with base change for all r ≥ 1.
Our purpose in this note is to prove this general statement. Kollár himself had
proved a special case of this result in a more restricted setting [Kol, Thm 78].
We pause to say precisely what is meant by generically free and commutes with
base change. Suppose H is a functor which, for every morphism of schemes X → S
and every quasi-coherent sheaf F on X, produces a quasi-coherent sheaf H(F ) on S.
We say H(F ) is generically free if there exists a dense open set S 0 of S over which
Date: July 4, 2016.
Thanks to János Kollár for encouraging me to write this paper, for his insistence that I establish
the more general version of the theorem, and for sharing his insight as to why my arguments should
go through more generally, especially the idea to reduce to the affine case in Section 2. I’d also
like to thank Mel Hochster for listening to my arguments, Karl Schwede for reading a draft and
suggesting the reference [DGI], and Barghav Bhatt for his interest. Financial support partially from
NSF DMS-1501625.
1
2
KAREN E. SMITH
p
the OS -module H(F ) is free. If in addition, for every change of base T −→ S 0 , the
natural map
p∗ H(F ) → H(p∗X F )
of quasi-coherent sheaves on T is an isomorphism (where pX is the induced morphism
X ×S T → X), then we say that H(F ) is generically free and commutes with base
change. See [Kol, §72].
Remark 1.2. We do not claim the r = 0 case of Theorem 1.1; in fact, it is false.
For a counterexample, consider the ring homomorphism splitting Z ֒→ Z × Q ։ Z.
The corresponding morphism of Noetherian schemes
Z = Spec(Z) ֒→ X = Spec(Z × Q) → S = Spec Z
satisfies the hypothesis of Theorem 1.1. The open set U = X \ Z is the component
Spec Q of X. The coherent sheaf determined by the module Q on U is not generically
free over Z, since there is no open affine subset Spec Z[ n1 ] over which Q is a free
module. [In this case, the map j is affine, so the higher direct image sheaves Rp j∗ F
all vanish for p > 0.]
On the other hand, if f is a map of finite type, then the r = 0 case of Theorem 1.1
can be deduced from Grothendiecks’s Lemma on Generic freeness; see Lemma 4.1.
For the commutative algebraists, we record the following version of the main result,
which is essentially just the statement in the affine case:
Corollary 1.3. Let A be a reduced Noetherian ring. Let R be a Noetherian Aalgebra with ideal I ⊂ R such that the induced map A → R/I is finite. Then for any
Noetherian R module M, the local cohomology modules HIi (M) are generically free
and commute with base change over A for all i ≥ 0. Explicitly, this means that there
exists element g not in any minimal prime of A such that the modules HIi (M) ⊗A Ag
are free over Ag , and that for any algebra L over Ag , the natural map
HIi (M) ⊗A L → HIi (M ⊗A L)
is an isomorphism.
Some version of Theorem 1.1 may follow from known statements in the literature,
but looking through works of Grothendieck ([RD], [LC], [SGA2]) and [Conrad], I am
not able to find it; nor presumably could Kollár. After this paper was written, I did
find a related statement due to Hochster and Roberts [HR, Thm 3.4] in a special case,
not quite strong enough to directly answer Kollár’s original question; furthermore,
my proof is different and possibly of independent interest. In any case, there may be
value in the self-contained proof here, which uses a relative form of Matlis duality
proved here using only basic results about local cohomology well-known to most
commutative algebraists.
LOCAL COHOMOLOGY AND BASE CHANGE
3
2. Restatement of the Problem
In this section, we show that Theorem 1.1 reduces to the following special statement:
Theorem 2.1. Let A be a Noetherian domain. Let R = A[[x1 , . . . , xn ]] be a power
series ring over A, and M a finitely generated R-module. Denote by I the ideal
(x1 , . . . , xn ) ⊂ R. Then the local cohomology modules
HIi (M)
are generically free over A and commute with base change for all i.
For basic definitions and properties of local cohomology modules, we refer to [LC].
For the remainder of this section, we show how Theorem 1.1 and Corollary 1.3
follow from Theorem 2.1.
First, Theorem 1.1 is local on the base. Because the scheme S is reduced, it is the
finite union of its irreducible components, each of which is reduced and irreducible,
so it suffices to prove the result on each of them. Thus we can immediately reduce
to the case where S = Spec A, for some Noetherian domain A.
We now reduce to the case where X is affine as well. The coherent sheaf F on
the open set U extends to a coherent sheaf on X, which we also denote by F . To
simplify notation, let us denote the sheaf Rr j∗ F by H, which we observe vanishes
outside the closed set Z. Each section is annihilated by a power of the ideal IZ of
Z, so that although the sheaf of abelian groups H on Z is not an OZ -module, it
does have the structure of a module over the sheaf of OS -algebras limt OI Xt , which we
←− Z
bZ
\
denote OX,Z ; put differently, H can be viewed as sheaf on the formal scheme X
0 i
[
[
over S. Observe that i∗ O
X,Z |X 0 = OX,Z , where X ֒→ X is the inclusion of any
open set X 0 ⊂ X containing the generic points of all the components of Z. This
means that H is generically free as an OS -module if and only if H|X 0 is, and there
is no loss of generality in replacing X by any such open set X0 . In particular, we
can choose such X 0 to be affine, thus reducing the proof of Theorem 1.1 to the case
where both X and S are affine.
We can now assume that X → S is the affine map of affine schemes corresponding
to a ring homomorphism A → T . In this case the closed set Z is defined by some
ideal I of T . Because Z is finite over S = Spec A, the composition A → T → T /I is a
finite integral extension of A. The coherent sheaf F on U extends to a coherent sheaf
F on X, which corresponds to a finitely generated T -module M. Since X = Spec T
is affine, we have natural identifications for r ≥ 1
Rr j∗ F = H r (X \ Z, F ) = HIr+1 (M)
4
KAREN E. SMITH
of modules over T [LC, Cor 1.9]. Thus we have reduced Theorem 1.1 to showing that
if T is a Noetherian ring over a Noetherian domain A and I is any ideal such that
T /I is finitely generated as an A-module, then for any finitely generated T -module
M, the modules HIr+1(M) are generically free and commute with base change over
A for r ≥ 1. In fact, we will be able to show this for all indices r ≥ −1.
To get to the power series case, we first observe that for all i, every element of
HIi (M) is annihilated by some power of I. This means that HIi (M) has the structure
of a module over the I-adic completion T̂ I . There is no loss of generality in replacing
T and M by their I-adic completions T̂ I and M̂ I —the module HIi (M) is canonically
identified with HIi T̂ I (M̂ I ). So with out loss of generality, we assume that T is Iadically complete.
Now, Lemma 2.2 below guarantees that T is a module-finite algebra over a power
series ring A[[x1 , . . . , xn ]]. So the finitely generated T -module M is also a finitely
generated module over A[[x1 , . . . , xn ]], and the computation of local cohomology is
the same viewed over either ring. This means that to prove Theorem 1.1, it would
suffice to prove Theorem 2.1. It only remains to prove Lemma 2.2.
Lemma 2.2. Let A be a Noetherian ring. Let T be a Noetherian A-algebra containing an ideal I such that the composition of natural maps A → T ։ T /I is finite.
Then there is a natural ring homomorphism from a power series ring
A[[x1 , . . . , xn ]] → T̂ I := lim T /I t
←−
t
which is also finite.
Proof of Lemma. Fix generators y1 , . . . , yn for the ideal I of T . Consider the Aalgebra homomorphism
A[x1 , . . . , xn ] → T
xi 7→ yi .
We will show that this map induces a ring homomorphism
A[[x1 , . . . , xn ]] → T̂ I
which is finite. First note that for each natural number t, there is a naturally induced
ring homomorphism
(1)
A[x1 , . . . , xn ]
T
→ t
t
(x1 , . . . , xn )
I
sending the class xi to the class yi .
Claim: For every t, the map (1) is finite. Indeed, if t1 , . . . , td are elements of T
whose classes modulo I are A-module generators for T /I, then the classes of t1 , . . . , td
modulo I t are generators for T /I t as a module over A[x1 , . . . , xn ]/(x1 , . . . , xn )t .
LOCAL COHOMOLOGY AND BASE CHANGE
5
The claim is straightforward to prove using induction on t and the exact sequence
0 → I t /I t+1 → T /I t+1 → T /I t → 0.
We leave these details to the reader.
Now to prove the lemma, we take the direct limit of the maps (1). Since at every
stage, the target is generated over the source by the classes of t1 , . . . , td , also in the
limit, T̂ I will be generated over A[[x1 , . . . , xn ]] by the images of t1 , . . . , td . So the
induced ring homomorphism A[[x1 , . . . , xn ]] → T is finite.
Having reduced the proof of the main results discussed in the introduction to
Theorem 2.1, the rest of the paper focuses on the local cohomology statement in the
special case. Our proof of Theorem 2.1 uses an A-relative version of Matlis duality
to convert the problem to an analagous one for finitely generated modules over a
power series ring, where it will follow from the theorem on generic freeness. This
relative version of Matlis duality might be of interest to commutative algebraists in
other contexts, and holds in greater generality than what we develop here. To keep
the paper as straightforward and readable as possible, we have chosen to present it
only in the case we need to prove the main result. Some related duality is worked
out in [DGI].
3. A Relative Matlis Dual Functor
3.1. Matlis Duality. We first recall the classical Matlis duality in the complete
local (Noetherian) case.
Let (R, m) be a complete local ring, and let E be an injective hull of its residue
field R/m. The Matlis dual functor HomR (−, E) is an exact contravariant functor on
the category of R-modules. It takes each Noetherian R-module (i.e., one satisfying
the ascending chain condition) to an Artinian R-module (i.e., one satisfying the
descending chain condition) and vice-versa. Moreover, for any Artinian or Noetherian
R-module H, we have a natural isomorphism H → HomR (HomR (H, E), E). That is,
the Matlis dual functor defines an equivalence of categories between the category of
Noetherian and the category of Artinian R-modules. See [LC], [BH, Thm 3.2.13] or
[Hoch] for more on Matlis duality.
3.2. A relative version of Matlis duality. Let A be a domain. Let R be an Aalgebra, with ideal I ⊂ R such that R/I is finitely generated as an A-module. Define
the relative Matlis dual functor to be the functor
{R − modules} → {R − modules}
M 7→ M ∨A := lim HomA (M/I t M, A).
−→
t
6
KAREN E. SMITH
We also denote M ∨A by Homcts
A (M, A), since it is the submodule of HomA (M, A)
consisting of maps continuous in the I-adic topology. That is, Homcts
A (M, A) is the
R-submodule of HomA (M, A) consisting of maps φ : M → A satisfying φ(I t M) = 0
for some t.
Proposition 3.1. Let R be a Noetherian algebra over a Noetherian domain A, with
ideal I ⊂ R such that R/I is finitely generated as an A-module.
(1) The functor Homcts
A (−, A) is naturally equivalent to the functor
M 7→ HomR (M, Homcts
A (R, A)).
(2) The functor preserves exactness of sequences
0 → M1 → M2 → M3 → 0
of finitely generated R-modules, provided that the modules M3 /I n M3 are (locally) free A-modules for all n ≫ 0.
Remark 3.2. If A = R/I is a field, then the relative Matlis dual specializes to the
usual Matlis dual functor HomR (−, E), where E is the injective hull of the residue
field of R at the maximal ideal I (denoted here now m). Indeed, one easily checks
that Homcts
A (R, A) is an injective hull of R/m. To wit, the R-module homomorphism
(
R→A
R/m → Homcts
sending
r mod m 7→
A (R, A)
s 7→ rs mod m
is a maximal essential extension of R/m.
Proof of Proposition. Statement (1) follows from the adjointness of tensor and Hom,
which is easily observed to restrict to the corresponding statement for modules of
continuous maps.
For (2), we need to show the sequence remains exact after applying the relative
Matlis dual functor. The functor HomA (−, A) preserves left exactness: that is,
(2)
0 → HomA (M3 , A) → HomA (M2 , A) → HomA (M1 , A)
is exact. We want to show that, restricting to the submodules of continuous maps,
we also have exactness at the right. That is, we need the exactness of
(3)
cts
cts
0 → Homcts
A (M3 , A) → HomA (M2 , A) → HomA (M1 , A) → 0.
The exactness of (3) at all spots except the right is easy to verify using the description
of a continuous map as one annihilated by a power of I.
To check exactness of (3) at the right, we use the Artin Rees Lemma [AM, 10.10].
n
Take φ ∈ Homcts
A (M1 , A). By definition of continuous, we φ is annihilated by I for
sufficiently large n. By the Artin-Rees Lemma, there exists t such that for all n ≥ t,
we have I n+t M2 ∩ M1 ⊂ I n M1 . This means we have a surjection
M1 /(I n+t M2 ∩ M1 ) ։ M1 /I n M1 .
LOCAL COHOMOLOGY AND BASE CHANGE
7
Therefore the composition
M1 /I n+t M2 ∩ M1 ։ M1 /I n M1 → A
gives a lifting of φ to an element φ′ in HomA (M1 /I n+t M2 ∩ M1 , A).
Now note that for n ≫ 0, we have exact sequences
0 → M1 /M1 ∩ I n+t M2 → M2 /I n+t M2 → M3 /I n+t M3 → 0,
which are split over A by our assumption that M3 /I n+t M3 is projective. Thus
(4) 0 → HomA (M3 /I n+t M3 , A) → HomA (M2 /I n+t M2 , A) → HomA (M1 /M1 ∩ I n+t M2 , A) → 0
is also split exact. This means we can pull φ′ ∈ HomA (M1 /I n+t M2 ∩ M1 , A) back to
φ
some element φ̃ in HomA (M2 /I n+t M2 , A). So our original continuous map M1 → A
φ̃
is the restriction of some map M2 → A which satisfies φ̃(I n+t M2 ) = 0. This exactly
says the continuous map φ on M1 extends to a continuous map φ̃ of M2 . That is,
the sequence (3) is exact.
Remark 3.3. If M3 is a Noetherian module over a Noetherian algebra R over the
Noetherian domain A, then the assumption that M3 /I n M3 is locally free for all n
holds generically on A—that is, after inverting a single element of A. See Lemma
4.2.
4. Generic Freeness
We briefly review Grothendieck’s idea of generic freeness, and use it to prove that
the relative Matlis dual of a Noetherian R-module is generically free over A (under
suitable Noetherian hypothesis on R and A).
Let M be a module over a commutative domain A. We say that M is generically
free over A if there exists a non-zero g ∈ A, such that M ⊗A Ag is a free Ag -module,
where Ag denotes the localization of A at the element g. Likewise, a collection M of
A-modules is simultaneously generically free over A if there exists a non-zero g ∈ A,
such that M ⊗A Ag is a free for all modules M ∈ M. Note that any finite collection
of generically free modules is always simultaneously generically free, since we can
take g to be the product the gi that work for each of the Mi .
Of course, finitely generated A-modules are always generically free. Grothendieck’s
famous Lemma on Generic Freeness ensures that many other modules are as well:
Lemma 4.1. [EGA, 6.9.2] Let A be a Noetherian domain. Let M be any finitely
generated module over a finitely generated A-algebra T . Then M is generically free
over A.
We need a version of Generic Freeness for certain infinite families of modules over
more general A-algebras:
8
KAREN E. SMITH
Lemma 4.2. Let A be any domain. Let T be any Noetherian A-algebra, and I ⊂ T
any ideal such that T /I is finite over A. Then for any Noetherian T -module M, the
family of modules
{M/I n M | n ≥ 1}
is simultaneously generically free over A. That is, after inverting a single element of
A, the modules M/I n M for all n ≥ 1 become free over A.
Remark 4.3. We will make use of Lemma 4.2 in the case where T = A[[x1 , . . . , xn ]].
Proof. If M is finitely generated over T , then the associated graded module
grI M = M/IM ⊕ IM/I 2 M ⊕ I 2 M/I 3 M ⊕ . . .
is finitely generated over the associated graded ring grI T = T /I ⊕ I/I 2 ⊕ I 2 /I 3 ⊕ . . . ,
which is a homomorphic image of a polynomial ring over T /I. Hence grI T is a finitely
generated A-algebra. Applying the Lemma of generic freeness to the grI T -module
grI M, we see that after inverting a single non-zero element of g of A, the module
grI M becomes A-free. Since grI M is graded over grI T and A acts in degree zero,
clearly its graded pieces are also free after tensoring with Ag . We can thus replace A
by Ag for suitable g, and assume that the I n M/I n+1 M are Ag -free for all n ≥ 0.
Now consider the short exact sequences
(5)
0 → I n M/I n+1 M → M/I n+1 M → M/I n M → 0,
for each n ≥ 1. We already know that M/I 1 M and I n M/I n+1 M for all n ≥ 1 are
free (over Ag ), so by induction, we conclude that the sequences (5) are all split over
Ag for every n. In particular, the modules M/I n M are also free over Ag for all n ≥ 1.
The lemma is proved.
Proposition 4.4. Let A be a Noetherian domain. Let R be any Noetherian Aalgebra with ideal I ⊂ R such that R/I is a finitely generated A-module. Then for
any Noetherian R-module M, the relative Matlis dual
Homcts
A (M, A)
is a generically free A-module. Furthermore, if g ∈ A is a non-zero element such
that Ag ⊗A Homcts
A (M, A) is free over Ag , then for any base change A → L factoring
through Ag , the natural map
cts
Homcts
A (M, A) ⊗A L → HomL (M ⊗A L, L)
is an isomorphism of R ⊗A L-modules, functorial in M.
Proof. We can invert one element of A so that each M/I t M is free over A; replace
A by this localization. We now claim that the A-module
M
cts
HomA (M, A) = lim HomA
,A
−→
I tM
t
LOCAL COHOMOLOGY AND BASE CHANGE
9
t
is free. Indeed,
since each M/I M is free and has finite rank over A, its A-dual
M
HomA I t M , A is also free of finite rank. The direct limit is also A-free because the
maps in the limit system are all split over A and injective. Indeed, if some finite
A-linear combination of fi ∈ Homcts
combination is
A (M, A) is zero, then that same
,
A
of
homomorphisms
zero considered as elements of the free-submodule HomA IM
tM
in Homcts
(M,
A)
killed
by
a
large
power
of
I.
It
follows
that
Homcts
A
A (M, A) is free
over A, as desired.
The result on base change follows as well, since tensor commutes with direct limits
and with dualizing a finitely generated free module.
Remark 4.5. We can interpret Proposition 4.4 as saying that generically on A,
the relative Matlis dual functor (applied to Noetherian R-modules) is exact and
commutes with base change.
5. Relative Local Duality and the Proof of the Main Theorem
The proof Theorem 2.1 and therefore of our main result answering Kollár’s question, follows from a relative version of Local Duality:
Theorem 5.1. Let R be a power series ring A[[x1 , . . . , xn ]] over a Noetherian domain A, and let M be a finitely generated R-module. Then, after replacing A by its
localization at one element, there is a functorial isomorphism for all i
H i (M) ∼
= [Extn−i (M, R)]∨A .
I
R
To prove Theorem 5.1, we need the following lemma.
Lemma 5.2. Let R be a power series ring A[[x1 , . . . , xn ]] over a ring A. There is a
∼ n
natural R-module isomorphism Homcts
A (R, A) = HI (R), where I = (x1 , . . . , xn ). In
particular, the relative Matlis dual functor can also be expressed
M 7→ HomR (M, HIn (R)).
Proof. We recall that HIi (R) is the i-th cohomology of the extended Čech complex
K • on the elements x1 , . . . , xn . This is the complex
M
δ1
δn
0→R→
Rxi xj → · · · −→
Rx1 ⊕ Rx2 · · · ⊕ Rxn →
Rx1 x2 ···xn → 0
i<j
where the maps are (sums of) suitably signed localization maps. In particular, HIn (R)
is the cokernel of δn , which can be checked to be the free A-module on (the classes
of) the monomials xa11 . . . xann with all ai < 0. 1
L
page 226 of [Hart], although I have included one extra map δ1 : R →
Rxi sending
f 7→ ( f1 , . . . , f1 ) in order to make the complex exact on the left, and my ring is a power series ring
over A instead of a polynomial rings over a field. This is also discussed in [LC] page 22.
1See
10
KAREN E. SMITH
Now define an explicit R-module isomorphism Φ from HIn (R) to Homcts
A (R, A) by
sending the (class of the) monomial xα to the map φα ∈ Homcts
(R,
A):
A
φα
R −→ A
(
xb11 +a1 +1 . . . xnbn +an +1
b1
bn
x1 . . . xn 7→
0 otherwise
mod I
if αi + βi ≥ −1 for all i
Since {xβ | β ∈ Nn } is an A-basis for R, the map φα is a well-defined A-module map
from R to A, and since it sends all but finitely many xβ to zero, it is I-adically continΦ
uous. Thus the map HIn (R) −→ Homcts
A (R, A) is is an A-module homomorphism; in
fact, Φ is an A-module isomorphism, since it defines a bijection between the A-basis
{xα | αi < 0} for HIn (R) and {πβ | βi ≥ 0} for Homcts
A (R, A) (meaning the dual basis
β
on the free basis x for R over A) matching the indices αi ↔ βi = −αi − 1. It is easy
to check that Φ is also R-linear, by thinking of it as “premultiplication by xα+1 ” on
the cokernel of δn . Thus Φ identifies the R-modules HIn (R) and Homcts
A (R, A).
Proof of Theorem 5.1. We proceed by proving that both modules are generically ison
morphic to a third, namely TorR
n−i (M, HI (R)).
First, recall how to compute HIi (M). Let K • be the extended Čech complex on
the elements x1 , . . . , xn
M
δ1
δn
0→R→
Rxi xj → · · · −→
Rx1 ⊕ Rx2 · · · ⊕ Rxn →
Rx1 x2 ···xn → 0.
i<j
This is a complex of flat R-modules, all free over A, exact at every spot except the
right end. Thus it is a flat R-module resolution of the local cohomology module
HIn (R). The local cohomology module HIi (M) is the cohomology of this complex
after tensoring over R with M, that is
n
HIi (M) = TorR
n−i (M, HI (R)).
n−i
On the other hand, let us compute the relative Matlis dual of ExtR
(M, R). Let P•
•
be a free resolution of M over R. The module ExtR (M, R) is the cohomology of the
complex HomR (P• , R). We would like to say that the computation of the cohomology
of this complex commutes with the relative Matlis dual functor, but the best we can
say is that this is true generically on A. To see this, we will apply Lemma 4.2 to the
following finite set of R-modules:
• For i = 0, . . . , n, the image Di of the i-th map of the complex HomR (P• , R);
n−i
• For i = 0, . . . , n, the cohomology ExtR
(M, R) of the same complex.
Lemma 4.2 guarantees that the modules
Di /I t Di
and
n−i
n−i
ExtR
(M, R)/I t ExtR
(M, R)
LOCAL COHOMOLOGY AND BASE CHANGE
11
are all simultaneously generically free over A for all t ≥ 1. This allows us to break up
the complex Ag ⊗A HomR (P• , R) into many short exact sequences, split over Ag , which
satisfy the hypothesis of Proposition 3.1(2) (using Ag in place of A and Ag ⊗A R in
place of R). It follows that the computation of cohomology of HomR (P• , R) commutes
with the relative Matlis dual functor (generically on A).
n−i
Thus, after replacing A by a localization at one element, ExtR
(M, R)]∨A is the
cohomology of the complex
HomR (HomR (P• , R), HIn (R)).
In general, for any finitely generated projective module P and any module H (over
any Noetherian ring R), the natural map
P ⊗ H → Hom(Hom(P, R), H)
sending a simple tensor x ⊗ h to the map which sends f ∈ Hom(P, R) to f (x)h, is
an isomorphism, functorially in P and H. Thus we have a natural isomorphism of
complexes
P• ⊗ HIn (R) ∼
= HomR (HomR (P• , R), HIn (R)),
and so [Extn−i (M, R)]∨A is identified with Torn−i (M, HIn (R)), which as we saw is
identified with HIi (M).
Since all identifications are functorial, we have proven the relative local duality
∼
= [Extn−i (M, R)]∨A , generically on A.
HIi (M)
We can finally finish the proof of Theorem 1.1, and hence the main result:
Proof of Theorem 2.1. Let R be a power series ring over a Noetherian domain A,
and let M be any Noetherian R-module. We need to show that the local cohomology
modules HIi (M) are generically free and commute with base change over A.
In light of Corollary 4.4 , we can accomplish this by showing that HIi (M) is the relative Matlis dual of a Noetherian R-module, generically on A. But this is guaranteed
by the relative local duality theorem Theorem 5.1, which guarantees that
H i (M) ∼
= Extn−i (M, R)∨A
I
R
generically on A.
Remark 5.3. One could obviously develop the theory of relative Matlis duality,
especially Theorem 5.1, further; I wrote down only the simplest possible case and
the simplest possible statements needed to answer Kollár’s question as directly as
possible.
12
KAREN E. SMITH
References
[AM] M. F. Atiyah and I. G. MacDonald, Introduction to Commutative Algebra, Addison Wesley
Series in Mathematics, Addison Wesley, London, (1969).
[BH] Winfred Burns and Jürgen Herzog, Cohen-Macaulay Rings, Cambridge series in Advanced
Mathematics, v 39. Cambridge University Press, (1993).
[Conrad] Brian Conrad, Grothendieck Duality and Base Change, Lecture Notes in Mathematics
1750, Springer (2001).
[DGI] W.G. Dwyer, J.P.C. Greenlees and S. Iyengar, Duality in algebra and topology, Advances in
Mathematics Volume 200, Issue 2, 1 March 2006, Pages 357D402.
[Eisen] David Eisenbud, Commutative Algebra with a view towards Algebraic Geometry, Graduate
Texts in Mathematics 150, Springer (1995).
[EGA] Alexander Grothendieck and Jean Dieudonné, Éléments de Geométrie Algébrique Chapter
IV, part 2, Inst. Hautes Études Sci. Pub. Math. 24 (1965).
[SGA2] Alexander Grothendieck and Michele Raynaud, Cohomologie locale des faisceaux cohérents
et théorèmes de Lefschetz locaux et globaux (SGA 2), Documents Mathèmatiques (Paris) 4,
Paris: Sociéité Mathèmatique de France, (2005) [1968], Laszlo, Yves, ed., arXiv:math/0511279,
ISBN 978-2-85629-169-6, MR 2171939
[LC] Robin Hartshorne, Local cohomology. A seminar given by A. Grothendieck, Harvard University, Fall, 1961, Lecture notes in mathematics 41, Berlin, New York: Springer-Verlag, (1967).
MR0224620
[RD] Robin Hartshorne, Residues and Duality: Lecture Notes of a Seminar on the Work of A.
Grothendieck, Given at Harvard 1963 /64, Springer Verlag Lecture Notes in Mathematics, vol
20 (1966).
[Hart] Robin Hartshorne, Algebraic Geometry Graduate Texts in Mathematics 52 Springer-Verlag,
(2006).
[Hoch] Mel Hochster, Lecture notes on Local Cohomology, unpublished, from his University of
Michigan website http://www.math.lsa.umich.edu/ hochster/615W11/loc.pdf
[HR] Mel Hochster and Joel Roberts, The Purity of Frobenius and Local Cohomology, Advances
in Mathematics 21 117–172 (1976).
[Kol] János Kollár, Maps between local Picard groups, arXiv:1407.5108v2 (preprint) 2014.
| 0math.AC
|
TOPOLOGICAL RIGIDITY FAILS FOR QUOTIENTS OF THE DAVIS
COMPLEX
arXiv:1610.08699v2 [math.GT] 13 Sep 2017
EMILY STARK
Abstract. A Coxeter group acts properly and cocompactly by isometries on the Davis complex
for the group; we call the quotient of the Davis complex under this action the Davis orbicomplex
for the group. We prove the set of finite covers of the Davis orbicomplexes for the set of oneended Coxeter groups is not topologically rigid. We exhibit a quotient of a Davis complex by a
one-ended right-angled Coxeter group which has two finite covers that are homotopy equivalent
but not homeomorphic. We discuss consequences for the abstract commensurability classification
of Coxeter groups.
1. Introduction
The notion of topological rigidity has its roots in the setting of manifolds. A closed manifold
M is called topologically rigid if every homotopy equivalence from M to another closed manifold
is homotopic to a homeomorphism. A well-known example of this phenomenon is the Poincaré
Conjecture, which states that the 3-sphere is topologically rigid and was proven by Perelman.
Many lens spaces are examples of 3-manifolds that are not topologically rigid. The Borel conjecture
states that closed aspherical manifolds are topologically rigid. The conjecture was proven for
manifolds of dimension at least five whose fundamental group is either Gromov hyperbolic or
CAT(0) by Bartels and Lück [BL12], building on the techniques of Farrell and Jones [FJ89].
The definition of topological rigidity extends from manifolds to orbifolds and to classes of
topological spaces. Background on orbifolds and orbifold homeomorphisms is given by Kapovich
[Kap09, Chapter 6] and Ratcliffe [Rat06, Chapter 13]. An orbicomplex is the union of orbifolds
identified along homeomorphic suborbifolds, and the notion of homeomorphism extends to these
spaces as well.
Definition 1.1. Let X be a class of topological spaces, orbifolds, or orbicomplexes. The class X
is said to be topologically rigid if for all X1 , X2 ∈ X , if π1 (X1 ) ∼
= π1 (X2 ), then X1 and X2 are
homeomorphic.
Simplicial graphs provide a simple example of a class of spaces that is not topologically rigid.
More generally, for graphs of spaces with one-ended universal covers the presence of topological
rigidity is more subtle. Lafont proved that simple, thick, n-dimensional hyperbolic piecewisemanifolds are topologically rigid for n ≥ 2 [Laf04][Laf06][Laf07]. In dimension two, these spaces
decompose as graphs of spaces with vertex spaces that are compact hyperbolic surfaces with
boundary, edge spaces that are circles, and edge-to-vertex space inclusions that identify the
Date: March 11, 2018.
2010 Mathematics Subject Classification. Primary 51F15; Secondary 20E07; 20F55; 20F67.
1
boundary components of the surfaces so that each boundary component is identified to at least
two others; the higher-dimensional analogues are similar. The orbicomplexes considered in this
paper also have hyperbolic fundamental groups and codimension-1 singularities along embedded
locally geodesic 1-complexes; we show that finite covers of these spaces do not exhibit topological
rigidity.
The spaces studied in this paper have fundamental groups of the following form. If Γ is a finite
simplicial graph with vertex set V Γ and edge set EΓ, the right-angled Coxeter group WΓ with
defining graph Γ has generating set V Γ and relations s2 = 1 for all s ∈ V Γ and st = ts whenever
{s, t} ∈ EΓ. If WΓ is a right-angled Coxeter group, then WΓ acts properly and cocompactly by
isometries on its Davis complex ΣΓ . The quotient of this space under the action of the right-angled
Coxeter group WΓ is called the Davis orbicomplex DΓ := WΓ \\ΣΓ . Background on Coxeter groups
and the Davis complex is given by Davis [Dav08]. A description of reflection orbicomplexes related
to those described here can be found in [Sta17, Section 5.2] and [DST16, Section 3].
A natural setting for questions of topological rigidity for spaces with fundamental groups rightangled Coxeter groups and their finite-index subgroups is the set of Davis orbicomplexes and their
finite-sheeted covers, as these spaces have a natural orbicomplex structure. If the graph Γ has no
edges, so that WΓ is a free product of groups isomorphic to Z/2Z, then the set of Davis orbicomplexes and their finite-sheeted covers is not topologically rigid: such a group is the fundamental
group of an orbicomplex which is finitely covered by a finite simplicial graph. So, one may ask
if topological rigidity holds if one restricts to one-ended right-angled Coxeter groups. The main
result of this paper is the following.
Theorem 1.2. Let X be the set of finite covers of the Davis orbicomplexes for the set of one-ended
right-angled Coxeter groups. The set X is not topologically rigid.
To prove Theorem 1.2, we construct an example of a one-ended right-angled Coxeter group WΓ
so that if D is its Davis orbicomplex, then there exist two finite covers of D that have the same
fundamental group but are not homeomorphic. The orbicomplex D contains a singular subspace
that is finitely covered by a graph; we find two non-homeomorphic covers of the graph that extend
to covers of D so that the covers have the same fundamental group. The construction of WΓ and
D is given in Section 2, and the finite covers are described in Section 3.
In Proposition 3.2, we employ similar ideas to produce further finite covers of D that are
quotients of the Davis complex by isomorphic torsion-free subgroups of WΓ and so that these
covers are not homeomorphic.
As shown by Crisp–Paoluzzi [CP08] using the work of Lafont [Laf07], there are one-ended rightangled Coxeter groups which are not virtually manifold groups for which the Davis orbicomplex
together with its finite-sheeted covers is topologically rigid. So, we state the following question.
Question 1.3. For which set W of Coxeter groups is the set of Davis orbicomplexes for groups
in W together with their finite-sheeted covers topologically rigid?
Relatedly, Xie [Xie06] proved the set of quotients of Fuchsian buildings by the action of a
cocompact lattice is topologically rigid. An interesting problem is to determine the set of lattices
2
in the isometry group of a hyperbolic building which have a set of quotients that is topologically
rigid.
Definition 1.4. A class of spaces X is said to be closed under finite covers if whenever X ∈ X ,
all finite-sheeted covering spaces of X are in X .
Topological rigidity of X , a class of spaces closed under finite covers, has applications in the
study of the abstract commensurability classes of the fundamental groups of spaces in X . Recall
that two groups are abstractly commensurable if they contain finite-index subgroups that are isomorphic. If X is a topologically rigid class of spaces closed under finite covers and X1 , X2 ∈ X ,
then π1 (X1 ) and π1 (X2 ) are abstractly commensurable if and only if X1 and X2 have homeomorphic finite-sheeted covering spaces. Thus, in this setting, topological invariants can be used
to distinguish abstract commensurability classes. For example, this technique was employed by
Crisp–Paoluzzi [CP08] and Dani-Stark-Thomas [DST16] for certain right-angled Coxeter groups
and by the author for related surface group amalgams [Sta17]. A survey on the use of orbifolds in the study of commensurability classes is given by Walsh [Wal11], and background on
commensurability classification is given by Paoluzzi [Pao13].
This paper was motivated by an interest in understanding the abstract commensurability classes
of Coxeter groups. Dani–Thomas [DT14] provide a quasi-isometry classification within a class
of hyperbolic one-ended right-angled Coxeter groups, and in joint work with Dani and Thomas
[DST16], we refine their work to give an abstract commensurability classification for a subclass of
these groups. In particular, it was of interest to determine whether topological rigidity holds for
finite covers of the Davis orbicomplex, which are natural spaces for right-angled Coxeter groups
and their finite-index subgroups.
Acknowledgments. The author thanks Pallavi Dani and Anne Thomas for enlightening conversations during our work on [DST16] and for comments on a draft of this paper. The author
thanks the anonymous referee for helpful comments. The author was partially supported by the
Azrieli Foundation.
2. The Davis orbicomplex
v3
v1
v2
Figure 2.1. The graph Γ defining the group WΓ
Definition 2.1 (The group W ). Let W = WΓ be the right-angled Coxeter group with defining
graph Γ given in Figure 2.1.
3
4
4
4
4
P
2
4
D
2
4
4
4
4
2
4 2
2
4
4
4
4
4
4
4
4
2
4
4
4
2
2
2
2
4
4
4
4
2
4
4
4
4
4
4
4
4
4
4
4
4
2
2
2
4
4
4
4
4
4
4
4
4
4
2
4
4
4
4
4
4
4
4
4
Figure 2.2. Pictured on the above right is the Davis orbicomplex D for the right-angled
Coxeter group with defining graph Γ given in Definition 2.1. On the left is illustrated the
collection P of six right-angled reflection orbifolds that are glued to each other by local
isometries to form the Davis orbicomplex. Each edge of these orbifolds is a reflection edge
except for the two edges which are glued to other orbifolds as indicated by the arrows.
The numbers indicate the order of the isotropy group at the orbifold point.
Construction 2.2 (The Davis orbicomplex D for W ). Let W be the right-angled Coxeter group
given in Definition 2.1. The Davis orbicomplex D for W has the following form, which is illustrated
in Figure 2.2. The space D is an orbicomplex whose underlying space is topologically the cone
on the defining graph Γ. The space D may be viewed as a graph of spaces with vertex spaces
2-dimensional right-angled reflection orbifolds with boundary, and these orbifolds are identified
along their boundary components as follows.
Let P be the following collection of orbifolds, which will be the vertex spaces of D. Define
a branch of the graph Γ to be an embedded path connecting two vertices of valence four. For
a branch β, let nβ be the number of vertices of the branch including the endpoints. So, Γ has
six branches, and if β is a branch of Γ, then nβ ∈ {5, 7}. The right-angled Coxeter group with
defining graph a branch β is the orbifold fundamental group of the following orbifold Pβ . The
orbifold Pβ has underlying space a right-angled hyperbolic (nβ + 1)-gon, nβ reflection edges, and
one non-reflection edge of length L > 0. Let P = {Pβ | β is a branch of Γ}; the six orbifolds in P
are illustrated on the left of Figure 2.2.
Identify the orbifolds in P along convex suborbifolds to form the orbicomplex D as follows.
For each orbifold in P, attach a 0-cell at the midpoint of each non-reflection edge, creating two
non-reflection edges of length L2 . Label these non-reflection edges as follows. First, label the three
vertices of Γ of valence four {v1 , v2 , v3 } as illustrated in Figure 2.1. Then, label a non-reflection
edge ei if the edge is incident to the reflection edge corresponding to the vertex vi ; this labeling
is indicated using arrows in Figure 2.2. To build the Davis orbicomplex D, identify the middle
4
2
2
2
2
2
2
π
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
4
2
4
2
2
4
2
4
2
4
2
2
2
2
4
Figure 3.1. Illustrated above are two orbifold covers. On the left, the group Z/2Z acts
by reflection, identifying the yellow and black vertices, and with quotient space a reflection
orbifold in P. On the right, Z/2Z acts by rotation by π.
vertex of the two non-reflection edges in each polygon in P and all non-reflection edges of the
same label as shown on the right of Figure 2.2.
It remains to check that the orbifold fundamental group of D is the right-angled Coxeter group
WΓ . Indeed, gluing non-reflection edges with the same label to form a single edge creates a single
reflection wall perpendicular to this edge; this gluing corresponds to identifying the endpoints of
the branches of Γ to form Γ. The claim then follows from arguments similar to those found in
[DST16, Section 3].
3. The set of finite-sheeted covers of D is not topologically rigid.
The covers of D restricted to the reflection orbifolds in P are given by the following two maps,
which are illustrated in Figure 3.1.
Lemma 3.1. Let D2 ( 2, . . . , 2 ) denote the orbifold with underlying space a disk and with ramifi| {z }
n
cation locus n cone points of order 2.
(a) Each orbifold in P with n reflection edges is covered by D2 ( 2, . . . , 2 ).
| {z }
n−1
(b) The orbifold D2 ( 2, . . . , 2 ) double covers D2 ( 2, . . . , 2 ).
| {z }
| {z }
2m
m+1
Proof. The first covering map is realized by reflection: arrange the cone points along a diameter
of the disk and reflect across this segment. The second covering map is realized by rotation:
arrange the 2m cone points symmetrically about a central (non-orbifold) point in the disk and
rotate by π.
5
2
2
2
2
2
2
2
Figure 3.2. Non-homeomorphic covers of the singular subspace, a 1-dimensional orbicomplex with ramification points of order two. The three graph covering maps can be
realized by rotation by π about a center point in the embedding of the graph in the plane,
and the orbifold covering map can be realized by reflection about a vertical line in the
Θ-graph.
Proof of Theorem 1.2. Let D ∈ X be the Davis orbicomplex defined in Construction 2.2 and
illustrated in Figure 2.2. To prove the class of spaces X is not topologically rigid, we will exhibit
two covers of D with the same fundamental group that are not homeomorphic.
The orbicomplex D has a singular subspace with underlying space a tripod formed by gluing
together the non-reflection edges of the orbifolds in P; to prove the finite covers constructed are
not homeomorphic, we prove the covers restricted to the singular subspace are not homeomorphic.
The singular subspace of D is a 1-dimensional orbicomplex with underlying space a star on three
vertices and so that each vertex of valence one is a ramification point of order two. The singular
subspace and the covers restricted to the singular subspace are drawn in Figure 3.2. We prove
that these graph coverings can be extended to finite coverings of the Davis orbicomplex D.
To construct two covers of D that are not homeomorphic, we first construct covers of degree
2
2
2
2
two, X2 →
− X1 →
− D, and then two further covers of degree two, Y →
− X2 and Z →
− X2 so that Y
and Z are not homeomorphic but are homotopy equivalent.
2
We begin by describing the cover X1 →
− D. This cover is illustrated in Figure 3.3. The space
2
X1 consists of two copies of D (2, 2, 2, 2, 2, 2), labeled D1 and D2 , and four copies of D2 (2, 2, 2, 2),
labeled D3 , D4 , D5 , D6 . To define the identification of these orbifolds, label the boundary circle
of Di by first subdividing it into two segments of equal length by adding vertices xi and yi .
Label one oriented edge {xi , yi } by di and the other by d0i . To form X1 , identify all vertices in
{xi | 1 ≤ i ≤ 6} to form a vertex x and identify all vertices in {yi | 1 ≤ i ≤ 6} to form a vertex y.
Identify edges {d1 , d2 , d3 , d4 } to form a single edge c1 ; identify edges {d01 , d02 , d05 , d06 } to form a
single edge c2 ; and, identify edges {d03 , d04 , d5 , d6 } to form a single edge c3 . Then, for each i, the
orbifold Di ⊂ X1 covers an orbifold Pβ ⊂ D, so that if the boundary of Di is labeled cj c−1
k , then
−1
this curve double-covers the non-reflection edge of Pβ labeled ej ek . These covering maps agree
along the intersection of the set {Di | 1 ≤ i ≤ 6} in X1 , and hence the union of these spaces covers
D, the union of the Pβ , by degree two.
The remaining covering maps may be realized by rotation in R3 ; these are illustrated in Figure
3.4. First, observe that the singular subspace of X1 is the Θ-graph, with two vertices of valence
6
2
X1
c3
2
2
2
2
c3
2
c3
c3
c1
2
c2
2
4
2
4
4
2
e3
e3
e1
4
2
4
e2
2
4
2
e3
4 2
4
2
D
2
e1
4
2
2
2
2
e3
c2
c1
4
4
4
4
2
2
2
4
4
2
2
c1 c2
2
2
2
2
2
c1
2
c2
2
2
4
4
4
2
2
2
2
e1 e2
4
2
4
e2
4
e1 e2
4
4
4
4
4
4
2
2
4
2
2
c1
c2
e1
2
e3
c3
2
e2
2
Figure 3.3. Illustrated on the top row is the degree-2 cover X1 → D. The space
X1 contains six orbifolds, each with underlying space a disk and with either four or six
cone points of order two. The orbifolds are glued along the boundary of the disks as
illustrated with the labeled arrows; all yellow vertices are identified, and all black vertices
are identified. The cover restricted to the singular subspaces is illustrated below; the
covering map is given by reflection through a vertical line through the left-hand graph.
three and the three directed edges {c1 , c2 , c3 } connecting the two vertices. This graph embeds in
−1
the plane and the boundary curves of the disk orbifolds Di are the three curves c1 c−1
2 , c1 c3 , and
3
3
c2 c−1
3 . Then, the space X1 embeds in the 3-ball B ⊂ R as illustrated in Figure 3.4 so that the
Θ-graph embeds in the equatorial xy-plane. The copies of D2 (2, 2, 2, 2) with boundary c1 c−1
3 may
2
be viewed as the two hemispheres of the unit sphere. The copies of D (2, 2, 2, 2) with boundary
c2 c−1
3 may be viewed as the two hemispheres of a sphere embedded inside the unit sphere; likewise
for the copies of D2 (2, 2, 2, 2, 2, 2). The remaining covering maps may be realized by rotations by
π about the z-axis in Euclidean space. The covering map restricted to each copy of D2 (2, . . . , 2)
is either exactly 2-to-1 or is the rotational covering map given in Lemma 3.1.
The two covers given in Figure 3.4 are not homeomorphic since their singular subspaces are not
homeomorphic. It remains to show that these spaces are homotopy equivalent and hence have the
same orbifold fundamental group. To see this, take a regular neighborhood of the singular locus
of Y and Z in the xy-plane to form a surface with genus zero and with six boundary components.
The homotopy from the embedded graph to its regular neighborhood in the plane extends to
homotopies from Y to a space Y 0 and from Z to a space Z 0 , where the homotopy is the identity
on the complement of the singular subspace. That is, the space Y 0 contains a surface of genus zero
7
Y
Z
2
2
X2
2
X1
c1
c2
c3
Figure 3.4. Two coverings of an orbicomplex8 that are not homeomorphic, but which have the
same orbifold fundamental group. The covering maps are each given by a Z/2Z action of rotation
about the z-axis. All of the blue points are orbifold points of order 2.
Λ
2
d3
2
d4
d1
d6
d5
d2
Figure 3.5. On the left are degree-2 covers given by rotation by π about a vertical axis
positioned through the surface or orbifold. The blue points represent cone points of order 2. On
the right is the singular subspace of the orbicomplex Y.
and with six boundary components B1 , . . . , B6 . For 1 ≤ i ≤ 4, two copies of D2 (2, 2, 2, 2, 2, 2) are
identified to Bi by homeomorphisms of the boundary curve of the disk orbifolds; for 5 ≤ i ≤ 6,
two copies of D2 (2, 2, 2, 2, 2, 2, 2, 2, 2, 2) are identified to Bi by a homeomorphisms of the boundary
curve of the disk orbifolds. The space Z 0 is homeomorphic to Y 0 . Thus, the orbicomplexes Y and
Z are homotopy equivalent, so their orbifold fundamental groups are isomorphic.
3.1. Torsion-free covers.
Proposition 3.2. Let Y and Z be the orbicomplexes described in the proof of Theorem 1.2 and
b and π1 (Z)
b are
shown in Figure 3.4. There exist finite covers Yb → Y and Zb → Z so that π1 (Y)
b ∼
b and Yb and Zb are not homeomorphic.
torsion-free, π1 (Y)
= π1 (Z),
Proof. We first describe the finite cover Yb → Y. The orbicomplex Y has a singular subspace the
planar graph Λ, shown in Figure 3.5. Let d1 . . . , d6 denote the boundary curves of the six planar
regions in the complement of Λ ⊂ R2 as marked in Figure 3.5. Glued to d1 and d2 are two copies of
D2 (2, 2, 2, 2, 2, 2, 2, 2, 2, 2), and glued to each of d3 , . . . , d6 are two copies of D2 (2, 2, 2, 2, 2, 2). As
shown in Figure 3.5, the surface Sg,4 of genus g and four boundary components forms a degree-4
cover of D2 ( 2, . . . , 2 ), where g = 3 if n = 6 and g = 7 if n = 10. Indeed, embed the surface in
| {z }
n
R3 so that the boundary components are arranged symmetrically in pairs about a vertical axis
and the holes of the surface lie along the axis. Rotate by π about the vertical axis to produce an
orbifold with underlying space an annulus and with 2g + 2 cone points of order 2. Arrange the
cone points symmetrically along a core curve of the annulus and rotate by π about an axis that
b take four
skewers the core curve in two non-singular points to obtain D2 ( 2, . . . , 2 ). To form Y,
| {z }
n
copies of the graph Λ; for i = 1, 2, glue to the four copies of di the four boundary curves of S7,4 by
homeomorphisms; and, for i = 3, . . . , 6, glue to the four copies of di the boundary curves of S3,4
by homeomorphisms. Since each boundary curve of Sg,4 covers the boundary of D2 ( 2, . . . , 2 ) by
| {z }
n
degree-1, Yb forms a degree-4 cover of Y.
The finite cover Zb → Z is constructed similarly. By analogous arguments to those in the proof
of Theorem 1.2, Yb and Zb are homotopy equivalent. The singular subspace of Yb is homeomorphic
to four copies of the singular subspace of Y; likewise, the singular subspace of Zb is homeomorphic
to four copies of the singular subspace of Z. Thus, Yb and Zb are not homeomorphic.
9
References
[BL12]
[CP08]
[Dav08]
[DST16]
[DT14]
[FJ89]
[Kap09]
[Laf04]
[Laf06]
[Laf07]
[Pao13]
[Rat06]
[Sta17]
[Wal11]
[Xie06]
Arthur Bartels and Wolfgang Lück. The Borel conjecture for hyperbolic and CAT(0)-groups. Ann. of
Math. (2), 175(2):631–689, 2012.
John Crisp and Luisa Paoluzzi. Commensurability classification of a family of right-angled Coxeter groups.
Proc. Amer. Math. Soc., 136(7):2343–2349, 2008.
Michael W. Davis. The geometry and topology of Coxeter groups, volume 32 of London Mathematical
Society Monographs Series. Princeton University Press, Princeton, NJ, 2008.
Pallavi Dani, Emily Stark, and Anne Thomas. Commensurability classification for certain right-angled
Coxeter groups and geometric amalgams of free groups. arXiv:1610.06245, 2016.
Pallavi Dani and Anne Thomas. Bowditch’s JSJ tree and the quasi-isometry classification of certain
Coxeter groups. arXiv:1402.6224, 2014.
F. T. Farrell and L. E. Jones. A topological analogue of Mostow’s rigidity theorem. J. Amer. Math. Soc.,
2(2):257–370, 1989.
Michael Kapovich. Hyperbolic manifolds and discrete groups. Modern Birkhäuser Classics. Birkhäuser
Boston, Inc., Boston, MA, 2009. Reprint of the 2001 edition.
Jean-François Lafont. Rigidity result for certain three-dimensional singular spaces and their fundamental
groups. Geom. Dedicata, 109:197–219, 2004.
Jean-François Lafont. Strong Jordan separation and applications to rigidity. J. London Math. Soc. (2),
73(3):681–700, 2006.
Jean-François Lafont. Diagram rigidity for geometric amalgamations of free groups. J. Pure Appl. Algebra,
209(3):771–780, 2007.
Luisa Paoluzzi. The notion of commensurability in group theory and geometry. RIMS Kkyroku,
1836:124137, 2013.
John G. Ratcliffe. Foundations of hyperbolic manifolds, volume 149 of Graduate Texts in Mathematics.
Springer, New York, second edition, 2006.
Emily Stark. Abstract commensurability and quasi-isometry classification of hyperbolic surface group
amalgams. Geom. Dedicata, 186:39–74, 2017.
Genevieve S. Walsh. Orbifolds and commensurability. In Interactions between hyperbolic geometry, quantum topology and number theory, volume 541 of Contemp. Math., pages 221–231. Amer. Math. Soc.,
Providence, RI, 2011.
Xiangdong Xie. Quasi-isometric rigidity of Fuchsian buildings. Topology, 45(1):101–169, 2006.
10
| 4math.GR
|
arXiv:1307.6894v1 [math.CT] 25 Jul 2013
THE OPERAD OF TEMPORAL WIRING DIAGRAMS: FORMALIZING A
GRAPHICAL LANGUAGE FOR DISCRETE-TIME PROCESSES
DYLAN RUPEL AND DAVID I. SPIVAK
Abstract. We investigate the hierarchical structure of processes using the mathematical
theory of operads. Information or material enters a given process as a stream of inputs,
and the process converts it to a stream of outputs. Output streams can then be supplied to
other processes in an organized manner, and the resulting system of interconnected processes
can itself be considered a macro process. To model the inherent structure in this kind of
system, we define an operad W of black boxes and directed wiring diagrams, and we define a
W-algebra P of processes (which we call propagators, after [RS]). Previous operadic models
of wiring diagrams (e.g. [Sp2]) use undirected wires without length, useful for modeling
static systems of constraints, whereas we use directed wires with length, useful for modeling
dynamic flows of information. We give multiple examples throughout to ground the ideas.
Contents
1. Introduction
2. W, the operad of directed wiring diagrams
3. P, the algebra of propagators on W
4. Future work
References
1
5
18
34
36
1. Introduction
Managing processes is inherently a hierarchical and self-similar affair. Consider the case of
preparing a batch of cookies, or if one prefers, the structurally similar case of manufacturing a
pharmaceutical drug. To make cookies, one generally follows a recipe, which specifies a process
that is undertaken by subdividing it as a sequence of major steps. These steps can be performed
in series or in parallel. The notion of self-similarity arises when we realize that each of these
major steps can itself be viewed as a process, and thus it can also be subdivided into smaller
steps. For example, procuring the materials necessary to make cookies involves getting oneself
to the appropriate store, selecting the necessary materials, paying for them, etc., and each of
these steps is itself a simpler process.
Perhaps every such hierarchy of nesting processes must touch ground at the level of atomic
detail. Hoping that the description of processes within processes would not continue ad infinitum may have led humanity to investigate matter and motion at the smallest level possible. This
investigation into atomic and quantum physics has yielded tremendous technological advances,
such as the invention of the microchip.
Working on the smallest possible scale is not always effective, however. It appears that the
planning and execution of processes benefits immensely from hierarchical chunking. To write
Spivak acknowledges support by ONR grant N000141310260.
1
2
DYLAN RUPEL AND DAVID I. SPIVAK
a recipe for cookies at the level of atomic detail would be expensive and useless. Still, when
executing our recipe, the decision to add salt will initiate an unconscious procedure, by which
signals are sent from the brain to the muscles of the arm, on to individual cells, and so on
until actual atoms move through space and “salt has been added”. Every player in the larger
cookie-making endeavor understands the current demand (e.g. to add salt) as a procedure that
makes sense at his own level of granularity. The decision to add salt is seen as a mundane
(low-level) job in the context of planning to please ones girlfriend by baking cookies; however
this same decision is seen as an abstract (high-level) concept in the context of its underlying
performance as atomic movements.
For designing complex processes, such as those found in manufacturing automobiles or in
large-scale computer programming, the architect and engineers must be able to change levels
of abstraction with ease. In fact, different engineers working on the same project are often
thinking about the same basic structures, but in different terms. They are most effective when
they can chunk the basic structures as they see fit.
A person who studies a supply chain in terms of the function played by each chain member
should be able to converse coherently with a person who studies the same supply chain in terms
of the contracts and negotiations that exist at each chain link. These are two radically different
viewpoints on the same system, and it is useful to be able to switch fluidly between them.
Similarly, an engineer designing a system’s hardware must be able to converse with an engineer
working on the system’s software. Otherwise, small perturbations made by one of them will be
unexpected by the other, and this can lead to major problems.
The same types of issues emerge whether one is concerned with manufacturers in a supply
chain, neurons in a functional brain region, modules in a computer program, or steps in a recipe.
In each case, what we call propagators (after [RS]) are being arranged into a system that is
itself a propagator at a higher level. The goal of this paper is to provide a mathematical basis
for thinking about this kind of problem. We offer a formalism that describes the hierarchical
and self-similar nature of a certain kind of wiring diagram.
A similar kind of wiring diagram was described in [Sp2], the main difference being that
the present one is built for time-based processes whereas the one in [Sp2] was built for static
relations. In the present work we take the notion of time (or one may say distance) seriously. We
go through considerable effort to integrate a notion of time and distance into the fundamental
architecture of our description, by emphasizing that communication channels have a length, i.e.
that communication takes time.
Design choices such as these greatly affect the behavior of our model, and ours was certainly
not the only viable choice. We hope that the basic idea we propose will be a basis upon which
future engineers and mathematicians will improve. For the time being, we may at least say that
the set of rules we propose for our wiring diagrams roughly conform with the IDEF0 standard
set by the National Institute of Standards and Technology [NIST]. The main differences are
that in our formalism,
• wires can split but not merge (each merging must occur within a particular box),
• feedback loops are allowed,
• the so-called control and mechanism arrows are subsumed into input and output arrows,
and
• the rules for and meaning of hierarchical composition is made explicit.
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
3
The basic picture to have in mind for our wiring diagrams is the following:
(1)
flour
sugar
dry mix
salt
milk
eggs
wet mix
cookie batter
egg yolks
egg yolks
In this picture we see an exterior box, some interior boxes, and a collection of directed wires.
These directed wires transport some type of product from the export region of some box to the
import region of some box. In (1) we have a supply chain involving three propagators, one of
whom imports flour, sugar, and salt and exports dry mixture, and another of whom imports
eggs and milk and exports egg yolks and wet mixture. The dry mixture and the wet mixture
are then transported to a third propagator who exports cookie batter. The whole system itself
constitutes a propagator that takes five ingredients and produces cookie batter and egg yolks.
The formalism we offer in this paper is based on a mathematical structure called an operad
(more precisely, a symmetric colored operad), chosen because they capture the self-similar nature
of wiring diagrams. The rough idea is that if we have a wiring diagram and we insert wiring
diagrams into each of its interior boxes, the result is a new wiring diagram.
(2)
We will make explicit what constitutes a box, what constitutes a wiring diagram (WD), and
how inserting WDs into a WD constitutes a new WD. Like Russian dolls, we may have a nesting
of WDs inside of WDs inside of WDs, etc. We will prove an associativity law that guarantees
that no matter how deeply our Russian dolls are nested, the resulting WD is well-defined. Once
all this is done, we will have an operad W.
To make this directed wiring diagrams operad W useful, we will take our formalism to the
next logical step and provide an algebra on W. This algebra P encodes our application to
process management by telling us what fits in the boxes and how to use wiring diagrams to
build more complex systems out of simpler components. More precisely, the algebra P makes
explicit
• the set of things that can go in every box, namely the set of propagators, and
• a method for taking a wiring diagram and a propagator for each of its interior boxes
and producing a propagator for the exterior box.
To prove that we have an algebra, we will show that no matter how one decides to group the
various internal propagators, the behavior of the resulting system is unchanged.
Operads were invented in the 1970s by [May] and [BV] in order to encode the relationship
between various operations they noticed taking place in the mathematical field of algebraic
topology. At the moment we are unconcerned with topological properties of our operads, but
the formalism grounds the picture we are trying to get across. For more on operads, see [Lei].
4
DYLAN RUPEL AND DAVID I. SPIVAK
1.1. Structure of the paper. In Section 2 we discuss operads. In Section 2.1 we give the
mathematical definition of operads and some examples. In Section 2.2 we propose the operad
of interest, namely W, the operad of directed wiring diagrams. We offer an example wiring
diagram in Section 2.3 that will run throughout the paper and eventually output the Fibonacci
sequence. In Section 2.4 we prove that W has the required properties so that it is indeed an
operad.
In Section 3 we discuss algebras on an operad. In Section 3.1 we give the mathematical
definition of algebras. In Sections 3.2 we discuss some preliminaries on lists and define our
notion of historical propagators, which we will then use in 3.3 where we propose the W-algebra of
interest, the algebra of propagators. In Section 3.5 we prove that P has the required properties
so that it is indeed a W-algebra.
We expect the majority of readers to be most interested in the running examples sections,
Sections 2.3 and 3.4. Readers who want more details, e.g. those who may wish to write code
for propagators, will need to read Sections 2.2, 3.3. The proof that our algebra satisfies the
necessary requirements is technical; we expect only the most dedicated readers to get through
it. Finally, in Section 4 we discuss some possibilities for future work in this area.
The remainder of the present section is devoted to our notational conventions (Section 1.2)
and our acknowledgments (1.3).
1.2. Notation and background. Here we describe our notational conventions. These are
only necessary for readers who want a deep understanding of the underlying mathematics. Such
readers are assumed to know some basic category theory. For mathematicians we recommend
[Awo] or [Mac], for computer scientists we recommend [Awo] or [BW], and for a general audience
we recommend [Sp1].
We will primarily be concerned only with the category of small sets, which we denote by Set,
and some related categories. We denote by Fin ⊆ Set the full subcategory spanned by finite
sets. We often use the symbol n ∈ Ob(Fin) to denote a finite set, and may speak of elements
i ∈ n. The cardinality of a finite set is a natural number, denoted |n| ∈ N. In particular, we
consider 0 to be a natural number.
Suppose given a finite set n and a function X : n → Ob(Set), and let ∐i∈I X(i) be the
disjoint union. Then there is a canonical function πX : ∐i∈n X(i) −→ n which we call the
component projection. We use almost the same symbol in a different context; namely, for any
function s : m → n we denote the s-coordinates projection by
Y
Y
X(s(j)).
X(i) −→
πs :
i∈n
j∈m
In particular,
if i ∈ n is an element, we consider it as a function i : {∗} → n and write
Q
πi : i∈n X(i) → X(i) for the usual ith coordinate projection.
A pointed set is a pair (S, s) where S ∈ Ob(Set) is a set and s ∈ S is a chosen element, called
the base point. In particular a pointed set cannot be empty. Given another pointed set (T, t),
a pointed function from (S, s) to (T, t) consists of a function f : S → T such that f (s) = t. We
denote the category of pointed sets by Set∗ . There is a forgetful functor Set∗ → Set which
forgets the basepoint; it has a left adjoint which adjoins a free basepoint X 7→ X ∐ {∗}. We
often find it convenient not to mention basepoints; if we speak of a set X as though it is pointed,
we are actually speaking of X ∐ {∗}. If S, S ′ are pointed sets then the product S × S ′ is also
naturally pointed, with basepoint (∗, ∗), again denoted simply by ∗.
We often speak of functions n → Ob(Set∗ ), where n is a finite set. Of course, Ob(Set∗ )
is not itself a small set, but using the theory of Grothendieck universes [Bou], this is not a
problem. It will be even less of a problem in applications.
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
5
1.3. Acknowledgements. David Spivak would like to thank Sam Cho as well as the NIST
community, especially Al Jones and Eswaran Subrahmanian. Special thanks go to Nat Stapleton
for many valuable conversations in which substantial progress was made toward subjects quite
similar to the ones we discuss here.
Dylan Rupel would like to thank Jason Isbell and Kiyoshi Igusa for many useful discussions.
2. W, the operad of directed wiring diagrams
In this section we will define the operad W of black boxes and directed wiring diagrams
(WDs). It governs the forms that a black box can take, the rules that a WD must follow, and
the formula for how the substitution of WDs into a WD yields a WD. There is no bound on
the depth to which wiring diagrams can be nested. That is, we prove an associative law which
roughly says that the substitution formula is well-defined for any degree of nesting, shallow or
deep.
We will use the operad W to discuss the hierarchical nature of processes. Each box in our
operad will be filled with a process, and each wiring diagram will effectively build a complex
process out of simpler ones. However, this is not strictly a matter of the operad W but of an
algebra on W. This algebra will be discussed in Section 3.
The present section is organized as follows. First, in Section 2.1 we give the technical
definition of the term operad and a few examples. In Section 2.2 we propose our operad
W of wiring diagrams. It will include drawings that should clarify the matter. In Section
2.3 we present an example that will run throughout the paper and end up producing the
Fibonacci sequence. This section is recommended especially to the more category-theoretically
shy reader. Finally, in Section 2.4 we give a technical proof that our proposal for W satisfies
the requirements for being a true operad, i.e. we establish the well-definedness of repeated
substitution as discussed above.
2.1. Definition and basic examples of operads. Before we begin, we should give a warning
about our use of the term “operad”.
Warning 2.1.1. Throughout this paper, we use the word operad to mean what is generally called
a symmetric colored operad or a symmetric multicategory. This abbreviated nomenclature is
not new, for example it is used in [Lur]. Hopefully no confusion will arise. For a full treatment
of operads, multicategories, and how they fit into a larger mathematical context, see [Lei].
Most of Section 2.1 is recycled material, taken almost verbatim from [Sp2]. We repeat it
here for the convenience of the reader.
Definition 2.1.2. An operad O is defined as follows: One announces some constituents (A.
objects, B. morphisms, C. identities, D. compositions) and proves that they satisfy some requirements (1. identity law, 2. associativity law). Specifically,
A. one announces a collection Ob(O), each element of which is called an object of O.
B. for each object y ∈ Ob(O), finite set n ∈ Ob(Fin), and n-indexed set of objects
x : n → Ob(O), one announces a set On (x; y) ∈ Ob(Set). Its elements are called
morphisms from x to y in O.
C. for every object x ∈ Ob(O), one announces a specified morphism denoted idx ∈
O1 (x; x) called the identity morphism on x.
D. Let s : m → n be a morphism in Fin. Let z ∈ Ob(O) be an object, let y : n → Ob(O)
be an n-indexed set of objects, and let x : m → Ob(O) be an m-indexed set of objects.
6
(3)
DYLAN RUPEL AND DAVID I. SPIVAK
For each element i ∈ n, write mi := s−1 (i) for the pre-image of s under i, and write
xi = x|mi : mi → Ob(O) for the restriction of x to mi . Then one announces a function
Y
Omi (xi ; y(i)) −→ Om (x; z),
◦ : On (y; z) ×
i∈n
called the composition formula for O.
Given an n-indexed set of objects x : n → Ob(O) and an object y ∈ Ob(O), we sometimes abuse
notation and denote the set of morphisms from x to y by O(x1 , . . . , xn ; y). 1 We may write
HomO (x1 , . . . , xn ; y), in place of O(x1 , . . . , xn ; y), when convenient. We can denote a morphism
φ ∈ On (x; y) by φ : x → y or by φ : (x1 , . . . , xn ) → y; we say that each xi is a domain object of
φ and that y is the codomain object of φ. We use infix notation for the composition formula,
e.g. writing ψ ◦ (φ1 , . . . , φn ).
These constituents (A,B,C,D) must satisfy the following requirements:
1. for every x1 , . . . , xn , y ∈ Ob(O) and every morphism φ : (x1 , . . . , xn ) → y, we have
φ ◦ (idx1 , . . . , idxn ) = φ
s
and
idy ◦ φ = φ;
t
2. Let m −
→ n −
→ p be composable morphisms in Fin. Let z ∈ Ob(O) be an object,
let y : p → Ob(O), x : n → Ob(O), and w : m → Ob(O) respectively be a p-indexed,
n-indexed, and m-indexed set of objects. For each i ∈ p, write ni = t−1 (i) for the
pre-image and xi : ni → Ob(O) for the restriction. Similarly, for each k ∈ n write
mk = s−1 (k) and wk : mk → Ob(O); for each i ∈ p, write mi,− = (t ◦ s)−1 (i) and
wi,− : mi,− → Ob(O); for each j ∈ ni , write mi,j := s−1 (j) and wi,j : mi,j → Ob(O).
Then the diagram below commutes:
Q
Q Q
Op (y; z) × i∈p Oni (xi ; y(i)) × i∈p,
❚❚❚❚j∈ni Omi,j (wi,j ; xi (j))
❥
❚❚❚❚
❥❥❥
❥
❥
❥
❚❚❚❚
❥❥
❥
❥
❥
t❥ Q
Q❚*
Q
Q
Omi,− (wi,− ; y(i))
O
(y;
z)
×
(w
;
x(k))
On (x; z) × k∈n Om
p
i∈p
❙❙k❙ k
❙❙❙
❦❦❦❦
❙❙❙
❦❦❦
❦
❙❙❙
❦
❦
u❦❦
)
Om (w; z)
Remark 2.1.3. In this remark we will discuss the abuse of notation in Definition 2.1.2 and
how it relates to an action of a symmetric group on each morphism set in our definition of
operad. We follow the notation of Definition 2.1.2, especially following the use of subscripts in
the composition formula.
Suppose that O is an operad, z ∈ Ob(O) is an object, y : n → Ob(O) is an n-indexed
set of objects, and φ : y → z is a morphism. If we linearly order n, enabling us to write
φ : (y(1), . . . , y(|n|)) → z, then changing the linear ordering amounts to finding an isomorphism
of finite sets σ : m −
→ n, where |m| = |n|. Let x = y ◦ σ and for each i ∈ n, note that
mi = σ −1 ({i}) = {σ −1 (i)}, so xi = x|σ−1 (i) = y(i). Taking idxi ∈ Omi (xi ; y(i)) for each
i ∈ n, and using the identity law, we find that the composition formula induces a bijection
On (y; z) −
→ Om (x; z), which we might denote by
σ : O(y(1), y(2), . . . , y(n); z) O y(σ(1)), y(σ(2)), . . . , y(σ(n)); z .
1There are three abuses of notation when writing O(x , . . . , x ; y), which we will fix one by one. First, it
n
1
confuses the set n ∈ Ob(Fin) with its cardinality |n| ∈ N. But rather than writing O(x1 , . . . , x|n| ; y), it would
be more consistent to write O(x(1), . . . , x(|n|); y), because we have assigned subscripts another meaning in D.
However, even this notation unfoundedly suggests that the set n has been endowed with a linear ordering, which
it has not. This may be seen as a more serious abuse, but see Remark 2.1.3.
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
7
In other words, there is an induced group action of Aut(n) on On (y(1), . . . , y(n); z), where
Aut(n) is the group of permutations of an n-element set.
Throughout this paper, we will permit ourselves to abuse notation and speak of morphisms
φ : (x1 , x2 , . . . , xn ) → y for a natural number n ∈ N, without mentioning the abuse inherent in
choosing an order, so long as it is clear that permuting the order of indices would not change
anything up to canonical isomorphism.
Example 2.1.4. We define the operad of sets, denoted Sets, as follows. We put Ob(Sets) :=
Ob(Set). Given a natural number n ∈ N and objects X1 , . . . , Xn , Y ∈ Ob(Sets), we define
Sets(X1 , X2 , . . . , Xn ; Y ) := HomSet (X1 × X2 × · · · × Xn , Y ).
For any X ∈ Ob(Sets) the identity morphism idX : X → X is the same identity as that in Set.
The composition formula is as follows. Suppose given a set Z ∈ Ob(Set), a finite set
n ∈ Ob(Fin), for each i ∈ n a set Yi ∈ Ob(Set) and a finite set mi ∈ Ob(Fin), and for each
j ∈ mi a setQXi,j ∈ Ob(Set). Suppose furthermore that we
Qhave composable morphisms: a
function g : i∈n Yi → Z and for each i ∈ n a function fi : j∈mi Xi,j → Yi . Let m = ∐i mi .
Q
We need a function j∈m Xj → Z, which we take to be the composite
Q
Y
Y Y
fi
g
i∈n
Yi −−−→ Z.
Xi,j −−−−−−−−→
i∈n j∈mi
i∈n
It is not hard to see that this composition formula is associative.
Example 2.1.5. The commutative operad E has one object, say Ob(E) = {N}, and for each n ∈ N
it has a single n-ary morphism, En (N, . . . , N; N) = {µn }.
2.2. The announced structure of the wiring diagrams operad W. To define our operad
W, we need to announce its structure, i.e.
• define what constitutes an object of W,
• define what constitutes a morphism of W,
• define the identity morphisms in W, and
• the formula for composing morphisms of W.
For each of these we will first draw and describe a picture to have in mind, then give a
mathematical definition. In Section 2.4 we will prove that the announced structure has the
required properties.
Objects are black boxes. Each object X will be drawn as a box with input arrows entering on
the left of the box and output arrows leaving from the right of the box. The arrows will be
called wires. All input and output wires will be drawn across the corresponding vertical wall
of the box.
in(X)
out(X)
(4)
Each wire is also assigned a set of values that it can carry, and this set can be written next to
the wire, or the wires may be color coded. See Example 2.2.2 below. As above, we often leave
off the values assignment in pictures for readability reasons.
Announcement 2.2.1 (Objects of W). An object X ∈ Ob(W) is called a black box, or box
for short. It consists of a tuple X := (in(X), out(X), vset), where
• in(X) ∈ Ob(Fin) is a finite set, called the set of input wires to X,
• out(X) ∈ Ob(Fin) is a finite set, called the set of output wires from X, and
8
DYLAN RUPEL AND DAVID I. SPIVAK
• vset(X) : in(X) ∐ out(X) → Ob(Set∗ ) is a function, called the values assignment for
X. For each wire i ∈ in(X) ∐ out(X), we call vset(i) ∈ Ob(Set∗ ) the set of values
assigned to wire i, and we call its basepoint element the default value on wire i.
♦
Example 2.2.2. We may take X = ({1}, {2, 3}, vset), where vset: {1, 2, 3} → Ob(Set∗ ) is
given by vset(1) = N, vset(2) = N, and vset(3) = {a, b, c}. 2 We would draw X as follows.
X
{a, b, c}
N
N
The input wire carries natural numbers, as does one of the output wires, and the other
output wire carries letters a, b, c.
Morphisms are directed wiring diagrams. Given black boxes Y1 , . . . , Yn ∈ Ob(W) and a black
box Z ∈ Ob(W), we must define the set Wn (Y ; Z) of wiring diagrams (WDs) of type Y1 , . . . , Yn →
Z. Such a wiring diagram can be taken to denote a way to wire black boxes Y1 , . . . , Yn together
to form a larger black box Z. A typical such wiring diagram is shown below:
ψ : (Y1 , Y2 , Y3 ) → Z
Y1
Y3
in(Z)
out(Z)
Y2
(5)
Here n = 3, and for example Y1 has two input wire and three outputs wires. Each wire in a
WD has a specified directionality. As it travels a given wire may split into separate wires, but
separate wires cannot come together. The wiring diagram also includes a finite set of delay
nodes; in the above case there are four.
One should think of a wiring diagram ψ : Y1 , . . . , Yn → Z as a rule for managing material (or
information) flow between the components of an organization. Think of ψ as representing this
organization. The individual components of the organization are the interior black boxes (the
domain objects of ψ) and the exterior black box (the codomain object of ψ). Each component
supplies material to ψ as well as demands material from ψ. For example component Z supplies
material on the left side of ψ and demands it on the right side of ψ. On the other hand, each Yi
supplies material on its right side and demands material on its left. Like the IDEF0 standard
for functional modeling diagrams [NIST], we always adhere to this directionality.
We insist on one perhaps surprising (though seemingly necessary rule), namely that the
wiring diagram cannot connect an output wire of Z directly to an input wire of Z. Instead,
each output wire of Z is supplied either by an output wire of some Y (i) or by a delay node.
2 The functor vset is supposed to assign pointed sets to each wire, but no base points are specified in the
description above. As discussed in Section 1.2, in this case we really have vset(1) = N ∐ {∗}, vset(2) = N ∐ {∗},
and vset(3) = {a, b, c} ∐ {∗}, where ∗ is the default value.
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
9
Announcement 2.2.3 (Morphisms of W). Let n ∈ Ob(Fin) be a finite set, let Y : n → Ob(W)
be an n-indexed set of black boxes, and let Z ∈ Ob(W) be another black box. We write
(6)
in(Y ) = ∐i∈n in(Y (i)),
out(Y ) = ∐i∈n out(Y (i)).
We take vset: in(Y ) ∐ out(Y ) → Ob(Set∗ ) to be the induced map.
A morphism
ψ : Y (1), . . . , Y (n) → Z
in Wn (Y ; Z) is called a temporal wiring diagram, a wiring diagram, or a WD for short. It
consists of a tuple (DNψ , vset, sψ ) as follows. 3
• DNψ ∈ Ob(Fin) is a finite set, called the set of delay nodes for ψ. At this point we
can define the following sets:
Dmψ := out(Z) ∐ in(Y ) ∐ DNψ
Spψ := in(Z) ∐ out(Y ) ∐ DNψ
the set of demand wires in ψ, and
the set of supply wires in ψ.
• vset: DNψ → Ob(Set) is a function, called the value-set assignment for ψ, such that
the diagram
idDNψ
/ Dmψ
●●
●●vset
●●
idDNψ
vset
●●
●#
Spψ vset / Set∗
DNψ
commutes (meaning that every delay node demands the same value-set that it supplies).
• sψ : Dmψ → Spψ is a function, called the supplier assignment for ψ. The supplier
assignment sψ must satisfy two requirements:
(1) The following diagram commutes:
Dmψ
●●
●●vset
●●
●●
#
Spψ vset / Set∗
sψ
meaning that whenever a demand wire is assigned a supplier, the set of values
assigned to these wires must be the same.
(2) If z ∈ out(Z) then sψ (z) < in(Z). Said another way,
sψ |out(Z) ⊆ out(Y ) ∐ DNψ ,
meaning that a global output cannot be directly supplied by a global input. We
call this the non-instantaneity requirement.
We have functions vset : in(Z) ∐ out(Z) → Set∗ , vset: in(Y (i)) ∐ out(Y (i)) → Set∗ , and
vset : DNψ → Set∗ . It should not cause confusion if we use the same symbol to denote the
induced functions vset : Dmψ → Set∗ and vset: Spψ → Set∗ .
♦
3A morphism ψ : Y → Z is in fact an isomorphism class of this data. That is, given two tuples (DN , vset, s )
ψ
ψ
′ , vset′ , s′ ) as above, with a bijection DN DN ′ making all the appropriate diagrams commute,
and (DNψ
ψ
ψ
ψ
we consider these two tuples to constitute the same morphism ψ : Y → Z.
10
DYLAN RUPEL AND DAVID I. SPIVAK
Remark 2.2.4. We have taken the perspective that W is an operad. One might more naturally
think of W as the underlying operad of a symmetric monoidal category whose objects are again
black boxes and whose morphisms are again wiring diagrams, though now a morphism connects
a single internal domain black box to the external codomain black box. From this perspective
one should merge the many isolated black boxes occurring in the domain of a multicategory
wiring diagram into a single black box as the domain of the monoidal category wiring diagram.
Though mathematically equivalent and though we make use of this perspective in the course
of our proofs, it is somewhat unnatural to perform this grouping in applications. For example,
though it makes some sense to view ourselves writing this paper and you reading this paper as
black boxes inside a single “information conveying” wiring diagram it would be rather strange
to conglomerate all of our collective inputs and outputs so that we become a single metainformation entity. For reasons of this sort we choose to take the perspective of the underlying
operad rather than of a monoidal category.
On the other hand, the notation of monoidal categories is convenient, so we introduce it
here. Given a finite set n and an n-indexed set of objects Y : n → Ob(W), we discussed in (6)
what should be seen as a tensor product
O
Y (i) = (∐i∈n in(Y (i)), ∐i∈n out(Y (i)), vset),
i∈n
which we write simply as Y = (in(Y ), out(Y ), vset).
Similarly, given an n-indexed set of morphisms φi : Xi → Y (i) in W, we can form their
tensor product
O
O
O
Y (i),
Xi →
φi :
i∈n
i∈n
i∈n
which we write simply as φ : X → Y , in a similar way. That is, we form a set of delay nodes
DNφ = ∐i∈n DNφi , supplies Spφ = ∐i∈n Spφi , demands Dmφ = ∐i∈n Dmφi , and a supplier
assignment sφ = ∐i∈n sφi , all by taking the obvious disjoint unions.
Example 2.2.5. In the example below, we see a big box with three little boxes inside, and
we see many wires with arrowheads placed throughout. It is a picture of a wiring diagram
φ : (X1 , X2 , X3 ) → Y . The big box can be viewed as Y , which has some number of input and
output wires; however, when we see the big box as a container of the little boxes wired together,
we are actually seeing the morphism φ.
φ : (X1 , X2 , X3 ) → Y
flour
sugar
salt
milk
eggs
X1
dry mix
X3
cookie batter
wet mix
X2
egg yolks
We aim to explain our terminology of demand and supply, terms which interpret the organization forced on us by the mathematics. Each wire has a demand side and a supply side; when
there are no feedback loops, as in the picture above, supplies are on the left side of the wire
and demands are to the right, but this is not always the case. Instead, the distinction to make
is whether an arrowhead is entering the big box or leaving it: those that enter the big box are
supplies to φ, and those that are leaving the big box are demands upon φ. The five left-most
arrowheads are entering the big box, so flour, sugar, etc. are being supplied. But flour, sugar,
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
11
and salt are demands when they leave the big box to enter X1 . Counting, one finds 9 supply
wires and 9 demand wires (though the equality of these numbers is just a coincidence due to
the fact that no wire splits or is wasted).
Identity morphisms are identity supplier assignments. Let Z = (in(Z), out(Z), vset). The
identity wiring diagram idZ : Z → Z might be drawn like this:
Z
Z
Even though the interior box is of a different size than the exterior box, the way they are wired
together is as straightforward as possible.
Announcement 2.2.6 (Identity morphisms in W). Let Z = (in(Z), out(Z), vsetZ ). The
identity wiring diagram idZ : Z → Z has DNidZ = ∅ with the unique function vset : ∅ →
Ob(Set), so that DmidZ = out(Z) ∐ in(Z) and SpidZ = in(Z) ∐ out(Z). The supplier
assignment sidZ : SpidZ → DmidZ is given by the identity function, which satisfies the noninstantaneity requirement.
♦
Composition of morphisms is achieved by removing intermediary boxes and associated arrowheads. We are interested in substituting a wiring diagram into each black box of a wiring
diagram, to produce a more detailed wiring diagram. The basic picture to have in mind is the
following:
φ1
ψ
φ2
ω = ψ ◦ (φ1 , φ2 )
12
DYLAN RUPEL AND DAVID I. SPIVAK
On the top we see a wiring diagram ψ in which each internal box, say Y (1) and Y (2), has
a corresponding wiring diagram φ1 and φ2 respectively. Dropping them into place and then
removing the intermediary boxes leaves a single wiring diagram ω. One can see that every
input of Y (i) plays a dual role. Indeed, it is a demand from the perspective of ψ, and it is a
supply from the perspective of φi . Similarly, every output of Y (i) plays a dual role as supply
in ψ and demand in φi .
In Announcement 2.2.8 we will provide the composition formula for W. Namely, we will be
given morphisms φi : Xi → Y (i) and ψ : Y → Z. Each of these has
N its own delay nodes, DNφi
and DNψ as well as its own supplier assignments. Write φ =
i φi : X → Y as in Remark
2.2.4. For the reader’s convenience, we now summarize the demands and supplies for each of the
given morphisms φi : Xi → Y (i) and ψ : Y → Z, as well as their (not-yet defined) composition
ω : X → Z. Let DNω = DNφ ∐ DNψ .
(7)
Morphism
φi
φ
ψ
ω
Summary of notation for composition in W
Dm−
Sp−
out(Y (i)) ∐ in(Xi ) ∐ DNφi in(Y (i)) ∐ out(Xi ) ∐ DNφi
out(Y ) ∐ in(X) ∐ DNφ
in(Y ) ∐ out(X) ∐ DNφ
out(Z) ∐ in(Y ) ∐ DNψ
in(Z) ∐ out(Y ) ∐ DNψ
out(Z) ∐ in(X) ∐ DNω
in(Z) ∐ out(X) ∐ DNω
φ
ψ
Lemma 2.2.7. Suppose given morphisms X −
→ Y and Y −
→ Z in W, as above. That is, we
are given sets of delay nodes, DNφ and DNψ , as well as supplier assignments
sφ : Dmφ → Spφ
sψ : Dmψ → Spψ
and
each of which is subject to a non-instantaneity requirement,
(8)
sφ
out(Y )
⊆ out(X) ∐ DNφ
and
sψ
out(Z)
⊆ out(Y ) ∐ DNψ .
Let sω be as in Table 7. It follows that the diagram below is a pushout
h
Spφ
O
x
/ Spω
O
g
f
in(Y ) ∐ out(Y )
/ Spψ
e
where
(9)
e = sψ
in(Y )
∐ idout(Y )
f = idin(Z) ∐ sφ
out(Y )
g = idin(Y ) ∐ sφ
out(Y )
h = (f ◦ sψ )
in(Y )
∐ idDNψ
∐ idout(X) ∐ idDNφ .
Moreover, each of e, f, g, and h commute with the appropriate functions vset.
Proof. We first show that the diagram commutes; here are the calculations on each component:
f ◦e
f ◦e
in(Y )
= f ◦ sψ
out(Y )
= sφ
in(Y )
out(Y )
=h◦g
=h◦g
in(Y )
out(Y )
.
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
13
We now show that the diagram is a pushout. Suppose given a set Q and a commutative
solid-arrow diagram (i.e. with h′ ◦ g = f ′ ◦ e):
❦/5 Q
h′
❦❦ J
❦
❦
❦ ❦α
❦
❦
h
/ in(Z) ∐ out(X) ∐ DNφ ∐ DNψ
in(Y ) ∐ out(X) ∐ DNφ
O
O
x
g
f
in(Y ) ∐ out(Y )
f′
/ in(Z) ∐ out(Y ) ∐ DNψ
e
Looking at components on which f and h are identities, we see that if we want the equations
α ◦ f = f ′ and α ◦ h = h′ to hold, there is at most one way to define α : Spω → Q. Namely,
α := f ′
in(Z)∐DNψ
∐ h′
out(X)∐DNφ
.
To see that this definition works, it remains to check that α ◦ f
α◦h
in(Y )
=h
′
in(Y )
α◦f
out(Y )
= f′
out(Y )
and that
. For the first we use a non-instantaneity requirement (8) to calculate:
out(Y )
= α ◦ sφ
out(Y )
=α
out(X)∐DNφ
◦ sφ
out(Y )
′
= h ◦ sφ
out(Y )
′
=h ◦g
out(Y )
= f′ ◦ e
out(Y )
= f′
out(Y )
Now we have shown that α ◦ f = f ′ and the second calculation follows:
α◦h
in(Y )
= α ◦ f ◦ sψ
in(Y )
= f ′ ◦ sψ
in(Y )
= f′ ◦ e
in(Y )
= h′ ◦ g
in(Y )
= h′
in(Y )
Each of e, f, g, h commute with the respective functions vset because each is built solely out
of identity functions and supplier assignments. This completes the proof.
Announcement 2.2.8 (Composition formula for W). Let m, n ∈ Ob(Fin) be finite sets and
let t : m → n be a function. Let Z ∈ Ob(W) be a black box, let Y : n → Ob(O) be an n-indexed
set of black boxes, and let X : m → Ob(O) be an m-indexed set of black boxes. For each element
i ∈ n, write mi := t−1 (i) for the pre-image of i under t, and write Xi = X mi : mi → Ob(O)
for the restriction of X to mi . Then the composition formula
Y
Wmi (Xi ; Y (i)) −→ Wm (X; Z),
◦ : Wn (Y ; Z) ×
i∈n
is defined as follows.
Suppose that we
N are given morphisms φi : Xi → Y (i) for each i ∈ n, which we gather into
a morphism φ = i φi : X → Y as in Remark 2.2.4, and that we are also given a morphism
ψ : Y → Z. Then we have finite sets of delay nodes DNφ and DNψ , and supplier assignments
sφ : Dmφ → Spφ
and
sψ : Dmψ → Spψ
as in Announcement 2.2.3.
We are tasked with defining a morphism ω := ψ ◦ φ : X → Z. The set of demand wires and
supply wires for ω are given in Table (7). Thus our job is to define a set DNω and a supplier
assignment sω : Dmω → Spω .
We put DNω = DNφ ∐ DNψ . It suffices to find a function
sω : out(Z) ∐ in(X) ∐ DNω −→ in(Z) ∐ out(X) ∐ DNω ,
14
DYLAN RUPEL AND DAVID I. SPIVAK
which satisfies the two requirements of being a supplier assignment. We first define the function
by making use of the following diagram, where the pushout is as in Lemma 2.2.7:
sφ
in(X) ∐ DNφ
(10)
in(X)∐DNφ
h
/ Spφ
O
x
g
/ Spω
O
f
in(Y ) ∐ out(Y )
e
/ Spψ
O
sψ
out(Z)∐DNψ
out(Z) ∐ DNψ
Thus we can define a function
sω = h ◦ sφ
(11)
in(X)∐DNφ
∐ f ◦ sψ
out(Z)∐DNψ
.
We need to show that sω satisfies the two requirements of being a supplier assignment (see
Announcement 2.2.3).
(1) The fact that sω commutes with the appropriate functions vset follows from the fact
that sφ , sψ , f, and h do so (by Lemma 2.2.7).
(2) The fact that the non-instantaneity requirement holds for sω , i.e. that sω (out(Z)) ⊆
out(X) ∐ DNω , follows from the fact that it holds for sψ and sψ (see (8)), as follows.
sω (out(Z)) = f ◦ sψ (out(Z))
⊆ f (out(Y ) ∐ DNψ )
= sφ (out(Y )) ∐ DNψ
⊆ out(X) ∐ DNφ ∐ DNψ = out(X) ∐ DNω .
♦
2.3. Running example to ground ideas and notation regarding W. In this section we
will discuss a few objects of W (i.e. black boxes), a couple morphisms of W (i.e. wiring
diagrams), and a composition of morphisms. We showed objects and morphisms in more
generality above (see Examples 2.2.2 and 2.2.5). Here we concentrate on a simple case, which
we will take up again in Section 3.4 and which will eventually result in a propagator that
outputs the Fibonacci sequence. First, we draw three objects, X, Y, Z ∈ Ob(W).
aX
bX
X
Y
cX
aY
Z
cY
cZ
(12)
These objects are not complete until the pointed sets associated to each wire are specified. Let
N := (N, 1) be the set of natural numbers with basepoint 1, and put
vset(aX ) = vset(bX ) = vset(cX ) = vset(aY ) = vset(cY ) = vset(cZ ) = N.
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
15
Now we draw two morphisms, i.e. wiring diagrams, φ : X → Y and ψ : Y → Z:
aY
X−
→Y
φ
Y −
→Z
Y
Z
aX
ψ
cX
X
cY
Y
aY
cY
cZ
dψ
bX
(13)
To clarify the notion of inputs, outputs, supplies, and demands, we provide two tables that lay
out those sets in the case of (13).
Objects shown above
Object in(−)
out(−)
X
{aX , bX } {cX }
Y
{aY }
{cY }
Z
{}
{cZ }
Morphisms shown above
Morphism DN− Dm−
Sp−
φ
{}
{cY , aX , bX } {aY , cX }
ψ
{dψ } {cZ , aY , dψ } {cY , dψ }
To specify the morphism φ : X → Y (respectively ψ : Y → Z), we are required not only to
provide a set of delay nodes DNφ , which we said was DNφ = ∅ (respectively, DNψ = {dψ }),
but also a supplier assignment function sφ : Dmφ → Spφ (resp., sψ : Dmψ → Spψ ). Looking at
the picture of φ (resp. ψ) above, the reader can trace backward to see how every demand wire
is attached to some supply wire. Thus, the supplier assignment sφ for φ : X → Y is
cY 7→ cX ,
aX 7→ aY ,
bX 7→ cX ,
and the supplier assignment sψ for ψ : Y → Z is
cZ 7→ dψ ,
aY 7→ dψ ,
dψ 7→ cY .
We now move on to the composition of ψ and φ. The idea is that we “plug the φ diagram
into the Y -box of the ψ diagram, then erase the Y -box”. We follow this in two steps below: on
the left, we shrink down a copy of φ and fit it into the Y -box of ψ. On the right, we erase the
Y -box:
ψ
φ
ψ◦φ
X−
→Z
→Y −
X −−→ Z
Z
Z
Y
X
X
The pushout (10) ensures that wires of Y connect wires inside (i.e. from φ) to wires outside
(i.e. from ψ). In other words, when we erase box Y , we do not erase the connections it made
for us. We compute the pushout of the diagram
aY 7→aY , cY 7→cX
aY 7→dψ , cY 7→cY
{aY , cX } ←−−−−−−−−−−− {aY , cY } −−−−−−−−−−−→ {cY , dψ },
16
DYLAN RUPEL AND DAVID I. SPIVAK
defining Spω , to be isomorphic to {dψ , cX }. The supplier assignment sω : Dmω = {cZ , aX , bX , dψ } →
{dψ , cZ } = Spω is given by
(14)
cZ 7→ dψ ,
aX 7→ dψ ,
bX 7→ cX ,
dψ 7→ cX .
We take this example up again in Section 3.4, where we show that installing a “plus” function
into box X yields the Fibonacci sequence.
2.4. Proof that the operad requirements are satisfied by W. We need to show that the
announced operad W satisfies the requirements set out by Definition 2.1.2. There are two such
requirements: the first says that composing with the identity morphism has no effect, and the
second says that composition is associative.
Proposition 2.4.1. The identity law holds for the announced structure of W.
Proof. Let X1 , . . . , Xn and Y be black boxes and let φ : X1 , . . . , Xn → Y be a morphism. We
need to show that the following equations hold:
?
φ ◦ (idx1 , . . . , idxn ) = φ
?
idy ◦ φ = φ.
and
We are given a set DNφ and a function vset : DNφ → Ob(Set). Let idX =
form in(X) and out(X) as in Remark 2.2.4. Thus we have
Spφ = in(Y ) ∐ out(X) ∐ DNφ
and
N
i∈n
idXi , and
Dmφ = out(Y ) ∐ in(X) ∐ DNφ
and a supplier assignment sφ : Dmφ → Spφ . For each i ∈ n we have SpidXi = DmidXi , and the
supplier assignments are the identity, so we have
SpidX = DmidX = in(X) ∐ out(X)
The supplier assignment sidX is the identity function. Similarly, SpidY = DmidY = in(Y ) ∐
out(Y ), and the supplier assignment sidY is the identity function.
Let ω = φ ◦ (idX1 , . . . , idXn ) and ω ′ = idY ◦ φ. Then the relevant pushouts become
id
in(X)
in(X)
sφ in(X)∐id
/ Spid
X
out(X)
x
/ Spω
O
/ Spφ
O
in(X) ∐ out(X)
sφ
out(Y )∐DNφ
out(Y ) ∐ DNφ
sφ
in(X) ∐ DNφ
in(X)∐DNφ
/ Spφ
O
x
/ Spω′
O
id
in(Y ) ∐ out(Y )
in(Y )
∐sφ
SpidY
O
id
out(Y )
out(Y )
out(Y )
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
17
The pushout of an isomorphism is an isomorphism so we have isomorphisms Spφ Spω and
Spφ Spω′ . 4 In both the case of ω and ω ′ , one checks using (9) that the induced supplier
assignments are also in agreement (up to isomorphism), sω = sφ = sω′ .
Proposition 2.4.2. The associativity law holds for the announced structure of W.
Proof. Suppose we are given morphisms τ : W → X, φ : X → Y and ψ : Y → Z. We must
check that (ψ ◦ φ) ◦ τ = ψ ◦ (φ ◦ τ ). With notation as in Lemma 2.2.7, pushout square defining
φ ◦ τ and then ψ ◦ (φ ◦ τ ) are these:
SpO τ
hφ,τ
x
/ Spφ◦τ
O
gφ,τ
eφ,τ
/ Spψ◦(φ◦τ )
O
x
gψ,φ◦τ
fφ,τ
in(X) ∐ out(X)
hψ,φ◦τ
Spφ◦τ
O
/ Spφ
fψ,φ◦τ
in(Y ) ∐ out(Y )
eψ,φ◦τ
/ Spψ
whereas the pushout square defining ψ ◦ φ and then (ψ ◦ φ) ◦ τ are these:
Spφ
O
hψ,φ
x
gψ,φ
/ Spψ◦φ
O
eψ,φ
/ Sp(ψ◦φ)◦τ
O
x
gψ◦φ,τ
fψ,φ
in(Y ) ∐ out(Y )
hψ◦φ,τ
SpO τ
/ Spψ
fψ◦φ,τ
in(X) ∐ out(X)
eψ◦φ,τ
/ Spψ◦φ
One checks directly from the formulas (9) that eψ◦φ,τ = hψ,φ ◦ eφ,τ as functions in(X) ∐
out(X) → Spψ◦φ , and that gψ,φ◦τ = fφ,τ ◦ gψ,φ as functions in(Y ) ∐ out(Y ) → Spφ◦τ .
We combine them into the following pushout diagram:
SpO τ
hφ,τ
x
gφ,τ
in(X) ∐ out(X)
/ Spφ◦τ
O
hψ,φ◦τ
x
/ Spψ◦φ◦τ
O
fφ,τ
eφ,τ
/ Spφ
O
fψ◦φ,τ
hψ,φ
x
/ Spψ◦φ
O
gψ,φ
in(Y ) ∐ out(Y )
fψ,φ
eψ,φ
/ Spψ
The pasting lemma for pushout squares ensures that the set labeled Spψ◦φ◦τ is isomorphic to
Spψ◦(φ◦τ ) and to Sp(ψ◦φ)◦τ , so these are indeed isomorphic to each other. It is also easy to
check using the formulas provided in (11) and (9) that the supplier assignments
Dmψ◦φ◦τ = out(Z) ∐ in(W ) ∐ DNτ ∐ DNφ ∐ DNψ −→ Spψ◦φ◦τ
agree regardless of the order of composition. This proves the result.
4Note that a morphism (e.g. ω) in W are defined only up to isomorphism class of tuples (DN , vset, s ), see
ω
ω
Announcement 2.2.3.
18
DYLAN RUPEL AND DAVID I. SPIVAK
3. P, the algebra of propagators on W
In this section we will introduce our algebra of propagators on W. This is where form meets
function: the form called “black box” is a placeholder for a propagator, i.e. a function, that
carries input streams to output streams, and the form called “wiring diagram” is a placeholder
for a circuit that links propagators together to form a larger propagator.
To formalize these ideas we introduce the mathematical notion of operad algebra in Section
3.1. In Section 3.2 we discuss some preliminaries on lists and streams, and define our notion
of historical propagator. In Section 3.3 we announce our algebra of these propagators and in
Section 3.4 we ground it in our running example. Finally in Section 3.5 we prove that the
announced structure really satisfies the requirements of being an algebra.
3.1. Definition and basic examples of algebras. In this section we give the formal definition for algebras over an operad.
Definition 3.1.1. Let O be an operad. An O-algebra, denoted F : O → Sets, is defined as
follows: One announces some constituents (A. map on objects, B. map on morphisms) and
proves that they satisfy some requirements (1. identity law, 2. composition law). Specifically,
A. one announces a function Ob(F ) : Ob(O) → Ob(Sets).
B. for each object y ∈ Ob(O), finite set n ∈ Ob(Fin), and n-indexed set of objects
x : n → Ob(O), one announces a function
Fn : On (x; y) → HomSets (F x; F y).
As in B. above, we often denote Ob(F ), and also each Fn , simply by F .
These constituents (A,B) must satisfy the following requirements:
1. For each object x ∈ Ob(O), the equation F (idx ) = idF x holds.
2. Let s : m → n be a morphism in Fin. Let z ∈ Ob(O) be an object, let y : n → Ob(O)
be an n-indexed set of objects, and let x : m → Ob(O) be an m-indexed set of objects.
Then, with notation as in Definition 2.1.2, the following diagram of sets commutes:
Q
◦
/ Om (x; z)
On (y; z) × i∈n Omi (xi ; y(i))
(15)
F
F
HomSets (F y; F z) ×
Q
i∈n
HomSets (F xi ; F y(i))
◦
/ HomSets (F x; F z)
Example 3.1.2. Let E be the commutative operad of Example 2.1.5. An E-algebra S : E → Sets
consists of a set M ∈ Ob(Set), and for each natural number n ∈ N a morphism µn : M n →
M . It is not hard to see that, together, the morphism µ2 : M × M → M and the element
µ0 : {∗} → M give M the structure of a commutative monoid. Indeed, the associativity and
unit axioms are encoded in the axioms for operads and their morphisms. The commutativity of
multiplication arises by applying the commutative diagram (15) in the case s : {1, 2} → {1, 2}
is the non-identity bijection, as discussed in Remark 2.1.3.
3.2. Lists, streams, and historical propagators. In this section we discuss some background on lists. We also develop our notion of historical propagator, which formalizes the idea
that a machine’s output at time t0 can depend only on what has happened previously, i.e. for
time t < t0 . While strictly not necessary for the development of this paper, we also discuss the
relation of historical propagators to streams.
Given a set S, an S-list is a pair (t, ℓ), where t ∈ N is a natural number and ℓ : {1, 2, . . . , t} →
S is a function. We denote the set of S-lists by List(S). We call t the length of the list; in
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
19
particular a list may be empty because we may have t = 0. Note that there is a canonical
bijection
a
S t.
List(S)
t∈N
We sometimes denote a list simply by ℓ and write |ℓ| to denote its length; that is we have the
component projection |·| : List(A) → N. We typically write-out an S-list as ℓ = [ℓ(1), ℓ(2), . . . , ℓ(t)],
where each ℓ(i) ∈ S. We denote the empty list by [ ]. Given a function f : S → S ′ , there is
an induced function List(f ) : List(S) → List(S ′ ) sending (t, ℓ) to (t, f ◦ ℓ); in the parlance of
computer science List(f ) is the function that “maps f over
Q ℓ”.
Given sets X1 , . . . , Xk ∈ Ob(Set), an element in List( 1≤i≤k Xi ) is a list of k-tuples. Given
sets A and B there is a bijection
: List(A) ×N List(B) −−−→ List(A × B),
|·|
|·|
where on the left we have formed the fiber product of the diagram List(A) −→ N ←− List(B).
We call this bijection zipwith, following the terminology from modern functional programming
languages. The idea is that an A-list ℓA can be combined with a B-list ℓB , as long as they have
the same length |ℓA | = |ℓB |; the result will be an (A × B)-list ℓA ℓB again of the same length.
We will usually abuse this distinction and freely identify List(A × B) List(A) ×N List(B) with
its image in List(A) × List(B). For example, we may consider the N × N-list
[(1, 2), (3, 4), (5, 6)] = [1, 3, 5] [2, 4, 6]
as an element of List(N) × List(N). Hopefully this will not cause confusion.
Let List≥1 (S) ⊆ List(S) denote the set ∐t≥1 S t . We write ∂S : List≥1 (S) → List(S) to denote
the function that drops off the last entry. More precisely, for any integer t ≥ 1 if we consider ℓ
as a function ℓ : {1, 2, . . . , t} → S, then the list ∂S ℓ is given by pre-composition with the subset
consisting of the first t − 1 elements,
ℓ
{1, 2, . . . , t − 1} ֒→ {1, 2, . . . , t} −
→ S.
For example we have ∂[0, 1, 4, 9, 16] = [0, 1, 4, 9].
Definition 3.2.1. Let R, S be pointed sets and let n ∈ N. A n-historical propagator f from R
to S is a function f : List(R) → List(S) satisfying the following conditions:
(1) If a list ℓ ∈ List(R) has length |ℓ| = t, then |f (ℓ)| = t + n,
(2) If ℓ ∈ List(R) is a list of length t ≥ 1, then
∂S f (ℓ) = f (∂R ℓ).
We denote the set of n-historical propagators from R to S by Histn (R, S). If f is n-historical
for some n ≥ 0 we say that f is historical.
We usually drop the subscript from the symbol ∂− , writing e.g. ∂f (ℓ) = f (∂ℓ).
Example 3.2.2. Let S be a pointed set and let n ∈ N be a natural number. Define an n-historical
propagator δ n ∈ Histn (S, S) as follows for ℓ ∈ List(S):
(
∗
if 1 ≤ i ≤ n
δ n (ℓ)(i) =
ℓ(i − n) if n + 1 ≤ i ≤ t + n
We call δ n the n-moment delay function. For example if n = 3, S = {a, b, c, d} ∐ {∗}, and
ℓ = [a, a, b, ∗, d] ∈ S 5 then δ 3 (S) = [∗, ∗, ∗, a, a, b, ∗, d] ∈ S 8 .
The following Lemma describes the behavior of historical functions.
20
DYLAN RUPEL AND DAVID I. SPIVAK
Lemma 3.2.3. Let S, S ′ , S ′′ , T, T ′ ∈ Set∗ be pointed sets.
(1) Let f : S → T be a function. The induced function List(f ) : List(S) → List(T ) is
0-historical.
(2) Given n-historical propagators q ∈ Histn (S, S ′ ) and r ∈ Histn (T, T ′ ), there is an induced
n-historical propagator q × r ∈ Histn (S × T, S ′ × T ′ ).
(3) Given q ∈ Histm (S, S ′ ) and q ′ ∈ Histn (S ′ , S ′′ ), then q ′ ◦ q : List(S) → List(S ′′ ) is
(m + n)-historical.
(4) If n ≥ 1 is an integer and q ∈ Histn (S, S ′ ) is n-historical then ∂q : List(S) → List(S ′ )
is (n − 1)-historical.
Proof. We show each in turn.
(1) Let ℓ ∈ List(S) be a list of length t. Clearly, List(f ) sends ℓ to a list of length
t. If t ≥ 1 then the fact that ∂List(f )(ℓ) = List(f )(∂ℓ) follows by associativity of
composition in Set. That is, List(f )(ℓ) is the right-hand composition and ∂ℓ is the
left-hand composition below:
ℓ
f
{1, . . . , t − 1} ֒→ {1, . . . , t} −
→S−
→ T.
q×r
(2) On the length t component we use the function (S × T )t = S t × T t −−→ S t+n × T t+n =
(S × T )t+n . As necessary, we have
∂ ◦ (q × r) = ∂q × ∂r = q∂ × r∂ = (q × r) ◦ ∂.
(3) This is straightforward; for example the second condition is checked
∂q ′ (q(ℓ)) = q ′ (∂q(ℓ)) = q ′ (q(∂ℓ)).
(4) On lengths we indeed have |∂q(ℓ)| = |q(ℓ)| − 1 = |ℓ| + n − 1. If |ℓ| = t ≥ 1 then
∂(∂q)(ℓ) = ∂(∂q(ℓ)) = ∂q(∂ℓ) because q is historical.
Definition 3.2.4. Let S be a pointed set. An S-stream is a function σ : N≥1 → S. We denote
the set of S-streams by Strm(S).
For any natural number t ∈ N, let σ [1,t] ∈ List(S) denote the list of length t corresponding
σ
→ S and call it the t-restriction of S.
to the composite {1, 2, . . . , t} ֒→ N≥1 −
Lemma 3.2.5. Let S be a pointed set, let {∗} be a pointed set with one element, and let n ∈ N
be a natural number. There is a bijection
Histn ({∗}, S) −
→ Strm(S).
Proof. For any natural number t ∈ N, let t = {1, 2, . . . , t} ∈ Ob(Set). Let [N] be the poset
(considered as a category) with objects {t | t ∈ N}, ordered by inclusion of subsets. For any
n ∈ N there is a functor [N] → Set sending t ∈ Ob([N]) to {1, 2, . . . , t + n} ∈ Ob(Set).
For any n ∈ N, there is a bijection N colimt∈[N] {1, 2, . . . , t + n}. Thus we have a bijection
Strm(S) = HomSet (N≥1 , S) lim HomSet ({1, 2, . . . , t + n}, S).
t∈[N]
On the other hand, an n-historical function f : List({∗}) → List(S) acts as follows. For each
t ∈ N and list [∗, . . . , ∗t ] of length t, it assigns a list f ([∗, . . . , ∗t ]) ∈ List(S) of length t + n,
i.e. a function {1, . . . , t + n} → S, such that f ([∗, . . . , ∗t−1 ]) is the restriction to the subset
{1, . . . , t + n − 1}.
The fact that these notions agree follows from the construction of limits in the category Set.
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
21
Below we define an awkward-sounding notion of n-historical stream propagator. The idea is
that a function carrying streams to streams is n-historical if, for all t ∈ N, its output up to time
t + n depends only on its input up to time t. In Proposition 3.2.7 we show that this notion of
historicality for streams is equivalent to the notion for lists given in Definition 3.2.1.
Definition 3.2.6. Let S and T be pointed sets, and let n ∈ N be a natural number. A function
f : Strm(S) → Strm(T ) is called an n-historical stream propagator if, given any natural number
t ∈ N and any two streams σ, σ ′ ∈ Strm(S), if σ [1,t] = σ ′ [1,t] then f (σ) [1,t+n] = f (σ ′ ) [1,t+n] .
Let Histnstrm (S, T ) denote the set of n-historical stream propagators Strm(S) → Strm(T ).
Proposition 3.2.7. Let S and T be pointed sets. There is a bijection
Histn (S, T ) −
→ Histnstrm (S, T ).
Proof. We construct two functions α : Histn (S, T ) → Histnstrm (S, T ) and β : Histnstrm (S, T ) →
Histn (S, T ) that are mutually inverse.
Given an n-historical function f : List(S) → List(T ) and a stream σ ∈ Strm(S), define the
stream α(f )(σ) : N≥1 → T to be the function whose (t + n)-restriction (for any t ∈ N) is given
by
α(f )(σ) [1,t+n] = f (σ [1,t] ).
Because f is historical, this construction is well defined.
Given an n-historical stream propagator F : Strm(S) → Strm(T ) and a list ℓ ∈ List(S) of
length |ℓ| = t, let ℓ∗ ∈ Strm(S) denote the stream N≥1 → S given on i ∈ N≥1 by
(
ℓ(i) if 1 ≤ i ≤ t
ℓ∗ (i) =
∗
if i ≥ t + 1.
Now define the list β(F )(ℓ) ∈ List(T ) by
β(F )(ℓ) = F (ℓ∗ )
One checks directly that for all F ∈
f ∈ Histn (S, T ) we have β ◦ α(f ) = f .
.
[1,t+n]
Histnstrm (S, T ) we have
α ◦ β(F ) = F and that for all
The above work shows that the notion of historical propagator is the same whether one
considers it as acting on lists or on streams. Throughout the rest of this paper we work solely
with the list version. However, we sometimes say the word “stream” (e.g. “a propagator takes
a stream of inputs and returns a stream of outputs”) for the image it evokes.
3.3. The announced structure of the propagator algebra P. In this section we will
announce the structure of our W-algebra of propagators, which we call P. That is, we must
specify
• the set P(Y ) of allowable “fillers” for each black box Y ∈ Ob(W),
• how a wiring diagram ψ : Y1 , . . . , Yn → Z and a filler for each Yi serves to produce a
filler for Z.
In this section we will explain in words and then formally announce mathematical definitions.
In Section 2.4 we will prove that the announced structure has the required properties.
As mentioned above, the idea is that each black box is a placeholder for (i.e. can be filled
with) those propagators which carry the specified local input streams to the specified local
output streams. Each wiring diagram with propagators installed in each interior black box will
constitute a new propagator for the exterior black box, which carries the specified global input
22
DYLAN RUPEL AND DAVID I. SPIVAK
streams to the specified global output streams. We now go into more detail and make these
ideas precise.
Black boxes are filled by historical propagators. Let Z = (in(Z), out(Z), vset) be an object in
W. Recall that each element w ∈ in(Z) is called an input wire, which carries a set vset(w)
of possible values, and that element w′ ∈ out(Z) is called an output wire, which also carries a
set vset(w′ ) of possible values. This terminology is suggestive of a machine, which we call a
historical propagator (or propagator for short), which takes a list of values on each input wire,
processes it somehow, and emits a list of values on each output wire. The propagator’s output
at time t0 can depend on the input it received for time t < t0 , but not on input that arrives
later.
Announcement 3.3.1 (P on objects). Let Z = (in(Z), out(Z), vset) be an object in W. For
any subset I ⊆ in(Z) ∐ out(Z) we define
Y
vsetI =
vset(i).
i∈I
In particular, if I = ∅ then vsetI is a one-element set.
We define P(Z) ∈ Ob(Set) to be the set of 1-historical propagators of type Z,
P(Z) := Hist1 (vsetin(Z) , vsetout(Z) ).
♦
Consider the propagator below, which has one input wire and one output wire, say both
carrying integers.
“Σ”
The name “Σ” suggests that this propagator takes a list of integers and returns their running
total. But for it to be 1-historical, its input up to time t determines its output up to time t + 1.
Thus for example it might send an input list ℓ := [1, 3, 5, 7, 10] of length 5 to the output list
“Σ”(ℓ) = [0, 1, 4, 9, 16, 26] of length 6.
Remark 3.3.2. As in Remark 2.2.4 the following notation is convenient. Given
a finite set
N
Y
n ∈ Ob(Fin) and black boxes Yi ∈ Ob(W) for i ∈ n, we can form Y =
i∈n i , with for
example in(Y ) = ∐i∈n in(Yi ). Similarly, given N
a 1-historical propagator gi ∈ P(Yi ) for each
i ∈ n weQcan form a 1-historical propagator g := i∈n gi ∈ Hist1 (vsetin(Y ) , vsetout(Y ) ) simply
by g = i∈n gi .
Wiring diagrams shuttle value streams between propagators. Let Z ∈ Ob(W) be a black box,
let n ∈ Ob(Fin) be a finite set, and let Y : n → Ob(W) be an n-indexed set of black boxes.
A morphism ψ : Y → Z in W is little more than a supplier assignment sψ : Dmψ → Spψ . In
other words, it connects each demand wire to a supply wire carrying the same set of values.
Therefore, if a propagator is installed in each black box Y (i), then ψ tells us how to take each
value stream being produced by some propagator and feed it into the various propagators that
it supplies.
Announcement 3.3.3 (P on morphisms). Let Z ∈ Ob(W) be a black box, let n ∈ Ob(Fin)
be a finite set, let Y : n → Ob(W) be an n-indexed set of black boxes, and let ψ : Y → Z be a
morphism in W. We must construct a function
P(ψ) : P(Y (1)) × · · · × P(Y (n)) → P(Z).
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
23
That is, given a historical propagator gi ∈ Hist1 (vsetin(Y (i)) , vsetout(Y (i)) ) for each i ∈ n,
we need to produce a historical propagator P(ψ)(g1 , . . . , gn ) ∈ Hist1 (vsetin(Z) , vsetout(Z) ).
N
Define g ∈ Hist1 (vsetin(Y ) , vsetout(Y ) ) by g := i∈n gi , as in Remark 3.3.2. Let inDmψ =
in(Y ) ∐ DNψ and inSpψ = out(Y ) ∐ DNψ , denote the set of internal demands of ψ and the
set of internal supplies of ψ, respectively.
We will define P(ψ)(g) by way of five helper functions:
Sψ ∈ Hist0 (vsetSpψ , vsetDmψ ),
Sψ′ ∈ Hist0 (vsetSpψ , vsetinDmψ ),
Sψ′′ ∈ Hist0 (vsetinSpψ , vsetout(Z) ),
Eψ,g ∈ Hist1 (vsetinDmψ , vsetinSpψ ),
Cψ,g ∈ Hist0 (vsetin(Z) , vsetSpψ ),
where we will refer to the Sψ , Sψ′ , Sψ′′ as “shuttle”, Eψ,g as “evaluate”, and Cψ,g as “cascade”.
We will abbreviate by in(Z) the set List(vsetin(Z) ), and similarly for Spψ , inDmψ , etc.
By Announcement 2.2.3, a morphism ψ : Y → Z in W is given by a tuple (DNψ , vset, sψ ),
where in particular we remind the reader of a commutative diagram
Dmψ
●●
●●vset
●●
●●
#
Spψ vset / Set∗
sψ
where we require sψ (out(Z)) ⊆ inSpψ . The function sψ : Dmψ → Spψ induces the coordinate
projection function πsψ : vsetSpψ → vsetDmψ (see Section 1.2). Applying the functor List
gives a 0-historical function (see Lemma 3.2.3), List(πsψ ) which we abbreviate as
Sψ : Spψ → Dmψ .
This is the function that shuttles a list of tuples from where they are supplied directly along a
wire to where they are demanded. We define a commonly-used projection,
Sψ′ := πinDmψ ◦ Sψ : Spψ → inDmψ .
The purpose of defining the set inDmψ of internal demands above is that the supplier assignment sends out(Z) into it, i.e. we have sψ out(Z) : out(Z) → inSpψ by the non-instantaneity
requirement. It induces π
: vsetinSpψ → vsetout(Z) . Applying List gives a 0-historical
sψ
function List(π
sψ
out(Z)
) which we abbreviate as
out(Z)
S ′′ : inSpψ → out(Z).
Thus S ′ and S ′′ first shuttle from supply lines to all demand lines, and then focus on only
a subset of them. Let δψ1 ∈ Hist1 (vsetDNψ , vsetDNψ ) be the 1-moment delay. Note that if
DNψ = ∅ then δψ1 : {∗} → {∗} carries no information and can safely be ignored.
We now define the remaining helper functions:
(16)
Eψ,g := (g × δψ1 ),
(
[]
Cψ,g (ℓ) :=
(ℓ, Eψ,g ◦ Sψ′ ◦ Cψ,g (∂ℓ))
if |ℓ| = 0
if |ℓ| ≥ 1.
24
DYLAN RUPEL AND DAVID I. SPIVAK
The last is an inductive definition, which we can rewrite for |ℓ| ≥ 1 as
Cψ,g = idin(Z) × (Eψ,g ◦ Sψ′ ◦ Cψ,g ◦ ∂) ◦ ∆,
where ∆ : in(Z) → in(Z) × in(Z) is the diagonal map. Intuitively it says that a list of length
t on the input wires will produces a list of length t on all supply wires. By Lemma 3.2.3 Eψ,g
is 1-historical and Cψ,g is 0-historical.
We are ready to define the 1-historical function
P(ψ)(g) = Sψ′′ ◦ Eψ,g ◦ Sψ′ ◦ Cψ,g .
(17)
♦
Remark 3.3.4. The definitions of Sψ′ and Eψ,g above implicitly make use of the “zipwith”
functions
: in(Z) ×N inDmψ −−−→ Dmψ
and
: in(Y ) ×N DNψ −−−→ inDmψ ,
respectively. In section 3.5 we will make similar abuses in the calculations; however, when
commutative diagrams are given, the zipwith is made “explicit” by writing an equality between
products of streams and streams of products when we mean that should be applied to a
product of streams.
3.4. Running example to ground ideas and notation regarding P. In this section we
compose elementary morphisms and apply them to a simple “addition” propagator to construct
a propagator that outputs the Fibonacci sequence. Let X, Y, Z ∈ Ob(W) and φ : X → Y and
ψ : Y → Z be as in (12) and (13). Let N = (N, 1) ∈ Set∗ denote the set of natural numbers
with basepoint 1. We recall the shapes of X, Y , and Z here, but draw them with different
labels:
aX
bX
“+”
“1 + Σ”
cX
aY
“F ib”
cY
cZ
We have replaced the symbol X with the symbol “+ ” because we are about to define an
X-shaped propagator “+” ∈ P(X). Given an incoming list of numbers on wire aX and another
incoming list of numbers on wire bX , it will create a list of their sums and output that on cX .
More precisely, we take “+” : List(N × N ) → List(N ) to be the 1-historical propagator defined
as follows. Suppose given a list ℓ ∈ List(N × N ) of length t, say
ℓ = ℓa (1), ℓa (2), . . . , ℓa (t), ℓb (1), ℓb (2), . . . , ℓb (t)
Define “+”(ℓ) ∈ List(N ) to be the list whose nth entry (for 1 ≤ n ≤ t + 1) is
(
1
if n = 1
“+”(ℓ)(n) =
ℓa (n − 1) + ℓb (n − 1) if 2 ≤ n ≤ t + 1
So for example “ + ”([4, 5, 6, 7] [1, 1, 3, 7]) = [1, 5, 6, 9, 14].
We will use only this “+” propagator to build our Fibonacci sequence generator. To do so,
we will use wiring diagrams φ and ψ, whose shapes we recall here from (13) above.
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
“1 + Σ” = P(φ)(“+”)
“F ib” = P(ψ)(“1 + Σ”)
“1 + Σ”
“F ib”
aX
aY
25
“+” cX
cY
aY
“1 + Σ”
cY
cZ
dψ
bX
The Y -shaped propagator “1 + Σ” = P(φ)(“+”) ∈ P(Y ) will have the following behavior: given
an incoming list of numbers on wire aY , it will return a list of their running totals, plus 1.
More precisely “1 + Σ” : List(N ) → List(N ) is the 1-historical propagator defined as follows.
Suppose given a list ℓ ∈ List(N ) of length t, say ℓ = [ℓ1 , ℓ2 , . . . , ℓt ]. Then “1 + Σ”(ℓ) will be the
list whose nth entry (for 1 ≤ n ≤ t + 1) is
(18)
“1 + Σ”(ℓ)(n) = 1 +
n−1
X
ℓi .
i=1
But this is not by fiat—it is calculated using the formula given in Announcement 3.3.3. We
begin with the following table.
Calculating “1 + Σ”
ℓ ∈ aY
[]
[ℓ1 ]
[ℓ1 , ℓ2 ]
[ℓ1 , ℓ2 , ℓ3 ]
[ℓ1 , . . . , ℓt ]
Cφ,“+” (ℓ) ∈ {aY , cX }
[]
[ℓ1 ] [1]
[ℓ1 , ℓ2 ] [1, 1 + ℓ1 ]
[. . . , ℓ3 ] [. . . , 1 + ℓ1 + ℓ2 ]
Pt−1
[. . . , ℓt ] [. . . , 1 + i=1 ℓi ]
Sφ′ Cφ,“+” (ℓ) ∈ {aX , bX }
[]
[ℓ1 ] [1]
[ℓ1 , ℓ2 ] [1, 1 + ℓ1 ]
[. . . , ℓ3 ] [. . . , 1 + ℓ1 + ℓ2 ]
Pt−1
[. . . , ℓt ] [. . . , 1 + i=1 ℓi ]
Eφ,“+” Sφ′ Cφ,“+” (ℓ) ∈ cX
[1]
[1, 1 + ℓ1 ]
[1, 1 + ℓ1 , 1 + ℓ1 + ℓ2 ]
[. . . , 1 + ℓ1 + ℓ2 + ℓ3 ]
Pt
[. . . , 1 + i=1 ℓi ]
where the last row can be established by induction. The ellipses (. . .) in the later boxes indicate
that the beginning part of the sequence is repeated from the row above, which is a consequence
of the fact that the formulas in Announcement 3.3.3 are historical. We need only calculate
“1 + Σ”(ℓ) = P(φ)(“+”)(ℓ) = Sφ′′ ◦ Eφ,“+” ◦ Sφ′ ◦ Cφ,“+” (ℓ)
"
= 1, 1 + ℓ1 , 1 + ℓ1 + ℓ2 , . . . , 1 +
t
X
i=1
#
ℓi ,
just as in (18).
The Z-shaped propagator “F ib” = P(ψ)(“1 + Σ”) ∈ P(Z) will have the following behavior:
with no inputs, it will output the Fibonacci sequence
“F ib”() = [1, 1, 2, 3, 5, 8, 13 . . .].
Again, this is calculated using the formula given in Announcement 3.3.3. We note first that
since in(Z) = ∅ we have vsetin(Z) = {∗}, so in(Z) = List(vsetin(Z) ) = List({∗}).
26
DYLAN RUPEL AND DAVID I. SPIVAK
As above we provide a table that shows the calculation given the formula in Announcement
3.3.3.
Cψ,“1+Σ” (ℓ)
∈ {cY , dψ }
[]
[1] [1]
[1, 2] [1, 1]
[1, 2, 3] [1, 1, 2]
[1, 2, 3, 5] [1, 1, 2, 3]
ℓ∈∅
[]
[∗]
[∗, ∗]
[∗, ∗, ∗]
[∗, ∗, ∗, ∗]
Calculating “F ib”
Sψ′ Cψ,“1+Σ” (ℓ)
∈ {aY , dψ }
[]
[1] [1]
[1, 1] [1, 2]
[1, 1, 2] [1, 2, 3]
[1, 1, 2, 3] [1, 2, 3, 5]
Eψ,“1+Σ” Sψ′ Cψ,“1+Σ” (ℓ)
∈ {cY , dψ }
[1] [1]
[1, 2] [1, 1]
[1, 2, 3] [1, 1, 2]
[1, 2, 3, 5] [1, 1, 2, 3]
[1, 2, 3, 5, 8] [1, 1, 2, 3, 5]
In the case of a list ℓ ∈ List({∗}) of length t, we have
"
“F ib”(n) = P(ψ)(“1 + Σ”)(ℓ) = 1, 1, 2, 3, . . . , 1 +
t−2
X
i=1
#
“F ib”(i) .
Thus we have achieved our goal. Note that, while unknown to the authors, the fact that
Pt−2
“F ib”(t) = 1 + i=1 “F ib”(i) was known at least as far back as 1891, [Luc]. For us it appeared
not by any investigation, but merely by cordoning off part of our original wiring diagram for
“F ib”,
“F ib”
+
Above in (14) we computed the supplier assignment for the composition WD, ω := ψ◦φ : X →
Z. In case the above tables were unclear, we make one more attempt at explaining how
propagators work by showing a sequence of images with values traversing the wires of ω applied
to “+”. The wires all start with the basepoint on their supply sides, at which point it is shuttled
to the demand sides. It is then processed, again giving values on the supply sides that are again
shuttled to the demand sides. This is repeated once more.
“F ib”–Supply (iter. 1)
+
1
1
“F ib”–Demand (iter. 1)
1
1
+
1
“F ib”–Supply (iter. 2)
1
+
2
1
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
“F ib”–Demand (iter. 2)
1
2
“F ib”–Supply (iter. 3)
1
+
+
3
27
“F ib”–Demand (iter. 3)
2
2
3
2
+
2
3
One sees the first three elements of the Fibonacci sequence [1, 1, 2], as demanded, emerging
from the output wire.
3.5. Proof that the algebra requirements are satisfied by P. Below we prove that P,
as announced, satisfies the requirements necessary for it to be a W-algebra. Unfortunately, the
proof is quite technical and not very enlightening. Given a composition ω = ψ ◦ φ, there is a
correspondence between the wires in ω with the wires in ψ and φ, as laid out in Announcement
2.2.8. The following proof essentially amounts to checking that, under this correspondence, the
way Announcement 3.3.3 instructs us to shuttle information along the wires of ω is in agreement
with the way it instructs us to shuttle information along the wires of ψ and φ.
Theorem 3.5.1. The function P : Ob(W) → Ob(Sets) defined in Announcement 3.3.1 and
the function P : W(Y ; Z) → HomSets (P(Y ); P(Z)) given in Announcement 3.3.3 satisfy the
requirements for P to be a W-algebra.
Proof. We must show that both the identity law and the composition law hold. This will require
several technical lemmas, which for the sake of flow we have included within the current proof.
We begin with the identity law. Let Z = (in(Z), out(Z), vsetZ ) be an object. The supplier
assignment for idZ : Z → Z is given by the identity function
id
sidZ : out(Z) ∐ in(Z) −−−→ in(Z) ∐ out(Z).
Let f ∈ P(Z) = Hist1 (vsetin(Z) , vsetout(Z) ) be a historical propagator. We need to show that
P(idZ )(f ) = f .
Recall the maps
SidZ : SpidZ → DmidZ ,
EidZ ,f : in(Z) → inSpidZ ,
′
Sid
: SpidZ → inDmidZ ,
Z
CidZ ,f : in(Z) → SpidZ ,
′′
Sid
: inSpidZ → out(Z),
Z
from Announcement 3.3.3, where inSpidZ = out(Z).
Lemma 3.5.2. Suppose given a list ℓ ∈ in(Z). We have
(
[]
CidZ ,f (ℓ) =
ℓ, f (∂ℓ)
if |ℓ| = 0,
if |ℓ| ≥ 1.
Proof. We work by induction. The result holds trivially for the empty list. Thus we may
assume that the result holds for ∂ℓ (i.e. that CidZ ,f (∂ℓ) = (∂ℓ, f (∂∂ℓ) holds) and deduce that
28
DYLAN RUPEL AND DAVID I. SPIVAK
′
it holds for ℓ. Note that Sid
([ ]) = [ ] and EidZ ,f ([ ]) = f ([ ]). By the formulas (16) we have
Z
′
CidZ ,f (ℓ) = idin(Z) × (EidZ ,f ◦ Sid
◦ CidZ ,f ◦ ∂) ◦ ∆(ℓ)
Z
′
= idin(Z) × (EidZ ,f ◦ Sid
◦ CidZ ,f ◦ ∂) (ℓ, ℓ)
Z
′
= idin(Z) (ℓ), EidZ ,f ◦ Sid
◦ CidZ ,f ◦ ∂(ℓ)
Z
′
(∂ℓ, f (∂∂ℓ))
= ℓ, EidZ ,f ◦ Sid
Z
= ℓ, EidZ ,f ◦ πinDmid ◦ SidZ (∂ℓ, f (∂∂ℓ))
Z
= ℓ, EidZ ,f ◦ πinDmid (∂ℓ, f (∂∂ℓ))
Z
= ℓ, EidZ ,f (∂ℓ)
= ℓ, f (∂ℓ) .
Expanding the definition of P(idZ )(f )(ℓ) we now complete the proof that the identity law
holds for P:
′
′′
◦ CidZ ,f (ℓ)
◦ EidZ ,f ◦ Sid
P(idZ )(f )(ℓ) = Sid
Z
Z
′
′′
(ℓ, f (∂ℓ))
◦ EidZ ,f ◦ Sid
= Sid
Z
Z
′′
◦ EidZ ,f ◦ πinDmid ◦ SidZ (ℓ, f (∂ℓ))
= Sid
Z
Z
′′
◦ EidZ ,f ◦ πinDmid (ℓ, f (∂ℓ))
= Sid
Z
Z
′′
◦ EidZ ,f (ℓ)
= Sid
Z
′′
= SidZ f (ℓ)
= f (ℓ).
We now move on to the composition law. Let s : m → n be a morphism in Fin. Let
Z ∈ Ob(W) be a black box, let Y : n → Ob(W) be an n-indexed set of black boxes, and let
x : m → Ob(W) be an m-indexed set of black boxes. We must show that the following diagram
of sets commutes:
Q
◦W
/ Wm (X; Z)
Wn (Y ; Z) × i∈n Wmi (Xi ; Y (i))
P
P
Setsn (P(Y ); P(Z)) × i∈n Setsmi (P(Xi ); P(Y (i))) ◦Sets / Setsm (P(X); P(Z))
N
Suppose given ψ : Y → Z and φi : Xi → Y (i) for each i, and let φ = i φi : X → Y . We can
trace through the diagram to obtain P(ψ)◦Sets P(φ) and P(ψ◦W φ), both in Setsm (P(X); P(Z)))
and we want to show they are equal as functions. From here on, we drop the subscripts on ◦− ,
i.e. we want to show P(ψ) ◦ P(φ) = P(ψ ◦ φ).
Let ω = ψ ◦ φ. An element f ∈ P(X) = Hist1 (vsetin(X) , vsetout(X) ) is a 1-historical
propagator, f : in(X) → out(X). We are required to show that the following equation holds
in P(Z):
Q
(19)
?
P(ψ) ◦ P(φ)(f ) = P(ω)(f ).
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
29
Expanding using the definition (17) of P(ψ)◦P(φ)(f ) and P(ω)(f ) we see that this translates
into proving the commutativity of the following diagram:
′′
Sψ
inSpψ
O
/ out(Z) o
′′
Sω
inSp
O ω
Eψ,g
Eω,f
inDmψ
O
inDm
O ω
?
′
Sψ
′
Sω
Cψ,g
Spψ o
in(Z)
Cω,f
/ Spω
where we abbreviated g = P(φ)(f ). To do so, we must prove some technical results (Lemmas
3.5.3, 3.5.4, and 3.5.5) which assert the equality of various demand and supply streams flowing
on the composed wiring diagram ω = ψ ◦ φ.
The ultimate proof of (19) will be inductive in nature. That is, to prove that the result
holds for a nonempty list ℓ of length t ≥ 1, we will assume that it holds for the list ∂ℓ of length
t − 1. More precisely, to prove (19) we will need to know the following equality of functions
in(Z) → inDmω
(20)
Sω′ ◦ Cω,f = (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ Cψ,g
and this is proven by induction on the length of ℓ ∈ in(Z). The base of the induction is clear
after recalling that definition (16) gives Cω,f ([ ]) = [ ], Cφ,f ([ ]) = [ ] and Cψ,g ([ ]) = [ ], and
that Sψ′ and s′ω are 0-historical.
The next three lemmas carry out the induction step and assume the following induction
hypothesis regarding the equality of functions in(Z) → inDmω
(21)
Sω′ ◦ Cω,f ◦ ∂ = (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ Cψ,g ◦ ∂.
Lemma 3.5.3. If we assume that equation (21) holds then the following diagram commutes:
in(Z)
Cψ,P(φ)(f )
Cω,f
/ Spω
in(Z) × inSpφ × DNψ
′′
id×Sφ
×id
Spψ
in(Z) × out(Y ) × DNψ
in other words, we have the following equality between functions in(Z) → Spψ :
(22)
Cψ,P(φ)(f ) = (id × Sφ′′ × id) ◦ Cω,f .
Proof. For convenience we will abbreviate g = P(φ)(f ). It follows from our induction hypothesis
(21), the internal square in the following diagram (when composed with (id × ∂) ◦ ∆ : in(Z) →
30
DYLAN RUPEL AND DAVID I. SPIVAK
in(Z) × in(Z)) commutes:
Cω,f
in(Z)
❖❖❖(id×∂)◦∆
❖❖❖
❖'
id×Cω,f
in(Z) × in(Z)
id×Cψ,g
Cψ,g
/
/
′
id×Sω
in(Z) × Spω
Spψ
id×Eψ,g
in(Z) × inDmφ × DNψ
id×S ′ ×id
φ
id×Cφ,f ×id
in(Z) × inDmψ
id×Eω,f
/
in(Z) × inDmω
in(Z) × inSpω
id×Eφ,f ×δ1
ψ
in(Z) × Spψ
id×S ′
ψ
/
Spω
in(Z) × in(Y ) × DNψ
/
O
in(Z) × Spφ × DNψ
/
in(Z) × inSpφ × DNψ
id×S ′′ ×id
φ
in(Z) × inSpψ
in(Z) × out(Y ) × DNψ
The top square and left square commute by definition of Cω,f and Cψ,f respectively, see
(16). The square Eω,f = Eφ,f × δψ1 commutes also by definition (16). The commutativity of
the bottom-right corner of the diagram translates into the following identity between functions
inDmψ → out(Y ) × DNψ :
Eψ,P(φ)(f ) = (Sφ′′ × id) ◦ (Eφ,f × δψ1 ) ◦ (Sφ′ × id) ◦ (Cφ,f × id).
But this is a direct consequence of the definitions Eψ,P(φ)(f ) = P(φ)(f ) × δψ1 and P(φ)(f ) =
Sφ′′ ◦ Eφ,f ◦ Sφ′ ◦ Cφ,f . It follows that the outer square commutes.
Lemma 3.5.4. If we assume that equation (21) holds, then so does the following equality of
functions in(Z) → inSpφ : 5
(23)
πinSpφ ◦ Cω,f = πinSpφ ◦ Cφ,f ◦ πin(Y ) ◦ Sψ′ ◦ Cψ,g .
Proof. We will use the following three “forgetful” equations,
(24)
(25)
(26)
(27)
Eφ,f ◦ πinDmφ ◦ Sω′ = πinSpφ ◦ (Eφ,f × δψ1 ) ◦ Sω′ ,
πinSpφ ◦ Eω,f ◦ Sω′ ◦ Cω,f ◦ ∂ = πinSpφ ◦ idin(Z) × (Eω,f ◦ Sω′ ◦ Cω,f ◦ ∂) ◦ ∆,
Eφ,f ◦ Sφ′ ◦ Cφ,f ◦ ∂ = πinSpφ ◦ idin(Y ) × (Eφ,f ◦ Sφ′ ◦ Cφ,f ◦ ∂) ◦ ∆,
Sφ′ ◦ Cφ,f ◦ πin(Y ) = πinDmφ ◦ (Sφ′ × id) ◦ (Cφ,f × id).
which are “obvious” in the sense that they are simply a matter of tracking coordinate projections. The proof will go as follows. We apply Eφ,f ◦ πinDmψ to both sides of the assumed
equality (21) and simplify. On the left-hand side we use (24) then the fact that by definition
we have
(28)
Eω,f = Eφ,f × δψ1 ,
then (25), then the definition of Cω,f which we reproduce here:
(29)
Cω,f = idin(Z) × (Eω,f ◦ Sω′ ◦ Cω,f ◦ ∂) ◦ ∆
5It is possible for one to draw a diagram representing this equation as we did in the preceding lemma, however
we did not find such a diagram enlightening in this case.
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
31
to obtain the following equality of functions in(Z) → inSpφ :
Eφ,f ◦ πinDmφ ◦ Sω′ ◦ Cω,f ◦ ∂
=(24) πinSpφ ◦ (Eφ,f × δψ1 ) ◦ Sω′ ◦ Cω,f ◦ ∂
=(28) πinSpφ ◦ Eω,f ◦ Sω′ ◦ Cω,f ◦ ∂
=(25) πinSpφ ◦ idin(Z) × (Eω,f ◦ Sω′ ◦ Cω,f ◦ ∂) ◦ ∆
=(29) πinSpφ ◦ Cω,f .
On the right hand side we use (27), then commute the ∂, then apply (26), and then the definition
of Cφ,f which we reproduce here:
(30)
Cφ,f = idin(Y ) × (Eφ,f ◦ Sφ′ ◦ Cφ,f ◦ ∂) ◦ ∆
to obtain the following equality of functions in(Z) → inSpφ :
Eφ,f ◦ πinDmφ ◦ (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ Cψ,g ◦ ∂
=(27) Eφ,f ◦ Sφ′ ◦ Cφ,f ◦ πin(Y ) ◦ Sψ′ ◦ Cψ,g ◦ ∂
= Eφ,f ◦ Sφ′ ◦ Cφ,f ◦ ∂ ◦ πin(Y ) ◦ Sψ′ ◦ Cψ,g
=(26) πinSpφ ◦ idin(Y ) × (Eφ,f ◦ Sφ′ ◦ Cφ,f ◦ ∂) ◦ ∆ ◦ πin(Y ) ◦ Sψ′ ◦ Cψ,g
=(30) πinSpφ ◦ Cφ,f ◦ πin(Y ) ◦ Sψ′ ◦ Cψ,g .
Combining these computations with the induction hypothesis (21) gives the result:
πinSpφ ◦ Cω,f = Eφ,f ◦ πinDmφ ◦ Sω′ ◦ Cω,f ◦ ∂
= Eφ,f ◦ πinDmφ ◦ (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ Cψ,g ◦ ∂
= πinSpφ ◦ Cφ,f ◦ πin(Y ) ◦ Sψ′ ◦ Cψ,g .
Lemma 3.5.5 (Main Induction Step). If we assume that equation (21), reproduced here
(21)
Sω′ ◦ Cω,f ◦ ∂ = (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ Cψ,g ◦ ∂,
holds, then equation (21) holds without the precomposed ∂, i.e. we have the following equality
of functions in(Z) → inDmω :
Sω′ ◦ Cω,f = (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ Cψ,g .
Proof. To keep the notation from becoming too cluttered we adopt the following convention:
an identity map written as the right hand term of a product will always mean idDNψ , while
an identity map written as the left hand term of a product will mean one of idin(Y ) , idout(Y ) ,
idin(Z) , or idout(Z) , which one should be clear from the context.
The proof will be by cases, we show for each j ∈ inDmω that
πj ◦ (Sφ′ × id ◦ (Cφ,f × id) ◦ Sψ′ ◦ Cψ,g = πj ◦ Sω′ ◦ Cω,f ,
i.e. we show that the two ways of producing internal demand streams agree by checking wire
by wire. Since inDmω = inDmφ ∐ DNψ , there are three main cases to consider: j ∈ DNψ ,
32
DYLAN RUPEL AND DAVID I. SPIVAK
j ∈ inDmφ with sφ (j) ∈ in(Y ), and j ∈ inDmφ with sφ (j) ∈ inSpφ . We go through these in
turn below. Most of the necessary equalities will use that shuttling streams between outputs
and inputs does not change the value stream.
(1) Suppose j ∈ DNψ . We use Lemma 3.5.3 and the fact that the right hand identity maps
are idDNψ to see
πj ◦ (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ Cψ,g
=(22) πj ◦ (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ (id × Sφ′′ × id) ◦ Cω,f
= πj ◦ Sψ′ ◦ (id × Sφ′′ × id) ◦ Cω,f
= πsψ (j) ◦ (id × Sφ′′ × id) ◦ Cω,f .
(∗)
Now there are two cases depending on what has supplied wire j.
• Suppose sψ (j) ∈ in(Z) ∐ DNψ . Notice that in this case (11) gives sψ (j) = sω (j).
Then (∗) above becomes
πsψ (j) ◦ (idin(Z) × Sφ′′ × idDNψ ) ◦ Cω,f = πsψ (j) ◦ Cω,f = πsω (j) ◦ Cω,f
= πj ◦ Sω′ ◦ Cω,f .
• Suppose sψ (j) ∈ out(Y ). In this case (11) gives sφ ◦ sψ (j) = sω (j). Because
Sφ′′ = π
: inSpφ → out(Y ), we see that (∗) simplifies as
sφ
out(Y )
πsψ (j) ◦ (id × Sφ′′ × id) ◦ Cω,f = πsψ (j) ◦ Sφ′′ ◦ πinSpφ ◦ Cω,f
= πsφ ◦sψ (j) ◦ Cω,f = πsω (j) ◦ Cω,f
= πj ◦ Sω′ ◦ Cω,f .
(2) Suppose j ∈ inDmφ and sφ (j) ∈ in(Y ). We will use Lemma 3.5.3 and the equation
πj ◦ (Sφ′ × id) = πsφ (j) .
We will also use the fact that πsφ (j) ◦ (Cφ,f × id) = πsφ (j) , which holds because sφ (j) ∈
in(Y ) and Cφ,f is the identity on in(Y ). With these in hand we compute:
πj ◦ (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ Cψ,g
=(22) πj ◦ (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ (id × Sφ′′ × id) ◦ Cω,f
= πsφ (j) ◦ (Cφ,f × id) ◦ Sψ′ ◦ (id × Sφ′′ × id) ◦ Cω,f
= πsφ (j) ◦ Sψ′ ◦ (id × Sφ′′ × id) ◦ Cω,f ,
(∗∗)
= πsψ ◦sφ (j) ◦ (id × Sφ′′ × id) ◦ Cω,f ,
There are again two cases to consider depending on what has supplied wire j:
• Suppose sψ ◦ sφ (j) ∈ in(Z) ∐ DNψ . Then we get
πsψ ◦sφ (j) ◦ (idin(Z) × Sφ′′ × idDNψ ) = πsψ ◦sφ (j) .
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
33
Now (11) implies the identity sψ ◦ sφ (j) = sω (j) and thus (∗∗) becomes
πsψ ◦sφ (j) ◦ (id × Sφ′′ × id) ◦ Cω,f = πsψ ◦sφ (j) ◦ Cω,f = πsω (j) ◦ Cω,f
= πj ◦ Sω′ ◦ Cω,f .
• Suppose sψ ◦ sφ (j) ∈ out(Y ). Then notice that by (11) we have sω (j) = sφ ◦ sψ ◦
sφ (j) and (∗∗) simplifies as
πsψ ◦sφ (j) ◦ (id × Sφ′′ × id) ◦ Cω,f = πsψ ◦sφ (j) ◦ Sφ′′ ◦ πinSpφ ◦ Cω,f
= πsφ ◦sψ ◦sφ (j) ◦ Cω,f = πsω (j) ◦ Cω,f
= πj ◦ Sω′ ◦ Cω,f .
(3) Suppose j ∈ inDmφ and sφ (j) ∈ inSpφ . As usual we have πj ◦ Sφ′ = πsφ (j) , but noting
that vsetj = vsetsφ (j) , the assumptions on j imply that we have
πj ◦ Sφ′ = πsφ (j) ◦ πinSpφ .
In this case (11) gives sω (j) = sφ (j) and thus by Lemma 3.5.4,
πj ◦ (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ Cψ,g
= πj ◦ Sφ′ ◦ Cφ,f ◦ πin(Y ) ◦ Sψ′ ◦ Cψ,g
= πsφ (j) ◦ πinSpφ ◦ Cφ,f ◦ πin(Y ) ◦ Sψ′ ◦ Cψ,g
=(23) πsφ (j) ◦ πinSpφ ◦ Cω,f
= πsω (j) ◦ Cω,f
= πj ◦ Sω′ ◦ Cω,f .
To complete the proof of Theorem 3.5.1 recall that we have been given morphisms φ : X → Y
and ψ : Y → Z and ω = ψ ◦ φ in W with notation as in Announcement 2.2.8. These have
corresponding supplier assignments sφ , sψ , and sω . Abbreviate g = P(φ)(f ) : in(Y ) → out(Y ).
Consider the following diagram of sets:
34
DYLAN RUPEL AND DAVID I. SPIVAK
′′
Sψ
inSpψ ◆
O
◆◆◆◆
◆◆◆
◆◆◆
Eψ,g
out(Y ) × DNψ o
O
/ out(Z) o
′′
Sφ
×id
′′
Sω
inSpω
♣♣ O
♣
♣
♣
♣♣♣
♣♣♣
Eω,f
inSpφ × DNψ
O
1
g×δψ
inDmψ ◆
O
◆◆◆◆
◆◆◆
◆◆◆
′
Sψ
Spψ o
1
Eφ,f ×δψ
inDmω
♣♣ O
♣
♣
♣♣♣
♣♣♣
♣
♣
S ′ ×id
Cφ,f ×id
′
/ Spφ × DNψ φ / inDmφ × DNψ
Sω
in(Y ) × DNψ
Cψ,g
in(Z)
Cω,f
/ Spω
Recall that our goal was to show that the outermost square commutes. We will see that
each inner square is commutative in the sense that the following equations hold:
Sω′′ = Sψ′′ ◦ (Sφ′′ × id) : inSpφ × DNψ −→ out(Z)
Eψ,g = g × δψ1 : inDmψ −→ out(Y ) × DNψ
g × δψ1 = (Sφ′′ × id) ◦ (Eφ,f × δψ1 ) ◦ (Sφ′ × id) ◦ (Cφ,f × id) : in(Y ) × DNψ → out(Y ) × DNψ
Eφ,f × δψ1 = Eω,f : inDmφ × DNψ −→ inSpω
Sω′ ◦ Cω,f = (Sφ′ × id) ◦ (Cφ,f × id) ◦ Sψ′ ◦ Cψ,g : in(Z) −→ inDmω
The first follows from Lemma 2.2.7, especially (9), and Announcement 2.2.8, especially (11).
The next three follow directly from definitions (16). The last equality has been proven in
Lemma 3.5.5.
It follows that the equation below holds for functions in(Z) −→ out(Z):
P(ω)(f ) = Sω′′ ◦ Eω,f ◦ Sω′ ◦ Cω,f = Sψ′′ ◦ Eψ,g ◦ Sψ′ ◦ Cψ,g = P(ψ)(g) = P(ψ) ◦ P(φ)(f )
Indeed, the left-hand equality and the second-to-last equality are by definition of P on morphisms, as given in (17). The second equality is found by a diagram chase using the six equations
above.
4. Future work
The authors hope that this work can be put to use rather directly in modeling and design
applications. The relationship between the operad W and its algebra P is quite explicitly a
relationship between form and function. The ability to zoom in and out, i.e. to change levels of
abstraction with ease is a facility which we believe is essential to any good theory of the brain,
computer programs, cyber-physical systems, etc.
Below we will discuss some possibilities for future work. We see three major directions in
which to go. The first is to connect this work to other work on wiring diagrams. The second is
to consider applications, e.g. to computer science and cognitive neuroscience. The third is to
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
35
investigate the notion of dependency, or cause and effect, in our formalism. We discuss these
in turn below.
4.1. Connecting to other work on wiring diagrams. While wiring diagrams have been
useful in engineering for many years, there are a few mathematical approaches that should
connect to our own, including [AADF], [BB], [DL], and [Sp2].
The work by [AADF] studies dynamics inside of strongly connected (transitive) networks of
identical units. Their main aim is to relate the dynamics on the network to properties of the
underlying network architecture. The underlying network should be viewed as analogous to a
morphism ψ in W, while the dynamics lying over the network should be viewed as analogous
to the morphism P(ψ). The cells in their networks are considered to have internal states which
collude with the inputs to produce the output of a cell. There exists an algebra over W of
“propagators with internal states” and a retract from this algebra to P, which should allow the
transfer of results of [AADF] to our framework. Arguably one of the main aims of [AADF] is
to introduce a notion of inflation for these networks. A careful comparison, see for example
[AADF, Figure 15] and [AADF, Figure 29], reveals that their inflation procedure is a special
case of the composition of morphisms in W where the black boxes being inserted into a wiring
diagram come from a special class called inflations.
In [BB], the authors investigate reaction networks and in particular stochastic Petri nets.
There, various species (e.g. chemicals or populations) interact in prescribed ways, and the
dynamics of their changing populations are studied. A similar but more complex situation is
studied in [DL]. Both of these papers work with continuous time processes, whereas we work
with discrete time processes. Still, we plan to investigate the relationship between these ideas
in the future.
The only other place, other than the present paper, where operads are explicitly mentioned in
the context of wiring diagrams seems to be [Sp2], where the author studies systems of interacting
relations using an operad T . One might think that an operad functor would appropriately relate
it to the present operad W, but that does not appear to be the case because of the delay nodes
that exist in W but not T . Instead, these two operads need to be compared via a third, in
which delay nodes do not occur, but wires are still directed. We hope to make this precise in
the future.
4.2. Applications, e.g. to computer science and cognitive neuroscience. The authors’
primary purposes in the above work was to formalize what we considered fundamental principles
in the relation of form and function in both computers and brains. On the operad/form level we
are speaking of hierarchical chunking; on the algebra/function level we are speaking of historical
propagators.
One can ask several interesting questions at this point. For example, can we create from W
and P a viable computer programming language? We would hope that the propagators given
by computable functions are closed in, i.e. form a subalgebra of, P. But perhaps one could ask
for more as well. For example, if each transistor in a computer acts like a NOR gate, one could
ask whether or not the subalgebra generated by NOR gates is Turing complete. We conjecture
that something like this is true. If so, we believe our language will provide a simple, grounded,
and useful perspective on the actual operation of computers.
There are also many interesting questions on the neuroscience side that motivated this work.
These essentially amount to a question of “what”. What is a neuron? What is a brain?
What is the relationship between the actions of individual neurons and the brain as a whole?
It is easy to imagine that a neuron is simply a black box where we assign certain multisets
of neurotransmitters to each input and output, the historical propagators would then record
36
DYLAN RUPEL AND DAVID I. SPIVAK
activity patterns of discretized neurons. If this turns out to be the case then the distinction
between neuron and brain becomes blurred, each is simply a black box with some specified
inputs and outputs. From this perspective the questions of how the activity of individual
neurons relates to the activity of a functional brain region or of the entire brain becomes
subsumed by the operad formalism where we can think of each as a different choice of chunking
within a single (massively complex) wiring diagram representing the connections occurring
within an entire brain. Deep questions regarding precisely how the actions of neurons in one
part of the brain influence the activity in other areas will rely on the work of neuroscientists’
understanding of the precise wiring pattern of the brain and remain to be understood. We will
speak more on these questions of dependency within our formalism in the next section.
4.3. Investigating the notion of dependency. Given a propagator with m-inputs and noutputs, one may ask about the relation of dependency between them. When one says that the
outcome of a process is dependent on the inputs, this should mean that changing the inputs
will cause a change in the outputs.
In one form or another, the ability to track changes as they propagate through a network
of processes is one of the basic questions in almost any field of research. Indeed, concern with
notions of cause and effect is an essential characteristic of human thought. Making mathematical
sense of this notion would presumably be immensely valuable. In particular, it should have
direct applications to neuroscience and computer programming disciplines.
It is not clear that there exists a reasonable notion of causality that is algebraic in nature,
i.e. one that can be formulated as a W-algebra receiving a morphism from P. In that case we
may look to other approaches, e.g. that of Bayesian networks as in [Pea] and [Fon]. Whether
Bayesian networks also form an algebra on W or a related operad, and how such an algebra
compares with P should certainly be investigated.
References
[Awo]
S. Awodey. (2010) Category theory. Second edition. Oxford Logic Guides, 52. Oxford University Press,
Oxford.
[AADF] Aguiar, M., Ashwin, P., Dias, A., Field, M. (2010) “Dynamics of coupled cell networks: synchrony,
heteroclinic cycles, and inflation”. Journal of nonlinear science, Springer.
[Bou]
Bourbaki, N. (1972) “Univers”. In M. Artin et al. eds. SGA 4 - vol 1, Lecture Notes in Mathematics
269 (in French). Springer-Verlag pp. 185–217.
[BB]
Baez, J.C., Biamonte, J. (2012). “A Course on Quantum Techniques for Stochastic Mechanics”. Available online, http://arxiv.org/abs/1209.3632.
[BV]
Boardman, M.; Vogt, R. (1973) “Homotopy invariant algebraic structures on topological spaces.” Lecture notes in mathematics 347. Springer-Verlag.
[BW]
Barr M., Wells, C. (1990) Category theory for computing science. Prentice Hall International Series in
Computer Science. Prentice Hall International, New York.
[DL]
Deville, L., Lerman, E. (2013) “Dynamics on networks of manifolds”. Available online:
http://arxiv.org/pdf/1208.1513v2.pdf.
[Fon]
Fong, B. (2013) “Causal Theories: A Categorical Perspective on Bayesian Networks”. Available online
http://arxiv.org/abs/1301.6201
[Lei]
Leinster, T. (2004) Higher Operads, Higher Categories. London Mathematical Society Lecture Note
Series 298, Cambridge University Press.
[Luc]
Lucas, É. (1891), Théorie des nombres (in French) 1, Gauthier-Villars.
[Lur]
Lurie, J. (2012) “Higher algebra”. http://www.math.harvard.edu/˜ lurie/papers/HigherAlgebra.pdf.
[Mac]
(1998) Mac Lane, S. Categories for the working mathematician. Second edition. Graduate Texts in
Mathematics, 5. Springer-Verlag, New York.
[Man]
Manzyuk, O. (2009) “Closed categories vs. closed multicategories”. http://arxiv.org/abs/0904.3137
[May]
May, P. (1972). The geometry of iterated loop spaces. Springer-Verlag.
[NIST] National Institute of Standards and Technology (1993). IDEF0: functional modeling method.
THE OPERAD OF TEMPORAL WIRING DIAGRAMS
[Pea]
[Pen]
[RS]
[Sp1]
[Sp2]
37
Pearl, J. (2009) Causality: Models, reasoning, and inference. Cambridge University Press.
Penrose, R. (2011). Cycles of time: An extraordinary new view of the universe. Random House.
Radul, A.; Sussman, G.J. (2009). “The art of the propagator”. MIT Computer science and artificial
intelligence laboratory technical report.
Spivak, D.I. (2013) Category theory for scientists. http://arxiv.org/abs/1302.6946
Spivak, D.I. (2013) “The operad of wiring diagrams: Formalizing a graphical language for databases,
recursion, and plug-and-play circuits.” ePrint available: http://arxiv.org/abs/1305.0297
Department of Mathematics, Northeastern University, Boston, MA 02115
E-mail address: [email protected]
Department of Mathematics, Massachusetts Institute of Technology, Cambridge MA 02139
E-mail address: [email protected]
| 6cs.PL
|
Simple, Scalable and Accurate Posterior Interval
Estimation
Cheng Li
arXiv:1605.04029v2 [stat.CO] 24 Dec 2016
1
∗1
, Sanvesh Srivastava
†2
, and David B. Dunson
‡3
Department of Statistics and Applied Probability, National University of Singapore
2
Department of Statistics and Actuarial Science, The University of Iowa
3
Department of Statistical Science, Duke University
Abstract
There is a lack of simple and scalable algorithms for uncertainty quantification. Bayesian
methods quantify uncertainty through posterior and predictive distributions, but it is difficult to rapidly estimate summaries of these distributions, such as quantiles and intervals.
Variational Bayes approximations are widely used, but may badly underestimate posterior
covariance. Typically, the focus of Bayesian inference is on point and interval estimates for
one-dimensional functionals of interest. In small scale problems, Markov chain Monte Carlo
algorithms remain the gold standard, but such algorithms face major problems in scaling
up to big data. Various modifications have been proposed based on parallelization and approximations based on subsamples, but such approaches are either highly complex or lack
theoretical support and/or good performance outside of narrow settings. We propose a simple and general posterior interval estimation algorithm, which is based on running Markov
chain Monte Carlo in parallel for subsets of the data and averaging quantiles estimated from
each subset. We provide strong theoretical guarantees and illustrate performance in several
applications.
Key words: Bayesian; Big data; Credible interval; Embarrassingly parallel; Markov chain
Monte Carlo; Quantile estimation; Wasserstein barycenter.
1
Introduction
We propose a posterior interval estimation algorithm for uncertainty quantification in massive data settings in which usual Bayesian sampling algorithms are too slow. Bayesian models
quantify uncertainty via the joint posterior distribution of the model parameters and predictive
distributions of new observations. As joint posteriors and predictives are difficult to visualize
and use in practice, the focus is almost always on posterior summaries of one-dimensional functionals. For example, it is typical to report 95% posterior credible intervals for a variety of onedimensional functionals of interest. In practice, by far the most common approach to estimate
∗
[email protected]
[email protected]
‡
[email protected]
†
1
credible intervals relies on running a Markov chain Monte Carlo algorithm to obtain samples
from the joint posterior, based on which estimating intervals for different one-dimensional functionals is trivial. Traditional Markov chain Monte Carlo algorithms are too slow to be practically
useful in massive data applications. However, given their rich history and broad use, it would
be appealing to be able to incorporate a simple fix-up, which would allow trivial modifications
of existing code, solve the computational bottleneck, and enable provably accurate estimation
of posterior quantiles for any one-dimensional functional of interest.
Current classes of analytic approximations, such as Gaussian/Laplace, variational Bayes
[12, 5, 24], and expectation propagation [29], clearly do not provide a generally useful alternative
to sampling methods in terms of accurate estimation of posterior credible intervals. Hence, in
comparing with the literature, we focus on scalable sampling algorithms. There has been a recent
interest in scaling up Bayesian sampling in general and Markov chain Monte Carlo algorithms
in particular, with many different threads considered. Three of the most successful include (i)
approximating expensive Markov chain Monte Carlo transition kernels with easier to sample
surrogates; (ii) running Markov chain Monte Carlo on a single machine but with different subsets
of the data used as sampling proceeds [28, 17]; and (iii) running Markov chain Monte Carlo in
parallel for different data subsets and then combining [21, 20, 18, 23, 27]. Motivated by our
goal of defining a very simple and theoretically supported algorithm, we focus on embarassingly
parallel Markov chain Monte Carlo following strategy (iii).
The key question in embarassingly parallel Markov chain Monte Carlo is how to combine
samples from the different subset posteriors. If each subset posterior were approximately Gaussian, then weighted averaging is well justified, motivating the consensus Monte Carlo algorithm
[21]. Outside of this restrictive setting, one can instead rely on the product equation representation to combine using kernel smoothing [20] or multi-scale histograms [27]. Such approaches
have theory support in terms of accuracy as the number of samples increases, but rely heavily on
the accuracy of density estimators for the subset posteriors, suffering badly when subset posteriors have even slightly non-overlapping supports. Moreover, the product equation representation
obtained by splitting the prior is not invariant to model reparameterization. An alternative approach is to use data subsamples to define noisy approximations to the full data posterior, and
then take an appropriate notion of geometric center, such as geometric median [18] or mean [23]
of these approximations. These later approaches are invariant to model reparameterization, but
they require a somewhat conceptually and computationally complex combining algorithm.
In this article, we propose a new scalable algorithm for posterior interval estimation. Our
algorithm first runs Markov chain Monte Carlo or any alternative posterior sampling algorithm
in parallel for each subset posterior, with the subset posteriors proportional to the prior multiplied by the subset likelihood raised to the full data sample size divided by the subset sample
size. To obtain an accurate estimate of a posterior quantile for any one-dimensional functional
of interest, we simply calculate the quantile estimates in parallel for each subset posterior and
then average these estimates. Hence, our combining step is completely trivial conceptually and
computationally. We also provide theory justifying the performance of the quantile estimates.
We emphasize that we are not proposing a new Markov chain Monte Carlo algorithm, but we
are instead developing a simple approach to scale up existing algorithms to datasets with large
numbers of observations.
2
Our approach is related to the frequentist Bag of Little Bootstraps [14] and provides a
Bayesian interpretation. Bag of Little Bootstraps divides massive data into small subsets and
obtains bootstrap confidence intervals for a one-dimensional parameter on every subset from
weighted bootstrap samples. Then the confidence interval of the one-dimensional parameter
based on the whole data is constructed by averaging lower and upper bounds of the bootstrap
confidence intervals across all subsets. Similarly, our algorithm averages quantiles from all
subset posteriors. Our theory leads to new insights into Bag of Little Bootstraps, showing that
its confidence intervals correspond to the confidence intervals of the Wasserstein barycenter of
bootstrap distributions across all subsets.
2
2.1
Preliminaries
Wasserstein Distance and Barycenter
Our algorithm is related to the concept of Wasserstein barycenter of subset posteriors [23],
which depends on the notions of Wasserstein distance and Wasserstein barycenter. Suppose
Θ ∈ Rd and kθ1 − θ2 k is the Euclidean distance between any θ1 , θ2 ∈ Θ. For any two measures
ν1 , ν2 on Θ, their Wasserstein-2 distance is defined as
W2 (ν1 , ν2 ) =
Z
1/2
kθ1 − θ2 k dγ(ν1 , ν2 )
,
2
inf
γ∈Γ(ν1 ,ν2 ) Θ×Θ
where Γ(ν1 , ν2 ) is the set of all probability measures on Θ × Θ with marginals ν1 and ν2 ,
R
respectively. If we let P2 (Θ) = ν : Θ kθk2 dν(θ) < ∞ , then the W2 distance is well defined
for every pair of measures in P2 (Θ). The topological space {Θ, P2 (Θ)} is a Polish space, and
the W2 distance metricizes the weak convergence of measures on P2 (Θ). Convergence in W2
distance on P2 (Θ) is equivalent to weak convergence plus convergence of the second moment;
see for example, Lemma 8.3 in [4]. Given N different measures ν1 , . . . , νN in P2 (Θ), their
Wasserstein barycenter is defined as the solution to the following optimization problem [2]:
ν = arg min
N
X
W22 (µ, νj ) ,
(1)
µ∈P2 (Θ) j=1
which can be viewed as the geometric center of the N measures ν1 , . . . , νN .
2.2
Wasserstein Posterior and Posterior Interval Estimation
Consider n observations that are conditionally independent given model parameters and can
be partitioned into K non-overlapping subsets. For ease of presentation, we assume that all
subsets have the same sample size m, such that n = Km. The data in the jth subset are denoted
Xj = {X1j , X2j , . . . , Xmj } for j = 1, . . . , K, and the whole dataset is denoted X = ∪K
j=1 Xj . The
model P (x | θ), or for short Pθ , describes the distribution of X, with parameter θ ∈ Θ ⊆ Rd ,
where d is the dimension of θ. Suppose P (x | θ) is absolutely continuous with respect to
dominating measure λ such that dP (x | θ) = p(x | θ)dλ(x). For theory development, we assume
the existence of a true parameter θ0 ∈ Θ, such that the data X are generated from Pθ0 . Given a
3
prior distribution Π(θ) over Θ with density π(θ), define the overall posterior density of θ given
X and the jth subset posterior density of θ given Xj , j = 1, . . . , K, as
Q K Qm
i=1 p(Xij
j=1
πn (θ | X) = R QK Qm
| θ)π(θ)dθ
j=1
i=1 p(Xij | θ)π(θ)dθ
Qm
{ i=1 p(Xij | θ)}K π(θ)dθ
πm (θ|Xj ) = R Q
,
K
m
Θ { i=1 p(Xij | θ)} π(θ)dθ
Θ
(2)
and we denote their corresponding distribution functions as Πn (θ | X) and Πm (θ | Xj ), respectively. In the definition of subset posterior density πm (θ | Xj ), we have raised the subset
likelihood function to the Kth power. As a stochastic approximation to the overall posterior
πn (θ | X), this modification rescales the variance of each subset posterior given Xj to be roughly
of the same order as the variance of the overall posterior Πn (θ | X), as in [18] and [23]. Based
on (2), [23] runs Markov chain Monte Carlo algorithms on the K subsets in parallel, producing
draws from each Πm (θ | Xj ), j = 1, . . . , K. Empirical estimates of Πm (θ | Xj ) for all K subsets are obtained from the Markov chain Monte Carlo draws, their Wasserstein barycenter is
estimated via a linear program, and used as an approximation of the overall posterior Πn (θ | X).
Suppose we are interested in a scalar parameter ξ = h(θ) ∈ Ξ with h : Θ 7→ Ξ ⊆ R. We
denote the overall posterior for ξ by Πn (ξ | X) and the jth subset posterior for ξ by Πm (ξ | Xj ).
For theory development, we mainly focus on the linear functional ξ = h(θ) = a> θ + b for
some fixed a ∈ Rd and b ∈ R, which includes the individual components in θ as special
cases. We can define the W2 distance and the set of measures P2 (Ξ) on the univariate space
Ξ. If Πm (ξ | Xj ) ∈ P2 (Ξ) for all j = 1, . . . , K, then the one-dimensional Wasserstein posterior
Πn (ξ | X) is defined as the Wasserstein barycenter of Πm (ξ | Xj ) as in (1):
Πn (ξ | X) = arg min
K
X
W22 {µ, Πm (ξ | Xj )} .
(3)
µ∈P2 (Ξ) j=1
In the one-dimensional case, the Wasserstein posterior has an explicit relation with the K subset
posteriors. Let F −1 (u) = inf{x : F (x) ≥ u} be the quantile function of a generic univariate
distribution function F (x). Let F1 and F2 be two univariate distributions in P2 (Ξ), with quantile
functions F1−1 (u) and F2−1 (u), for any u ∈ (0, 1), respectively. Then the W2 distance between
F1 and F2 has an explicit expression by Lemma 8.2 of [4]:
Z
W2 (F1 , F2 ) =
0
1
F1−1 (u) − F2−1 (u)
2
1/2
du
.
Therefore, Πn (ξ | X) in (3) is explicitly related to the subset posteriors Πm (ξ | Xj ) by
−1
Πn (u | X) =
K
1 X −1
Πm (u | Xj ) ,
K
j=1
−1
where Π−1
m (u | Xj ) and Πn (u | X) are the quantile functions of Πm (ξ | Xj ) and Πn (ξ | X),
respectively. This expression for the one-dimensional W2 barycenter has been derived in [2]
from an optimal transport perspective. The relation indicates that for a scalar functional ξ,
the average of subset posterior quantiles produces another quantile function that corresponds
4
exactly to the one-dimensional Wasserstein posterior. Therefore, in our algorithm, to evaluate
the Wasserstein posterior of ξ, we simply take the empirical quantiles based on posterior draws
from each Πm (ξ | Xj ) and then average them over j = 1, . . . , K. Our algorithm is summarized
in Algorithm 1.
Algorithm 1 Posterior Interval Estimation
Input: K subsets of data X1 , . . . , XK , each with sample size m.
Output: Posterior credible intervals [q α/2 , q 1−α/2 ], for α ∈ (0, 1).
For j = 1 to K
# Parallel in K subsets
For t = 1 to T
Draw θtj from Πm (θ | Xj ), using an appropriate posterior sampler.
Calculate ξtj = h(θtj ).
End for
Sort {ξ1j , . . . , ξT j } into ξ(1)j ≤ . . . ≤ ξ(T )j ;
Obtain the empirical α/2 and 1 − α/2 quantiles qα/2,j = ξ(bT α/2c)j
and q1−α/2,j = ξ(bT (1−α/2)c)j , where bxc denotes the integer part of x.
End for
P
1 PK
Set q α/2 = K1 K
j=1 qα/2,j and q 1−α/2 = K
j=1 q1−α/2,j .
Return: [q α/2 , q 1−α/2 ].
3
Main Results
In this section, we develop theory supporting our approach. Under mild regularity conditions, we show that the one-dimensional Wasserstein posterior Πn (ξ | X) is an accurate
approximation to the overall posterior Πn (ξ | X). Essentially, as the subset sample size m
increases, the W2 distance between them diminishes at a faster than parametric rate op (m−1/2 ).
Their biases, variances and quantiles are only different in high orders of m. This rate can be
improved to op (n−1/2 ) when the maximum likelihood estimator of ξ is unbiased. Our results
are improved relative to previous papers relying on combining subset posteriors, such as [18]
and [23], with more detailed description of the limiting behavior of the estimated posterior and
weaker restrictions on the growth rate of the number of subsets K.
Our theory relies on the parametric Bernstein-von Mises theorem. The consensus Monte
Carlo algorithm in [21] also leverages approximate normality in their asymptotic justification
and can be viewed as a different way of averaging subset posteriors. They used weighted averages of subset posterior samples as an approximate sample from the true posterior, where the
weights were taken as the inverse covariance matrices based on each subset posterior samples.
Their weighting strategy relies more heavily on the normality assumption than our strategy of
averaging quantiles. In contrast to the heuristic arguments in [21], we provide formal justification for using normal approximations on a large number of subsets, and quantify the asymptotic
orders of the induced approximation errors.
P
We first define some useful notation. Let `j (θ) = m
i=1 log p(Xij | θ) be the log-likelihood
PK
0
in the j subset, and `(θ) =
j=1 `j (θ) be the overall log-likelihood. Let `j (θ) = ∂`j (θ)/∂θ
5
and `00j (θ) = −∂ 2 `j (θ)/∂θ∂θ> be the first and second derivatives of `j (θ) with respect to θ.
Let θ̂j = arg maxθ∈Θ `j (θ) be the maximum likelihood estimator of θ based on the jth subset
Xj , j = 1, . . . , K. Similarly let θ̂ = arg maxθ∈Θ `(θ) be the maximum likelihood estimator of
P
θ based on the full dataset X. Let θ = K
j=1 θ̂j /K denote the average of maximum likelihood
estimators across subsets.
We make the following assumptions on the data generating process, the prior and the posterior.
Assumption 1. θ0 is an interior point of Θ ∈ Rd , where d is a fixed positive integer and does
not depend on n. Pθ = Pθ0 almost everywhere if and only if θ = θ0 . X contains independent
and identically distributed observations generated from Pθ0 .
Assumption 2. The support of p(x | θ) is the same for all θ ∈ Θ.
Assumption 3. log p(x | θ) is three times differentiable with respect to θ in a neighborhood
Bδ0 (θ0 ) ≡ {θ ∈ Θ : kθ−θ0 k ≤ δ0 } of θ0 , for some constant δ0 > 0. EPθ0 {p0 (X | θ0 )/p(X | θ0 )} =
0. Furthermore, there exists an envelope function M (x) such that supθ∈Bδ (θ0 ) |∂ log p(x | θ)/∂θl1 | ≤
0
M (x), supθ∈Bδ (θ0 ) ∂ 2 log p(x | θ)/∂θl1 ∂θl2 ≤ M (x), supθ∈Bδ (θ0 ) ∂ 3 log p(x | θ)/∂θl1 ∂θl2 ∂θl3 ≤
0
0
M (x) for all l1 , l2 , l3 = 1, . . . , d, for all values of x, and EPθ0 M (X)4 < ∞.
Assumption 4. I(θ) = EPθ0 {−∂ 2 p(X | θ)/∂θ∂θ> } = EPθ0 {∂p(X | θ)/∂θ}{∂p(X | θ)/∂θ}> .
−`001 (θ)/m is positive definite with eigenvalues bounded from below and above by constants, for
all θ ∈ Θ, all values of X1 , and all sufficiently large m.
Assumption
h 5. For any δ > 0, there exists an i> 0 such that
limm→∞ Pθ0 sup|θ−θ0 |≥δ {`1 (θ) − `1 (θ0 )} /m ≤ − = 1.
Assumption 6. The prior density π(θ) is continuous, bounded from above in Θ and bounded
R
below at θ0 . The prior has finite second moment Θ kθk2 π(θ)dθ < ∞.
Assumption 7. Let ψ(X1 ) = EΠm (θ|X1 ) Kmkθ − θ̂1 k2 , where EΠm (θ|X1 ) is the expectation with
respect to θ under the posterior Πm (θ | X1 ). Then there exists an integer m0 ≥ 1, such that
{ψ(X1 ) : m ≥ m0 , K ≥ 1} is uniformly integrable under Pθ0 . In other words,
limC→+∞ supm≥m0 ,K≥1 EPθ0 ψ(X1 )I{ψ(X1 ) ≥ C} = 0, where I(·) is the indicator function.
Assumptions 1-5 are standard and mild regularity conditions on the model P (x | θ), which
are similar to the assumptions of Theorem 8.2 in Chapter 6 of [16] and Theorem 4.2 in [11]
for showing the asymptotic normality of posteriors. Assumption 6 requires the prior to have
a finite second moment, such that with high probability all the posterior distributions are in
the P2 (Θ) space and the W2 distance is well defined. In models with heavy tailed priors, such
as our example in Section D.1, one can replace Assumption 6 by assuming that the posterior
distribution conditional on a fixed number of initial observations has finite second moment;
see Example 8.5 in Chapter 6 of [16] and our Proposition 3 in the Appendix. The uniform
integrability of subset posteriors in Assumption 7 is an extra mild technical assumption that
helps us to generalize the usual Bernstein-von Mises result on the subsets from the convergence
in probability to the convergence in L1 distance. We verify Assumption 7 for normal linear
models and some exponential family distributions in the Appendix. A stronger condition that
6
can replace Assumption 7 is supm≥m0 ,K≥1 EX1 EΠm (θ|X1 ) Kmkθ − θ̂1 k2 < +∞. The following
theorems hold for the one-dimensional Wasserstein posterior defined in (3).
Theorem 1. Suppose Assumptions 1–7 hold and ξ = a> θ + b for some fixed a ∈ Rd and b ∈ R.
−1
Let Iξ (θ0 ) = a> I −1 (θ0 )a
. Let ξ = a> θ + b, ξˆ = a> θ̂ + b. Let Φ(·; µ, Σ) be the normal
distribution with mean µ and variance Σ.
(i) As m → ∞,
h
i
n1/2 W2 Πn (ξ | X) , Φ ξ; ξ, {nIξ (θ0 )}−1 → 0,
h
i
−1
1/2
ˆ
n
W2 Πn (ξ | X) , Φ ξ; ξ, {nIξ (θ0 )}
→ 0,
m1/2 W2 Πn (ξ | X) , Πn (ξ | X) → 0,
where the convergence is in Pθ0 -probability.
(ii) If θ̂1 is an unbiased estimator for θ, so EPθ0 θ̂1 = θ0 , then as m → ∞,
n1/2 W2 Πn (ξ | X) , Πn (ξ | X) → 0
in Pθ0 -probability.
Theorem 1 shows that both the one-dimensional Wasserstein posterior of ξ from combining
K subset posteriors and the overall posterior of ξ based on the full dataset are asymptotically
close in the W2 distance to their respective limiting normal distributions, with slightly different
means and the same variance. Such convergence in the W2 distance implies weak convergence
and convergence of the second moment. Furthermore, the W2 distance between the Wasserstein
and full posteriors converges to zero in probability with rates m1/2 and n1/2 , depending on the
behavior of the maximum likelihood estimator θ̂1 .
Previous asymptotic justifications for embarrassingly parallel Markov chain Monte Carlo
approaches focus on consistency [23] or convergence rates [18], while the above theorem is
stronger in providing a limiting distribution. In addition, our conditions are much weaker in
only requiring the subset sample size m to increase, while imposing no restrictions on the growth
rates of m and K. Hence, the number of subsets K can grow polynomially in n, mimicking
the case in which many computers are available but computational resources per computer are
limited. For example, the theorem allows K = O(nc ), m = O(n1−c ) for any c ∈ (0, 1). Under
this setup, the one-dimensional Wasserstein posterior, the overall posterior and their normal
limits will all converge to θ0 at the same rate of Op (n−1/2 ), and their mutual difference is of
order op (m−1/2 ).
When the maximum likelihood estimator θ̂1 is unbiased, Part (ii) of the theorem provides a
sharper convergence rate of Op (n−1/2 ) compared to the Op (m−1/2 ) rate in Part (i), still with no
explicit restrictions on the growth rates of m and K. When K increases very fast, for example
K ≈ n1/2 and m ≈ n1/2 , the Op (n−1/2 ) rate in Part (ii) is much faster than the Op (n−1/4 ) rate
from Part (i). Moreover, Op (m−1/2 ) is suboptimal since it is the parametric rate based on only
the subset data with size m, while Op (n−1/2 ) is the optimal parametric rate based on the full
data with size n. The reason for the improvement in Part (ii) lies in the high order difference
between the two means ξ and ξˆ of the limiting normal distributions of the one-dimensional
Wasserstein posterior and the overall posterior. When the unbiasedness assumption does not
hold and K increases with n, the difference between the averaged maximum likelihood estimator
7
ξ and the overall maximum likelihood estimator ξˆ is typically of order op (m−1/2 ), which does
not scale in the number of subsets K. However, when all subset maximum likelihood estimators
are unbiased, this difference is reduced by a factor of K 1/2 due to the averaging effect over K
subset posteriors and decreases faster as op (n−1/2 ). Hence, in models having unbiased maximum
likelihood estimators, the one-dimensional Wasserstein posterior achieves high order accuracy
in approximating the overall posterior with a difference op (n−1/2 ).
Independently, [22] has considered a nonparametric generalized linear model and shown a
related Bernstein-von Mises theorem. Besides the difference between the form of models, we
emphasize that our result in Theorem 1 does not rely on the strong requirement of a uniform
normal approximation for all subset posteriors, as used in Shang and Cheng’s paper. Instead,
to show Theorem 1, it is only necessary for the normal approximation to work well on average
among all subset posteriors. As a result, we have no explicit constraint on the growth rate
on the number of subsets K, while their paper needs to control K explicitly depending on the
posterior convergence rate.
Theorem 2. Suppose Assumptions 1–7 hold. Let ξ0 = a> θ0 + b and ξˆ be the same as defined
in Theorem 1. For a generic distribution F on Ξ, let bias(F ) = EF (ξ) − ξ0 and var(F ) be the
variance of F . Let u1 and u2 be two arbitrary fixed numbers such that 0 < u1 < u2 < 1. Then
the following relations hold:
(i) bias Πn (ξ | X) = ξ − ξ0 + op n−1/2 , bias {Πn (ξ | X)} = ξˆ − ξ0 + op n−1/2 ;
1
1
(ii) var Πn (ξ | X) = Iξ−1 (θ0 ) + op n−1 , var {Πn (ξ | X)} = Iξ−1 (θ0 ) + op n−1 ;
n
n
(iii)
sup
−1
−1/2
,
Πn (u | X) − Π−1
n (u | X) = op m
u∈[u1 ,u2 ]
where Op and op are in Pθ0 -probability. Furthermore, if θ̂1 is an unbiased estimator of θ0 , then
bias Πn (ξ | X) − bias {Πn (ξ | X)} = op n−1/2 ;
−1
−1/2
(u
|
X)
=
o
n
.
sup Πn (u | X) − Π−1
p
n
u∈[u1 ,u2 ]
Theorem 2 provides the order for the differences for the bias, the variance and the quantiles
between the one-dimensional Wasserstein posterior and the overall posterior. Essentially the
one-dimensional Wasserstein posterior has an asymptotic bias ξ − ξˆ from the overall posterior,
which is generally of order op (m−1/2 ) and has higher order op (n−1/2 ) when the subset maximum
likelihood estimators are unbiased. The variances of the one-dimensional Wasserstein posterior
and the overall posterior agree in the leading order. Similar to the biases, the difference between
their quantiles has order op (m−1/2 ) in the general case, and improves to a higher order op (n−1/2 )
when the subset maximum likelihood estimators are unbiased. In our algorithm, when we take
K different subset posterior credible intervals and average them, the averages of the lower and
upper quantiles are asymptotically close to the quantiles from the overall posterior in the leading
order. Therefore, Theorem 2 also validates our algorithm in the sense of posterior uncertainty
quantification. We can also account for Monte Carlo errors in approximating subset posteriors
using samples under mild mixing conditions on the subset Markov chains; see Theorem 3 in the
Appendix.
8
4
Experiments
We applied the proposed algorithm in a variety of cases, using consensus Monte Carlo [21],
Wasserstein posterior [23], semiparametric density product [20], and variational Bayes as our
competitors. Posterior summaries from Markov chain Monte Carlo applied to the full data
served as the benchmark for all the comparisons. As our theory guarantees good performance
for very large samples, we focused on simulations with moderate sample sizes. All Markov chain
Monte Carlo algorithms were run for 10,000 iterations. After discarding the first 5000 samples
as burn-in, we retained every fifth sample in all the chains; convergence diagnostics suggested
that every chain had converged to its stationary distribution. We used the combination step
implemented in R package parallelMCMCcombine [19] for consensus Monte Carlo and semiparametric density product methods. We implemented the combination step of our algorithm
in R and of Srivastava et al.’s algorithm in Matlab. All experiments were run on an Oracle Grid
Engine cluster with 2.6GHz 16 core compute nodes. Memory resources were capped at 8GB
for all the methods, except for Markov chain Monte Carlo based on the full data, which had a
maximum memory limit of 16GB.
The accuracy of a density q(θ | X) approximating πn (θ | X) was evaluated using the metric
Z
|q(θ | X) − πn (θ | X)| dθ.
(4)
accuracy {q(θ | X)} = 1 − 21
Θ
This accuracy metric lies in [0, 1], with larger values indicating better performance of q in
approximating πn [8]. In our experiments, we first estimated q(θ | X) and πn (θ | X) based
on the posterior samples using the bkde or bkde2D functions in R package KernSmooth, with
automatic bandwidth selection via dpik [26]. The density estimates were used to compute a
numerical approximation of the integral in (4).
4.1
Linear model with varying dimension
We first evaluated the performance of our proposed algorithm under varying sample size,
dimension, and number of subsets in Bayesian linear models. Let the response, design matrix,
regression coefficients, and random error be denoted as y, X, β, and , where y, ∈ Rn ,
β ∈ Rp×1 , and X ∈ Rn×p . The model assumes that
y = Xβ + , ∼ N (0n×1 , σ 2 In ), β ∼ gdP, σ ∼ Half-t,
(5)
where gdP denotes the generalized double Pareto shrinkage prior of [3] and Half-t is chosen to be
weakly-informative [10]. See Section D.1 in the Appendix for detailed specifications. The priors
on β and σ in (5) are both heavy-tailed with infinite second moments, and therefore do not
satisfy Assumption 6. However, one can verify that conditional on the initial m0 observations
with m0 ≥ p + 4, every subset posterior has finite second moments for both β and σ. The result
is summarized in Proposition 3 in the Appendix.
We applied our approach for inference on β in (5) compared with an asymptotic normal
approximation. We calculated the accuracy of approximations using a full data Gibbs sampler
as the benchmark (Table 1). The first 10% of entries of β were set to ±1 with the remaining 0.
The entries of X were randomly set to ±1 and σ 2 was fixed at 1. We ran 10 replications for n ∈
9
{104 , 105 } and p ∈ {10, 100, 200, 300, 400, 500}. We varied K ∈ {10, 20} and applied Algorithm
1 after running a modification of the Gibbs sampler in (2) for each subset. We considered two
versions of normal approximations for the full posterior. The first version used N (m,
b Vb ) to
approximate the posterior of β, where m
b and Vb are the maximum likelihood estimates of β and
its estimated asymptotic covariance matrix in (5). For the second version, we first obtained
the asymptotic normal approximation of the jth subset posterior as N (m
b j , Vbj ), where m
b j and
b
Vj (j = 1, . . . , K) are the maximum likelihood estimates of β and its estimated asymptotic
covariance matrix for the jth subset. Then we found the W2 barycenter of the K subset normal
approximations, which is again a normal distribution N (m∗ , V ∗ ) [2]. This provides an empirical
verification of Theorem 1. See Section D.1 in the Appendix for details of the Gibbs sampler
and the form of m∗ and V ∗ .
The performance of all the approaches was fairly similar across all simulations and agreed
with our asymptotic theory (Table 1). The results in Table 1 show that the proposed algorithm closely matched the Gibbs sampling results for the full data in terms of uncertainty
quantification. It also performed better than the asymptotic normal approximations in some
cases. When the subset sample size was too small compared to the dimension, such as when
n = 104 , p = 400, K = 20 which has a subset size of only m = 500, we observe poor performance
for both the asymptotic approximations and the proposed approach.
Table 1: Accuracy of approximate posteriors for the non-zero and zero elements of β in (5). The
accuracies are averaged over 10 simulation replications. Normal, the asymptotic normal approximation
based on the full data; PIE, our posterior interval estimation algorithm; NB, the W2 barycenter of K
asymptotic normal approximations of subset posteriors.
p = 10
n=
NB
PIE
NB
PIE
Normal
(K=10)
(K=10)
(K=20)
(K=20)
104
p = 100
n=
105
n=
104
n=
n = 105
non-0s
0s
non-0s
0s
non-0s
0s
non-0s
0s
non-0s
0s
non-0s
0.95
0.94
0.95
0.93
0.94
0.89
0.91
0.97
0.92
0.97
0.96
0.96
0.97
0.96
0.97
0.96
0.95
0.97
0.95
0.97
0.95
0.89
0.90
0.84
0.85
0.90
0.87
0.92
0.84
0.87
0.96
0.95
0.96
0.94
0.95
0.95
0.94
0.96
0.93
0.95
0.95
0.84
0.85
0.75
0.77
0.89
0.83
0.85
0.76
0.78
0.96
0.94
0.95
0.92
0.93
0.95
0.94
0.95
0.92
0.93
n=
NB
PIE
NB
PIE
n=
104
0s
p = 300
Normal
(K=10)
(K=10)
(K=20)
(K=20)
p = 200
105
104
p = 400
n=
105
n=
104
p = 500
n=
105
n=
104
n = 105
0s
non-0s
0s
non-0s
0s
non-0s
0s
non-0s
0s
non-0s
0s
non-0s
0.95
0.80
0.82
0.65
0.67
0.89
0.79
0.81
0.67
0.68
0.96
0.93
0.94
0.91
0.92
0.95
0.93
0.94
0.91
0.91
0.94
0.75
0.77
0.51
0.52
0.89
0.75
0.78
0.52
0.53
0.96
0.93
0.93
0.90
0.91
0.95
0.92
0.93
0.90
0.91
0.94
0.71
0.73
0.31
0.89
0.71
0.74
0.31
0.96
0.92
0.93
0.89
0.90
0.95
0.91
0.93
0.88
0.89
10
4.2
Linear mixed effects model
Linear mixed effects models are widely used to characterize dependence in longitudinal and
nested data structures. Let ni be the number of observations associated with the ith individual,
for i = 1, . . . , s. Let yi ∈ Rni be the responses of the ith individual, Xi ∈ Rni ×p and Zi ∈ Rni ×q
be matrices including predictors having coefficients that are fixed across individuals and varying
across individuals, respectively. Let β ∈ Rp and ui ∈ Rq , respectively, represent the fixed effects
and ith random effect. The linear mixed effects model lets
yi ∼ N (Xi β + Zi ui , σ 2 Ini ),
ui ∼ N (0q , Σ),
i = 1, . . . , s.
(6)
Many software packages are available for Markov chain Monte Carlo-based Bayesian inference
P
in (6), but current implementations become intractable for data with large s and n = si=1 ni .
We applied our algorithm for inference on β and Σ in (6) and compared its performance
with maximum likelihood, consensus Monte Carlo, semiparametric density product, Wasserstein
posterior, and variational Bayes. We set s = 5000, ni = 20 for i = 1, . . . , s, n = 105 , p = 4,
q = 3, β = (−1, 1, −1, 1)> , and σ = 1. The random effects covariance Σ had Σii = i, i = 1, 2, 3,
Σ12 = − 0.56, Σ31 = 0.52, and Σ23 = 0.0025. This matrix included negative, positive, and
small to moderate strength correlations [13]. The simulation was replicated 10 times. The
approximate posterior distributions were obtained using consensus Monte Carlo, semiparametric
density product, Wasserstein posterior, and our algorithm in three steps. First, full data were
randomly partitioned into 20 subsets such that data for each individual were in the same subset.
Second, the Markov chain Monte Carlo sampler for β and Σ in (6) was modified following (2)
and Equation (2) in [21] and implemented in Stan (Version 2.5.0). Finally, the posterior samples
from all the subsets were combined. We used the streamlined algorithm for variational Bayes
[15]. Maximum likelihood produced a point estimate and asymptotic covariance for β, and only
a point estimate for Σ.
We compared the performance of the seven methods for inference on the fixed effects
β, the variances of random effects Σii (i = 1, 2, 3), and the correlations of random effects
ρij = Σij /(Σii Σjj )1/2 (1 ≤ i < j ≤ 3). The correlations are nonlinear functionals of the model
parameters Σ. Maximum likelihood estimator, consensus Monte Carlo, semiparametric density
product, Wasserstein posterior, and our algorithm had excellent performance in estimation of
β (Figure 1), as well as the variances and correlations (Tables 2 and 3 and Figure 2). Uncertainty quantification using consensus Monte Carlo, semiparametric density product, Wasserstein
posterior, and our algorithm closely agreed with Markov chain Monte Carlo based on the full
data. As shown in Figure 1 of the Appendix, variational Bayes was computationally most efficient, but it showed poor accuracy in approximating the posterior of β and the variances, with
underestimation of posterior uncertainty.
4.3
United States natality data
We applied our algorithm to United States natality data on birth weight of infants and
variables related to their mothers’ health [1]. Linear mixed effects models were used for the
covariance in birth weights among siblings. Following the example in [15], we selected the data
for mothers who smoked, had two infants, and had some college education but not a college
11
Table 2: 90% credible intervals for variances and correlations of random effects in simulation for linear
mixed effects model. The upper and lower bounds are averaged over 10 replications.MLE, maximum
likelihood estimator; MCMC, Markov chain Monte Carlo based on the full data; VB, variational Bayes;
CMC, consensus Monte Carlo; SDP, semiparametric density product; WASP, the algorithm in Srivastava
et al. [23]; PIE, our posterior interval estimation algorithm.
MLE
MCMC
VB
CMC
SDP
WASP
PIE
Σ11
Σ22
Σ33
ρ12
ρ13
ρ23
0.99
(0.96, 1.03)
(0.90, 0.96)
(0.96, 1.03)
(0.97, 1.03)
(0.96, 1.03)
(0.96, 1.03)
2.00
(1.94, 2.07)
(1.88, 2.01)
(1.94, 2.08)
(1.95, 2.09)
(1.94, 2.07)
(1.94, 2.07)
3.00
(2.90, 3.10)
(2.84, 3.04)
(2.91, 3.13)
(2.95, 3.14)
(2.90, 3.10)
(2.90, 3.10)
-0.40
(-0.42, -0.38)
(-0.44, -0.41)
(-0.42, -0.38)
(-0.42, -0.38)
(-0.42, -0.38)
(-0.42, -0.38)
0.30
(0.28, 0.32)
(0.29, 0.34)
(0.28, 0.32)
(0.28, 0.32)
(0.28, 0.32)
(0.28, 0.32)
0.00
(-0.03, 0.02)
(-0.03, 0.02)
(-0.02, 0.02)
(-0.03, 0.02)
(-0.03, 0.02)
(-0.03, 0.02)
Table 3: Accuracy of approximate posteriors for variances and correlations of random effects in simulation for linear mixed effects model. The standard deviation of accuracy across 10 replications is in
parentheses. VB, variational Bayes; CMC, consensus Monte Carlo; SC, semiparametric density product;
WASP, the algorithm in Srivastava et al. [23]; PIE, our posterior interval estimation algorithm.
Σ11
VB
CMC
SDP
WASP
PIE
0.11
0.92
0.90
0.95
0.95
(0.02)
(0.03)
(0.04)
(0.02)
(0.02)
Σ22
0.38
0.95
0.92
0.94
0.95
(0.02)
(0.03)
(0.04)
(0.02)
(0.02)
Σ33
0.59
0.93
0.89
0.95
0.96
ρ12
(0.03)
(0.03)
(0.07)
(0.02)
(0.02)
0.45
0.92
0.85
0.94
0.96
(0.02)
(0.04)
(0.09)
(0.02)
(0.02)
ρ13
0.96
0.96
0.95
0.96
0.97
(0.01)
(0.01)
(0.03)
(0.01)
(0.01)
ρ23
0.61
0.90
0.74
0.95
0.96
(0.02)
(0.03)
(0.07)
(0.01)
(0.01)
degree. Detailed information about the variables are in the Appendix. The data set contained
s = 3809 mothers and n = 7618 births. There were 13 variables related to mother’s health.
All these covariates and an intercept were used as fixed effects in (6), so p = 14. The random
effects included mother’s age, gestation period, and number of living infants, so q = 3. We
performed 10 fold cross-validation and randomly split the data into 10 data sets such that
data for siblings belonged to the same training data. We estimated fixed effects and covariance
matrix for random effects as in Section 4.2 using K = 20.
The seven methods in the previous section generally agreed in the inference on fixed effects
(Figure 3), with variational Bayes deviating the furthest. Our algorithm and the algorithm of
[23] differed significantly from variational Bayes, consensus Monte Carlo, and semiparametric
density product in the inference on variances and correlations of random effects (Tables 4
and 5 and Figure 4). Our algorithm and the algorithm of [23] showed better agreement with
Markov chain Monte Carlo based on the full data in estimating the correlations. The 90%
credible intervals from our algorithm included the maximum likelihood estimates of correlations.
Variational Bayes posterior concentrated very close to 0 for every element of the covariance
matrix and significantly underestimated posterior uncertainty. Consensus Monte Carlo and
semiparametric density product methods performed poorly in the inference on random effects
12
β1
●
●
●
●
−0.99
●
●
●
●
●
●
●
●
●
●
●
●
●
●
1.01
●
●
●
1.005
−0.995
β2
●
●
●
●
●
β3
●
●
●
●
●
●
●
●
●
−0.985
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
β4
●
●
●
●
●
●
●
1.01
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
1.005
−0.99
1
−1
0.995
−1.005
●
●
−1.01
●
●
●
●
●
●
●
1
−1
0.995
0.99
●
●
0.985
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
−1.015
−0.995
●
●
●
●
●
MCMC MLE VB CMC SDPWASP PIE
●
●
MCMC MLE VB CMC SDPWASP PIE
MCMC MLE VB CMC SDPWASP PIE
0.985
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.98
0.99
●
−1.005
●
●
●
●
●
MCMC MLE VB CMC SDPWASP PIE
Figure 1: Boxplots of posterior samples for fixed effects in simulation for linear mixed effects model.
MCMC, Markov chain Monte Carlo based on the full data; ML, maximum likelihood estimator; VB,
variational Bayes; CMC, consensus Monte Carlo; SDP, semiparametric density product; WASP, the
algorithm in Srivastava et al. [23]; PIE, our posterior interval estimation algorithm.
●
1.05
●
●
●
●
Σ11
1
●
●
●
●
●
●
●
2.2
●
Σ22
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
ρ21
CMC
SDP WASP
●
●
●
●
●
●
PIE
●
●
●
●
2.8
●
●
●
●
●
●
●
●
●
●
●
MCMC
0.38
●
●
●
●
−0.38
●
0.36
●
●
●
●
●
●
●
●
1.9
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
VB
Σ33
2.9
2
●
MCMC
●
●
●
3
●
0.9
●
3.1
2.1
●
●
●
3.2
●
●
●
●
●
●
●
●
0.95
−0.36
●
●
●
●
●
●
●
●
●
●
ρ31
VB
CMC
●
●
●
●
●
●
●
SDP WASP
●
●
●
●
●
●
●
●
●
PIE
0.04
●
●
●
●
●
●
●
●
MCMC
ρ23
VB
CMC
●
●
●
●
●
●
●
●
0.02
SDP WASP
●
●
●
●
●
●
PIE
●
●
●
●
●
●
●
−0.4
−0.42
0.34
0
0.32
−0.02
●
●
−0.44
●
●
●
●
●
●
●
−0.46
●
●
●
0.3
0.28
●
●
VB
●
●
●
●
●
●
●
●
●
●
MCMC
−0.04
●
●
●
●
−0.06
●
●
●
●
●
●
●
●
VB
CMC
CMC
SDP WASP
PIE
MCMC
VB
CMC
SDP WASP
●
●
●
●
●
●
●
PIE
MCMC
SDP WASP
PIE
Figure 2: Boxplots of posterior samples for variances and correlations of random effects in simulation for
the linear mixed effects model. MCMC, Markov chain Monte Carlo based on the full data; ML, maximum
likelihood estimator; VB, variational Bayes; CMC, consensus Monte Carlo; SDP, semiparametric density
product; WASP, the algorithm in Srivastava et al. [23]; PIE, our posterior interval estimation algorithm.
but were better than variational Bayes. Markov chain Monte Carlo based on the full data
was extremely slow compared to the other methods (see Figure 1 in the Appendix). Taking
into account both the approximation accuracy and the computational efficiency, we concluded
that our proposed algorithm performs better than the competing algorithms in estimating the
covariance matrix of random effects.
4.4
Extension to multi-dimensional parameters
Although Algorithm 1 only applies to one-dimensional functionals, we provide a simple extension to the multi-dimensional case with a numerical illustration. Suppose our goal is to find
the joint posterior of the d-dimensional parameter θ. First, we center and scale the posterior
samples of θ in every subset. Let m
b j and Vbj be the empirical mean and covariance matrix for
P
P
b −1
the jth subset posterior samples {θ1j , . . . , θT j }. Let m
b = K −1 K
b j , Vb −1 = K −1 K
j=1 m
j=1 Vj .
0 = V
b −1/2 (θij − m).
We transform every subset draw θij to θij
b If every subset posterior of θ is
asymptotically normal, then the centered and rescaled version θ0 will be asymptotically standard normal with approximately independent components, since T is large in practice. For
every component of θ0 , we apply Algorithm 1 to combine its K subset posterior samples and
obtain approximations of posterior quantiles for a fine grid of [0, 1]. This leads to accurate approximations of the marginal posteriors of θ0 ; we repeatedly draw samples from these marginals,
13
and then transform back to the original parameter using θ = Vb 1/2 θ0 + m.
b This yields approximate samples from the full posterior of θ, and credible regions can be estimated based on these
samples.
We implemented this generalized algorithm for combining subset posterior samples of all
pairs of variances and covariances in the simulation from Section D.2, and compared the results with consensus Monte Carlo, semiparametric density product, Wasserstein posterior, and
variational Bayes. The accuracies of our algorithm and the algorithm of [23] were higher than
the accuracies of the other three methods for all pairs of variances and covariances (Table 6).
Variational Bayes performed poorly in the estimation of posterior distributions for all the pairs
of variances. We obtained kernel density estimates of the three pairs of covariances in (6) using
the combined posterior samples and the bkde2D function in the KernSmooth R package with a
bandwidth of 0.01 (Figure 5). The kernel density estimates centered very close to the true values
of the covariance pairs. Compared to the algorithm of [23], our algorithm was more efficient,
easier to implement, and robust to the grid-size of quantiles, while having similar accuracy and
stability across all simulation replications.
Table 4: 90% credible intervals for variances and correlations of random effects in United States natality
data analysis. The upper and lower bounds are averaged over 10 folds of cross-validation. MLE, maximum likelihood estimator; MCMC, Markov chain Monte Carlo based on the full data; VB, variational
Bayes; CMC, consensus Monte Carlo; SDP, semiparametric density product; WASP, the algorithm in
Srivastava et al. [23]; PIE, our posterior interval estimation algorithm.
MLE
MCMC
VB
CMC
SDP
WASP
PIE
(dmage, dmage)
(nlbnl, nlbnl)
(gestat, gestat)
(dmage, nlbnl)
(dmage, gestat)
(nlbnl, gestat)
0.135
(0.086, 0.152)
(0.000, 0.000)
(0.010, 0.029)
(0.015, 0.032)
(0.100, 0.163)
(0.098, 0.163)
0.006
(0.003, 0.021)
(0.000, 0.000)
(0.019, 0.051)
(0.018, 0.049)
(0.042, 0.088)
(0.042, 0.088)
0.004
(0.002, 0.004)
(0.000, 0.000)
(0.000, 0.001)
(0.000, 0.001)
(0.002, 0.004)
(0.002, 0.004)
-0.628
(-0.637, 0.283)
(-0.027, 0.028)
(-0.292, 0.043)
(-0.298, -0.029)
(-0.526, 0.066)
(-0.526, 0.073)
-0.955
(-0.959, -0.912)
(-0.028, 0.028)
(-0.574, -0.252)
(-0.656, -0.372)
(-0.928, -0.688)
(-0.927, -0.691)
0.750
(-0.194, 0.728)
(-0.027, 0.027)
(-0.181, 0.159)
(-0.054, 0.228)
(-0.145, 0.464)
(-0.145, 0.465)
Table 5: Accuracy of approximate posteriors for variances and correlations of random effects in United
States natality data analysis. The standard deviation of accuracy across 10 folds of cross-validation is
in parentheses. VB, variational Bayes; CMC, consensus Monte Carlo; SDP, semiparametric density
product; WASP, the algorithm in Srivastava et al. [23]; PIE, our posterior interval estimation algorithm.
VB
CMC
SDP
WASP
PIE
(dmage, dmage)
(nlbnl, nlbnl)
0.00
0.00
0.00
0.72
0.73
0.00
0.15
0.16
0.03
0.03
(0.00)
(0.00)
(0.00)
(0.21)
(0.21)
(0.01)
(0.10)
(0.11)
(0.04)
(0.04)
(gestat, gestat)
0.00
0.00
0.00
0.78
0.78
(0.00)
(0.00)
(0.00)
(0.15)
(0.15)
14
(dmage, nlbnl)
0.08
0.42
0.39
0.72
0.73
(0.02)
(0.07)
(0.08)
(0.11)
(0.11)
(dmage, gestat)
0.00
0.05
0.06
0.22
0.22
(0.00)
(0.16)
(0.18)
(0.14)
(0.14)
(nlbnl, gestat)
0.07
0.33
0.33
0.63
0.63
(0.02)
(0.11)
(0.08)
(0.14)
(0.14)
0.28
gestat
0.35
●
male
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
−0.2
black
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.3
●
●
●
●
●
●
●
●
●
−0.4
0.24
●
0.22
0.25
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.2
●
●
●
●
●
●
−0.5
●
●
●
●
●
●
●
●
●
●
●
●
●
MLE
VB
adeqcode2
CMC
●
●
WASP
PIE
●
●
●
●
●
●
●
●
MCMC
●
●
●
●
●
●
●
SDP
−0.1
MLE
adeqcode3
●
●
●
●
●
●
●
●
●
●
●
−0.1
VB
CMC
SDP
●
●
●
WASP
●
●
●
●
●
●
PIE
●
●
●
●
0.2
−0.3
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
MCMC
MLE
●
●
●
●
●
●
●
CMC
SDP
WASP
SDP
WASP
PIE
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.1
●
●
●
●
●
●
−0.5
VB
CMC
0.05
−0.4
●
●
pretri2
VB
0.15
−0.3
●
●
●
●
MLE
●
●
−0.2
−0.2
MCMC
0.25
●
●
●
−0.6
MCMC
●
●
●
●
0.2
0
●
●
●
●
●
●
●
0.26
−0.3
PIE
MCMC
MLE
0
−0.05
●
●
●
●
●
VB
CMC
SDP
WASP
PIE
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
SDP
WASP
PIE
●
●
MCMC
MLE
VB
CMC
Figure 3: Boxplots of posterior samples for six fixed effects in the United States natality data analysis.
MCMC, Markov chain Monte Carlo based on the full data; ML, maximum likelihood estimator; VB,
variational Bayes; CMC, consensus Monte Carlo; SDP, semiparametric density product; WASP, the
algorithm in Srivastava et al. [23]; PIE, our posterior interval estimation algorithm.
0.2
●
●
Σ(dmage, dmage)
●
●
●
●
●
●
●
●
0.15
●
●
●
●
●
●
●
●
●
●
●
●
0.05
●
●
●
●
●
●
●
●
●
●
0
0.004
●
0.003
●
●
●
●
●
●
●
VB
CMC
ρ(dmage, nbnl)
SDP WASP
●
●
●
●
●
●
MCMC
VB
CMC
0
●
●
●
●
●
●
●
●
●
●
−0.001
SDP WASP
ρ(dmage, gestat)
PIE
MCMC
VB
CMC
ρ(nlbnl, gestat)
0.8
SDP WASP
PIE
●
0.6
●
●
●
●
●
●
●
●
●
●
●
−0.2
●
●
●
●
●
●
●
●
●
●
●
●
●
0
●
●
●
●
●
●
●
●
●
0.001
●
●
●
●
●
●
0
0
PIE
●
●
●
●
●
●
●
●
●
0.002
●
●
●
●
●
●
0.02
●
●
●
●
●
MCMC
0.04
0.005
●
●
●
●
●
●
●
●
0.06
●
●
●
Σ(gestat, gestat)
●
●
0.1
0.08
0.1
0.2
Σ(nlbnl, nlbnl)
0.12
●
●
●
●
0.4
−0.2
−0.4
●
●
●
−0.4
−0.6
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
−0.8
−0.2
●
●
●
●
●
●
●
●
●
−0.8
MCMC
VB
CMC
SDP WASP
PIE
−1
●
●
●
−0.4
MCMC
VB
CMC
SDP WASP
PIE
●
●
●
●
●
●
0
●
●
−0.6
●
0.2
●
MCMC
VB
CMC
SDP WASP
PIE
Figure 4: Boxplots of posterior samples for variances and correlations of random effects in the United
States natality data analysis. MCMC, Markov chain Monte Carlo based on the full data; ML, maximum
likelihood estimator; VB, variational Bayes; CMC, consensus Monte Carlo; SDP, semiparametric density
product; WASP, the algorithm in Srivastava et al. [23]; PIE, our posterior interval estimation algorithm.
5
Conclusion
We have proposed a simple posterior interval estimation algorithm to rapidly and accurately
estimate quantiles of the posterior distributions for different one-dimensional functionals of interest. The algorithm is simple and efficient relative to existing competitors, just averaging
quantile estimates for each subset posterior based on applying existing sampling algorithms in
an embarrassingly parallel manner. There is a fascinating mathematical relationship with the
Wasserstein barycenter of subset posteriors: our algorithm calculates quantiles of the Wasserstein posterior without the optimization step in [23]. The credible intervals from our algorithm
asymptotically approximate those from the full posterior in the leading parametric order. The
quality of approximation is the same even if the subset sample size increases slowly and the
number of subsets increases polynomially fast. Our experiments have demonstrated excellent
performance for linear mixed effects models and linear models with varying dimension.
Although our current theory focuses on parametric models and one-dimensional linear func-
15
Σ12 , Σ13
Σ12 , Σ32
Σ13 , Σ23
0.60
20
20
20
20
20
20
40
40
40
40
60
40
60
60
0.05
80 80
100
100
100
60
60
80
100
100
140
140140
160
200
0.00
220
220
120
160
160160
180
180
180
180
40
0.05
60
80
100
120
140
140
0.55
MCMC
WASP
PIE
0.10
0.10
140
0.00
140
160
140
120
120
220
120 120
100
100
200
200180
80
−0.05
160
160 140
80
60
40
40
20
80
−0.10
20
0.45
80
40
120 120
0.50
100
−0.05
80
60
−0.60
−0.55
−0.50
20
−0.10
−0.60
−0.55
−0.50
0.50
0.55
0.60
Figure 5: Kernel density estimates of the posterior densities for all the covariance pairs in (6) and their
true values (black triangle). var1, var2 represents the two-dimensional posterior density of (var1, var2).
MCMC, Markov chain Monte Carlo based on the full data; WASP, the algorithm in Srivastava et al.
[23]; PIE, our posterior interval estimation algorithm.
Table 6: Accuracy of approximate posteriors for all pairs of variances and covariances in (6). The
standard deviation of accuracy across 10 folds of cross-validation is in parentheses. VB, variational
Bayes; CMC, consensus Monte Carlo; SDP, semiparametric density product; WASP, the algorithm in
Srivastava et al. [23]; PIE, our posterior interval estimation algorithm.
VB
CMC
SDP
WASP
PIE
(Σ11 , Σ22 )
(Σ11 , Σ33 )
(Σ22 , Σ33 )
(Σ12 , Σ13 )
(Σ12 , Σ23 )
(Σ13 , Σ23 )
0.09
0.86
0.82
0.92
0.91
0.10
0.84
0.73
0.90
0.91
0.36
0.84
0.74
0.88
0.92
0.89
0.83
0.82
0.93
0.91
0.90
0.84
0.85
0.94
0.92
0.90
0.88
0.83
0.93
0.91
(0.01)
(0.04)
(0.05)
(0.02)
(0.02)
(0.01)
(0.03)
(0.10)
(0.01)
(0.01)
(0.02)
(0.03)
(0.11)
(0.01)
(0.01)
(0.01)
(0.05)
(0.05)
(0.02)
(0.02)
(0.01)
(0.05)
(0.05)
(0.01)
(0.02)
(0.01)
(0.03)
(0.06)
(0.01)
(0.01)
tionals, the proposed algorithm can be practically implemented for general one-dimensional
functionals for semiparametric and nonparametric models. For example, in simulations not
shown in the paper, we found that our algorithm shows excellent performance for Dirichlet process mixture models for multivariate categorical data [7], and Gaussian process nonparametric
regression. Furthermore, we have provided an extension to the multi-dimensional case. It would
be appealing to develop theory justification in these more complex settings, and to develop guarantees on approximation accuracy for fixed subset sizes and growing numbers of subsets. Also
of interest in future work is to consider algorithms that do not require non-overlapping subsets,
potentially relying on subsampling. Although such modifications can be implemented trivially,
our proof techniques for the combining step in Theorem 1 do not apply directly. Other important extensions include optimal design of subsampling algorithms and extensions beyond
product likelihoods.
16
Appendix
In Section A we provide the detailed technical proofs of Theorem 1 and Theorem 2 in the
main paper. In Section B, we present a theorem that quantifies the Monte Carlo errors in subset
posterior sampling. In Section C, we verify Assumption 7 in the main paper for the normal
linear model and some exponential family distributions. Section D includes further details about
the data analysis in the main paper. In particular, for the heavy tailed priors used in Example
1 in the main paper, we verify a relaxed version of Assumption 6 in Section D.1.
A
A.1
Proofs of Theorem 1 and Theorem 2
Technical Lemmas
Lemma 1. (Villani [25] Theorem 6.15) For two measures P1 , P2 ∈ P2 (Θ), or similarly P2 (Ξ),
W22 (P1 , P2 ) ≤ 2 T V2 (P1 , P2 ),
where the total variation of moments distance [6, 25] is defined as
Z
T V2 (P1 , P2 ) = (1 + kθk2 )d|P1 (θ) − P2 (θ)|.
Θ
Lemma 2. Suppose that Assumptions 1–7 hold. Let θ̂j be a weakly consistent estimator of
θ0 based on the subset Xj such that it solves the score equation `0j (θ̂j ) = 0; θ̂j → θ0 in Pθ0 probability. Let θ̂ be a weakly consistent estimator of θ0 based on the whole X such that it
solves the score equation `0 (θ̂) = 0; θ̂ → θ0 in Pθ0 -probability. Let t = n1/2 (θ − θ̂j ) be the local
parameter for the jth subset, and s = n1/2 (θ − θ̂). Let Πm,t (t | Xj ) be the jth subset posterior
induced by Πm (θ | Xj ) and Πn,s (s | X) be the posterior of s induced by the overall posterior
Πn (θ | X). Then
lim EPθ0 T V2 Πm,t (t | Xj ), Φ t; 0, I −1 (θ0 ) = 0,
(A.1)
m→∞
lim EPθ0 T V2 Πn,s (s | X), Φ s; 0, I −1 (θ0 ) = 0.
(A.2)
m→∞
Remark 1. In comparison, the usual parametric Bernstein-von Mises theorem on the subset
Xj without raising the likelihood to the Kth power gives
lim T V2 Πm (z | Xj ), Φ z; 0, I −1 (θ0 ) = 0,
m→∞
in Pθ0 -probability, where z = m1/2 (θ − θ̂j ). See, for example, Theorem 8.2 in [16] and Theorem
4.2 in [11].
Proof of Lemma 2:
The relation (A.2) in Lemma 2 is the usual Bernstein-von Mises theorem for the overall posterior
Πn (θ | X). The proof of (A.2) follows a related line to the proof of Theorem 4.2 in [11], and
can be treated as a special case of (A.1) with m = n and K = 1. In the following we focus on
the proof of (A.1) in Lemma 2.
Given the independent and identically distributed assumption, we only need to show the
result for a fixed index j. To emphasize the different roles played by the subset sample size m
17
and the number of subsets K, in the following proofs we will write the total sample size n as
Km. We complete the proof in 3 steps. For a generic matrix A or a 3-dimensional array A, we
use kAk to denote its Frobenius norm.
Step 1: Show the existence of weakly consistent estimator θ̂j for θ0 that solves `0j (θ̂j ) = 0.
Given Assumption 3, `0j (θ) is continuously differentiable in an open neighborhood of θ0 and
EPθ0 {p0 (X | θ0 )/p(X | θ0 )} = 0. Therefore, with large Pθ0 -probability, there exists a root for
the equation `0j (θ) = 0 inside the neighborhood that attains the maximum of `(θ). Denote the
root as θ̂j and then θ̂j → θ0 in Pθ0 -probability is a clear consequence of Assumption 5.
Step 2: Show the following convergence as m → ∞ in Pθ0 -probability:
T V2 Πm,t (t | Xj ), Φ t; 0, I −1 (θ0 ) → 0.
(A.3)
We prove this result for a fixed subset Xj , since the data are independent and identically
distributed, and the conclusion is identical for any j = 1, . . . , K. Define the following quantities
t
− `j (θ̂j )
w(t) = `j θ̂j +
(Km)1/2
Z
z
Kw(z)
Cm = e
π θ̂j +
dz.
(Km)1/2
Then based on the expression of Πm (θ | Xj ), with the likelihood raised to the Kth power,
exp {K`j (θ)} π(θ)dθ
.
Θ exp {K`j (θ)} π(θ)dθ
πm (θ | Xj ) = R
The induced posterior density on t = (Km)1/2 (θ − θ̂j ) can be written as
n
o
t
exp{Kw(t)}π θ̂j + (Km)
1/2
πm (t | Xj ) =
.
Cm
Let T = {t = (Km)1/2 (θ − θ̂j ) : θ ∈ Θ}. Define
Kw(t)
t
1 >
2
gm (t) = 1 + ktk
e
π θ̂j +
− exp − t I(θ0 )t π(θ0 ) .
2
(Km)1/2
R
Pθ0
If we can show that T |gm (z)|dz −−→
0 as m → ∞ in Pθ0 -probability, then
Z
1 >
Cm →
exp − z I(θ0 )z π(θ0 )dz = (2π)d/2 {det I(θ0 )}−1/2 π(θ0 )
2
d
R
as m → ∞ in Pθ0 -probability. Hence, for the difference in (A.3), we obtain that
T V2 Πm,t (t | Xj ), Φ t; 0, I −1 (θ0 )
n
o
z
Kw(z) π θ̂ +
Z
e
j
1
1 >
(Km)1/2
2
=
1 + kzk
−
exp − z I(θ0 )z dz
Cm
2
(2π)d/2 {det I(θ0 )}−1/2
T
Z
1
(2π)d/2 {det I(θ0 )}−1/2 π(θ0 )
≤
|gm (z)|dz +
−1 ×
Cm T
Cm
18
1 + kzk2
Z
Rd
(2π)d/2 {det I(θ0 )}−1/2
1 >
exp − z I(θ0 )z dz → 0
2
(A.4)
as m → ∞ in Pθ0 -probability and (A.3) is proved. Therefore it suffices to show that
0 as m → ∞ in Pθ0 -probability.
R
T
|gm (z)|dz →
Divide the domain of the integral into 3 parts: A1 = {z : kzk ≥ δ1 (Km)1/2 }, A2 = {z : δ2 ≤
kzk < δ1 (Km)1/2 }, A3 = {z : kzk < δ2 }, where the constants δ1 , δ2 will be chosen later. Then
Z
Z
Z
Z
|gm (z)|dz.
(A.5)
|gm (z)|dz +
|gm (z)|dz +
|gm (z)|dz ≤
T
We have that
Z
|gm (z)|dz
A1
Z
Kw(z)
2
1 + kzk e
π θ̂j +
≤
A1
A3
A2
A1
z
(Km)1/2
Z
dz +
1 >
1 + kzk2 e− 2 z I(θ0 )z π(θ0 )dz,
(A.6)
A1
and
Z
2
1 + kzk
A1
− 21 z > I(θ0 )z
e
Z
π(θ0 )dz = π(θ0 )
1 >
1 + kzk2 e− 2 z I(θ0 )z dz → 0
kzk≥δ1 (Km)1/2
as m → ∞, because the integral on the whole z ∈ Rd is finite, π(θ0 ) is bounded from above
according to Assumption 6, and K ≥ 1.
Next we bound the first term in (A.6). By Assumption 5 and the weak consistency of θ̂j ,
there exists a constant 1 that depends on δ1 , such that for any z ∈ A1 and all sufficiently large
m, with Pθ0 -probability approaching 1,
`j {θ̂j + z/(Km)1/2 } − `j (θ̂j ) ≤ −m1 .
Furthermore, the weak consistency of θ̂j implies that for all sufficiently large m, with Pθ0 probability approaching 1, kθ̂j k ≤ kθ̂j − θ0 k + kθ0 k ≤ δ0 + kθ0 k. Therefore, as m → ∞ in
Pθ0 -probability,
Z
Kw(z)
z
2
dz
1 + kzk e
π θ̂j +
(Km)1/2
A1
Z
z
2
≤ exp(−Km1 )
1 + kzk π θ̂j +
dz
(Km)1/2
Z
d/2
2
2
≤ exp(−Km1 ) 1 + (Km)
2(kθk + kθ̂j k )π(θ)dθ
Θ
Z
d/2
2
2
2
≤ exp(−Km1 ) 1 + 2(Km)
2kθ0 k + 2δ0 +
kθk π(θ)dθ
→ 0,
(A.7)
Θ
where we have used the finite second moment of π(θ) from Assumption 6 in the last step. Hence,
we have proved that the first integral in (A.5) goes to zero in Pθ0 -probability.
For the second integral in (A.5), by the Taylor series expansion and `0j (θ̂j ) = 0,
z
1 >
− `j (θ̂j ) = −
z I(θ̂j )z + Rm (z)
w(z) = `j θ̂j +
1/2
2K
(Km)
19
(A.8)
1 ∂ 3 `j (θ̃)
Rm (z) ≡
6 ∂θ3
z
z
z
,
,
1/2
1/2
(Km)
(Km)
(Km)1/2
,
where ∂ 3 `j (θ̃)/∂θ3 is a 3-dimensional array and θ̃ satisfies kθ̃ − θ̂j k ≤ z/(Km)1/2 . Since θ̂j → θ0
in Pθ0 -probability, we have kθ̂j − θ0 k < δ0 /3 for all large m with Pθ0 -probability approaching
1, and we choose δ1 ≤ δ0 /3 such that kθ̃ − θ0 k < δ0 for all large m given z ∈ A2 . For every
fixed z ∈ A2 , Rm (z) in (A.8) converges to zero as m → ∞ in Pθ0 -probability, which implies that
on z ∈ A2 , gm (z) → 0 in Pθ0 -probability. Moreover, by Assumption 3, Rm (z) can be further
bounded by
|Rm (z)| ≤
z
d3
6 (Km)1/2
m
3X
M (Xij )
i=1
m
≤
d3 δ1
1 X
d3 δ1
kzk2
M (Xij ) →
kzk2 EPθ0 M (X11 ),
6K
m
6K
i=1
where the last convergence is almost surely in Pθ0 by the strong law of large numbers. Therefore,
we can choose δ1 as
"
#
δ0 3 minθ∈Bδ0 (θ0 ) λ1 {I(θ)}
δ1 = min
,
,
3
4d3 EPθ0 M (X11 )
where λ1 (A) denotes the smallest eigenvalue of a generic matrix A. Assumption 4 indicates that
minθ∈Bδ0 (θ0 ) λ1 {I(θ)} is bounded below by a constant. Thus, in (A.8), the choice of δ1 implies
that for every z ∈ A2 , for all large m with Pθ0 -probability approaching 1,
1 >
z I(θ̂j )z,
4K
1 >
exp{Kw(z)} ≤ exp K −
z I(θ̂j )z + |Rm (z)|
2K
1 >
1 >
≤ exp − z I(θ̂j )z ≤ exp − z I(θ0 )z .
4
8
|Rm (z)| ≤
Therefore for z ∈ A2 , for all large m with Pθ0 -probability approaching 1,
z
1 >
1 >
2
|gm (z)| ≤ 1 + kzk
exp − z I(θ0 )z π θ̂j +
+ exp − z I(θ0 )z π(θ0 )
8
2
(Km)1/2
1
≤ sup π(θ) × 2 1 + kzk2 exp − z > I(θ0 )z .
8
θ∈Θ
R
Hence A2 |gm (z)|dz < +∞, since supθ∈Θ π(θ) < ∞ by Assumption 6. We can choose the conR
stant δ2 sufficiently large, such that A2 |gm (z)|dz is arbitrarily small in Pθ0 -probability.
For the third integral in (A.5), we fix a constant δ2 > 0 and can use the similar Taylor series
expansion above, and notice that when kzk < δ2 , as m → ∞,
sup K|Rm (z)| ≤
kzk<δ2
m
m
d3 δ23
1 X
Kd3 δ23 X
M
(X
)
≤
×
M (Xij ) → 0,
ij
6(Km)3/2 i=1
6(Km)1/2 m i=1
(A.9)
where the last convergence is almost surely in Pθ0 . It follows from (A.8), (A.9), the weak
consistency of θ̃j and the continuity of I(θ) in Bδ0 (θ0 ) that as m → ∞ in Pθ0 -probability,
1
δ2
sup Kw(z) − z > I(θ0 )z ≤ 2 I(θ̂j ) − I(θ0 ) + sup K|Rm (z)| → 0.
2
2
kzk<δ2
kzk<δ2
20
(A.10)
By the continuity of π(θ) in Assumption 6 and the weak consistency of θ̃j , we also have that
t
− π(θ0 ) → 0,
sup π θ̂j +
(A.11)
(Km)1/2
kzk<δ2
as m → ∞ in Pθ0 -probability. Therefore, (A.10) and (A.11) together imply that as m → ∞ in
Pθ0 -probability,
1 >
t
Kw(z)
− e− 2 z I(θ0 )z π(θ0 ) → 0.
sup e
π θ̂j +
1/2
(Km)
kzk<δ2
Hence by the definition of gm (z), as m → ∞ in Pθ0 -probability,
Z
|gm (z)|dz
A3
Z
t
− 12 z > I(θ0 )z
≤
(1 + kzk2 )dz × sup eKw(z) π θ̂j +
−
e
π(θ0 ) → 0.
(Km)1/2
kzk≤δ2
kzk<δ2
This has proved that the right-hand side of (A.5) converges to zero in Pθ0 -probability, and
also completes the proof of (A.3).
Step 3: Show the convergence in L1 as m → ∞. It is clear from the derivation of (A.4) that
T V2 Πm,t (t | Xj ), Φ t; 0, I −1 (θ0 )
o
n
z
Z
eKw(z) π θ̂j + (Km)
1/2
1
1 >
2
−
exp − z I(θ0 )z dz
=
1 + kzk
Cm
2
(2π)d/2 {det I(θ0 )}−1/2
T
Z
Z
2
1 + kzk2
1 >
1/2
≤
1 + n (θ − θ̂j )
π(θ|Xj )dθ +
exp − z I(θ0 )z dz
−1/2
2
Θ
Rd (2π)d/2 {det I(θ0 )}
Z
2
1 + kzk2
1 >
= 1 + EΠm (θ|Xj ) Km θ − θ̂j +
exp − z I(θ0 )z dz
−1/2
2
Rd (2π)d/2 {det I(θ0 )}
In this display, the last term is a finite constant. The middle term is ψ(Xj ) defined in Assumption 7. According to Assumption 7, for any fixed j, {ψ(Xj ) : m ≥ m0 , K ≥ 1} is uniformly inte
grable under Pθ0 . Now since T V2 Πm,t (t | Xj ), Φ t; 0, I −1 (θ0 ) is upper bounded by ψ(Xj )+C
for all m, K and some constant C > 0, we obtain that T V2 Πm,t (t | Xj ), Φ t; 0, I −1 (θ0 ) : m ≥ m0 , K ≥ 1
is also uniformly integrable. This uniform integrability together with the convergence in Pθ0
probability from Step 2 implies the L1 convergence of T V2 Πm,t (t | Xj ), Φ t; 0, I −1 (θ0 ) to
zero.
Similar to the W2 distance, for any l ≥ 1, we can define the Wasserstein-l (Wl ) distance: for
any two measures ν1 , ν2 on Θ, their Wl distance is defined as
Wl (ν1 , ν2 ) =
Z
1/l
kθ1 − θ2 k dγ(ν1 , ν2 )
,
l
inf
γ∈Γ(ν1 ,ν2 ) Θ×Θ
where Γ(ν1 , ν2 ) is the set of all probability measures on Θ × Θ with marginals ν1 and ν2 ,
respectively. The Wl distance on the space Ξ can be similarly defined. The Wl distance between
21
two univariate distributions F1 and F2 is the same as the Ll distance between their quantile
functions (see Lemma 8.2 of [4]):
Z 1
1/l
−1
l
−1
Wl (F1 , F2 ) =
F1 (u) − F2 (u) du
.
0
Lemma 3. Let ξˆj = a> θ̂j + b. Then for any l ≥ 1,
K
h
i
h
i
1 X
Wl Πm (ξ | Xj ), Φ ξ; ξˆj , {nIξ (θ0 )}−1 .
Wl Πn (ξ | X), Φ ξ; ξ, {nIξ (θ0 )}−1 ≤
K
j=1
Proof of Lemma 3:
We use Φ(·) and Φ−1 (·) to denote the cumulative distribution function and the quantile function
of standard normal distribution N (0, 1). From [2], the univariate Wasserstein-2 barycenter
satisfies that for any u ∈ (0, 1),
−1
Πn (u | X) =
K
1 X −1
Πm (u | Xj ) .
K
j=1
Therefore,
h
i
Wl Πn (ξ | X), Φ ξ; ξ, {nIξ (θ0 )}−1
Z 1
h
i l 1/l
−1
−1
−1
u; ξ, {nIξ (θ0 )}
Πn (u | X) − Φ
=
du
0
Z
=
1/l
l
K
X
1
−1/2 −1
Π−1
Φ (u) du
m (u | Xj ) − ξ − {nIξ (θ0 )}
K
1
0
j=1
1
Z
=
0
1
K
K h
X
1/l
l
i
−1/2 −1
ˆ
Π−1
Φ (u) du .
m (u | Xj ) − ξj − {nIξ (θ0 )}
(A.12)
j=1
Define
−1/2 −1
ˆ
rj (u) = Π−1
Φ (u).
m (u | Xj ) − ξj − {nIξ (θ0 )}
then
Wl
h
Πm (ξ | Xj ), Φ ξ; ξˆj , {nIξ (θ0 )}−1
i
Z
=
1
(A.13)
1/l
|rj (u)| du
.
l
0
Since l ≥ 1, we apply Minkowski inequality to the right-hand side of (A.12) and obtain that
1/l
l
K
Z 1 1 X
h
i
−1
Wl Πn (ξ | X), Φ ξ; ξ, {nIξ (θ0 )}
(A.14)
=
rj (u) du
0 K j=1
1/l
K Z 1
K
h
i
1 X
1 X
l
≤
|rj (u)| du
=
Wl Πm (ξ | Xj ), Φ ξ; ξˆj , {nIξ (θ0 )}−1 ,
K
K
0
j=1
j=1
which concludes the proof.
22
Lemma 4. Suppose Assumptions 1–7 hold. Then
ξ − ξˆ = op m−1/2 ,
where op is in Pθ0 probability. Furthermore, if θ̂1 is an unbiased estimator of θ0 , then
ξ − ξˆ = op n−1/2 .
Proof of Lemma 4:
Because of the linearity ξ = a> θ + b, it suffices to show
θ̂ − θ = op m−1/2 ,
(A.15)
under Assumptions 1–7, and
θ̂ − θ = op n−1/2 ,
(A.16)
with the further assumption that θ̂1 is an unbiased estimator of θ0 .
We use the first order Taylor expansion of `0j (θ̂j ) (j = 1, . . . , K) and θ̂ around θ0 :
0 = `0j (θ̂j ) = `0j (θ0 ) + `00j (θ̃j )(θ̂j − θ0 ),
0 = `0 (θ̂) = `0 (θ0 ) + `00 (θ̃)(θ̂ − θ0 ),
where θ̃j is between θ̂j and θ0 , θ̃ is between θ̂ and θ0 , and `00 (θ) =
lead to
PK
00
j=1 `j (θ).
These expansions
`0j (θ0 )
1
θ̂j = θ0 + I −1 (θ0 )`0j (θ0 ) + Zj
,
m
m
−1
1
Zj ≡ − `00j (θ̃j )
− I −1 (θ0 ),
m
1
`0 (θ0 )
,
θ̂ = θ0 + I −1 (θ0 )`0 (θ0 ) + Z
n
n
−1
1
Z ≡ − `00 (θ̃j )
− I −1 (θ0 ).
n
Therefore by the equality `0 (θ0 ) =
PK
0
j=1 `j (θ0 ),
θ − θ̂ =
(A.17)
the difference between θ and θ̂ is
K
1 X `0j (θ0 )
`0 (θ0 )
Zj
−Z
.
K
m
n
(A.18)
j=1
For the second term in (A.18), by the central limit theorem n−1/2 `0 (θ0 ) converges in distribution
to N (0, I(θ0 )), so k`0 (θ0 )/nk = Op (n−1/2 ). Z converges in Pθ0 -probability to zero given the
consistency of θ̂ to θ0 in Lemma 2, so kZk = op (1). Therefore by the Slutsky’s theorem,
Z
`0 (θ0 )
= op n−1/2 .
n
(A.19)
Next we show the first term in (A.18) is of order op (m−1/2 ) under Assumptions 1–7, and is
of order op (n−1/2 ) if furthermore EPθ0 θ̂1 = θ0 .
23
Let Wj = Zj `0j (θ0 )/m1/2 . Then {Wj : j = 1, . . . , K} are independent and identically
P
1/2 ). Since Z → 0
distributed random vectors and the first term in (A.18) is K
j
j=1 Wj /(Km
in Pθ0 -probability as m → ∞, and m−1/2 `0j (θ0 ) = Op (1) as m → ∞, by the Slutsky’s theorem
again, Wj → 0 in Pθ0 -probability. Furthermore, we will show at the end of this proof that
EPθ0 (kW1 k2 ) → 0 as m → ∞. Assuming this is true, by the Markov’s inequality, for any c > 0,
K
mEPθ0
X
W
1
j
−1/2
Pθ0
≥
cm
≤
K
m1/2
j=1
≤
1
K
Wj 2
j=1 m1/2
PK
c2
K
EPθ0 kW1 k2
1 X
2
kW
k
=
→ 0.
E
j
Pθ0
c2 K
c2
j=1
Hence,
(A.15).
PK
= op m−1/2 . This together with (A.18) and (A.19) leads to
1/2 )
j=1 Wj /(Km
If we further assume unbiasedness EPθ0 θ̂1 = θ0 , then from (A.17) we can obtain that
EPθ0 Wj = m−1/2 EPθ0 Zj `0j (θ0 )
n
o
= m1/2 EPθ0 θ̂j − θ0 − m−1 I −1 (θ0 )`0j (θ0 )
= m1/2 EPθ0 θ̂j − θ0 − m−1/2 I −1 (θ0 )EPθ0 `0j (θ0 )
= 0,
for all j = 1, . . . , K. In other words, Wj ’s are centered at zero. Since Xj ’s (j = 1, . . . , K) are
all independent and Wj only depends on Xj , we have EPθ0 Wj>1 Wj2 = 0 for any j1 6= j2 .
We can again apply Markov’s inequality to the first term in (A.18) and obtain that for any
constant c > 0,
K
K
X
X
Wj
1
1
> cn−1/2 = Pθ0
Wj > cK −1/2
Pθ0
1/2
K
K
m
j=1
≤
=
=
KEPθ0
1
K
PK
2
c
K
EP
K 2 c2 θ0
EPθ0 (kW1
c2
j=1
2
j=1 Wj
K
X
j=1
k2 )
kWj k2 +
X
Wj>1 Wj2
j1 6=j2
.
Therefore, assuming that EPθ0 θ̂1 = θ0 and EPθ0 (kW1 k2 ) → 0 as m → ∞, which will be proven
PK
1/2 ) = o n−1/2 . This together with
below, the display above implies that
p
j=1 Wj /(Km
(A.18) and (A.19) leads to (A.16).
Proof of EPθ0 (kW1 k2 ) → 0 as m → ∞:
By Assumption 4, we let λ > 0 be a constant lower bound of the eigenvalues of −`001 (θ)/m for
24
all θ ∈ Θ, all X1 and all sufficiently large m. Then we have
−1
1 00
≤ d1/2 λ−1 ,
− `1 (θ̃1 )
m
I(θ0 )−1 ≤ d1/2 λ−1 ,
(A.20)
where d is the dimension of θ. We have used the property of the Frobenius norm: for a generic
d × d symmetric positive definite matrix A, kA−1 k ≤ d1/2 λ(A−1 ) = d1/2 {λ(A)}−1 , where
λ(A) and λ(A) denotes the largest and the smallest eigenvalues of the matrix A, respectively.
Furthermore, the envelop function condition in Assumption 3 implies that
−
1 00
` (θ̃1 )
m 1
2
≤
m
d2 X
M (Xi1 )2 .
m
(A.21)
i=1
It follows from (A.20) and (A.21) that for all large m,
)−1 (
)
1 ∂ 2 `1 (θ̃1 )
1 ∂ 2 `1 (θ̃1 )
kZ1 k =
−
−
− I(θ0 ) I −1 (θ0 )
m ∂θ∂θ>
m ∂θ∂θ>
(
)
m
1 −2 d2 X
2
2
M (Xi1 ) + kI(θ0 )k dλ−2
≤ dλ
2
m
(
2
2
i=1
m
1 X
≤ c1
M (Xi1 )2 + c2 ,
m
i=1
where c1 , c2 are positive constants that only depend on d, λ, kI(θ0 )k2 .
P
2
2
Now define V1 = c1 m
m−1/2 `01 (θ0 ) . Then we have kW1 k2 ≤ V1 .
i=1 M (Xi1 ) /m + c2
We are going to show that EPθ0 (V1 ) < ∞ and then apply the dominated convergence theorem
to kW1 k2 and conclude that EPθ0 kW1 k2 → 0 since we already have W1 → 0 in Pθ0 -probability.
To see why EPθ0 (V1 ) < ∞, we first apply the Cauchy-Schwarz inequality to V1 and obtain that
m
(
1 X
c1
M (Xi1 )2 + c2
m
EPθ0 (V1 ) = EPθ0
)
m−1/2 `01 (θ0 )
2
i=1
m
1 X
M (Xi1 )2 + c2
c1
m
(
≤ EPθ0
)2 1/2
EPθ0 m−1/2 `01 (θ0 )
4
1/2
.
(A.22)
i=1
Due to Assumption 3, the first term in (A.22) is bounded by
)2
m
1 X
2
EPθ0 c1
M (Xi1 ) + c2
m
i=1
(
)
m
X
1
≤ 2c21 EPθ0
M (Xi1 )4 + 2c22 = 2c21 E M (X)4 + 2c22 < ∞.
m
(
i=1
Pm
0
Now recall that `01 (θ0 ) =
i=1 p (Xi1 | θ0 )/p(Xi1 | θ0 ). Denote the lth component in the
random vector p0 (Xi1 | θ0 )/p(Xi1 | θ0 ) as Uil , such that p0 (Xi1 | θ0 )/p(Xi1 | θ0 ) = (Ui1 , ..., Uid )> .
Then Ui1 l and Ui2 l are independent if i1 6= i2 , due to the independence between Xij ’s. By
EPθ0 {p0 (X | θ0 )/p(X | θ0 )} = 0 in Assumption 3, we have EPθ0 Uil = 0 for all i = 1, . . . , m and
25
4
4 ≤ E
l = 1, . . . , d. From Assumption 3 we have for all l = 1, . . . , d, EPθ0 U1l
Pθ0 M (X) < ∞.
Therefore, the second term in (A.22) can be bounded as
!2 2
!4
d
m
d
m
X
X
X
X
4
1
d
EPθ0 m−1/2 `01 (θ0 ) = 2 EPθ0
Uil
≤ 2 EPθ0
Uil
m
m
i=1
i=1
l=1
l=1
d X
m
X
X
d
2
2
4
EPθ0 Ui1 l EPθ0 Ui2 l
= 2
EPθ0 Uil + 3
m
d
= 2
m
d
≤ 2
m
i1 6=i2
i=1
l=1
d
X
4
mEPθ0 U1l
l=1
d n
X
4
mEPθ0 U1l
+ 3m(m − 1)
+ 3m(m −
2
EPθ0 U1l
4
1)EPθ0 U1l
o
2
≤ 3d
d
X
4
< ∞.
EPθ0 U1l
l=1
l=1
Thus we have shown that both terms on the right-hand side of (A.22) are finite. Therefore,
EPθ0 (V1 ) < ∞ and by the dominated convergence theorem, EPθ0 kW1 k2 → 0.
A.2
Proof of Theorem 1
Proof of Theorem 1(i):
Since ξ = a> θ + b, we can derive the following for subset posteriors in terms of ξ using a change
of variable from θ to ξ in (A.1) of Lemma 2:
h
n
oi
lim EPθ0 T V2 Πm,t (t | Xj ), Φ t; 0, Iξ−1 (θ0 ) = 0,
m→∞
where t = n1/2 (ξ − ξˆj ) is now the local parameter for the jth subset. From the relation between
norms W2 and T V2 in Lemma 1, this directly implies
oi
h
n
lim EPθ0 W22 Πm,t (t | Xj ), Φ t; 0, Iξ−1 (θ0 ) = 0.
m→∞
We further use the rescaling property of the W2 distance and obtain the equivalent form in
terms of the original parameter ξ:
h
i
lim nEPθ0 W22 Πm (ξ | Xj ) , Φ ξ; ξˆj , {nIξ ( θ0 )}−1 = 0.
(A.23)
m→∞
From Lemma 3, we have that for any constant c > 0, as m → ∞,
n
h
i
o
Pθ0 W2 Πn (ξ | X) , Φ ξ; ξ, {nIξ (θ0 )}−1 ≥ cn−1/2
K
1 X
h
i
(i)
≤ Pθ0
W2 Πm (ξ | Xj ), Φ ξ; ξˆj , {nIξ (θ0 )}−1 ≥ cn−1/2
K
j=1
2
K
1 X
h
i
(ii) n
≤ 2 EPθ0
W2 Πm (ξ | Xj ), Φ ξ; ξˆj , {nIξ (θ0 )}−1
c
K
j=1
K
h
i
n X
2
ˆj , {nIξ (θ0 )}−1
E
W
Π
(ξ
|
X
),
Φ
ξ;
ξ
m
j
P
2
θ0
c2 K
j=1
h
i (iv)
n
≤ 2 EPθ0 W22 Πm (ξ | X1 ), Φ ξ; ξˆ1 , {nIξ (θ0 )}−1
→ 0,
c
(iii)
≤
26
where (i) follows from Lemma 3 with l = 2, (ii) uses Markov’s inequality, (iii) comes from the
relation between l1 norm and l2 norm, and (iv) follows from (A.23). This result indicates that
h
i
W2 Πn (ξ | X) , Φ ξ; ξ, {nIξ (θ0 )}−1 = op n−1/2 ,
(A.24)
which shows the first relation in Part (i) of Theorem 1. The second relation in Theorem 1 (i)
h
i
ˆ {nIξ (θ0 )}−1 = op n−1/2 .
W2 Πn (ξ | X) , Φ ξ; ξ,
(A.25)
follows from a similar argument using (A.2) in Lemma 2 for the overall posterior.
From Lemma 4, we have ξ − ξˆ = op m−1/2 . Therefore,
i
h
h
i
ˆ {nIξ (θ0 )}−1 ≤ ξ − ξˆ = op m−1/2 ,
W2 Φ ξ; ξ, {nIξ (θ0 )}−1 , Φ ξ; ξ,
(A.26)
where the first inequality follows because of the definition of W2 distance and the same variance
shared by the two normal distributions.
Finally, by (A.24), (A.25), (A.26) and the triangular inequality, we have
W2 Πn (ξ | X) , Πn (ξ | X)
h
i
h
i
h
i
ˆ {nIξ (θ0 )}−1
≤ W2 Πn (ξ | X) , Φ ξ; ξ, {nIξ (θ0 )}−1 + W2 Φ ξ; ξ, {nIξ (θ0 )}−1 , Φ ξ; ξ,
h
i
ˆ {nIξ (θ0 )}−1 , Πn (ξ | X)
+ W2 Φ ξ; ξ,
≤ op n−1/2 + op m−1/2 + op n−1/2 = op m−1/2 ,
which is equivalent to the third relation in Part (i).
Proof of Theorem 1(ii):
If θ̂1 is an unbiased estimator of θ0 , then by Lemma 4 and the definition of W2 distance, it
follows that
i
h
i
h
−1
−1
−1/2
ˆ
ˆ
.
(A.27)
, Φ ξ; ξ, {nIξ (θ0 )}
≤ ξ − ξ = op n
W2 Φ ξ; ξ, {nIξ (θ0 )}
Applying the triangular inequality to (A.24), (A.25) and (A.27), we obtain that as m → ∞,
W2 Πn (ξ | X) , Πn (ξ | X)
h
i
h
i
h
i
ˆ {nIξ (θ0 )}−1
≤ W2 Πn (ξ | X) , Φ ξ; ξ, {nIξ (θ0 )}−1 + W2 Φ ξ; ξ, {nIξ (θ0 )}−1 , Φ ξ; ξ,
h
i
ˆ {nIξ (θ0 )}−1 , Πn (ξ | X)
+ W2 Φ ξ; ξ,
= op n−1/2 + op n−1/2 + op n−1/2 = op n−1/2 .
Thus the conclusion of Part (ii) follows.
A.3
Proof of Theorem 2
Proof of Theorem 2(i):
[2] have shown that the barycenter Πn (ξ | X) is related to the K subset posteriors Πm (ξ | Xj )
(j = 1, . . . , K) through the quantile function:
−1
Πn (u | X) =
K
1 X −1
Πm (u | Xj ) ,
K
j=1
27
for any u ∈ (0, 1). Also the expectation of a generic univariate distribution F can be calculated
through its quantile functions: if a random variable Y has the cumulative distribution function
R1
F , EF (Y ) = 0 F −1 (u)du. Therefore
bias Πn (ξ | X) = EΠn (ξ|X) (ξ) − ξ0
Z 1
Z 1
K
1 X −1
−1
=
Πn (u | X)du − ξ0 =
Πm (u | Xj ) − ξ0
0
0 K
j=1
1
K
(i)
=
(ii)
=
K Z 1h
X
i
ξˆj + {nIξ (θ0 )}−1/2 Φ−1 (u) + rj (u) du − ξ0
j=1
0
K
K Z
K Z
1 Xˆ
1 X 1
1 X 1
rj (u)du − ξ0 = ξ − ξ0 +
rj (u)du,
ξj +
K
K
K
0
0
j=1
j=1
j=1
R1
where (i) follows from the definition of rj (u) in (A.13), and (ii) makes use of the fact 0 Φ−1 (u)du =
0. |ξˆ − ξ0 | = Op (n−1/2 ) from the central limit theorem. It remains to be shown that
K Z
1 X 1
rj (u)du = op n−1/2 .
K
0
(A.28)
j=1
To see why this is true, we notice that we have derived the following relation in the proof of
Theorem 1:
K
h
i
1 X
W2 Πm (ξ | Xj ), Φ ξ; ξˆj , {nIξ (θ0 )}−1 = op n−1/2 .
K
j=1
But according to the definition of rj (u) in (A.13), by Cauchy-Schwarz inequality,
1/2
K Z
K Z 1
1 X
1 X 1
2
rj (u)du ≤
rj (u)du
K
K
0
0
j=1
1
=
K
K
X
j=1
W2
h
i
−1
−1/2
ˆ
Πm (ξ | Xj ), Φ ξ; ξj , {nIξ (θ0 )}
= op n
,
j=1
which proves (A.28).
On the other hand, for the bias of the overall posterior Πn (ξ | X), we follow a similar
argument as above and obtain that
Z 1
bias {Πn (ξ | X)} = EΠn (ξ|X) (ξ) − ξ0 =
Π−1
n (u | X)du − ξ0
0
Z 1h
Z 1
Z 1
i
−1/2
−1
ˆ
ˆ
=
ξ + {nIξ (θ0 )}
Φ (u) − ξ0 +
r(u)du = ξ − ξ0 +
r(u)du,
0
0
0
−1/2 −1
ˆ
where r(u) = Π−1
Φ (u). Moreover we have
n (u|X) − ξ − {nIξ (θ0 )}
Z 1 h
i2 1/2
ˆ − {nIξ (θ0 )}−1/2 Φ−1 (u) du
r(u)du ≤
Π−1
(u
|
X)
−
ξ
n
0
0
h
i
ˆ {nIξ (θ0 )}−1 = op n−1/2
= W2 Πn (ξ | X), Φ ξ; ξ,
Z
1
28
by Theorem 1. This completes the proof of Part (i).
Proof of Theorem 2(ii):
Similar to the expectation, the variance of a generic univariate distribution F can be calculated
through its quantile functions: if Y ∼ F ,
2
Z
2
∞
2
var(Y ) = E(Y ) − (EY ) =
1
F −1 (u)
=
2
∞
y dF (y) −
−∞
Z
Z
−∞
Z
du −
0
2
ydF (y)
1
2
−1
F (u)du .
0
Therefore,
Z 1
2
Z 1n
o2
−1
−1
Πn (u | X) du −
Πn (u | X)du
var Πn (ξ | X) =
0
0
2
2
Z 1
Z 1 X
K
K
X
1
1
du −
=
Π−1
Π−1
m (u | X)
m (u | X)du
0 K
0 K j=1
j=1
2
Z 1
K
X
1
ξ + {nIξ (θ0 )}−1/2 Φ−1 (u) +
=
rj (u) du
K
0
j=1
2
Z 1
K
X
1
ξ + {nIξ (θ0 )}−1/2 Φ−1 (u) +
rj (u) du
−
K
0
j=1
2
2
Z 1
Z 1 X
K
K Z 1
X
1
1
1
2
−1
=
Φ (u) du +
rj (u) du −
rj (u)du
K
nIξ (θ0 ) 0
0 K
0
j=1
+ 2 {nIξ (θ0 )}−1/2
Z
1
Φ−1 (u)
0
1
K
K
X
j=1
rj (u)du
j=1
2
R1
R1
where we have used the fact 0 Φ−1 (u)du = 0 and 0 Φ−1 (u) du = 1. It remains to be shown
that the other three terms in the display above are of order op (n−1 ).
From (A.14) (with l = 2) and the conclusion of Theorem 1, we have
2
K
h
i
X
1
−1
2
rj (u) du = W2 Πn (ξ | X), Φ ξ; ξ, {nIξ (θ0 )}
= op n−1 .
K
Z
1
0
j=1
By Cauchy-Schwarz inequality, we have
2
2
2
Z 1 X
K Z 1
K
K
1 X
Z 1 1 X
1
rj (u)du
=
rj (u)du
≤
rj (u) du = op n−1 .
0 K
K
0
0 K
j=1
j=1
j=1
Again by Cauchy-Schwarz inequality, we have
2 {nIξ (θ0 )}−1/2
Z
0
1
Φ−1 (u) ×
K
1 X
rj (u)du
K
j=1
29
≤ 2 {nIξ (θ0 )}−1/2
Z
1/2 Z
2
Φ−1 (u) du
1
0
0
= O n−1/2 × op n−1/2 = op n
−1
2 1/2
K
1 X
rj (u) du
K
1
j=1
.
Therefore, we have shown that var Πn (ξ | X) = {nIξ (θ0 )}−1 + op n−1 .
For the variance of Πn (ξ | X), we use the same definition of r(u) as in Part (i) and derive
that
Z 1
2
Z 1
−1
2
−1
Πn (u | X) du −
Πn (u | X)du
var {Πn (ξ | X)} =
0
0
Z 1h
i2
=
ξˆ + {nIξ (θ0 )}−1/2 Φ−1 (u) + r(u) du
0
Z
−
i 2
−1/2 −1
ˆ
ξ + {nIξ (θ0 )}
Φ (u) + r(u) du
1h
0
=
1
+
nIξ (θ0 )
Z
1
r(u)2 du −
Z
0
1
2
Z
r(u)du + 2 {nIξ (θ0 )}−1/2
1
Φ−1 (u)r(u)du.
0
0
Based on the conclusion of Theorem 1 and Cauchy-Schwarz inequality, we have
Z 1
h
i
ˆ {nIξ (θ0 )}−1 = op n−1 ,
r(u)2 du = W22 Πn (ξ | X), Φ ξ; ξ,
0
1
Z
2 Z
r(u)du ≤
0
1
r(u)2 du = op n−1 ,
0
and also
−1/2
1
Z
2 {nIξ (θ0 )}
Φ−1 (u)r(u)du
0
−1/2
Z
1
−1
2
1/2 Z
du
≤ 2 {nIξ (θ0 )}
Φ (u)
0
= O n−1/2 op n−1/2 = op n−1 ,
1
1/2
r(u) du
2
0
which proves var {Πn (ξ | X)} = {nIξ (θ0 )}−1 + op n−1 .
Proof of Theorem 2(iii):
The convergence in W2 distance implies weak convergence. Therefore, it follows from Theorem
1 that in Pθ0 probability, both Πn,s (s | X) and Πn,s (s | X) converge in distribution to normal distributions as m → ∞. The weak convergence also implies the convergence of quantile
functions at any continuous point. Since both Πn (ξ | X) and Πn (ξ | X) are continuous distributions with posterior densities, their quantiles also converge pointwise to the quantiles of their
limiting normal distributions. For any fixed u ∈ (0, 1), as m → ∞, Theorem 1 implies that for
s = n1/2 (ξ − ξ),
−1
Πn,s (u | X) − Φ−1 {u; 0, Iξ−1 (θ0 )} = op (1).
We can make this convergence uniform over all quantiles u ∈ [u1 , u2 ] ⊂ (0, 1). Divide [u1 , u2 ]
into L equally spaced subintervals [u(j) , u(j+1) ] for j = 0, . . . , L − 1 and u(j) = u1 + j(u2 − u1 )/L.
30
For any > 0, since Φ−1 {u; 0, Iξ−1 (θ0 )} is uniformly continuous on [u1 , u2 ], we can pick L
sufficiently large such that
Φ−1 {u(j+1) ; 0, Iξ−1 (θ0 )} − Φ−1 {u(j) ; 0, Iξ−1 (θ0 )} < /2,
for all j = 0, . . . , L − 1. Furthermore, because Φ−1 (·) is continuous everywhere, we can find a
sufficiently large n0 , such that for all n > n0 , all j = 0, . . . , L − 1 with the L chosen above,
−1
Πn,s (u(j) | X) − Φ−1 {u(j) ; 0, Iξ−1 (θ0 )} < /2.
For any u ∈ [u1 , u2 ], we can find a j0 ∈ {0, . . . , L − 1} such that u ∈ [u(j0 ) , u(j0 +1) ]. Therefore
using the monotonicity of quantile functions,
−1
−1
Πn,s (u | X) − Φ−1 {u; 0, Iξ−1 (θ0 )} ≤ Πn,s (u(j0 +1) | X) − Φ−1 {u; 0, Iξ−1 (θ0 )}
−1
≤ Πn,s (u(j0 +1) | X) − Φ−1 {u(j0 +1) ; 0, Iξ−1 (θ0 )}
+ Φ−1 {u(j0 +1) ; 0, Iξ−1 (θ0 )} − Φ−1 {u(j0 ) ; 0, Iξ−1 (θ0 )}
< /2 + /2 < .
and
−1
−1
Πn,s (u | X) − Φ−1 {u; 0, Iξ−1 (θ0 )} ≥ Πn,s (u(j0 ) | X) − Φ−1 {u; 0, Iξ−1 (θ0 )}
−1
≥ Πn,s (u(j0 ) | X) − Φ−1 {u(j0 ) ; 0, Iξ−1 (θ0 )}
+ Φ−1 {u(j0 ) ; 0, Iξ−1 (θ0 )} − Φ−1 {u(j0 +1) ; 0, Iξ−1 (θ0 )}
> − /2 − /2 > −.
Therefore, we have shown that
sup
u∈[u1 ,u2 ]
−1
Πn,s (u | X) − Φ−1 {u; 0, Iξ−1 (θ0 )} = op (1),
which implies that for the quantiles in terms of ξ,
−1
Πn (u | X) − ξ − {nIξ (θ0 )}−1/2 Φ−1 (u) = op n−1/2 .
sup
u∈[u1 ,u2 ]
Similarly for the overall posterior
ˆ − {nIξ (θ0 )}−1/2 Φ−1 (u) = op n−1/2 .
Π−1
(u
|
X)
−
ξ
n
sup
u∈[u1 ,u2 ]
Therefore, by the triangular inequality,
−1
ˆ + op n−1/2 .
Πn (u | X) − Π−1
(u
|
X)
≤
ξ
−
ξ
n
sup
u∈[u1 ,u2 ]
By plugging in the order ξ − ξˆ = op (m−1/2 ) from the proof of Theorem 1, we have
sup
−1
−1/2
.
Πn (u | X) − Π−1
n (u | X) = op m
u∈[u1 ,u2 ]
31
(A.29)
If we further assume that θ̂1 is an unbiased estimator of θ0 , then Lemma 4 says that ξ − ξˆ =
op n−1/2 . Therefore, using the results from Part (i), we have
bias Πn (ξ | X) − bias {Πn (ξ | X)} = ξ − ξˆ + op n−1/2 = op n−1/2 .
Then (A.29) leads to
sup
−1
−1/2
−1/2
−1/2
Πn (u | X) − Π−1
(u
|
X)
≤
o
n
+
o
n
=
o
n
,
p
p
p
n
u∈[u1 ,u2 ]
which completes the proof.
B
Theoretical Results for the Posterior Monte Carlo Errors
In practice, the credible intervals are calculated from the averages of empirical quantiles
from subset posterior samples. In Algorithm 1, suppose that for each j = 1, . . . , K, Π◦j (θ)
and κj (θ, θ0 ) for θ, θ0 ∈ Θ are the initial distribution and the transition kernel for the Markov
chain of the jth subset posterior. {θ1j , . . . , θT j } with sample size T are drawn sequentially with
θ1j ∼ Π◦j (·) and θl+1,j ∼ κj (θlj , ·) for l = 1, . . . , T − 1. ξlj = a> θlj + b for l = 1, . . . , T and
b m (ξ | Xj ) be the empirical distribution of {ξ1j , . . . , ξT j } for j = 1, . . . , K.
j = 1, . . . , K. Let Π
b n (ξ | X) be the Wasserstein barycenter of Π
b m (ξ | X1 ), . . . , Π
b m (ξ | XK ), which can be
Let Π
P
b −1
b −1 (u | X) = K Π
calculated through its quantile function Π
n
j=1 m (u | Xj )/K for all u ∈ (0, 1).
Let L2 {Πm (· | Xj )} for j = 1, . . . , K be the L2 space of functions on Θ such that for any
f ∈ L2 {Πm (· | Xj )}, kf (θ)k2L2 ,j = EΠm (·|Xj ) f 2 (θ) < ∞ almost surely in Pθ0 . We need three
additional assumptions as follows.
Assumption 8. The jth subset posterior Πm (θ | Xj ) is the unique stationary distribution that
R
satisfies the balance condition πm (θ0 | Xj ) = Θ πm (θ | Xj )κj (θ, θ0 )dθ for any θ, θ0 ∈ Θ, where
πm (θ | Xj ) is the density of Πm (θ | Xj ). Furthermore, the Markov chain of each subset posterior
is reversible with the detailed balance condition πm (θ | Xj )κj (θ, θ0 ) = πm (θ0 | Xj )κj (θ0 , θ) for
any θ, θ0 ∈ Θ and all j = 1, . . . , K.
Assumption 9. max1≤j≤K EΠm (·|Xj ) kθk7 is upper bounded by a constant almost surely in Pθ0 .
max1≤j≤K EΠm (·|Xj ) {πj◦ (θ)/πm (θ | Xj )}3 is upper bounded by a constant almost surely in Pθ0 ,
where πj◦ (θ) is the density of Π◦j (θ) for j = 1, . . . , K.
Assumption 10. Every subset posterior Πm (θ | Xj ) (j = 1, . . . , K) is ρ-mixing: there exists
P
a nonnegative constant sequence {ρl }l≥1 decreasing to zero and ∞
l=1 ρl < ∞, such that almost
surely in Pθ0 , for any integer l ≥ 1, any f ∈ L2 {Πm (· | Xj )} and all j = 1, . . . , K,
Eκl (·|θ1j =θ) f (θl+1,j ) − EΠm (·|Xj ) f (θ)
j
L2 ,j
≤ ρl f (θ) − EΠm (·|Xj ) f (θ)
L2 ,j
,
where θl+1,j is the lth draw in the Markov chain with initial draw θ1j , and Eκl (·|θ1j =θ) is the
j
conditional distribution of θl+1,j given θ1j = θ.
Then the following theorem accounts for the Monte Carlo error in the empirical version of
Wasserstein posterior due to finite sample approximations.
32
Theorem 3. Suppose Assumptions 1–10 hold. Then for two arbitrary fixed numbers 0 < u1 <
u2 < 1,
n
o
b n (ξ | X) , Πn (ξ | X) = Op m−1/2 + Op T −1/4 ;
W2 Π
n
o
b n (ξ | X) − bias {Πn (ξ | X)} = op m−1/2 + Op T −1/2 ;
bias Π
o
n
b n (ξ | X) − var {Πn (ξ | X)} = op (n−1 ) + Op T −1/2 ;
var Π
b −1 (u | X) − Π−1 (u | X) = op m−1/2 + Op T −1/2 ,
sup Π
n
n
u∈[u1 ,u2 ]
where Op and op are in Pθ0 -probability. Furthermore, if θ̂1 is an unbiased estimator of θ0 , then
n
o
b n (ξ | X) , Πn (ξ | X) = Op n−1/2 + Op T −1/4 ;
W2 Π
n
o
b n (ξ | X) − bias {Πn (ξ | X)} = op n−1/2 + Op T −1/2 ;
bias Π
−1
−1/2
−1/2
b −1
sup Π
(u
|
X)
−
Π
(u
|
X)
=
o
n
+
O
T
.
p
p
n
n
u∈[u1 ,u2 ]
Proof of Theorem 3:
b m (ξ | Xj )
In this proof, we first establish the key relations between the empirical distribution Π
and the exact continuous subset posterior Πm (ξ | Xj ), using the recent results from [9]. Given
the linear relation ξ = a> θ + b and all the assumptions in Theorem 3,
h
n
oi1+δ
b m (ξ | Xj ), Πm (ξ | Xj )
≤ C1 T −1/2
EΠ◦j W1+δ Π
(A.30)
almost surely in Pθ0 for all j = 1, . . . , K, where 0 ≤ δ ≤ 1, C1 is a constant that only depends on
the sequence {ρl }l≥1 , the constant upper bound of max1≤j≤K EΠm (·|Xj ) kθk7 , and the constant
upper bound of max1≤j≤K EΠm (·|Xj ) {πj◦ (θ)/πm (θ | Xj )}3 in Assumption 9. The expectation
in (A.30) is taken with respect to Π◦j because the first posterior sample θ1j is drawn from the
initial distribution Π◦j . Given Assumptions 8-10, the inequality (A.30) is the consequence of
Theorem 15 of [9] by setting their d = 1, p = 1 + δ, r = 3, q = 7.
b n (ξ | X), we can establish a similar inequality to
For the empirical Wasserstein barycenter Π
Lemma 3: for any l ≥ 1,
Wl
K
h
i
h
i
1 X
−1
b
b m (ξ | Xj ), Φ ξ; ξˆj , {nIξ (θ0 )}−1 ,
Πn (ξ | X), Φ ξ; ξ, {nIξ (θ0 )}
≤
Wl Π
K
j=1
(A.31)
where ξˆ is defined in Lemma 3. Therefore, taking l = 2 in (A.31), we obtain that
h
i
b n (ξ | X), Φ ξ; ξ, {nIξ (θ0 )}−1
EP EΠ◦ ,...,Π◦ W 2 Π
θ0
(i)
≤
1
K
2
K
h
i
1 X
b m (ξ | Xj ), Φ ξ; ξˆj , {nIξ (θ0 )}−1
EPθ0 EΠ◦j W22 Π
K
j=1
(ii)
≤
K
n
o
2 X
b m (ξ | Xj ), Πm (ξ | Xj )
EPθ0 EΠ◦j W22 Π
K
j=1
33
K
h
i
2 X
+
EPθ0 W22 Πm (ξ | Xj ), Φ ξ; ξˆj , {nIξ (θ0 )}−1
K
j=1
(iii)
= O(T −1/2 ) + o(n−1 ),
where (i) is from the relation between l1 and l2 norms, (ii) is from the triangular inequality of the
W2 distance and (x1 + x2 )2 ≤ 2(x21 + x22 ) for x1 , x2 ∈ R, and (iii) follows from
i
h (A.23) and (A.30)
b
with δ = 1. By Markov’s inequality, it is clear that W2 Πn (ξ | X), Φ ξ; ξ, {nIξ (θ0 )}−1
=
op (n−1/2 ) + Op (T −1/4 ).
We can also take l = 1 in (A.31) and obtain that
h
i
b n (ξ | X), Φ ξ; ξ, {nIξ (θ0 )}−1
W1 Π
1
Z
=
0
j=1
1
Z
K
i
1 X h b −1
Πm (u | Xj ) − ξˆj − {nIξ (θ0 )}−1/2 Φ−1 (u) du
K
=
0
K
1 X
{r̂j (u) + rj (u)} du,
K
(A.32)
j=1
−1
b −1
where rj (u) is defined in (A.13) and r̂j (u) = Π
m (u | Xj ) − Πm (u | Xj ).
b n (ξ | X), we have
For the bias of Π
b n (ξ | X)} − bias{Πn (ξ | X)} = E b
(ξ) − EΠn (ξ|X) (ξ)
bias{Π
Πn (·|X)
Z
=
1
b −1 (u | X)du −
Π
n
0
≤
1
K
1
Z
Z
−1
Πn (u | X)du ≤
0
K Z 1
X
j=1
|r̂j (u)| du =
0
1
K
0
K
X
1
K
1 X
r̂j (u) du
K
j=1
n
o
b m (ξ | Xj ), Πm (ξ | Xj ) .
W1 Π
(A.33)
j=1
By Markov’s inequality and (A.30) with δ = 0, for any c > 0,
K
n
o
X
1
b m (ξ | Xj ), Πm (ξ | Xj ) > cT −1/2
P
W1 Π
K
j=1
n
o
P
K
1
b
◦
E
E
W
Π
(ξ
|
X
),
Π
(ξ
|
X
)
1
m
j
m
j
P
Πj
θ0
j=1
K
C1
≤
.
≤
−1/2
c
cT
Therefore, we have shown that
b n (ξ | X)} − bias{Πn (ξ | X)} = Op (T −1/2 ),
bias{Π
K Z
1 X 1
|r̂j (u)| du = Op (T −1/2 ).
K
0
j=1
Together with Theorem 2, we conclude that
b n (ξ | X)} − bias{Πn (ξ | X)}
bias{Π
b n (ξ | X)} − bias{Πn (ξ | X)}
≤ bias{Πn (ξ | X)} − bias{Πn (ξ | X)} + bias{Π
34
(A.34)
= Op (m−1/2 ) + Op (T −1/2 ).
Furthermore, if θ̂1 is unbiased for θ, then
b n (ξ | X)} − bias{Πn (ξ | X)} = Op (n−1/2 ) + Op (T −1/2 ).
bias{Π
The results for quantiles can be derived similarly and therefore the proofs are omitted here.
b n (ξ | X). Similar to the derivation in
Next we derive the rate for the posterior variance of Π
the proof of Theorem 2(ii), we can obtain the following equality:
n
o
b n (ξ | X)
var Π
2
2
Z 1
K
K Z 1
X
X
1
1
1
=
+
{r̂j (u) + rj (u)} du −
{r̂j (u) + rj (u)}du
nIξ (θ0 )
K
K
0
0
j=1
1
Z
+ 2 {nIξ (θ0 )}−1/2
j=1
Φ−1 (u)
0
K
1 X
{r̂j (u) + rj (u)}du.
K
(A.35)
j=1
We bound the last three terms in the display above. It is clear that by Cauchy-Schwarz inequality, the third term is upper bounded by the second term. For the second term, we have
2
Z 1
K
X
1
{r̂j (u) + rj (u)} du
K
0
j=1
2
2
Z 1 X
Z 1 X
K
K
1
1
≤2
r̂j (u) du + 2
rj (u) du
0 K j=1
0 K j=1
o
h
i
n
b m (ξ | Xj ), Πm (ξ | Xj ) + 2W 2 Πn (ξ | X), Φ ξ; ξ, {nIξ (θ0 )}−1
≤ 2W22 Π
2
= Op (T −1/2 ) + op (n−1 ),
(A.36)
where the last relation follows from (A.24) and applying Markov’s inequality to (A.30).
For the last term in (A.35), we have the following bound:
2 {nIξ (θ0 )}
−1/2
1
Z
Φ−1 (u)
0
≤ 2 {nIξ (θ0 )}−1/2
≤ 2 {nIξ (θ0 )}
−1/2
j=1
Z
1
0
K
1 X
{r̂j (u) + rj (u)}du
K
Z 1
K
K
X
X
1
1
r̂j (u) du +
Φ−1 (u)
rj (u) du
Φ−1 (u)
K
K
0
j=1
Z
1
Φ−1 (u)
1+1/δ
j=1
δ/(1+δ) ( Z
du
0
0
1
K
1 X
r̂j (u)
K
j=1
)1/2 !
K
X
2
1
+
Φ−1 (u)
rj (u) du
K
0
0
j=1
(
n
o
2(1+δ)/2
1+δ
−1/2
b n (ξ | X), Πn (ξ | X)
√
≤ 2 {nIξ (θ0 )}
Γ
W1+δ Π
2
π
)
h
i
+ W2 Πn (ξ | X), Φ ξ; ξ, {nIξ (θ0 )}−1
Z
1
2
1/2 ( Z
du
1
35
1+δ
)1/(1+δ)
du
2
= √ {nIξ (θ0 )}−1/2 2(1+δ)/2 Γ
π
1+δ
2
n
o
b n (ξ | X), Πn (ξ | X) + op (n−1 ),
W1+δ Π
(A.37)
where in the second inequality we used the Hölder’s inequality and the Cauchy-Schwarz inequality, and the last step is from Theorem 1. By (A.30), almost surely in Pθ0 ,
K
o
o
n
n
X
b n (ξ | X), Πn (ξ | X) ≤ 1
b m (ξ | Xj ), Πm (ξ | Xj )
EΠ◦1 ,...,Π◦K W1+δ Π
EΠ◦j W1+δ Π
K
j=1
(i)
≤
1
K
K
X
EΠ◦j
h
n
oi1+δ 1/(1+δ)
b m (ξ | Xj ), Πm (ξ | Xj )
W1+δ Π
≤ C1 T −1/{2(1+δ)} ,
(A.38)
j=1
where (i) is from 0 ≤ δ ≤ 1 and Jensen’s inequality. Now we set δ = min{1, log n/(2 log T )} and
derive from (A.39) that
n
o
2
1+δ
b n (ξ | X), Πn (ξ | X)
W1+δ Π
EΠ◦1 ,...,Π◦K √ {nIξ (θ0 )}−1/2 2(1+δ)/2 Γ
2
π
2
≤ √ {nIξ (θ0 )}−1/2 × 2Γ(1) × C1 T −1/{2(1+δ)}
π
1
1
−1/2
≤ 4C1 {Iξ (θ0 )}
exp − log n −
log T
2
2(1 + δ)
1
δ
1
−1/2
log T
= 4C1 {Iξ (θ0 )}
exp − log T − log n +
2
2
2(1 + δ)
1
δ
≤4C1 {Iξ (θ0 )}−1/2 T −1/2 exp − log n + log T
2
2
1
≤4C1 {Iξ (θ0 )}−1/2 T −1/2 exp − log n = o(T −1/2 ).
4
Hence, by Markov’s inequality, the right-hand side of (A.37) can be bounded by
2 {nIξ (θ0 )}−1/2
Z
0
1
Φ−1 (u)
K
1 X
{r̂j (u) + rj (u)}du = op (T −1/2 ) + op (n−1 ).
K
(A.39)
j=1
Now we combine (A.35), (A.36) and (A.39) and conclude that
n
o
1
b n (ξ | X) =
+ Op (T −1/2 ) + op (n−1 ) + op (T −1/2 ) + op (n−1 )
var Π
nIξ (θ0 )
1
=
+ op (n−1 ) + Op (T −1/2 ).
nIξ (θ0 )
If we compare this with the results in Theorem 2, we obtain that
n
o
b n (ξ | X) − var {Πn (ξ | X)} = op (n−1 ) + Op T −1/2 .
var Π
This concludes the proof of Theorem 3.
C
Justification of Assumption 7
In this section, we verify Assumption 7 for two special examples: the normal linear model
and some exponential family distributions. Without loss of generality, all the samples considered in this section refer to the first subset sample X1 in Assumption 7.
36
1. Normal Linear Model
We consider the following normal linear model based on independent and identically distributed observations:
yi = Zi> β + εi , εi ∼ N 0, σ 2 , i = 1, . . . , m,
(A.40)
where dim(β) = p and εi ’s are independent. We write y = (y1 , . . . , ym )> , Z = (Z1 , . . . , Zm )> ,
ε = (ε1 , . . . , εm )> , and the true parameter is θ0 = (β0> , σ02 )> . We impose the following conjugate
prior on the parameter θ = (β > , σ 2 )> :
β σ 2 , µ∗ , Ω ∼ N µ∗ , σ 2 Ω ,
σ 2 a, b ∼ Inverse-Gamma (a/2, b/2) ,
where a > 4, b > 0 is to guarantee a finite variance for the prior of σ 2 , and Ω is a positive
definite matrix. The subset posterior after the stochastic approximation is given by
K(y − Zβ)> (y − Zβ)
2
2 −Km/2
πm β, σ y, Z ∝ (σ )
exp −
×
2σ 2
(β − µ∗ )> Ω−1 (β − µ∗ )
b
2 −a/2−1
exp −
× (σ )
exp − 2
2σ 2
2σ
We have the following proposition, which shows that the ψ(·) function in Assumption 7 is L1 integrable uniformly for all m and K, which implies the uniform integrability condition.
Proposition 1. In the normal linear model (A.40), assume that kµ∗ k is upper bounded by
a constant. Assume that the eigenvalues of Ω and Z > Z/m are lower and upper bounded by
constants for all m ≥ 2. Assume that the error εi in (A.40) has finite 4th moment. Let βb and
c2 be the maximum likelihood estimators of β and σ 2 respectively. Then
σ
sup
m≥2,K≥1
sup
m≥2,K≥1
b 2 < +∞,
EPθ0 EΠm (·|y,Z) Kmkβ − βk
(A.41)
c2 k2 < +∞.
EPθ0 EΠm (·|y,Z) Kmkσ 2 − σ
(A.42)
Proof of Proposition 1:
Let kβ0 k, kµ∗ k ≤ c1 < +∞. Let the eigenvalues of Ω and Z > Z/m be lower bounded by c2 > 0
and upper bounded by c3 > 0. Let E(ε4i ) = c4 < +∞. The subset posterior distributions of β
and σ 2 are given by
−1
b∗
∗
∗
>
−1
KZ Z + Ω
,
β y, Z, µ , Ω, a, b ∼ Multi-ta+Km+p β ,
a + Km
a + Km b∗
σ 2 y, Z, µ∗ , Ω, a, b ∼ Inverse-Gamma
,
,
2
2
−1
β ∗ = KZ > Z + Ω−1
KZ > y + Ω−1 µ∗ ,
−1
b∗ = b + µ∗> Ω−1 µ∗ + Ky > Im − KZ KZ > Z + Ω−1
Z > y,
37
where Multi-tν (µ, Σ) denotes the multivariate-t distribution with mean µ, variance matrix Σ,
and ν degrees of freedom.
The maximum likelihood estimators of β and σ 2 are given by
βb = (Z > Z)−1 Z > y,
n
o
−1
>
2
−1 >
>
−1 >
c
2
σ = m ky − Z βk = m y
Im − Z(Z Z) Z
y.
We first prove (A.41). It is clear that
b 2,
b 2 = KmEP tr varπ (·|y,Z) (β) + KmEP kEΠ (·|y,Z) β − βk
EPθ0 EΠm (·|y,Z) Kmkβ − βk
θ0
θ0
m
m
(A.43)
where tr(A) denotes the trace of a generic square matrix A. The posterior variance of β can be
bounded as
KmEPθ0 tr varπm (·|y,Z) (β)
−1
a + Km + p
b∗
>
−1
= Km
× tr EPθ0
KZ Z + Ω
a + Km + p − 2
a + Km
−1
>
>
−1
2 −1
≤ 2 tr EPθ0 b + c1 c2 + Ky y KZ Z + Ω
n
−1 o
−1
2
2
I
+
Kmc
c
+
Kmσ
Kmc
I
+
c
≤ 2 tr EPθ0 b + c21 c−1
p
2
2
p
1
0
3
2
= 2p
Km(c21 c2 + σ02 ) + b + c21 c−1
2p(c21 c2 + σ02 )
2
→
as m → ∞.
c2
Kmc2 + c−1
3
(A.44)
The second term in (A.43) can be bounded as
b 2
KmEPθ0 kEΠm (·|y,Z) β − βk
−1
−1
−1
= KmEPθ0
KZ > Z + Ω−1
− KZ > Z
(KZ > y) + KZ > Z + Ω−1
Ω−1 µ∗
≤ 2KmEPθ0
+ 2Km
>
KZ Z + Ω
>
KZ Z + Ω
≤ 2KmEPθ0
+ 2Km
>
−1
−1
−1
KZ Z + Ω
−1
Kmc2 Ip + c−1
3 Ip
−1
>
− KZ Z
−1
2
2
>
(KZ y)
2
Ω
−1 ∗
µ
−1
−1
2
Ω
−1
c−1
2 c1
>
(Z Z)
−1
>
(Z y)
2
2
−1 −1 > −1 >
2c21
c2 (Z Z) (Z y) +
Kmc2 + c−1
3
Kmc42
2
2Km
2c21
2
>
−1
>
+
≤
−1 2 2 kβ0 k + EPθ0 (Z Z) (Z ε)
Kmc42
(Kmc2 + c3 ) c2
2Km
2c21
−2
2
2
≤
c
+
c
c
σ
+
→ 0 as m → ∞.
3
1
0
2
2 2
Kmc42
(Kmc2 + c−1
3 ) c2
≤ 2KmEPθ0
(A.45)
Since (A.44) and (A.45) have finite limits as m → ∞, they are both bounded by constants,
regardless of the value of K. They together with (A.43) lead to (A.41).
38
Next we prove (A.42). We have the similar decomposition
2
c2 k2 = KmEP var
EPθ0 EΠm (·|y,Z) Kmkσ 2 − σ
πm (·|y,Z) (σ )
θ0
c2 k2 .
+ KmEPθ0 kEΠm (·|y,Z) σ 2 − σ
(A.46)
We show an useful bound for the square of y > y:
2
2
EPθ0 y > y = EPθ0 kZβ0 + εk2
n
o2
2
≤ 4EPθ0 kZβ0 k2 + kεk2 ≤ 4EPθ0 β0> (Z > Z)β0 + kεk2
!2
m
2
X
2
2
2 4 2
2
≤ 4EPθ0 mc1 c3 + kεk
≤ 8m c1 c3 + 8EPθ0
εi
i=1
≤ 8m2 c41 c23 + 8mEPθ0
m
X
ε4i ≤ 8m2 (c41 c23 + c4 ).
(A.47)
i=1
By using (A.47), the first term in (A.46) can be bounded as
b∗2 /4
{(Km + a)/2 − 2}3
2
−1
∗> −1 ∗
>
>
−1
>
b + µ Ω µ + Ky
Im − KZ KZ Z + Ω
Z
y
KmEPθ0 varπm (·|y,Z) (σ 2 ) ≤ KmEPθ0
2
EP
(Km)2 θ0
2
2
>
2 −1
+
Ky
y
b
+
c
c
E
≤
P
1 2
(Km)2 θ0
2
2
4(b + c21 c−1
4
>
2 )
+
E
y
y
≤
P
(Km)2
m 2 θ0
≤
≤
2
4(b + c21 c−1
2 )
+ 32(c41 c23 + c4 ) → 32(c41 c23 + c4 ) as m → ∞.
(Km)2
(A.48)
And the second term in (A.46) can be bounded as
c2
KmEPθ0 EΠm (·|y,Z) σ 2 − σ
2
= KmEPθ0
b∗ /2
c2
−σ
(a + Km)/2 − 1
2
b + µ∗> Ω−1 µ∗ (a − 2)y > Im − Z(Z > Z)−1 Z > y
−
= KmEPθ0
Km + a − 2
(Km + a − 2)m
n
−1 > o 2
Ky > Z(Z > Z)−1 Z > − KZ KZ > Z + Ω−1
Z
y
+
Km + a − 2
2
h
n
o i2
3Km b + µ∗> Ω−1 µ∗
3(a − 2)
−1 >
>
−1 >
≤
+
E
m
y
I
−
Z(Z
Z)
Z
y
m
P
(Km + a − 2)2
(Km + a − 2)2 θ0
n
o2
3Km
>
>
−1
−1 −1
>
−1 >
+
E
y
Z(Z
Z
+
Ω
/K)
Ω
(Z
Z)
Z
y
P
(Km + a − 2)2 θ0
2
2
3(b + c21 c−1
3(a − 2)
−1 >
2 )
≤
+
E
m
y
y
P
Km
(Km + a − 2)2 θ0
2
3
>
>
+
E
y
ZZ
y
2 Pθ0
2 2
Km(mc2 + c−1
3 /K) c2 mc2
2
2
3(b + c21 c−1
24(a − 2)(c41 c23 + c4 )
3pmc3
>
2 )
E
y
y
≤
+
+
P
θ
0
Km
(Km + a − 2)2
Km4 c62
39
≤
2
3(b + c21 c−1
24(a − 2)(c41 c23 + c4 ) 24pc3 (c41 c23 + c4 )
2 )
+
+
→ 0 as m → ∞,
Km
(Km + a − 2)2
Kmc62
(A.49)
where we have used the relation λ(ZZ > ) ≤ tr(ZZ > ) = tr(Z > Z) ≤ pλ(Z > Z) ≤ pmc3 , and λ(A)
denotes the largest eigenvalue of a generic matrix A. Since (A.48) and (A.49) have finite limits
as m → ∞, they are both bounded by constants, regardless of the value of K. They together
with (A.46) lead to (A.42).
2. Some Exponential Family Models
In this section, we verify Assumption 7 for the following three commonly used exponential
family distributions: Poisson, exponential, and binomial.
Proposition 2. (i) Suppose the data yi (i = 1, . . . , m) are independent and identically distributed as Poisson(θ) with the probability mass function p(y|θ) = θy e−θ /y! and the true parameter θ0 . Suppose the prior on θ is Gamma(a, b) for some constants a > 0, b > 0. Let
P
θb = m
i=1 yi /m be the maximum likelihood estimator of θ. Then
2
sup
m≥1,K≥1
EPθ0 EΠm (·|y) Km θ − θb < +∞;
(ii) Suppose the data yi (i = 1, . . . , m) are independent and identically distributed as Exp(θ) with
the probability density function p(y|θ) = θe−θy and the true parameter θ0 . Suppose the prior
P
on θ is Gamma(a, b) for some constants a > 0, b > 0. Let θb = m/ m
i=1 yi be the maximum
likelihood estimator of θ. Then
2
sup
m≥3,K≥1
EPθ0 EΠm (·|y) Km θ − θb < +∞;
(iii) Suppose the data yi (i = 1, . . . , m) are {0, 1} binary data independent and identically
distributed as Bernoulli(θ) with the probability density function p(y|θ) = θy (1 − θ)1−y and the
true parameter θ0 ∈ (0, 1). Suppose the prior on θ is Beta(a, b) for some constants a > 0, b > 0.
P
Let θb = m
i=1 yi /m be the maximum likelihood estimator of θ. Then
2
sup
m≥1,K≥1
EPθ0 EΠm (·|y) Km θ − θb < +∞;
Proof of Proposition 2:
P
(i) The subset posterior distribution of θ is Gamma(K m
i=1 yi + a, Km + b). Therefore
2
EPθ0 EΠm (·|y) Km θ − θb
2
b
= EPθ0 Km EΠm (·|y) (θ) − θ + Kmvarπm (·|y) (θ)
)
(
Pm
P
P
2
Km (K m
K m
i=1 yi + a)
i=1 yi
i=1 yi + a
+
−
= EPθ0 Km
Km + b
m
(Km + b)2
Pm
P
K(b i=1 yi − am)2 Km (K m
i=1 yi + a)
= EPθ0
+
m(Km + b)2
(Km + b)2
P
P
2
2 2
2Kb2 m m
Km (K m
i=1 yi + 2Km a
i=1 yi + a)
≤ EPθ0
+
m(Km + b)2
(Km + b)2
40
=
2Km2 b2 (θ02 + θ0 ) + 2Km2 a2 Km (Kmθ0 + a)
+
→ θ0
m(Km + b)2
(Km + b)2
as m → ∞.
Hence, the conclusion holds.
P
(ii) The subset posterior distribution of θ is Gamma(Km + a, K m
i=1 yi + b), and notice that
Pm
W ≡ 1/ i=1 yi follows Inverse-Gamma(m, θ0 ) with E(W ) = θ0 /(m − 1), E(W 2 ) = θ02 /{(m −
1)(m − 2)}, E(W 3 ) = θ03 /{(m − 1)(m − 2)(m − 3)}. Therefore
2
EPθ0 EΠm (·|y) Km θ − θb
2
= EPθ0 Km EΠm (·|y) (θ) − θb + Kmvarπm (·|y) (θ)
(
)
2
Km + a
Km (Km + a)
m
P
P
= EPθ0 Km
+
− Pm
2
K m
(K m
i=1 yi + b
i=1 yi
i=1 yi + b)
Pm
K(a i=1 yi − bm)2
Km (Km + a)
Pm
Pm
P
+
= EPθ0
2
2
( i=1 yi )(K i=1 yi + b)
(K m
i=1 yi + b)
(
)
P
2a2 ( m
y )2 + 2b2 m2
m (Km + a)
i=1
Pmi
≤ EPθ0
+
P
2
3
K( i=1 yi )
K( m
i=1 yi )
≤
2a2 θ0
2b2 m2 θ03
m (Km + a) θ02
+
+
→ θ02
K(m − 1) K(m − 1)(m − 2)(m − 3) K(m − 1)(m − 2)
as m → ∞.
Therefore, the conclusion holds.
(iii) The subset posterior distribution of θ is Beta {K
fore
Pm
i=1 yi
+ a, K
Pm
i=1 (1
− yi ) + b}. There-
2
EPθ0 EΠm (·|y) Km θ − θb
2
= EPθ0 Km EΠm (·|y) (θ) − θb + Kmvarπm (·|y) (θ)
"
#
P
Pm
Pm
Pm
2
K m
y
+
a
Km
(K
y
+
a)
{K
(1
−
y
)
+
b}
y
i
i
i
i
i=1
i=1
i=1
= EPθ0 Km
− i=1
+
Km + a + b
m
(Km + a + b)2 (Km + a + b + 1)
"
P
2
2Km(a + b)2 ( m
2Kma2
i=1 yi )
+
≤ EPθ0
(Km + a + b)2
m2 (Km + a + b)2
#
P
Pm
Km (K m
y
+
a)
{K
(1
−
y
)
+
b}
i
i=1 i
i=1
+
(Km + a + b)2 (Km + a + b + 1)
2Km(a + b)2 m2 θ02 + mθ0 (1 − θ0 )
2Kma2
=
+
(Km + a + b)2
m2 (Km + a + b)2
2 2
Km K (m − m)θ0 (1 − θ0 ) + Kam(1 − θ0 ) + Kbmθ0 + ab
+
(Km + a + b)2 (Km + a + b + 1)
→ θ0 (1 − θ0 )
as m → ∞.
Therefore, the conclusion holds.
41
D
Data Analysis
D.1
Simulated data analysis: Linear model with varying dimension
The prior distributions of β and σ are specified as follows:
β ∼ generalized double Pareto(α, η), σ ∼ Half-t(ν, A).
The prior density of β = (β1 , . . . , βp )> given α and η is given by
p
Y
|βj | −(α+1)
α
π(β | α, η) =
1+
.
2η
η
j=1
The prior mean and variance of β are set to be 0 and 2η 2 (α − 1)−1 (α − 2)−1 . α and η have
independent hyperpriors with densities π(α) = 1/(1 + α)2 and π(η) = 1/(1 + η)2 . The Half-t
prior has a convenient parameter expanded form in terms of Inverse-Gamma(a, b) distribution,
where a and b are shape and scale parameters: if σ 2 | ρ ∼ Inverse-Gamma(ν/2, ν/ρ) and ρ ∼
Inverse-Gamma(1/2, 1/A2 ), then σ ∼ Half-t(ν, A). We fixed the hyperparameters ν and A at
recommended default values 2 and 100. We used griddy Gibbs for generating samples of α and
η from their posterior distribution; see Section 3 in [3] for details. The Gibbs sampler in [3] is
modified by changing the sample size, n, in their sampler to mK, where m is sample size for
the subset and K is the number of subsets.
Let N (m̂1 , V̂1 ), . . . , N (m̂K , V̂K ) represent the asymptotic approximations of K subset posteriors, then [2] has shown that their barycenter in Wasserstein-2 space is also Gausssian with
mean m∗ and covariance matrix V ∗ , where
m∗ = K −1
K
X
m̂j
and V ∗ satisfies
j=1
K
X
1/2
V∗
1/2
V̂j V ∗
1/2
= KV ∗ .
j=1
Therefore, we use the formula above to calculate the W2 barycenter of K normal approximations
to the K subset posteriors. Given V̂1 , . . . , V̂K , we can find V ∗ efficiently using fixed-point
iteration.
Although the priors of β and σ specified above are heavy-tailed with infinite second moments,
in the following proposition and its proof, we verify that every subset posterior after conditioning
on the first m0 observations has finite second moment in both β and σ, for some fixed integer
m0 .
Proposition 3. Suppose the form of a linear model and its priors are specified in Section 4.1 of
the main paper with fixed ν > 0 and A > 0. Assume that in the model X and are independent.
Let ỹ and X̃ be the response vector and the design matrix of the first m0 observations (m0 ≥ 1).
Suppose that the true parameters are θ0 = (β0> , σ0 )> with σ0 > 0. Assume that the eigenvalues of
X̃ > X̃ are bounded from above and below by positive constants almost surely. Then the posterior
distribution of θ = (β > , σ)> conditional on ỹ and X̃ has finite second moment almost surely in
Pθ0 , if m0 satisfies m0 ≥ p + 4.
42
Proof of Proposition 3:
For convenience we define the quadratic term S(β, ỹ, X̃) = (ỹ − X̃β)> (ỹ − X̃β), which has
+ (β − β̃)> X̃ > X̃(β − β̃) with ˜ = ỹ − X̃β0 ,
the decomposition S(β, ỹ, X̃) = ˜> (Im0 − H̃)˜
H̃ = X̃(X̃ > X̃)−1 X̃ > , Im0 being the m0 -dimensional identity matrix, and β̃ = (X̃ > X̃)−1 X̃ > ỹ.
Since m0 ≥ p + 4 and X̃ > X̃ is nonsingular, Im0 − H̃ is idempotent with rank m0 − p > 0. Since
is then almost surely positive. Let the smallest
σ0 > 0, the residual sum of squares ˜> (Im0 − H̃)˜
>
eigenvalue of X̃ X̃ be lower bounded by c1 > 0. Then S(β, ỹ, X̃) ≥ ˜> (Im0 − H̃)˜
+ c1 kβ − β̃k2 .
The subset posterior of the model parameter θ = (β > , σ)> given only ỹ, X̃ has the following
expression
πm0 (β, σ | ỹ, X̃, ν, A)
o
n
K
S(β,
ỹ,
X̃)
π(β)π(σ | ν, A)
(2π)−Km0 /2 σ −Km0 exp − 2σ
2
n
o
= RR
K
(2π)−Km0 /2 σ −Km0 exp − 2σ
π(β)π(σ | ν, A)dβdσ
2 S(β, ỹ, X̃)
n
on
o−(ν+1)/2
K
π(β)
σ −Km0 exp − 2σ
1 + ν −1 (σ/A)2
2 S(β, ỹ, X̃)
=
n
o
n
o
−(ν+1)/2
R R∞
2
K
−Km
−1
0 exp
− 2σ2 S(β, ỹ, X̃) 1 + ν (σ/A)
dσ π(β)dβ
0 σ
(A.50)
where the likelihood has been raised to the power of K according to our stochastic approximation. In the following, we bound EΠm (·|ỹ,X̃,ν,A) kβk2 and EΠm (·|ỹ,X̃,ν,A) (σ 2 ) respectively.
0
Step 1: Show that EΠm
0 (·|ỹ,X̃,ν,A)
0
kβk2 is finite almost surely in Pθ0 .
In the following, we use (A.50) to calculate EΠm (·|ỹ,X̃,ν,A) kβk2 and bound its numerator
0
and denominator respectively. For the numerator part, we have
n
Z ∞Z
o−(ν+1)/2
K
2 −Km0
kβk σ
exp − 2 S(β, ỹ, X̃)
1 + ν −1 (σ/A)2
π(β)dβdσ
2σ
0
Rp
Z ∞
Z
K
1/2 ν+1
2
−(Km0 +ν+1)
≤ (Aν )
kβk π(β)
σ
exp − 2 S(β, ỹ, X̃) dσ dβ
2σ
Rp
0
Km
+
ν
0
≤ 2(Km0 +ν)/2−1 K −(Km0 +ν)/2 (Aν 1/2 )ν+1 Γ
2
Z
kβk2 × π(β)
×
(A.51)
n
o(Km0 +ν)/2 dβ.
Rp
˜> (Im0 − H̃)˜
+ c1 kβ − β̃k2
The last integral of (A.51) can be further bounded by
Z
kβk2 × π(β)
n
o(Km0 +ν)/2 dβ
Rp
>
2
˜ (Im0 − H̃)˜
+ c1 kβ − β̃k
Z
2 kβ − β̃k2 + kβ̃k2 × π(β)
≤
n
o(Km0 +ν)/2 dβ
Rp
>
2
˜ (Im0 − H̃)˜
+ c1 kβ − β̃k
Z
o−(Km0 +ν)/2+1
n
≤
2c−1
˜> (Im0 − H̃)˜
+ c1 kβ − β̃k2
π(β)dβ
1
Rp
Z
n
o−(Km0 +ν)/2
+
2kβ̃k2 ˜> (Im0 − H̃)˜
π(β)dβ
Rp
43
o−(Km0 +ν)/2
on
n
2
>
>
−
H̃)˜
.
−
H̃)˜
+
k
β̃k
˜
(I
≤ 2 c−1
˜
(I
m0
m0
1
(A.52)
Next we provide a lower bound for the denominator of (A.50). The integral of σ can be
lower bounded by using a change of variable u = KS(β, ỹ, X̃)/(2σ 2 ):
n
Z ∞
o−(ν+1)/2
K
−Km0
σ
exp − 2 S(β, ỹ, X̃)
1 + ν −1 (σ/A)2
dσ
2σ
0
Z
o−(Km0 −1)/2 ∞ n
o−(ν+1)/2
1n
=
u(Km0 −3)/2 e−u du
KS(β, ỹ, X̃)/2
1 + KS(β, ỹ, X̃)/(2A2 νu)
2
0
o−(Km0 −1)/2 n
o−(ν+1)/2 Z ∞
1n
u(Km0 −3)/2 e−u du
≥
KS(β, ỹ, X̃)/2
1 + KS(β, ỹ, X̃)/(2A2 ν)
2
1
o−(Km0 −1)/2 n
o−(ν+1)/2
1n
≥
KS(β, ỹ, X̃)/2
1 + KS(β, ỹ, X̃)/(2A2 ν)
2
Km0 − 1
Km0 + 1 (Km0 −3)/2
−1
×e Γ
,
(A.53)
2
Km0 − 1
where we have used the fact Km0 ≥ 4 and the lower bound for the incomplete gamma function
R ∞ s−1 −u
e du ≥ e−1 Γ(s)(1 + 1/s)s−1 for s ≥ 1. Now to evaluate the denominator of (A.50),
1 u
we need to integrate the lower bound in (A.53) with respect to β. Consider the set A3 =
{β ∈ Rp : kβk ≤ 1}. Clearly the prior of β has positive probability mass on A3 . Define the
R
R RR
constant c3 = A3 π(β)dβ = A3
π(β | α, η)π(α)π(η)dαdηdβ > 0 which only depends on the
dimension p. Let the largest eigenvalue of X̃ > X̃ be upper bounded by c2 > 0. Then on A3 ,
S(β, ỹ, X̃) ≤ c2 kβ − β̃k2 + ˜> (Im0 − H̃)˜
≤ 2c2 (kβ̃k2 + 1) + ˜> (Im0 − H̃)˜
. This and (A.53) imply
that the denominator of (A.50) can be lower bounded by
n
Z Z ∞
o−(ν+1)/2
K
2
−1
−Km0
1 + ν (σ/A)
dσ π(β)dβ
σ
exp − 2 S(β, ỹ, X̃)
2σ
0
hn
o
i−(Km0 −1)/2
c3
Km0 − 1
≥ Γ
2c2 (kβ̃k2 + 1) + ˜> (Im0 − H̃)˜
K/2
2e
2
h
n
o
i−(ν+1)/2
× 1 + 2c2 (kβ̃k2 + 1) + ˜> (Im0 − H̃)˜
K/(2A2 ν)
Km0 − 1
≥ 2(Km0 −3)/2 e−1 c3 (Aν 1/2 )ν+1 K −(Km0 +ν)/2 Γ
2
n
o−(Km0 +ν)/2
× 2c2 (kβ̃k2 + 1) + ˜> (Im0 − H̃)˜
,
(A.54)
where the last inequality follows if we choose c2 > Aν 2 /K.
We can combine (A.51), (A.52), (A.54) and obtain that
EΠm
0 (·|ỹ,X̃,ν,A)
kβk2
n
on
o−(Km0 +ν)/2
>
2
>
≤ c4 K (ν+1)/2 c−1
˜
(I
−
H̃)˜
+
k
β̃k
˜
(I
−
H̃)˜
m
m
0
0
1
n
o(Km0 +ν)/2
× 2c2 (kβ̃k2 + 1) + ˜> (Im0 − H̃)˜
(A.55)
for some constant c4 > 0 that only depends on m0 , p, ν, A, c1 , c2 , c3 . Conditional on ỹ, X̃, both
kβ̃k2 and ˜> (Im0 − H̃)˜
are almost surely positive constants. Therefore, we have proved that
2
EΠm (·|ỹ,X̃,ν,A) kβk < ∞ almost surely in Pθ0 .
0
44
Step 2: Show that EΠm (·|ỹ,X̃,ν,A) (σ 2 ) is finite almost surely in Pθ0 .
0
To calculate EΠm (·|ỹ,X̃,ν,A) σ 2 , we integrate σ 2 with respect to the posterior density of (A.50).
0
We start with upper bounding the numerator:
n
Z Z ∞
o−(ν+1)/2
K
2
−Km0
σ ×σ
exp − 2 S(β, ỹ, X̃)
1 + ν −1 (σ/A)2
π(β)dσdβ
2σ
Z ∞
Z 0
K
σ −(Km0 +ν−1) exp − 2 S(β, ỹ, X̃) dσ π(β)dβ
≤ (Aν 1/2 )(ν+1)/2
2σ
0
Z
n
o
−(Km
+ν)/2+1
0
1
Km0 + ν
1/2 (ν+1)/2
≤
Γ
(Aν )
KS(β, ỹ, X̃)/2
− 1 π(β)dβ
2
2
Km0 + ν
−1
≤ 2(Km0 +ν)/2−2 (Aν 1/2 )(ν+1)/2 K −(Km0 +ν)/2+1 Γ
2
o−(Km0 +ν)/2+1
n
,
(A.56)
× ˜> (Im0 − H̃)˜
. If we
where we have used the fact that (Km0 + ν)/2 ≥ 2 and S(β, ỹ, X̃) ≥ ˜> (Im0 − H̃)˜
combine (A.54) and (A.56), then we can obtain that
EΠm
n
o−(Km0 +ν)/2+1
>
2
(ν+1)/2
˜
(I
(σ
)
≤
c
K
−
H̃)˜
m
5
0
0 (·|ỹ,X̃,ν,A)
n
o(Km0 +ν+1)/2
× 2c2 (kβ̃k2 + 1) + ˜> (Im0 − H̃)˜
,
(A.57)
for some constant c5 > 0 that only depends on m0 , p, ν, A, c2 , c3 . Conditional on ỹ, X̃, both
kβ̃k2 and ˜> (Im0 − H̃)˜
are almost surely positive constants. Therefore, we have proved that
EΠm (·|ỹ,X̃,ν,A) (σ 2 ) < ∞ almost surely in Pθ0 .
0
Remark 2. The second moment of the prior has only appeared in (A.7) in the proof of Lemma
2. We need the right-hand side of (A.7) to go to zero as n → ∞. Since the gdp prior and half-t
prior are both heavy tailed, Proposition 3 proposes to replace the prior in (A.7) by the posterior
conditional on the first m0 observations in the linear model example. It is straightforward to
check that the finite upper bounds for EΠm (·|ỹ,X̃,ν,A) kβk2 and EΠm (·|ỹ,X̃,ν,A) (σ 2 ) in (A.55) and
0
0
(A.57) increase at most exponentially fast in K. Even if K → ∞, we can see that the exponential
term exp(−Km1 ) in the right-hand side of (A.7) decays faster than any exponential rate in
K since m → ∞. Therefore, the conclusions of Lemma 2 and all subsequent theorems remain
valid conditional on the first m0 observations.
D.2
Simulated data analysis: Linear mixed effects model
Stochastic approximation for subset posteriors can be easily implemented in Stan. The
sampling model for linear mixed effects models implies that likelihood of β and Σ is
s Z
s
Y
Y
L(β, Σ) =
p(yi | Xi , Zi , β, ui )p(ui | Σ) dui =
φ(yi | Xi β, Zi ΣZi> ),
i=1
Rq
i=1
where φ(· | µ, Σ) is the multivariate normal density with mean µ and covariance matrix Σ. The
likelihood after stochastic approximation is
LK (β, Σ) = {L(β, Σ)}
K
s n
oK
Y
=
φ(yi | Xi β, Zi ΣZi> )
.
i=1
45
(A.58)
The generative model is completed by imposing default priors for β and Σ in Stan. We take
advantage of the increment log prob function in Stan to specify that
yi | β, Σ ∼ fK (yi | β, Σ),
where fK is the density that leads to the term for yi in the likelihood LK (β, Σ) in (A.58). In
general fK would be analytically intractable, but in the present case it corresponds to {φ(· |
µ, Σ)}K . The computation time of different methods is summarized in Figure 6a.
Table 7: 90% credible intervals for fixed effects in simulated data analysis. The upper and lower bounds
are averaged over 10 replications. MLE, maximum likelihood estimator; MCMC, Markov chain Monte
Carlo based on the full data; VB, variational Bayes; CMC, consensus Monte Carlo; SDP, semiparametric
density product; WASP, the algorithm in Srivastava et al. [23]; PIE, our posterior interval estimation
algorithm.
β1
MLE
MCMC
VB
CMC
SDP
WASP
PIE
β2
(-1.01,
(-1.01,
(-1.01,
(-1.01,
(-1.01,
(-1.01,
(-1.01,
-1.00)
-1.00)
-1.00)
-0.99)
-0.99)
-0.99)
-0.99)
(0.99,
(0.99,
(0.99,
(0.99,
(0.99,
(0.99,
(0.99,
β3
1.01)
1.01)
1.01)
1.01)
1.01)
1.01)
1.01)
β4
(-1.01,
(-1.01,
(-1.01,
(-1.01,
(-1.01,
(-1.01,
(-1.01,
-0.99)
-0.99)
-1.00)
-0.99)
-0.99)
-0.99)
-0.99)
(1.00,
(1.00,
(1.00,
(1.00,
(1.00,
(1.00,
(1.00,
1.01)
1.01)
1.01)
1.01)
1.01)
1.01)
1.01)
Table 8: Accuracy of approximate posteriors for fixed effects in simulated data analysis. The standard
deviation of accuracy across 10 replications is in parentheses. MLE, maximum likelihood estimator; VB,
variational Bayes; CMC, consensus Monte Carlo; SDP, semiparametric density product; WASP, the
algorithm in Srivastava et al. [23]; PIE, our posterior interval estimation algorithm.
β1
MLE
VB
CMC
SDP
WASP
PIE
D.3
0.96
0.86
0.96
0.96
0.95
0.95
(0.01)
(0.11)
(0.01)
(0.01)
(0.02)
(0.01)
β2
0.96
0.87
0.96
0.94
0.95
0.95
(0.01)
(0.10)
(0.01)
(0.02)
(0.02)
(0.01)
β3
0.96
0.86
0.95
0.95
0.96
0.94
β4
(0.01)
(0.10)
(0.02)
(0.03)
(0.01)
(0.01)
0.96
0.87
0.96
0.95
0.94
0.94
(0.01)
(0.10)
(0.01)
(0.02)
(0.02)
(0.02)
Real data analysis: United States natality data
We selected thirteen variables from the United States natality data summarized in Table 9
and analyzed in [1] and [15]. These data are available at http://qed.econ.queensu.ca/jae/
datasets/abrevaya001. The computation time of different methods is summarized in Figure
6b.
46
Table 9: Variables used in the United States natality data
Variable
Description
dmage
nlbnl
gestat
male
married
hsgrad
agesq
black
novisit
adeqcode2
adeqcode3
pretri2
pretri3
age of mother in years
number of live births now living
length of gestation in weeks
indicator variable for baby gender
indicator variable for marital status
high-school graduate indicator
age of mother squared
indicator variable for black race
indicator of no prenatal care visit
indicator that Kessner index 2
indicator that Kessner index 3
indicator that first prenatal visit occurred in 2nd trimester
indicator that first prenatal visit occurred in 3nd trimester
Table 10: 90% credible intervals for fixed effects in United States natality data analysis. The upper
and lower bounds are averaged over 10 folds of cross-validation. MLE, maximum likelihood estimator;
MCMC, Markov chain Monte Carlo based on the full data; VB, variational Bayes; CMC, consensus
Monte Carlo; SDP, semiparametric density product; WASP, the algorithm in Srivastava et al. [23]; PIE,
our posterior interval estimation algorithm.
MLE
MCMC
VB
CMC
SDP
WASP
PIE
MLE
MCMC
VB
CMC
SDP
WASP
PIE
Intercept
dmage
(-1.31, -0.18)
(-0.48, 0.11)
(-1.96, -0.94)
(-1.13, -0.04)
(-0.97, -0.10)
(-0.93, 0.03)
(-0.93, 0.04)
(0.09, 0.57)
(-0.04, 0.26)
(0.17, 0.56)
(0.06, 0.51)
(0.05, 0.42)
(0.04, 0.45)
(0.03, 0.45)
married
hsgrad
(-0.03,
(-0.03,
(-0.02,
(-0.03,
(-0.03,
(-0.02,
(-0.02,
0.06)
0.07)
0.06)
0.07)
0.07)
0.07)
0.07)
ageqcode3
MLE
MCMC
VB
CMC
SDP
WASP
PIE
(-0.42, -0.22)
(-0.37, -0.18)
(-0.43, -0.22)
(-0.39, -0.19)
(-0.38, -0.20)
(-0.39, -0.2)
(-0.39, -0.2)
(0.02,
(0.03,
(0.02,
(0.02,
(0.02,
(0.03,
(0.03,
0.12)
0.13)
0.11)
0.12)
0.12)
0.13)
0.13)
novisit
(-0.12,
(-0.14,
(-0.12,
(-0.13,
(-0.13,
(-0.14,
(-0.14,
0.16)
0.11)
0.18)
0.17)
0.13)
0.15)
0.15)
nlbnl
(0.02,
(0.02,
(0.01,
(0.02,
(0.02,
(0.02,
(0.02,
0.08)
0.08)
0.06)
0.08)
0.08)
0.08)
0.08)
agesq
(0.00,
(0.00,
(0.00,
(0.00,
(0.00,
(0.00,
(0.00,
0.00)
0.00)
0.00)
0.00)
0.00)
0.00)
0.00)
gestat
(0.22,
(0.22,
(0.25,
(0.22,
(0.22,
(0.22,
(0.22,
0.25)
0.24)
0.27)
0.24)
0.24)
0.24)
0.24)
male
(0.25,
(0.24,
(0.24,
(0.24,
(0.25,
(0.25,
(0.25,
0.32)
0.31)
0.31)
0.32)
0.32)
0.31)
0.31)
black
ageqcode2
(-0.46, -0.3)
(-0.44, -0.28)
(-0.43, -0.29)
(-0.45, -0.28)
(-0.45, -0.30)
(-0.45, -0.28)
(-0.44, -0.28)
(-0.24, -0.11)
(-0.22, -0.1)
(-0.26, -0.13)
(-0.24, -0.11)
(-0.23, -0.11)
(-0.23, -0.11)
(-0.23, -0.11)
petri2
pertri3
(0.03, 0.16)
(0.01, 0.14)
(0.04, 0.18)
(0.02, 0.16)
(0.02, 0.15)
(0.02, 0.15)
47
(0.02, 0.15)
(0.1, 0.33)
(0.06, 0.28)
(0.09, 0.34)
(0.08, 0.32)
(0.09, 0.30)
(0.08, 0.31)
(0.08, 0.31)
Table 11: Accuracy of approximate posteriors for fixed effects in US natality data analysis. The standard
deviation of accuracy across 10 replications is in parentheses. MLE, maximum likelihood estimator; VB,
variational Bayes; CMC, consensus Monte Carlo; SDP, semiparametric density product; WASP, the
algorithm in Srivastava et al. [23]; PIE, our posterior interval estimation algorithm.
Intercept
MLE
VB
CMC
SDP
WASP
PIE
0.28
0.02
0.40
0.39
0.55
0.55
(0.06)
(0.01)
(0.10)
(0.23)
(0.12)
(0.12)
married
MLE
VB
CMC
SDP
WASP
PIE
0.92
0.91
0.90
0.90
0.92
0.92
0.74
0.67
0.86
0.87
0.84
0.84
0.35
0.22
0.42
0.47
0.53
0.52
(0.06)
(0.04)
(0.09)
(0.19)
(0.10)
(0.11)
nlbnl
0.93
0.77
0.91
0.91
0.90
0.91
hsgrad
(0.02)
(0.03)
(0.06)
(0.06)
(0.05)
(0.06)
ageqcode3
MLE
VB
CMC
SDP
WASP
PIE
dmage
(0.04)
(0.04)
(0.08)
(0.09)
(0.11)
(0.11)
0.90
0.80
0.82
0.87
0.92
0.92
(0.02)
(0.03)
(0.08)
(0.08)
(0.06)
(0.06)
(0.04)
(0.06)
(0.08)
(0.07)
(0.09)
(0.08)
0.36
0.23
0.44
0.48
0.54
0.54
0.82
0.67
0.84
0.88
0.89
0.89
0.93
0.91
0.95
0.91
0.91
0.92
(0.05)
(0.04)
(0.10)
(0.19)
(0.11)
(0.11)
0.84
0.91
0.88
0.86
0.87
0.87
(0.02)
(0.05)
(0.03)
(0.05)
(0.06)
(0.05)
ageqcode2
(0.01)
(0.02)
(0.10)
(0.08)
(0.08)
(0.08)
0.83
0.66
0.85
0.87
0.88
0.88
(0.03)
(0.06)
(0.06)
(0.10)
(0.08)
(0.08)
pertri3
(0.02)
(0.04)
(0.10)
(0.09)
(0.11)
(0.11)
0.73
0.72
0.83
0.84
0.83
0.82
(0.03)
(0.05)
(0.08)
(0.10)
(0.06)
(0.07)
5.5
5.0
4.5
3.5
●
●
●
3.0
2.5
●
1.5
log10 Seconds
log10 Seconds
(0.06)
(0.01)
(0.11)
(0.16)
(0.11)
(0.11)
●
4.0
2.0
0.77
0.03
0.83
0.79
0.87
0.87
male
black
petri2
5.0
4.5
(0.03)
(0.06)
(0.02)
(0.07)
(0.04)
(0.05)
agesq
novisit
0.81
0.77
0.80
0.89
0.84
0.84
gestat
4.0
●
3.5
●
●
3.0
2.5
1.0
2.0
0.5
1.5
0.0
●
●
1.0
MCMC VB
CMC
SDP WASP PIE
(a) Linear mixed effects model
MCMC VB
CMC
SDP WASP PIE
(b) United States natality data
Figure 6: Computation time for the methods used in simulated and real data analysis. MCMC, Markov
chain Monte Carlo based on the full data; VB, variational Bayes; CMC, consensus Monte Carlo; SDP,
semiparametric density product; WASP, the algorithm in Srivastava et al. [23]; PIE, our posterior interval
estimation algorithm.
48
References
[1] Abrevaya, J. (2006). Estimating the effect of smoking on birth outcomes using a matched
panel data approach. Journal of Applied Econometrics 21, 489–519.
[2] Agueh, M. and Carlier, G. (2011). Barycenters in the Wasserstein space. SIAM Journal
on Mathematical Analysis 43, 904–924.
[3] Armagan, A. and Dunson, D. B. and Lee, J. (2013). Generalized double Pareto
shrinkage. Statistica Sinica, 23, 119–143.
[4] Bickel, P. J. & Freedman, D. A. (1981). Some asymptotic theory for the bootstrap.
The Annals of Statistics 9, 1196–1217.
[5] Broderick, T., Boyd, N.,Wibisono, A., Wilson, A. C. and Jordan, M. I. (2013).
Streaming variational Bayes. Advances in Neural Information Processing Systems 26
(NIPS), 1727–1735.
[6] Chernozhukov, V. and Hong, H. (2003). An MCMC approach to classical estimation.
Journal of Econometrics 115, 293-346.
[7] Dunson, D. B. and Xing, C. (2009). Nonparametric Bayes modeling of multivariate
categorical data. Journal of the American Statistical Association 104, 1042–1051.
[8] Faes, C., Ormerod, J. T. & Wand, M. P. (2012). Variational bayesian inference for parametric and nonparametric regression with missing data. Journal of the American Statistical
Association 106, 959–971.
[9] Fournier, N. & Guillin, A. (2015). On the rate of convergence in Wasserstein distance
of the empirical measure. Probability Theory and Related Fields 162, 707–738.
[10] Gelman, A. (2006). Prior distributions for variance parameters in hierarchical models
(comment on article by Browne and Draper) Bayesian Analysis, 3, 515–534.
[11] Ghosh, J. K., Delampady, M. and Samanta, T. (2006). An introduction to Bayesian
analysis: theory and methods. Springer-Verlag New York.
[12] Hoffman, M. D., Blei, D. M., Wang, C. and Paisley, J. (2013). Stochastic variational
inference. Journal of Machine Learning Research 14, 1303–1347.
[13] Kim, Y., Choi, Y.-K. and Emery, S. (2013). Logistic regression with multiple random
effects: a simulation study of estimation methods and statistical packages. The American
Statistician 67, 171–182.
[14] Kleiner, A., Talwalkar, A., Sarkar, P. and Jordan, M. I. (2014). A scalable
bootstrap for massive data. Journal of the Royal Statistical Society: Series B 76, 795–816.
[15] Lee, Y. Y. C. & Wand, M. P. (2016). Streamlined mean field variational Bayes for
longitudinal and multilevel data analysis. Biometrical Journal 58, 868–895.
49
[16] Lehmann, E. L. and Casella, G. (1998). Theory of point estimation. Springer-Verlag
New York.
[17] Maclaurin, D. and Adams, R. P. (2014). Firefly Monte Carlo: Exact MCMC with
subsets of data. Proceedings of the 30th Conference on Uncertainty in Artificial Intelligence
(UAI), 543–552.
[18] Minsker, S., Srivastava, S., Lin, L. and Dunson, D. B. (2014). Scalable and robust
Bayesian inference via the median posterior. Proceedings of the 31st International Conference
on Machine Learning (ICML) 32, 1656–1664.
[19] Miroshnikov, A. & Conlon, E. M. (2014). parallelMCMCcombine: an R package for
Bayesian methods for big data and analytics. PloS ONE 9, e108425.
[20] Neiswanger, W., Wang, C. and Xing, E. (2014). Asymptotically exact, embarrassingly parallel MCMC. Proceedings of the 30th International Conference on Uncertainty in
Artificial Intelligence (UAI), 623–632.
[21] Scott, S. L., Blocker, A. W., Bonassi, F. V., Chipman, H. A., George, E. I., &
McCulloch, R. E. (2016). Bayes and big data: the consensus Monte Carlo algorithm.
International Journal of Management Science and Engineering Management, to appear.
[22] Shang, Z. and Cheng, G. (2015). A Bayesian splitotic theory for nonparametric models.
arXiv preprint arXiv:1508.04175.
[23] Srivastava, S., Cevher, V., Dinh, Q. and Dunson, D. B. (2015). WASP: Scalable
Bayes via barycenters of subset posteriors. Proceedings of the 18th International Conference
on Artificial Intelligence and Statistics (AISTATS) 38, 912–920.
[24] Tan, L. S. L. and Nott, D. J. (2014). A stochastic variational framework for fitting and
diagnosing generalized linear mixed models. Bayesian Analysis 9, 963–1004.
[25] Villani, C. (2008). Optimal transport: old and new. Springer-Verlag Berlin Heidelberg.
[26] Wand, M. (2015). KernSmooth: Functions for Kernel Smoothing. R package version
2.23-15.
[27] Wang, X., Guo, F., Heller, K. and Dunson, D. B. (2015). Parallelizing MCMC with
random partition trees. Advances in Neural Information Processing Systems 28 (NIPS),
arXiv preprint arXiv:1506.03164.
[28] Welling, M. and Teh, Y. W. (2011). Bayesian learning via stochastic gradient Langevin
dynamics. Proceedings of the 28th International Conference on Machine Learning (ICML),
681–688.
[29] Xu, M., Lakshminarayanan, B.,Teh, Y. W., Zhu, J. and Zhang, B. (2014). Distributed Bayesian posterior sampling via moment sharing. Advances in Neural Information
Processing Systems 27 (NIPS), 1656–1664.
50
| 10math.ST
|
arXiv:1606.02431v1 [math.GR] 8 Jun 2016
Finite groups with small number of cyclic subgroups
Wei Zhou
School of Mathematics and Statistics,
Southwest University, Chongqing 400715, P. R. CHINA
zh [email protected]
Abstract
In this note, we study the finite groups with the number of cylic subgroups no
greater than 5.
1
Introduction
For a finite group G, let C(G) be the poset of cyclic subgroups of G. Sometimes C(G) can
decide the structure of G. For example, |C(G)| = |G| if and only if G is an elementary
abelian 2-group. The groups G such that |G| − |C(G)| ≤ 3 is classified in [2, 3] and [4].
It is well-known that a finite p-group G has eactly one cyclic subgroup of order p if and
only if G is cyclic or generalized quaternion group. Hence the groups with small number
of cyclic subgroups will be interesting. In this note, we shall study the finite group G with
|C(G)| ≤ 5.
It is esay to see that |C(G)| = 1 if and only G = 1, and |C(G)| = 2 if and only if G ∼
= Cp
for some prime p. In this note, we will focus on the group G such that 3 ≤ |C(G)| ≤ 5.
For a finite group G, denote by πe (G) the set of all element orders of G, and by π(G)
the set of all prime divisors of |G|. For any i ∈ πe (G), denote by Ci (G) the set of all cylic
subgroups of order i in G. Throughout this note, let ci = |Ci (G)|.
0
Support by National Natural Science Foundation of China (Grant No. 11471266).
AMS Subject Classification: 20D15, 20D25
Key words and phrases: finite groups, cyclic subgroups, 2-groups
1
2
2
The main result
For a finite group G, we know
|G| =
X
ck · φ(k),
k∈πe (G)
|C(G)| =
X
(2.1)
ck ,
k∈πe (G)
where φ is the Eucler function.
Lemma 2.1 If |C(G)| = 3, then G ∼
= Cp2 for some prime p.
Proof: We claim that |π(G)| = 1. Otherwise, by Cauchy theorem, |π(G)| = 2. Let
|G| = pa q b , where a, b ≥ 1. By equation 2.1, cp = cq = 1. Hence there exist A ✁ G and
∼ Cp and B ∼
B ✁ G such that A =
= Cq . Thus AB ∼
= Cpq . Note that |C(Cpq )| = 4, a
contradiction.
Hence G is a p-groups. Let |G| = pn . By equation 2.1, cp = 1 or 2. If cp = 2, we get
that pn = 2p + 1, a contradiction. Hence cp = 1. Then cp2 = 1. From equation 2.1, n = 2
and G ∼
✷
= Cp 2 .
Lemma 2.2 If |C(G)| = 4, then G ∼
= Cpq , Cp3 .
Proof: From equation 2.1, we see that |π(G)| ≤ 3. If π(G) = {p, q, r}, then cp = cq = cr = 1
and there exist A, B, C ✁ G such that A ∼
= Cp , B ∼
= Cq and C ∼
= Cr . Thus ABC ∼
= Cpqr .
Note that |C(Cpqr )| > 4, a contradiction. Now we get the following two cases:
Case 1. π(G) = {p, q}. Let |G| = pa q b . From equation 2.1, we have that cp = 2
and cq = 1 or cp = 1 = cq . If cp = 2 and cq = 1, then pa q b = 2p + q − 2. Since
pa q b − 2p − q + 2 ≥ pq − 2p − q + 2 = (p − 1)(q − 2), we get q = 2 and a = b = 1. Since
c2 = 1, we get G ∼
= C2 × Cq , contrary to cp = 2. Therefore cp = cq = 1, and we can find
A, B ✁ G such that A ∼
= Cp and B ∼
= Cq . Note that AB ∼
= Cpq and |C(AB)| = 4. It
follows that G = AB ∼
= Cpq .
Case 2. π(G) = {p}. Let |G| = pn . We know that cp ≤ 3. If cp = 3, then pn =
1 + 3(p − 1). Hence p = 2 and n = 2. This is impossible. If cp = 2, then cq2 = 1, and
pn = p2 + p − 1. Note that pn − p2 − p + 1 ≥ p3 − p2 − p + 1 = (p − 1)(p2 − 1) > 0. This
is impossible. So we get that cp = 1. By [1, Satz 3.8.2], G is a cyclic or p = 2 and G is a
generalized quaternion group. Clearly, we get G ∼
✷
= Cp3 in this case.
3
Lemma 2.3 If |C(G)| = 5, then G ∼
= S 3 , Cp 4 , C3 × C3 , Q 8 .
Proof: Let π(G) = {p1 , · · · , pt }. By Cauchy theorem, t ≤ 4. If t = 4, then cpi = 1 for
i = 1, · · · , 4. Thus G has a normal cyclic subgroup N ∼
= Cp1 p2 p3 p4 . We get a contradiction
for |C(Cp1 p2 p3 p4 )| > 5. If t = 3, from equation 2.1, we get that cp1 = 1 or 2, and
cp = cp = 1. Let V, W ✁ G such that V ∼
= Cp and W ∼
= Cp . Thus |C(V W )| = 4, which
2
3
implies that cp1
2
3
= 1. Let U ✁ G such that U ∼
= Cp1 . Note that U V W ∼
= Cp1 p2 p3 and
|C(Cp1 p2 p3 )| > 5. This is impossible. So we get the following two cases:
Case 1. t = 2. Let |G| = pa q b . Let P ∈ Sylp (G) and Q ∈ Sylq (G). Suppose that
neither P nor Q are cylcic subgroups of prime order. Then |C(P )| ≥ 3 and |C(Q)| ≥ 3.
Thus P ∪ Q contains all the 5 cyclic subgroups of G, and G = P ∪ Q. This is impossible.
So we can assume that P ∼
= Cp , and |G| = pq b .
If b = 1, by symmetry, we can assume P ✁ G. Note |C(Cpq )| = 4. We get that Q is
not normal in G. So cq = |G : NG (Q)| = p. Then p = 3, and q = 2. We get G ∼
= S3 .
Now assume b > 1. If P is not normal in G, then cp = |G : NG (P )| ≥ 1 + p. We get
p = 2 and cp = 3. This imples b = 1, a contradiction. Hence P ✁ G and cp = 1. Clearly,
|C(Q)| ≤ 3. By lemma 2.1, Q ∼
= Cq2 . Now we get cq2 = |G : NG (Q)| ≥ 3, contrary to
|C(G)| = 5.
Case 2. t = 1. Then |G| = pa . If p ≥ 5, then each cyclic subgroup of G must be normal
in G for |C(G)| = 5, which implies G is a Dedekind group. Thus G is abelian p-groups.
Note that |Cp (Cp × Cp )| = p + 1 > 5. It follows that G is cyclic, and then G ∼
= Cp 4 .
Next assume p = 3. Note |C(C3 × C3 )| = 5. Thus G can not have a proper subgroup
isomorphic to C3 × C3 . If G 6∼
= C3 × C3 , then all the subgroup of order 9 are cyclic, which
implies G is cyclic by [1, Satz 3.8.4]. So we get G ∼
= C3 × C3 or C34 .
Finally, we consider the case that p = 2. Let |G| = 2n . Since |C(G)| = 5, we get
exp(G) = 2s ≤ 24 . If s = 1, then G is an elementary abelian 2-group of order 2n and
|C(G)| = 2n , a contradicton. If s = 3, let x ∈ G such that |x| = 23 . Since |C(hxi)| = 4,
there is only one cyclic subgroup in G−hxi. Hence all the elements in G−hxi are contained
in a cyclic subgroup and have the same order, 2r say. So we get |G − hxi| = φ(2r ), and
2n − 8 = 2r−1 , where r ≤ 3. This is impossible.
Clearly, if s = 4, then G ∼
= C24 . We need only to consider the case that s = 2.
Let x ∈ G with |x| = 4. Hence we can get exactly two nontrivial cyclic subgroups from
G − hxi. Let 2r and 22 be the order of the two cyclic subgroups, where r, s ≤ 2. Then
2n − 4 = 2r−1 + 2s−1 . So we get n = 3 and r, s = 2. Thus G = Q8 .
✷
4
Combining lemma 2.1, 2.2 and 2.3, we get the following.
Theorem 2.4 For a finite group G, |C(G)| ≤ 5 if and only G is a subgroup of Cp4 or
G∼
= S3 , Q8 , C3 × C3 , or Cpq , where p and q are two different primes.
References
[1] B. Huppert, Endliche Gruppen I, Springer-Verlag, Berlin 1982.
[2] M. Tǎrnǎuceanu, Finite groups with a certain number of cyclic subgroups, Amer.
Math. Monthly 122(2015): 275-276.
[3] M. Tǎrnǎuceanu, Finite groups with a certain number of cyclic subgroups II,
http://arxiv.org/abs/1604.04974.
[4] W.
Zhou,
On
the
number
http://arxiv.org/abs/1605.00193.
of
cyclic
subgroups
in
finite
groups,
| 4math.GR
|
arXiv:1410.6843v2 [math.ST] 22 Apr 2016
Posteriors, conjugacy, and exponential families
for completely random measures
Tamara Broderick
Ashia C. Wilson
Michael I. Jordan
April 25, 2016
Abstract
We demonstrate how to calculate posteriors for general Bayesian nonparametric priors and likelihoods based on completely random measures
(CRMs). We further show how to represent Bayesian nonparametric priors as a sequence of finite draws using a size-biasing approach—and how
to represent full Bayesian nonparametric models via finite marginals. Motivated by conjugate priors based on exponential family representations
of likelihoods, we introduce a notion of exponential families for CRMs,
which we call exponential CRMs. This construction allows us to specify
automatic Bayesian nonparametric conjugate priors for exponential CRM
likelihoods. We demonstrate that our exponential CRMs allow particularly straightforward recipes for size-biased and marginal representations
of Bayesian nonparametric models. Along the way, we prove that the
gamma process is a conjugate prior for the Poisson likelihood process and
the beta prime process is a conjugate prior for a process we call the odds
Bernoulli process. We deliver a size-biased representation of the gamma
process and a marginal representation of the gamma process coupled with
a Poisson likelihood process.
1
Introduction
An important milestone in Bayesian analysis was the development of a general
strategy for obtaining conjugate priors based on exponential family representations of likelihoods [DeGroot, 1970]. While slavish adherence to exponentialfamily conjugacy can be criticized, conjugacy continues to occupy an important
place in Bayesian analysis, for its computational tractability in high-dimensional
problems and for its role in inspiring investigations into broader classes of priors (e.g., via mixtures, limits, or augmentations). The exponential family is,
however, a parametric class of models, and it is of interest to consider whether
similar general notions of conjugacy can be developed for Bayesian nonparametric models. Indeed, the nonparametric literature is replete with nomenclature
that suggests the exponential family, including familiar names such as “Dirichlet,” “beta,” “gamma,” and “Poisson.” These names refer to aspects of the
1
random measures underlying Bayesian nonparametrics, either the Lévy measure used in constructing certain classes of random measures or properties of
marginals obtained from random measures. In some cases, conjugacy results
have been established that parallel results from classical exponential families; in
particular, the Dirichlet process is known to be conjugate to a multinomial process likelihood [Ferguson, 1973], the beta process is conjugate to a Bernoulli
process [Kim, 1999, Thibaux and Jordan, 2007] and to a negative binomial
process [Broderick et al., 2015]. Moreover, various useful representations for
marginal distributions, including stick-breaking and size-biased representations,
have been obtained by making use of properties that derive from exponential
families. It is striking, however, that these results have been obtained separately,
and with significant effort; a general formalism that encompasses these individual results has not yet emerged. In this paper, we provide the single, holistic
framework so strongly suggested by the nomenclature. Within this single framework, we show that it is straightforward to calculate posteriors and establish
conjugacy. Our framework includes the specification of a Bayesian nonparametric analog of the finite exponential family, which allows us to provide automatic
and constructive nonparametric conjugate priors given a likelihood specification
as well as general recipes for marginal and size-biased representations.
A broad class of Bayesian nonparametric priors—including those built on the
Dirichlet process [Ferguson, 1973], the beta process [Hjort, 1990], the gamma
process [Ferguson, 1973, Lo, 1982, Titsias, 2008], and the negative binomial
process [Zhou et al., 2012, Broderick et al., 2015]—can be viewed as models
for the allocation of data points to traits. These processes give us pairs of
traits together with rates or frequencies with which the traits occur in some
population. Corresponding likelihoods assign each data point in the population
to some finite subset of traits conditioned on the trait frequencies. What makes
these models nonparametric is that the number of traits in the prior is countably
infinite. Then the (typically random) number of traits to which any individual
data point is allocated is unbounded, but also there are always new traits to
which as-yet-unseen data points may be allocated. That is, such a model allows
the number of traits in any data set to grow with the size of that data set.
A principal challenge of working with such models arises in posterior inference. There is a countable infinity of trait frequencies in the prior which we
must integrate over to calculate the posterior of trait frequencies given allocations of data points to traits. Bayesian nonparametric models sidestep the full
infinite-dimensional integration in three principal ways: conjugacy, size-biased
representations, and marginalization.
In its most general form, conjugacy simply asserts that the prior is in the
same family of distributions as the posterior. When the prior and likelihood are
in finite-dimensional conjugate exponential families, conjugacy can turn posterior calculation into, effectively, vector addition. As a simple example, consider a
model with beta-distributed prior, θ ∼ Beta(θ|α, β), for some fixed hyperparameters α and β. For the likelihood, let each observation xn with n ∈ {1, . . . , N } be
iid
iid Bernoulli-distributed conditional on parameter θ: xn ∼ Bern(x|θ). Then the
2
posterior is simply another beta distribution, Beta(θ|αpost , βpost ), with paramePN
PN
ters updated via addition: αpost := α + n=1 xn and βpost := β + N − n=1 xn .
While conjugacy is certainly useful and popular in the case of finite parameter cardinality, there is arguably a stronger computational imperative for its
use in the infinite-parameter case. Indeed, the core prior-likelihood pairs of
Bayesian nonparametrics are generally proven [Hjort, 1990, Kim, 1999, Lo, 1982,
Thibaux and Jordan, 2007, Broderick et al., 2015], or assumed to be [Titsias,
2008, Thibaux, 2008], conjugate. When such proofs exist, though, thus far
they have been specialized to specific pairs of processes. In what follows, we
demonstrate a general way to calculate posteriors for a class of distributions that
includes all of these classical Bayesian nonparametric models. We also define
a notion of exponential family representation for the infinite-dimensional case
and show that, given a Bayesian nonparametric exponential family likelihood,
we can readily construct a Bayesian nonparametric conjugate prior.
Size-biased sampling provides a finite-dimensional distribution for each of
the individual prior trait frequencies [Thibaux and Jordan, 2007, Paisley et al.,
2010]. Such a representation has played an important role in Bayesian nonparametrics in recent years, allowing for either exact inference via slice sampling [Damien et al., 1999, Neal, 2003]—as demonstrated by Teh et al. [2007],
Broderick et al. [2015]—or approximate inference via truncation [Doshi et al.,
2009, Paisley et al., 2011]. This representation is particularly useful for building
hierarchical models [Thibaux and Jordan, 2007]. We show that our framework
yields such representations in general, and we show that our construction is
especially straightforward to use in the exponential family framework that we
develop.
Marginal processes avoid directly representing the infinite-dimensional prior
and posterior altogether by integrating out the trait frequencies. Since the
trait allocations are finite for each data point, the marginal processes are finite for any finite set of data points. Again, thus far, such processes have
been shown to exist separately in special cases; for example, the Indian buffet
process [Griffiths and Ghahramani, 2006] is the marginal process for the beta
process prior paired with a Bernoulli process likelihood [Thibaux and Jordan,
2007]. We show that the integration that generates the marginal process from
the full Bayesian model can be generally applied in Bayesian nonparametrics
and takes a particularly straightforward form when using conjugate exponential
family priors and likelihoods. We further demonstrate that, in this case, a basic, constructive recipe exists for the general marginal process in terms of only
finite-dimensional distributions.
Our results are built on the general class of stochastic processes known as
completely random measures (CRMs) [Kingman, 1967]. We review CRMs in
Section 2.1 and we discuss what assumptions are needed to form a full Bayesian
nonparametric model from CRMs in Section 2.3. Given a general Bayesian
nonparametric prior and likelihood (Section 2.2), we demonstrate in Section 3
how to calculate the posterior. Although the development up to this point is
more general, we next introduce a concept of exponential families for CRMs
3
(Section 4.1) and call such models exponential CRMs. We show that we can
generate automatic conjugate priors given exponential CRM likelihoods in Section 4.2. Finally, we show how we can generate recipes for size-biased representations (Section 5) and marginal processes (Section 6), which are particularly
straightforward in the exponential CRM case (Corollary 5.2 in Section 5 and
Corollary 6.2 in Section 6). We illustrate our results on a number of examples
and derive new conjugacy results, size-biased representations, and marginal processes along the way.
We note that some similar results have been obtained by Orbanz [2010] and
James [2014]. In the present work, we focus on creating representations that
allow tractable inference.
2
Bayesian models based on completely random
measures
As we have discussed, we view Bayesian nonparametric models as being composed of two parts: (1) a collection of pairs of traits together with their frequencies or rates and (2) for each data point, an allocation to different traits. Both
parts can be expressed as random measures. Recall that a random measure is
a random element whose values are measures.
We represent each trait by a point ψ in some space Ψ of traits. Further,
let θk be the frequency, or rate, of the trait represented by ψk , where k indexes
the countably many traits. In particular, θk ∈ R+ . Then (θk , ψk ) is a tuple
consisting of the frequency of the kth trait together with its trait descriptor.
We can represent the full collection of pairs of traits with their frequencies by
the discrete measure on Ψ that places weight θk at location ψk :
Θ=
K
X
θk δ ψ k ,
(1)
k=1
where the cardinality K may be finite or infinity.
Next, we form data point Xn for the nth individual. The data point Xn is
viewed as a discrete measure. Each atom of Xn represents a pair consisting of
(1) a trait to which the nth individual is allocated and (2) a degree to which
the nth individual is allocated to this particular trait. That is,
Xn =
Kn
X
xn,k δψn,k ,
(2)
k=1
where again ψn,k ∈ Ψ represents a trait and now xn,k ∈ R+ represents the degree
to which the nth data point belongs to trait ψn,k . Kn is the total number of
traits to which the nth data point belongs.
Here and in what follows, we treat X1:N = {Xn : n ∈ [N ]} as our observed
data points for [N ] := {1, 2, 3, . . . , N }. In practice X1:N is often incorporated
4
into a more complex Bayesian hierarchical model. For instance, in topic modeling, ψk represents a topic; that is, ψk is a distribution over words in a vocabulary [Blei et al., 2003, Teh et al., 2006]. θk might represent the frequency with
which the topic ψk occurs in a corpus of documents. xn,k might be a positive
integer and represent the number of words in topic ψn,k that occur in the nth
P Kn
document. So the nth document has a total length of k=1
xn,k words. In this
case, the actual observation consists of the words in each document, and the
topics are latent. Not only are the results concerning posteriors, conjugacy, and
exponential family representations that we develop below useful for inference in
such models, but in fact our results are especially useful in such models—where
the traits and any ordering on the traits are not known in advance.
Next, we want to specify a full Bayesian model for our data points X1:N .
To do so, we must first define a prior distribution for the random measure Θ
as well as a likelihood for each random measure Xn conditioned on Θ. We let
ΣΨ be a σ-algebra of subsets of Ψ, where we assume all singletons are in ΣΨ .
Then we consider random measures Θ and Xn whose values are measures on Ψ.
Note that for any random measure Θ and any measurable set A ∈ ΣΨ , Θ(A) is
a random variable.
2.1
Completely random measures
We can see from Eqs. (1) and (2) that we desire a distribution on random
measures that yields discrete measures almost surely. A particularly simple
form of random measure called a completely random measure can be used to
generate a.s. discrete random measures [Kingman, 1967].
A completely random measure Θ is defined as a random measure that satisfies one additional property; for any disjoint, measurable sets A1 , A2 , . . . , AK ∈
ΣΨ , we require that Θ(A1 ), Θ(A2 ), . . . , Θ(AK ) be independent random variables. Kingman [1967] showed that a completely random measure can always
be decomposed into a sum of three independent parts:
Θ = Θdet + Θf ix + Θord .
(3)
Here, Θdet is the deterministic component, Θf ix is the fixed-location component,
and Θord is the ordinary component. In particular, Θdet is any deterministic
measure. We define the remaining two parts next.
The fixed-location component is called the “fixed component” by Kingman
[1967], but we expand the name slightly here to emphasize that Θf ix is defined
to be constructed from a set of random weights at fixed (i.e., deterministic)
locations. That is,
Kf ix
X
θf ix,k δψf ix,k ,
(4)
Θf ix =
k=1
where the number of fixed-location atoms, Kf ix , may be either finite or infinity;
ψf ix,k is deterministic, and θf ix,k is a non-negative, real-valued random variable
(since Φ is a measure). Without loss of generality, we assume that the locations
5
ψf ix,k are all distinct. Then, by the independence assumption of CRMs, we must
have that θf ix,k are independent random variables across k. Although the fixedlocation atoms are often ignored in the Bayesian nonparametrics literature, we
will see that the fixed-location component has a key role to play in establishing
Bayesian nonparametric conjugacy and in the CRM representations we present.
The third and final component is the ordinary component. Let #(A) denote
the cardinality of some countable set A. Let µ be any σ-finite, deterministic
measure on R+ × Ψ, where R+ is equipped with the Borel σ-algebra and ΣR+ ×Ψ
is the resulting product σ-algebra given ΣΨ . Recall that a Poisson point process
with rate measure µ on R+ × Ψ is a random countable subset Π of R+ × Ψ such
that two properties hold [Kingman, 1992]:
1. For any A ∈ ΣR+ ×Ψ , #(Π ∩ A) ∼ Poisson(µ(A)).
2. For any disjoint A1 , A2 , . . . , AK ∈ ΣR+ ×Ψ , #(Π∩A1 ), #(Π∩A2 ), · · · , #(Π∩
AK ) are independent random variables.
To generate an ordinary component, start with a Poisson point process on R+ ×
Ψ, characterized by its rate measure µ(dθ×dψ). This process yields Π, a random
ord
and countable set of points: Π = {(θord,k , ψord,k )}K
k=1 , where Kord may be finite
or infinity. Form the ordinary component measure by letting θord,k be the weight
of the atom located at ψord,k :
Θord =
K
ord
X
θord,k δψord,k .
(5)
k=1
Recall that we stated at the start of Section 2.1 that CRMs may be used to
produce a.s. discrete random measures. To check this assertion, note that Θf ix
is a.s. discrete by construction (Eq. (4)) and Θord is a.s. discrete by construction
(Eq. (5)). Θdet is the one component that may not be a.s. atomic. Thus the
prevailing norm in using models based on CRMs is to set Θdet ≡ 0; in what
follows, we adopt this norm. If the reader is concerned about missing any atoms
in Θdet , note that it is straightforward to adapt the treatment of Θf ix to include
the case where the atom weights are deterministic. When we set Θdet ≡ 0, we
are left with Θ = Θf ix + Θord by Eq. (3). So Θ is also discrete, as desired.
2.2
Prior and likelihood
The prior that we place on Θ will be a fully general CRM (minus any deterministic component) with one additional assumption on the rate measure of the
ordinary component. Before incorporating the additional assumption, we say
that Θ has a fixed-location component with Kf ix atoms, where the kth atom
indep
has arbitrary distribution Ff ix,k : θf ix,k ∼ Ff ix,k (dθ). Kf ix may be finite
or infinity, and Θ has an ordinary component characterized by rate measure
µ(dθ × dψ). The additional assumption we make is that the distribution on
6
the weights in the ordinary component is assumed to be decoupled from the
distribution on the locations. That is, the rate measure decomposes as
µ(dθ × dψ) = ν(dθ) · G(dψ),
(6)
where ν is any σ-finite, deterministic measure on R+ and G is any proper
distribution on Ψ. While the distribution over locations has been discussed
extensively elsewhere [Neal, 2000, Wang and Blei, 2013], it is the weights that
affect the allocation of data points to traits.
Given the factorization of µ in Eq. (6), the ordinary component of Θ can
ord
be generated by letting {θf ix,k }K
k=1 be the points of a Poisson point process
1
ord
generated on R+ with rate ν. We then draw the locations {ψf ix,k }K
k=1 iid
iid
according to G(dψ): ψf ix,k ∼ G(dψ). Finally, for each k, θf ix,k δψf ix,k is an
atom in Θord. This factorization will allow us to focus our attention on the
trait frequencies, and not the trait locations, in what follows. Moreover, going
forward, we will assume G is diffuse (i.e., G has no atoms) so that the ordinary
component atoms are all at a.s. distinct locations, which are further a.s. distinct
from the fixed locations.
Since we have seen that Θ is an a.s. discrete random measure, we can write
it as
K
X
Θ=
θk δ ψ k ,
(7)
k=1
where K := Kf ix + Kord may be finite or infinity, and every ψk is a.s. unique.
That is, we will sometimes find it helpful notationally to use Eq. (7) instead of
separating the fixed and ordinary components. At this point, we have specified
the prior for Θ in our general model.
Next, we specify the likelihood; i.e., we specify how to generate the data
points Xn given Θ. We will assume each Xn is generated iid given Θ across the
data indices n. We will let Xn be a CRM with only a fixed-location component
given Θ. In particular, the atoms of Xn will be located at the atom locations
of Θ, which are fixed when we condition on Θ:
Xn :=
K
X
xn,k δψk .
k=1
Here, xn,k is drawn according to some distribution H that may take θk , the
weight of Θ at location ψk , as a parameter; i.e.,
indep
xn,k ∼ H(dx|θk ) independently across n and k.
(8)
Note that while every atom of Xn is located at an atom of Θ, it is not
necessarily the case that every atom of Θ has a corresponding atom in Xn . In
particular, if xn,k is zero for any k, there is no atom in Xn at ψk .
1 Recall that K
ord may be finite or infinity depending on ν and is random when taking
finite values.
7
We highlight that the model above stands in contrast to Bayesian nonparametric partition models, for which there is a large literature. In partition models
(or clustering models), Θ is a random probability measure [Ferguson, 1974]; in
this case, the probability constraint precludes Θ from being a completely random measure, but it is often chosen to be a normalized completely random
measure [James et al., 2009, Lijoi and Prünster, 2010]. The choice of Dirichlet process (a normalized gamma process) for Θ is particularly popular due
to a number of useful properties that coincide in this single choice [Doksum,
1974, Escobar, 1994, Escobar and West, 1995, 1998, Ferguson, 1973, Lo, 1984,
MacEachern, 1994, Perman et al., 1992, Pitman, 1996a,b, Sethuraman, 1994,
West and Escobar, 1994]. In partition models, Xn is a draw from the probability distribution described by Θ. If we think of such Xn as a random measure,
it is a.s. a single unit mass at a point ψ with strictly positive probability in Θ.
One potential connection between these two types of models is provided
by combinatorial clustering [Broderick et al., 2015]. In partition models, we
might suppose that we have a number of data sets, all of which we would like
to partition. For instance, in a document modeling scenario, each document
might be a data set; in particular each data point is a word in the document.
And we might wish to partition the words in each document. An alternative
perspective is to suppose that there is a single data set, where each data point
is a document. Then the document exhibits traits with multiplicities, where the
multiplicities might be the number of words from each trait; typically a trait
in this application would be a topic. In this case, there are a number of other
names besides feature or trait model that may be applied to the overarching
model—such as admixture model or mixed membership model [Airoldi et al.,
2014].
2.3
Bayesian nonparametrics
So far we have described a prior and likelihood that may be used to form a
Bayesian model. We have already stated above that forming a Bayesian nonparametric model imposes some restrictions on the prior and likelihood. We
formalize these restrictions in Assumptions A0, A1, and A2 below.
Recall that the premise of Bayesian nonparametrics is that the number of
traits represented in a collection of data can grow with the number of data
points. More explicitly, we achieve the desideratum that the number of traits is
unbounded, and may always grow as new data points are collected, by modeling
a countable infinity of traits. This assumption requires that the prior have
a countable infinity of atoms. These must either be fixed-location atoms or
ordinary component atoms. Fixed-location atoms represent known traits in
some sense since we must know the fixed locations of the atoms in advance.
Conversely, ordinary component atoms represent unknown traits, as yet to be
discovered, since both their locations and associated rates are unknown a priori.
Since we cannot know (or represent) a countable infinity of traits a priori, we
cannot start with a countable infinity of fixed-location atoms.
8
A0. The number of fixed-location atoms in Θ is finite.
Since we require a countable infinity of traits in total and they cannot come
from the fixed-location atoms by Assumption A0, the ordinary component must
contain a countable infinity of atoms. This assumption will be true if and only
if the rate measure on the trait frequencies has infinite mass.
A1. ν(R+ ) = ∞.
Finally, an implicit part of the starting premise is that each data point be
allocated to only a finite number of traits; we do not expect to glean an infinite
amount of information from finitely represented data. Thus, we require that
the number of atoms in every Xn be finite. By Assumption A0, the number
of atoms in Xn that correspond to fixed-location atoms in Θ is finite. But by
Assumption A1, the number of atoms in Θ from the ordinary component is
infinite. So there must be some restriction on the distribution of values of X at
the atoms of Θ (that is, some restriction on H in Eq. (8)) such that only finitely
many of these values are nonzero.
In particular, note that if H(dx|θ) does not contain an atom at zero for any
θ, then a.s. every one of the countable infinity of atoms of X will be nonzero.
Conversely, it follows that, for our desiderata to hold, we must have that H(dx|θ)
exhibits an atom at zero. One consequence of this observation is that H(dx|θ)
cannot be purely continuous for all θ. Though this line of reasoning does not
necessarily preclude a mixed continuous and discrete H, we henceforth assume
that H(dx|θ) is discrete, with support Z∗ = {0, 1, 2, . . .}, for all θ.
In what follows, we write h(x|θ) for the probability mass function of x given
θ. So our requirement that each data point be allocated to only a finite number of traits translates into a requirement that the number of atoms of Xn
with values in Z+ = {1, 2, . . .} be finite. Note that, by construction, the pairs
ord
{(θord,k , xord,k )}K
k=1 form a marked Poisson point process with rate measure
µmark (dθ × dx) := ν(dθ)h(x|θ). And the pairs with xord,k equal to any particular value x ∈ Z+ further form a thinned Poisson point process with rate
measure νx (dθ) := ν(dθ)h(x|θ). In particular, the number of atoms of X with
weight x is Poisson-distributed with mean νx (R+ ). So the number of atoms of
X is finite if and only if the following assumption holds.2
P∞
A2.
x=1 νx (R+ ) < ∞ for νx := ν(dθ)h(x|θ).
Thus Assumptions A0, A1, and A2 capture our Bayesian nonparametric
desiderata. We illustrate the development so far with an example.
Example 2.1. The beta process [Hjort, 1990] provides an example distribution
for Θ. In its most general form, sometimes called the three-parameter beta
2 When we have the more general case of a mixed continuous and discrete H, Assumption A2
becomes
R
R
A2b. x>0 θ∈R ν(dθ)H(dx|θ) < ∞.
+
9
process [Teh and Görür, 2009, Broderick et al., 2012], the beta process has an
ordinary component whose weight rate measure has a beta distribution kernel,
ν(dθ) = γθ−α−1 (1 − θ)c+α−1 dθ,
(9)
with support on (0, 1]. Here, the three fixed hyperparameters are γ, the mass
parameter ; c, the concentration parameter ; and α, the discount parameter.3
Moreover, each of its Kf ix fixed-location atoms, θk δψk , has a beta-distributed
weight [Broderick et al., 2015]:
θf ix,k ∼ Beta(θ|ρf ix,k , σf ix,k ),
(10)
where ρf ix,k , σf ix,k > 0 are fixed hyperparameters of the model.
By Assumption A0, Kf ix is finite. By Assumption A1, ν(R+ ) = ∞. To
achieve this infinite-mass restriction, the beta kernel in Eq. (9) must be improper; i.e., either −α ≤ 0 or c + α ≤ 0. Also, note that we must have γ > 0
since ν is a measure (and the case γ = 0 would be trivial).
Often the beta process is used as a prior paired with a Bernoulli process
likelihood [Thibaux
and Jordan, 2007]. The Bernoulli process specifies that,
P∞
given Θ = k=1 θk δψk , we draw
indep
xn,k ∼ Bern(x|θk ),
which is well-defined since every atom weight θk of Θ is in (0, 1] by the beta
process construction. Thus,
Xn =
∞
X
xn,k δψk .
k=1
The marginal distribution of the X1:N in this case is often called an Indian buffet process [Griffiths and Ghahramani, 2006, Thibaux and Jordan, 2007]. The
locations of atoms in Xn are thought of as the dishes sampled by the nth customer.
We take a moment to highlight the fact that continuous distributions for
H(dx|θ) are precluded based on the Bayesian nonparametric desiderata by considering an alternative likelihood. Consider instead if H(dx|θ) were continuous
here. Then X1 would have atoms at every atom of Θ. In the Indian buffet process analogy, any customer would sample an infinite number of dishes,
which contradicts our assumption that our data are finite. Indeed, any customer would sample all of the dishes at once. It is quite often the case in
practical applications, though, that the Xn are merely latent variables, with
the observed variables chosen according to a (potentially continuous) distribution given Xn [Griffiths and Ghahramani, 2006, Thibaux and Jordan, 2007];
3 In [Teh and Görür, 2009, Broderick et al., 2012], the ordinary component features the
beta distribution kernel in Eq. (9) multiplied not only by γ but also by a more complex, positive, real-valued expression in c and α. Since all of γ, c, and α are fixed hyperparameters, and
γ is an arbitrary positive real value, any other constant factors containing the hyperparameters
can be absorbed into γ, as in the main text here.
10
consider, e.g., mixture and admixture models. These cases are not precluded
by our development.
Finally, then, we may apply Assumption A2, which specifies that the number
of atoms in each observation Xn is finite; in this case, the assumption means
Z
∞ Z
X
ν(dθ) · h(1|θ)
ν(dθ) · h(x|θ) =
x=1
θ∈(0,1]
θ∈R+
since θ is supported on (0, 1] and x is supported on {0, 1}
Z
Z
θ1−α−1 (1 − θ)c+α−1 dθ < ∞.
γθ−α−1 (1 − θ)c+α−1 dθ · θ = γ
=
θ∈(0,1]
θ∈(0,1]
The integral here is finite if and only if 1 − α and c + α are the parameters of a
proper beta distribution: i.e., if and only if α < 1 and c > −α. Together with
the restrictions above, these restrictions imply the following allowable parameter
ranges for the beta process fixed hyperparameters:
γ > 0,
α ∈ [0, 1),
c > −α,
ρf ix,k , σf ix,k > 0 for all k ∈ [Kf ix ].
(11)
These correspond to the hyperparameter ranges previously found in [Teh and Görür,
2009, Broderick et al., 2012].
3
Posteriors
In Section 2, we defined a full Bayesian model consisting of a CRM prior for Θ
and a CRM likelihood for an observation X conditional on Θ. Now we would
like to calculate the posterior distribution of Θ|X.
Theorem 3.1 (Bayesian nonparametric posteriors). Let Θ be a completely random measure that satisfies Assumptions A0 and A1; that is, Θ is a CRM with
Kf ix fixed atoms such that Kf ix < ∞ and such that the kth atom can be written
θf ix,k δψf ix,k with
indep
θf ix,k ∼ Ff ix,k (dθ)
for proper distribution Ff ix,k and deterministic ψf ix,k . Let the ordinary component of Θ have rate measure
µ(dθ × dψ) = ν(dθ) · G(dψ),
P∞
where G is a proper distribution and ν(R+ ) = ∞. Write ΘP=
k=1 θk δψk ,
∞
and let X be generated conditional on Θ according to X = k=1 xk δψk with
indep
xk ∼ h(x|θk ) for proper, discrete probability mass function h. And suppose
X and Θ jointly satisfy Assumption A2 so that
∞ Z
X
ν(dθ)h(x|θ) < ∞.
x=1
θ∈R+
Then let Θpost be a random measure with the distribution of Θ|X. Θpost is
a completely random measure with three parts.
11
1. For each k ∈ [Kf ix ], Θpost has a fixed-location atom at ψf ix,k with weight
θpost,f ix,k distributed according to the finite-dimensional posterior Fpost,f ix,k (dθ)
that comes from prior Ff ix,k , likelihood h, and observation X({ψf ix,k }).
2. Let {xnew,k δψnew,k : k ∈ [Knew ]} be the atoms of X that are not at fixed
locations in the prior of Θ. Knew is finite by Assumption A2. Then Θpost
has a fixed-location atom at xnew,k with random weight θpost,new,k , whose
distribution Fpost,new,k (dθ) is proportional to
ν(dθ)h(xnew,k |θ).
3. The ordinary component of Θpost has rate measure
νpost (dθ) := ν(dθ)h(0|θ).
Proof. To prove the theorem, we consider in turn each of the two parts of
the prior: the fixed-location component and the ordinary component. First,
consider any fixed-location atom, θf ix,k δψf ix,k , in the prior. All of the other
fixed-location atoms in the prior, as well as the prior ordinary component, are
independent of the random weight θf ix,k . So it follows that all of X except
xf ix,k := X({ψf ix,k }) is independent of θf ix,k . Thus the posterior has a fixed
atom located at ψf ix,k whose weight, which we denote θpost,f ix,k , has distribution
Fpost,f ix,k (dθ) ∝ Ff ix,k (dθ)h(xf ix,k |θ),
which follows from the usual finite Bayes Theorem.
Next, consider the ordinary component in the prior. Let
Ψf ix = {ψf ix,1 , . . . , ψf ix,Kf ix }
be the set of fixed-location atoms in the prior. Recall that Ψf ix is deterministic,
and since G is continuous, all of the fixed-location atoms and ordinary component atoms of Θ are at a.s. distinct locations. So the measure Xf ix defined
by
Xf ix (A) := X(A ∩ Ψf ix )
can be derived purely from X, without knowledge of Θ. It follows that the
measure Xord defined by
Xord (A) := X(A ∩ (Ψ\Ψf ix ))
can be derived purely from X without knowledge of Θ. Xord is the same as
the observed data measure X but with atoms only at atoms of the ordinary
component of Θ and not at the fixed-location atoms of Θ.
Now for any value x ∈ Z+ , let
{ψnew,x,1 , . . . , ψnew,x,Knew,x }
be all of the locations of atoms of size x in Xord. By Assumption A2, the number
of such atoms, Knew,x , is finite. Further let θnew,x,k := Θ({ψnew,x,k }). Then
12
K
new,x
are generated from a thinned Poisson point process
the values {θnew,x,k }k=1
with rate measure
νx (dθ) := ν(dθ)h(x|θ).
(12)
And since νx (R+ ) < ∞ by assumption, each θnew,x,k has distribution equal to
the normalized rate measure in Eq. (12). Note that θnew,x,k δψnew,x,k is a fixedlocation atom in the posterior now that its location is known from the observed
Xord .
By contrast, if a likelihood draw at an ordinary component atom in the prior
returns a zero, that atom is not observed in Xord . Such atom weights in Θpost
thus form a marked Poisson point process with rate measure
ν(dθ)h(0|θ),
as was to be shown.
In Theorem 3.1, we consider generating Θ and then a single data point
X conditional on Θ. Now suppose we generate Θ and then N data points,
X1 , . . . , XN , iid conditional on Θ. In this case, Theorem 3.1 may be iterated to
find the posterior Θ|X1:N . In particular, Theorem 3.1 gives the ordinary component and fixed atoms of the random measure Θ1 := Θ|X1 . Then, using Θ1 as
the prior measure and X2 as the data point, another application of Theorem 3.1
gives Θ2 := Θ|X1:2 . We continue recursively using Θ|X1:n for n between 1 and
N − 1 as the prior measure until we find Θ|X1:N . The result is made explicit in
the following corollary.
Corollary 3.2 (Bayesian nonparametric posteriors given multiple data points).
Let Θ be a completely random measure that satisfies Assumptions A0 and A1;
that is, Θ is a CRM with Kf ix fixed atoms such that Kf ix < ∞ and such that
the kth atom can be written θf ix,k δψf ix,k with
indep
θf ix,k ∼ Ff ix,k (dθ)
for proper distribution Ff ix,k and deterministic ψf ix,k . Let the ordinary component of Θ have rate measure
µ(dθ × dψ) = ν(dθ) · G(dψ),
P∞
where G is a proper distribution and ν(R+ ) = ∞. Write Θ = P
k=1 θk δψk , and
∞
let X1 , . . . , Xn be generated conditional on Θ according to X = k=1 xn,k δψn,k
indep
with xn,k ∼ h(x|θk ) for proper, discrete probability mass function h. And
suppose X1 and Θ jointly satisfy Assumption A2 so that
∞ Z
X
ν(dθ)h(x|θ) < ∞.
x=1
θ∈R+
It is enough to make the assumption for X1 since the Xn are iid conditional on
Θ.
Then let Θpost be a random measure with the distribution of Θ|X1:N . Θpost
is a completely random measure with three parts.
13
1. For each k ∈ [Kf ix ], Θpost has a fixed-location atom at ψf ix,k with weight
θpost,f ix,k distributed according to the finite-dimensional posterior Fpost,f ix,k (dθ)
that comes from prior Ff ix,k , likelihood h, and observation X({ψf ix,k }).
2. Let {ψnew,k : k ∈ [Knew ]} be the union of atom locations across X1 , X2 , . . . , XN
minus the fixed locations in the prior of Θ. Knew is finite. Let xnew,n,k
be the weight of the atom in Xn located at ψnew,k . Note that at least one
of xnew,n,k across n must be non-zero, but in general xnew,n,k may equal
zero. Then Θpost has a fixed-location atom at xnew,k with random weight
θpost,new,k , whose distribution Fpost,new,k (dθ) is proportional to
ν(dθ)
N
Y
h(xnew,n,k |θ).
n=1
3. The ordinary component of Θpost has rate measure
n
νpost,n (dθ) := ν(dθ) [h(0|θ)] .
Proof. Corollary 3.2 follows from recursive application of Theorem 3.1. In order
to recursively apply Theorem 3.1, we need to verify that Assumptions A0, A1,
and A2 hold for the posterior Θ|X1:(n+1) when they hold for the prior Θ|X1:n .
Note that the number of fixed atoms in the posterior is the number of fixed atoms
in the prior plus the number of new atoms in the posterior. By Theorem 3.1,
these counts are both finite as long as Θ|X1:n satisfies Assumptions A0 and A2,
which both hold for n = 0 by assumption and n > 0 by the recursive assumption.
So Assumption A0 holds for Θ|X1:(n+1) .
Next we notice that since Assumption A1 implies that there is an infinite
number of ordinary component atoms in Θ|X1:n and only finitely many become
fixed atoms in the posterior by Assumption A2, it must be that Θ|X1:(n+1)
has infinitely many ordinary component atoms. So Assumption A1 holds for
Θ|X1:(n+1) .
Finally, we note that
∞ Z
X
x=1
=
νpost,n (dθ)h(x|θ)
θ∈R+
∞ Z
X
x=1
n
ν(dθ) [h(0|θ)] h(x|θ) ≤
θ∈R+
∞ Z
X
x=1
ν(dθ)h(x|θ) < ∞,
θ∈R+
where the penultimate inequality follows since h(0|θ) ∈ [0, 1] and where the
inequality follows by Assumption A2 on the original Θ (conditioned on no data).
So Assumption A2 holds for Θ|X1:(n+1) .
We now illustrate the results of the theorem with an example.
Example 3.3. Suppose we again start with a beta process prior for Θ as
in Example 2.1. This time we consider a negative binomial process likelihood
14
[Zhou et al., 2012, Broderick
et al., 2015]. The negative
binomial process specP∞
P∞
ifies that, given Θ = k=1 θk δψk , we draw X = k=1 xk δψk with
indep
xk ∼ NegBin(x|r, θk ),
for some fixed hyperparameter r > 0. So
Xn =
∞
X
xn,k δψk .
k=1
In this case, Assumption A2 translates into the following restriction.
∞ Z
X
x=1
=
Z
ν(dθ) · h(x|θ)
θ∈R+
ν(dθ) · [1 − h(0|θ)] =
Z
γθ−α−1 (1 − θ)c+α−1 dθ · [1 − (1 − θ)r ] < ∞,
θ∈(0,1]
θ∈R+
where the penultimate equality follows since the support of ν(dθ) is (0, 1].
By a Taylor expansion, we have 1 − (1 − θ)r = rθ + o(θ) as θ → 0, so we
require
Z
θ1−α−1 (1 − θ)c+α−1 dθ < ∞,
θ∈(0,1]
which is satisfied if and only if 1 − α and c + α are the parameters of a proper
beta distribution. Thus, we have the same parameter restrictions as in Eq. (11).
Now we calculate the posterior given the beta process prior on Θ and the
negative binomial process likelihood for X conditional on Θ. In particular,
the posterior has the distribution of Θpost , a CRM with three parts given by
Theorem 3.1.
First, at each fixed atom ψf ix,k of the prior with weight θf ix,k given by
Eq. (10), there is a fixed atom in the posterior with weight θpost,f ix,k . Let
xpost,f ix,k := X({ψf ix,k }). Then θpost,f ix,k has distribution
Fpost,f ix,k (dθ) ∝ Ff ix (dθ) · h(xpost,f ix,k |θ)
= Beta(θ|ρf ix,k , σf ix,k ) dθ · NegBin(xpost,f ix,k |r, θ)
∝ θρf ix,k −1 (1 − θ)σf ix,k −1 dθ · θxpost,f ix,k (1 − θ)r
(13)
∝ Beta (θ |ρf ix,k + xpost,f ix,k , σf ix,k + r ) dθ.
Second, for any atom xnew,k δψnew,k in X that is not at a fixed location in the
prior, Θpost has a fixed atom at ψnew,k whose weight θpost,new,k has distribution
Fpost,new,k (dθ) ∝ ν(dθ) · h(xnew,k |θ)
= ν(dθ) · NegBin(xnew,k |r, θ)
∝ θ−α−1 (1 − θ)c+α−1 dθ · θxnew,k (1 − θ)r
∝ Beta (θ |−α + xnew,k , c + α + r ) dθ,
15
(14)
which is a proper distribution since we have the following restrictions on its
parameters. For one, by assumption, xnew,k ≥ 1. And further, by Eq. (11), we
have α ∈ [0, 1) as well as c + α > 0 and r > 0.
Third, the ordinary component of Θpost has rate measure
ν(dθ)h(0|θ) = γθ−α−1 (1 − θ)c+α−1 dθ · (1 − θ)r = γθ−α−1 (1 − θ)c+r+α−1 dθ.
Not only have we found the posterior distribution Θpost above, but now
we can note that the posterior is in the same form as the prior with updated
ordinary component hyperparameters:
γpost = γ,
αpost = α,
cpost = c + r.
The posterior also has old and new beta-distributed fixed atoms with beta distribution hyperparameters given in Eq. (13) and Eq. (14), respectively. Thus,
we have proven that the beta process is, in fact, conjugate to the negative binomial process. An alternative proof was first given by Broderick et al. [2015].
As in Example 3.3, we can use Theorem 3.1 not only to calculate posteriors but also, once those posteriors are calculated, to check for conjugacy. This
approach unifies existing disparate approaches to Bayesian nonparametric conjugacy. However, it still requires the practitioner to guess the right conjugate
prior for a given likelihood. In the next section, we define a notion of exponential families for CRMs, and we show how to automatically construct a conjugate
prior for any exponential family likelihood.
4
Exponential families
Exponential families are what typically make conjugacy so powerful in the finite
case. For one, when a finite likelihood belongs to an exponential family, then existing results give an automatic conjugate, exponential family prior for that likelihood. In this section, we review finite exponential families, define exponential
CRMs, and show that analogous automatic conjugacy results can be obtained
for exponential CRMs. Our development of exponential CRMs will also allow
particularly straightforward results for size-biased representations (Corollary 5.2
in Section 5) and marginal processes (Corollary 6.2 in Section 6).
In the finite-dimensional case, suppose we have some (random) parameter θ
and some (random) observation x whose distribution is conditioned on θ. We
say the distribution Hexp,like of x conditional on θ is in an exponential family if
Hexp,like (dx|θ) = hexp,like (x|θ) dx = κ(x) exp {hη(θ), φ(x)i − A(θ)} µ(dx),
(15)
where η(θ) is the natural parameter, φ(x) is the sufficient statistic, κ(x) is the
base density, and A(θ) is the log partition function. We denote the density
16
of Hexp,like here, which exists by definition, by hexp,like . The measure µ—
with respect to which the density hexp,like exists—is typically Lebesgue measure
when Hexp,like is diffuse or counting measure when Hexp,like is atomic. A(θ) is
determined by the condition that Hexp,like (dx|θ) have unit total mass on its
support.
It is a classic result [Diaconis and Ylvisaker, 1979] that the following distribution for θ ∈ RD constitutes a conjugate prior:
Fexp,prior (dθ) = fexp,prior (θ) dθ = exp {hξ, η(θ)i + λ [−A(θ)] − B(ξ, λ)} dθ.
(16)
Fexp,prior is another exponential family distribution, now with natural parameter (ξ ′ , λ)′ , sufficient statistic (η(θ)′ , −A(θ))′ , and log partition function B(ξ, λ).
Note that the logarithms of the densities in both Eq. (15) and Eq. (16) are linear
in η(θ) and −A(θ). So, by Bayes Theorem, the posterior Fexp,post also has these
quantities as sufficient statistics in θ, and we can see Fexp,post must have the
following form.
Fexp,post (dθ|x) = fexp,post (θ|x) dθ
= exp {hξ + φ(x), η(θ)i + (λ + 1) [−A(θ)] − B(ξ + φ(x), λ + 1)} dθ.
(17)
Thus we see that Fexp,post belongs to the same exponential family as Fexp,prior
in Eq. (16), and hence Fexp,prior is a conjugate prior for Hexp,like in Eq. (15).
4.1
Exponential families for completely random measures
In the finite-dimensional case, we saw that for any exponential family likelihood,
as in Eq. (15), we can always construct a conjugate exponential family prior,
given by Eq. (16).
In order to prove a similar result for CRMs, we start by defining a notion of
exponential families for CRMs.
Definition 4.1. We say that a CRM Θ is an exponential CRM if it has the
following two parts. First, let Θ have Kf ix fixed-location atoms, where Kf ix
may be finite or infinite. The kth fixed-location atom is located at any ψf ix,k ,
unique from the other fixed locations, and has random weight θf ix,k , whose
distribution has density ff ix,k :
ff ix,k (θ) = κ(θ) exp {hη(ζk ), φ(θ)i − A(ζk )} ,
for some base density κ, natural parameter function η, sufficient statistic φ,
and log partition function A shared across atoms. Here, ζk is an atom-specific
parameter.
Second, let Θ have an ordinary component with rate measure µ(dθ × dψ) =
ν(dθ) · G(dψ) for some proper distribution G and weight rate measure ν of the
form
ν(dθ) = γ exp {hη(ζ), φ(θ)i} .
In particular, η and φ are shared with the fixed-location atoms, and fixed hyperparameters γ and ζ are unique to the ordinary component.
17
4.2
Automatic conjugacy for completely random measures
With Definition 4.1 in hand, we can specify an automatic Bayesian nonparametric conjugate prior for an exponential CRM likelihood.
P∞
Theorem 4.2 (Automatic conjugacy). Let Θ =
k=1 θk δψk , in accordance
with Assumption A1. Let X be generated conditional on Θ according to an
exponential CRM with fixed-location atoms at {ψk }∞
k=1 and no ordinary component. In particular, the distribution of the weight xk at ψk of X has the
following density conditional on the weight θk at ψk of Θ:
h(x|θk ) = κ(x) exp {hη(θk ), φ(x)i − A(θk )} .
Then a conjugate prior for Θ is the following exponential CRM distribution.
First, let Θ have Kprior,f ix fixed-location atoms, in accordance with Assumption A0. The kth such atom has random weight θf ix,k with proper density
fprior,f ix,k (θ) = exp {hξf ix,k , η(θ)i + λf ix,k [−A(θ)] − B(ξf ix,k , λf ix,k )} ,
where (η ′ , −A)′ here is the sufficient statistic and B is the log partition function.
ξf ix,k and λf ix,k are fixed hyperparameters for this atom weight.
Second, let Θ have ordinary component characterized by any proper distribution G and weight rate measure
ν(dθ) = γ exp {hξ, η(θ)i + λ [−A(θ)]} ,
where γ, ξ, and λ are fixed hyperparameters of the weight rate measure chosen
to satisfy Assumptions A1 and A2.
Proof. To prove the conjugacy of the prior for Θ with the likelihood for X, we
calculate the posterior distribution of Θ|X using Theorem 3.1. Let Θpost be
a CRM with the distribution of Θ|X. Then, by Theorem 3.1, Θpost has the
following three parts.
First, at any fixed location ψf ix,k in the prior, let xf ix,k be the value of X
at that location. Then Θpost has a fixed-location atom at ψf ix,k , and its weight
θpost,f ix,k has distribution
Fpost,f ix,k (dθ) ∝ fprior,f ix,k (θ) dθ · h(xf ix,k |θ)
∝ exp {hξf ix,k , η(θ)i + λf ix,k [−A(θ)]} dθ · exp {hη(θ), φ(xf ix,k )i − A(θ)} dθ
= exp {hξf ix,k + φ(xf ix,k ), η(θ)i + (λf ix,k + 1) [−A(θ)]} dθ.
It follows, from putting in the normalizing constant, that the distribution of
θpost,f ix,k has density
fpost,f ix,k (θ) = exp {hξf ix,k + φ(xf ix,k ), η(θ)i + (λf ix,k + 1) [−A(θ)]
− B(ξf ix,k + φ(xf ix,k ), λf ix,k + 1)} .
Second, for any atom xnew,k δψnew,k in X that is not at a fixed location in the
prior, Θpost has a fixed atom at ψnew,k whose weight θpost,new,k has distribution
Fpost,new,k (θ) ∝ ν(dθ) · h(xnew,k |θ)
18
∝ exp {hξ, η(θ)i + λ [−A(θ)]} · exp {hη(θ), φ(xnew,k )i − A(θ)} dθ
= exp {hξ + φ(xnew,k ), η(θ)i + (λ + 1) [−A(θ)]} dθ
and hence density
fpost,new,k (θ) = exp {hξ + φ(xnew,k ), η(θ)i + (λ + 1) [−A(θ)] − B(ξ + φ(xnew,k ), λ + 1)} .
Third, the ordinary component of Θpost has weight rate measure
ν(dθ) · h(0|θ)
= γ exp {hξ, η(θ)i + λ [−A(θ)]} · κ(0) exp {hη(θ), φ(0)i − A(θ)}
= γκ(0) · exp {hξ + φ(0), η(θ)i + (λ + 1) [−A(θ)]} .
Thus, the posterior rate measure is in the same exponential CRM form as
the prior rate measure with updated hyperparameters:
γpost = γκ(0),
ξpost = ξ + φ(0),
λpost = λ + 1.
Since we see that the posterior fixed-location atoms are likewise in the same
exponential CRM form as the prior, we have shown that conjugacy holds, as
desired.
We next use Theorem 4.2 to give proofs of conjugacy in cases where conjugacy has not previously been established in the Bayesian nonparametrics literature.
4
Example 4.3. Let X be generated
P∞according to a Poisson likelihood
P∞process
conditional on Θ. That is, X = k=1 xk δψk conditional on Θ = k=1 θk δψk
has an exponential CRM distribution with only a fixed-location component.
The weight xk at location ψk has support on Z∗ and has a Poisson density with
parameter θk ∈ R+ :
h(x|θk ) =
1 x −θk
1
θk e
=
exp {x log(θk ) − θk } .
x!
x!
(18)
The final line is rewritten to emphasize the exponential family form of this
density, with
κ(x) =
1
,
x!
φ(x) = x,
η(θ) = log(θ),
A(θ) = θ.
By Theorem 4.2, this Poisson likelihood process has a Bayesian nonparametric
conjugate prior for Θ with two parts.
First, Θ has a set of Kprior,f ix fixed-location atoms, where Kprior,f ix < ∞
by Assumption A0. The kth such atom has random weight θf ix,k with density
fprior,f ix,k (θ) = exp {hξf ix,k , η(θ)i + λf ix,k [−A(θ)] − B(ξf ix,k , λf ix,k )}
4 We use the term “Poisson likelihood process” to distinguish this specific Bayesian nonparametric likelihood from the Poisson point process.
19
= θξf ix,k e−λf ix,k θ exp {−B(ξf ix,k , λf ix,k )}
= Gamma(θ |ξf ix,k + 1, λf ix,k ),
(19)
where Gamma(θ|a, b) denotes the gamma density with shape parameter a > 0
and rate parameter b > 0. So we must have fixed hyperparameters ξf ix,k > −1
and λf ix,k > 0. Further,
ξ
ix,k
exp {−B(ξf ix,k , λf ix,k )} = λffix,k
+1
/Γ(ξf ix,k + 1)
to ensure normalization.
Second, Θ has an ordinary component characterized by any proper distribution G and weight rate measure
ν(dθ) = γ exp {hξ, η(θ)i + λ [−A(θ)]} dθ = γθξ e−λθ dθ.
(20)
Note that Theorem 4.2 guarantees that the weight rate measure will have the
same distributional kernel in θ as the fixed-location atoms.
Finally, we need to choose the allowable hyperparameter ranges for γ, ξ, and
λ. First, γ > 0 to ensure ν is a measure. By Assumption A1, we must have
ν(R+ ) = ∞, so ν must represent an improper gamma distribution. As such, we
require either ξ + 1 ≤ 0 or λ ≤ 0. By Assumption A2, we must have
Z
Z
∞ Z
X
γθξ e−λθ dθ · 1 − e−θ < ∞.
ν(dθ) · [1 − h(0|θ)] =
ν(dθ) · h(x|θ) =
x=1
θ∈R+
θ∈R+
θ∈R+
To ensure the integral over [1, ∞) is finite, we must have λ > 0. To ensure the
integral over (0, 1) is finite, we note that 1 − e−θ = θ + o(θ) as θ → 0. So we
require
Z
γθξ+1 e−λθ dθ < ∞,
θ∈(0,1)
which is satisfied if and only if ξ + 2 > 0.
Finally, then the hyperparameter restrictions can be summarized as:
γ > 0,
ξ ∈ (−2, −1],
λ > 0;
ξf ix,k > −1 and λf ix,k > 0
for all k ∈ [Kprior,f ix ].
The ordinary component of the conjugate prior for Θ discovered in this
example is typically called a gamma process. Here, we have for the first time
specified the distribution of the fixed-location atoms of the gamma process and,
also for the first time, proved that the gamma process is conjugate to the Poisson
likelihood process. We highlight this result as a corollary to Theorem 4.2.
Corollary 4.4. Let the Poisson likelihood process be a CRM with fixed-location
atom weight distributions as in Eq. (18). Let the gamma process be a CRM with
fixed-location atom weight distributions as in Eq. (19) and ordinary component
weight measure as in Eq. (20). Then the gamma process is a conjugate Bayesian
nonparametric prior for the Poisson likelihood process.
20
Example 4.5. Next, let X be generated according to a new process we call
an odds Bernoulli process. We have previously seen a typical Bernoulli process
likelihood in Example 2.1. In the odds Bernoulli process, we say that X, conditional on Θ, has an exponential CRM distribution. In this case, the weight of
the kth atom, xk , conditional on θk has support on {0, 1} and has a Bernoulli
density with odds parameter θk ∈ R+ :
h(x|θk ) = θkx (1 + θk )−1
= exp {x log(θk ) − log(1 + θk )} .
(21)
That is, if ρ is the probability of a successful Bernoulli draw, then θ = ρ/(1 − ρ)
represents the odds ratio of the probability of success over the probability of
failure.
The final line of Eq. (21) is written to emphasize the exponential family form
of this density, with
κ(x) = 1,
φ(x) = x,
η(θ) = log(θ),
A(θ) = log(1 + θ).
By Theorem 4.2, the likelihood for X has a Bayesian nonparametric conjugate
prior for Θ. This conjugate prior has two parts.
First, Θ has a set of Kprior,f ix fixed-location atoms. The kth such atom has
random weight θf ix,k with density
fprior,f ix,k (θ) = exp {hξf ix,k , η(θ)i + λf ix,k [−A(θ)] − B(ξf ix,k , λf ix,k )}
= θξf ix,k (1 + θ)−λf ix,k exp {−B(ξf ix,k , λf ix,k )}
= BetaPrime (θ |ξf ix,k + 1, λf ix,k − ξf ix,k − 1 ) ,
(22)
where BetaPrime(θ|a, b) denotes the beta prime density with shape parameters
a > 0 and b > 0. Further,
exp {−B(ξf ix,k , λf ix,k )} =
Γ(λf ix,k )
Γ(ξf ix,k + 1)Γ(λf ix,k − ξf ix,k − 1)
to ensure normalization.
Second, Θ has an ordinary component characterized by any proper distribution G and weight rate measure
ν(dθ) = γ exp {hξ, η(θ)i + λ [−A(θ)]} dθ = γθξ (1 + θ)−λ dθ.
(23)
We need to choose the allowable hyperparameter ranges for γ, ξ, and λ.
First, γ > 0 to ensure ν is a measure. By Assumption A1, we must have
ν(R+ ) = ∞, so ν must represent an improper beta prime distribution. As such,
we require either ξ + 1 ≤ 0 or λ − ξ − 1 ≤ 0. By Assumption A2, we must have
∞ Z
X
x=1
θ∈R+
ν(dθ) · h(x|θ) =
Z
ν(dθ) · h(1|θ)
θ∈R+
since the support of x is {0, 1}
21
=
Z
γθξ (1 + θ)−λ dθ · θ1 (1 + θ)−1 = γ
θ∈R+
Z
θξ+1 (1 + θ)−λ−1 dθ < ∞.
θ∈R+
Since the integrand is the kernel of a beta prime distribution, we simply require
that this distribution be proper; i.e., ξ + 2 > 0 and λ − ξ − 1 > 0.
The hyperparameter restrictions can be summarized as:
γ > 0, ξ ∈ (−2, −1], λ > ξ + 1; ξf ix,k > −1 and λf ix,k > ξf ix,k + 1 for all k ∈ [Kprior,f ix ].
We call the distribution for Θ described in this example the beta prime process. Its ordinary component has previously been defined by Broderick et al.
[2015]. But this result represents the first time the beta prime process is described in full, including parameter restrictions and fixed-location atoms, as well
as the first proof of its conjugacy with the odds Bernoulli process. We highlight
the latter result as a corollary to Theorem 4.2 below.
Corollary 4.6. Let the odds Bernoulli process be a CRM with fixed-location
atom weight distributions as in Eq. (21). Let the beta process be a CRM with
fixed-location atom weight distributions as in Eq. (22) and ordinary component
weight measure as in Eq. (23). Then the beta process is a conjugate Bayesian
nonparametric prior for the odds Bernoulli process.
5
Size-biased representations
We have shown in Section 4.2 that our exponential CRM (Definition 4.1) is
useful in that we can find an automatic Bayesian nonparametric conjugate prior
given an exponential CRM likelihood. We will see in this section and the next
that exponential CRMs allow us to build representations that allow tractable
inference despite the infinite-dimensional nature of the models we are using.
The best-known size-biased representation of a random measure in Bayesian
nonparametrics is the stick-breaking representation of the Dirichlet process ΘDP
[Sethuraman, 1994]:
ΘDP =
∞
X
θDP,k δψk ;
k=1
For k ∈ Z∗ , θDP,k = βk
k−1
Y
(24)
(1 − βj ),
iid
βk ∼ Beta(1, c),
iid
ψk ∼ G,
j=1
where c is a fixed hyperparameter satisfying c > 0.
The name “stick-breaking” originates from thinking of the unit interval as
a stick of length one. At each round k, only some of the stick remains; βk
describes the proportion of the remaining stick that is broken off in round k,
and θDP,k describes the total amount of remaining stick that is broken off in
22
round k. By construction, not only is each θDP,k ∈ (0, 1) but in fact the θDP,k
add to one (the total stick length) and thus describe a distribution.
Eq. (24) is called a size-biased representation for the following reason. Since
the weights {θDP,k }∞
k=1 describe a distribution, we can make draws from this
distribution; each such draw is sometimes thought of as a multinomial draw with
a single trial. In that vein, typically we imagine that our data points Xmult,n are
described as iid draws conditioned on ΘDP , where Xmult,n is a random measure
with just a single atom:
Xmult,n = δψmult,n ;
ψmult,n = ψk with probability θDP,k .
(25)
Then the limiting proportion of data points Xmult,n with an atom at ψmult,1
(the first atom location chosen) is θDP,1 . The limiting proportion of data points
with an atom at the next unique atom location chosen will have size θDP,2 , and
so on [Broderick et al., 2013].
The representation in Eq. (24) is so useful because there is a familiar, finitedimensional distribution for each of the atom weights θDP,k of the random measure ΘDP . This representation allows approximate inference via truncation
[Ishwaran and James, 2001] or exact inference via slice sampling [Walker, 2007,
Kalli et al., 2011].
Since the weights {θDP,k }∞
k=1 are constrained to sum to one, the Dirichlet
process is not a CRM.5 Indeed, there has been much work on size-biased representations for more general normalized random measures, which include the
Dirichlet process as just one example [Perman et al., 1992, Pitman, 1996a,b,
2003].
By contrast, we here wish to explore size-biasing for non-normalized CRMs.
In the normalized CRM case, we considered which atom of a random discrete
probability measure was drawn first and what is the distribution of that atom’s
size. In the non-normalized CRM case considered in the present work, when
drawing X conditional on Θ, there may be multiple atoms (or one atom or no
atoms) of Θ that correspond to non-zero atoms in X. The number will always
be finite though by Assumption A2. In this non-normalized CRM case, we
wish to consider the sizes of all such atoms in Θ. Size-biased representations
have been developed in the past for particular CRM examples, notably the beta
process [Paisley et al., 2010, Broderick et al., 2012]. And even though there is
typically no interpretation of these representations in terms of a single stick
representing a unit probability mass, they are sometimes referred to as stickbreaking representations as a nod to the popularity of Dirichlet process stickbreaking.
In the beta process case, such size-biased representations have already been
shown to allow approximate inference via truncation [Doshi et al., 2009, Paisley et al.,
2011] or exact inference via slice sampling [Teh et al., 2007, Broderick et al.,
2015]. Here we provide general recipes for the creation of these representations and illustrate our recipes by discovering previously unknown size-biased
5 In fact, the Dirichlet process is a normalized gamma process (cf. Example 4.3) [Ferguson,
1973].
23
representations.
We have seen that a general CRM Θ takes the form of an a.s. discrete random
measure:
∞
X
θk δ ψ k .
(26)
k=1
The fixed-location atoms are straightforward to simulate; there are finitely many
by Assumption A0, their locations are fixed, and their weights are assumed to
come from finite-dimensional distributions. The infinite-dimensionality of the
Bayesian nonparametric CRM comes from the ordinary component (cf. Section 2.3 and Assumption A1). So far the only description we have of the ordinary component is its generation from the countable infinity of points in a
Poisson point process. The next result constructively demonstrates that we
can represent the distributions of the CRM weights {θk }∞
k=1 in Eq. (26) as a
sequence of finite-dimensional distributions, much as in the familiar Dirichlet
process case.
Theorem 5.1 (Size-biased representations). Let Θ be a completely random
measure that satisfies Assumptions A0 and A1; that is, Θ is a CRM with Kf ix
fixed atoms such that Kf ix < ∞ and such that the kth atom can be written
θf ix,k δψf ix,k . The ordinary component of Θ has rate measure
µ(dθ × dψ) = ν(dθ) · G(dψ),
P∞
where G is a proper distribution and ν(R+ ) = ∞. Write P
Θ =
k=1 θk δψk ,
∞
and let Xn be generated iid given Θ according to Xn =
k=1 xn,k δψk with
indep
xn,k ∼ h(x|θk ) for proper, discrete probability mass function h. And suppose
Xn and Θ jointly satisfy Assumption A2 so that
∞ Z
X
x=1
ν(dθ)h(x|θ) < ∞.
θ∈R+
Then we can write
Θ=
m,x
∞ X
∞ ρX
X
θm,x,j δψm,x,j
m=1 x=1 j=1
iid
ψm,x,k ∼ G iid across m, x, j
Z
indep
ρm,x ∼ Poisson ρ
ν(dθ)h(0|θ)m−1 h(x|θ) across m, x
(27)
θ
indep
θm,x,j ∼ Fsize,m,x (dθ) ∝ ν(dθ)h(0|θ)m−1 h(x|θ)
iid across j and independently across m, x.
Proof. By construction, Θ is an a.s. discrete random measure with a countable
infinity of atoms. Without loss of generality, suppose that for every (non-zero)
value of an atom weight θ, there is a non-zero probability of generating an atom
24
with non-zero weight x in the likelihood. Now suppose we generate X1 , X2 , . . ..
Then, for every atom θδψ of Θ, there exists some finite n with an atom at ψ.
Therefore, we can enumerate all of the atoms of Θ by enumerating
• Each atom θδψ such that there is an atom in X1 at ψ.
• Each atom θδψ such that there is an atom in X2 at ψ but there is not an
atom in X1 at ψ.
..
.
• Each atom θδψ such that there is an atom in Xm at ψ but there is not an
atom in any of X1 , X2 , . . . , Xm−1 at ψ.
..
.
Moreover, on the mth round of this enumeration, we can further break down
the enumeration by the value of the observation Xm at the atom location:
• Each atom θδψ such that there is an atom in Xm of weight 1 at ψ but
there is not an atom in any of X1 , X2 , . . . , Xm−1 at ψ.
• Each atom θδψ such that there is an atom in Xm of weight 2 at ψ but
there is not an atom in any of X1 , X2 , . . . , Xm−1 at ψ.
..
.
• Each atom θδψ such that there is an atom in Xm of weight x at ψ but
there is not an atom in any of X1 , X2 , . . . , Xm−1 at ψ.
..
.
Recall that the values θk that form the weights of Θ are generated according
to a Poisson point process with rate measure ν(dθ). So, on the first round, the
values of θk such that x1,k = x also holds are generated according to a thinned
Poisson point process with rate measure
ν(dθ)h(x|θ).
In particular, since the rate measure has finite total mass by Assumption A2,
we can define
Z
M1,x := ν(dθ)h(x|θ),
θ
which will be finite. Then the number of atoms θk for which x1,k = x is
ρ1,x ∼ Poisson(ρ|M1,x ).
And each such θk has weight with distribution
Fsize,1,x (dθ) ∝ ν(dθ)h(x|θ).
25
Finally, note from Theorem 3.1 that the posterior Θ|X1 has weight rate measure
ν1 (dθ) := ν(dθ)h(0|θ).
Now take any m > 1. Suppose, inductively, that the ordinary component of
the posterior Θ|X1 , . . . , Xm−1 has weight rate measure
νm−1 (dθ) := ν(dθ)h(0|θ)m−1 .
The atoms in this ordinary component have been selected precisely because they
have not appeared in any of X1 , . . . , Xm−1 . As for m = 1, we have that the
atoms θk in this ordinary component with corresponding weight in Xm equal to
x are formed by a thinned Poisson point process, with rate measure
νm−1 (dθ)h(x|θ) = ν(dθ)h(0|θ)m−1 h(x|θ).
Since the rate measure has finite total mass by Assumption A2, we can define
Z
Mm,x := ν(dθ)h(0|θ)m−1 h(x|θ),
θ
which will be finite. Then the number of atoms θk for which x1,k = x is
ρm,x ∼ Poisson(ρ|Mm,x ).
And each such θk has weight
Fsize,m,x ∝ ν(dθ)h(0|θ)m−1 h(x|θ).
Finally, note from Theorem 3.1 that the posterior Θ|X1:m , which can be thought
of as generated by prior Θ|X1:(m−1) and likelihood Xm |Θ, has weight rate measure
ν(dθ)h(0|θ)m−1 h(0|θ) = νm (dθ),
confirming the inductive hypothesis.
Recall that every atom of Θ is found in exactly one of these rounds and that
x ∈ Z+ . Also recall that the atom locations may be generated independently
and identically across atoms, and independently from all the weights, according
to proper distribution G (Section 2.2). To summarize, we have then
Θ=
m,x
∞ X
∞ ρX
X
θm,x,j δψm,x,j ,
m=1 x=1 j=1
where
iid
ψm,x,k ∼ G iid across m, x, j
Z
Mm,x = ν(dθ)h(0|θ)m−1 h(x|θ) across m, x
θ
26
indep
ρm,x ∼ Poisson(ρ|Mm,x ) across m, x
Fsize,m,x (dθ) ∝ ν(dθ)h(0|θ)m−1 h(x|θ) across m, x
indep
θm,x,j ∼ Fsize,m,x (dθ) iid across j and independently across m, x,
as was to be shown.
The following corollary gives a more detailed recipe for the calculations in
Theorem 5.1 when the prior is in a conjugate exponential CRM to the likelihood.
Corollary 5.2 (Exponential CRM size-biased representations). Let Θ be an
exponential CRM with no fixed-location atoms (thereby trivially satisfying Assumption A0) such that Assumption A1 holds.
Let X be generated conditional on Θ according to an exponential CRM with
fixed-location atoms at {ψk }∞
k=1 and no ordinary component. Let the distribution
of the weight xn,k at ψk have probability mass function
h(x|θk ) = κ(x) exp {hη(θk ), φ(x)i − A(θk )} .
Suppose that Θ and X jointly satisfy Assumption A2. And let Θ be conjugate
to X as in Theorem 4.2. Then we can write
Θ=
m,x
∞ X
∞ ρX
X
θm,x,j δψm,x,j
m=1 x=1 j=1
iid
ψm,x,j ∼ G
iid across m, x, j
Mm,x = γ · κ(0)m−1 κ(x) · exp {B(ξ + (m − 1)φ(0) + φ(x), λ + m)}
indep
ρm,x ∼ Poisson (ρ|Mm,x )
independently across m, x
(28)
indep
θm,x,j ∼ fsize,m,x (θ) dθ
= exp {hξ + (m − 1)φ(0) + φ(x), η(θ)i + (λ + m)[−A(θ)]
− B(ξ + (m − 1)φ(0) + φ(x), λ + m)}
iid across j and independently across m, x.
Proof. The corollary follows from Theorem 5.1 by plugging in the particular
forms for ν(dθ) and h(x|θ).
In particular,
Z
ν(dθ)h(0|θ)m−1 h(x|θ)
Mm,x =
θ∈R+
Z
γ exp {hξ, η(θ)i + λ [−A(θ)]}
=
θ∈R+
m−1
· [κ(0) exp {hη(θ), φ(0)i − A(θ)}]
· κ(x) exp {hη(θ), φ(x)i − A(θ)} dθ
27
= γκ(0)m−1 κ(x) exp {B (ξ + (m − 1)φ(0) + φ(x), λ + m)} ,
Corollary 5.2 can be used to find the known size-biased representation of
the beta process [Thibaux and Jordan, 2007]; we demonstrate this derivation in
detail in Example B.1 in Appendix B. Here we use Corollary 5.2 to discover a
new size-biased representation of the gamma process.
Example 5.3. Let Θ be a gamma process, and let Xn be iid Poisson likelihood
processes conditioned on Θ for each n as in Example 4.3. That is, we have
ν(dθ) = γθξ e−λθ dθ.
And
h(x|θk ) =
1 x −θk
θ e
x! k
with
γ > 0,
ξ ∈ (−2, −1],
ξf ix,k > −1 and λf ix,k > 0
λ > 0;
for all k ∈ [Kprior,f ix ]
by Example 4.3.
We can pick out the following components of h:
κ(x) =
1
,
x!
φ(x) = x,
η(θ) = log(θ),
A(θ) = θ.
Thus, by Corollary 5.2, we have
fsize,m,x (θ) ∝ θξ+x e−(λ+m)θ ∝ Gamma (θ |ξ + x + 1, λ + m ) .
We summarize the representation that follows from Corollary 5.2 in the following
result.
Corollary 5.4. Let the gamma process be a CRM Θ with fixed-location atom
weight distributions as in Eq. (19) and ordinary component weight measure as
in Eq. (20). Then we may write
Θ=
m,x
∞ X
∞ ρX
X
θm,x,j δψm,x,j
m=1 x=1 j=1
iid
ψm,x,j ∼ G
Mm,x = γ ·
iid across m, x, j
1
· Γ(ξ + x + 1) · (λ + m)−(ξ+x+1) across m, x
x!
indep
ρm,x ∼ Poisson (ρ|Mm,x) across m, x
indep
θm,x,j ∼ Gamma (θ |ξ + x + 1, λ + m )
iid across j and independently across m, x.
28
6
Marginal processes
In Section 5, although we conceptually made use of the observations {X1 , X2 , . . .},
we focused on a representation of the prior Θ: cf. Eqs. (27) and (28). In this
section, we provide a representation of the marginal of X1:N , with Θ integrated
out.
The canonical example of a marginal process again comes from the Dirichlet
process (DP). In this case, the full model consists of the DP-distributed prior
on ΘDP (as in Eq. (24)) together with the likelihood for Xmult,n conditional on
ΘDP (iid across n) described by Eq. (25). Then the marginal distribution of
Xmult,1:N is described by the Chinese restaurant process. This marginal takes
the following form.
For each n = 1, 2, . . . , N ,
K
n−1
be the union of atom locations in Xmult,1 , . . . , Xmult,n−1 .
1. Let {ψk }k=1
Then Xmult,n |Xmult,1 , . . . , Xmult,n−1 has a single atom at ψ, where
PKn−1
ψk
with probability ∝
k=1 Xmult,m ({ψk })
ψ=
ψnew with probability ∝ c
ψnew ∼ G
In the case of CRMs, the canonical example of a marginal process is the Indian buffet process [Griffiths and Ghahramani, 2006]. Both the Chinese restaurant process and Indian buffet process have proven popular for inference since
the underlying infinite-dimensional prior is integrated out in these processes
and only the finite-dimensional marginal remains. By Assumption A2, we know
that the marginal will generally be finite-dimensional for our CRM Bayesian
models. And thus we have the following general marginal representations for
such models.
Theorem 6.1 (Marginal representations). Let Θ be a completely random measure that satisfies Assumptions A0 and A1; that is, Θ is a CRM with Kf ix
fixed atoms such that Kf ix < ∞ and such that the kth atom can be written
θf ix,k δψf ix,k . The ordinary component of Θ has rate measure
µ(dθ × dψ) = ν(dθ) · G(dψ),
P∞
where G is a proper distribution and ν(R+ ) = ∞. Write P
Θ =
k=1 θk δψk ,
∞
and let Xn be generated iid given Θ according to Xn =
x
δψk with
n,k
k=1
indep
xn,k ∼ h(x|θk ) for proper, discrete probability mass function h. And suppose
Xn and Θ jointly satisfy Assumption A2 so that
∞ Z
X
ν(dθ)h(x|θ) < ∞.
x=1
θ∈R+
Then the marginal distribution of X1:N is the same as that provided by the
following construction.
For each n = 1, 2, . . . , N ,
29
K
n−1
be the union of atom locations in X1 , . . . , Xn−1 . Let xm,k :=
1. Let {ψk }k=1
Xm ({ψk }). Let xn,k denote the weight of Xn |X1 , . . . , Xn−1 at ψk . Then
xn,k has distribution described by the following probability mass function:
R
Qn−1
m=1 h(xm,k |θ)
θ∈R+ ν(dθ)h(x|θ)
.
hcond xn,k = x x1:(n−1),k =
R
Qn−1
m=1 h(xm,k |θ)
θ∈R+ ν(dθ)
2. For each x = 1, 2, . . .
ρ
n,x
,
• Xn has ρn,x new atoms. That is, Xn has atoms at locations {ψn,x,j }j=1
where
ρn,x
Kn−1
= ∅ a.s.
∩ {ψk }k=1
{ψn,x,j }j=1
Moreover,
Z
indep
ρn,x ∼ Poisson ρ
ν(dθ)h(0|θ)n−1 h(x|θ) across n, x
θ
iid
ψn,x,j ∼ G(dψ) across n, x, j.
Proof. We saw in the proof of Theorem 5.1 that the marginal for X1 can be
expressed as follows. For each x ∈ Z+ , there are ρ1,x atoms of X1 with weight
x, where
Z
indep
ν(dθ)h(x|θ) across x.
ρ1,x ∼ Poisson
These atoms have locations
θ
ρ1,x
{ψ1,x,j }j=1 ,
where
iid
ψ1,x,j ∼ G(dψ) across x, j.
P∞
1
For the upcoming induction, let K1 := x=1 ρ1,x . And let {ψk }K
k=1 be the (a.s.
ρ1,x
disjoint by assumption) union of the sets {ψ1,x,j }j=1 across x. Note that K1 is
finite by Assumption A2.
We will also find it useful in the upcoming induction to let Θpost,1 have the
distribution of Θ|X1 . Let θpost,1,x,j = Θpost,1 ({ψ1,x,j }). By Theorem 3.1 or the
proof of Theorem 5.1, we have that
indep
θpost,1,x,j ∼ Fpost,1,x,j (dθ) ∝ ν(dθ)h(x|θ)
independently across x and iid across j.
K
n−1
is the union
Now take any n > 1. Inductively, we assume {ψn−1,k }k=1
of all the atom locations of X1 , . . . , Xn−1 . Further assume Kn−1 is finite. Let
Θpost,n−1 have the distribution of Θ|X1 , . . . , Xn−1 . Let θn−1,k be the weight of
Θpost,n−1 at ψn−1,k . And, for any m ∈ [n − 1], let xm,k be the weight of Xm at
ψn−1,k . We inductively assume that
indep
θn−1,k ∼ Fn−1,k (dθ) ∝ ν(dθ)
n−1
Y
m=1
independently across k.
30
h(xm,k |θ)
(29)
Now let ψn,k equal ψn−1,k for k ∈ [Kn−1 ]. Let xn,k denote the weight of Xn
at ψn,k for k ∈ [Kn−1 ]. Conditional on the atom weight of Θ at ψn,k , the atom
weights of X1 , . . . , Xn−1 , Xn are independent. Since the atom weights of Θ are
independent as well, we have that xn,k |X1 , . . . , Xn−1 has the same distribution
as xn,k |x1,k , . . . , xn−1,k . We can write the probability mass function of this
distribution as follows.
hcond (xn,k = x |x1,k , . . . , xn−1,k )
Z
Fn−1,k (dθ)h(x|θ)
=
θ∈R+
=
R
θ∈R+
R
h
ν(dθ)
θ∈R+
Qn−1
m=1
ν(dθ)
i
h(xm,k |θ) · h(x|θ)
Qn−1
m=1
h(xm,k |θ)
,
where the last line follows from Eq. (29).
We next show the inductive hypothesis in Eq. (29) holds for n and k ∈
[Kn−1 ]. Let xn,k denote the weight of Xn at ψn,k for k ∈ [Kn−1 ]. Let Fn,k (dθ)
denote the distribution of xn,k and note that
Fn,k (dθ) ∝ Fn−1,k (dθ) · h(xn,k |θ)
n
Y
= ν(dθ)
h(xm,k |θ),
m=1
which agrees with Eq. (29) for n when we assume the result for n − 1.
The previous development covers atoms that are present in at least one of
X1 , . . . , Xn−1 . Next we consider new atoms in Xn ; that is, we consider atoms in
Xn for which there are no atoms at the same location in any of X1 , . . . , Xn−1 .
We saw in the proof of Theorem 5.1 that, for each x ∈ Z+ , there are ρn,x
new atoms of Xn with weight x such that
Z
indep
ρn,x ∼ Poisson ρ
ν(dθ)h(0|θ)n−1 h(x|θ) across x.
θ
ρ
n,x
with
These new atoms have locations {ψn,x,j }j=1
iid
ψn,x,j ∼ G(dψ) across x, j.
By Assumption A2,
P∞
x=1
ρn,x < ∞. So
Kn := Kn−1 +
∞
X
ρn,x
x=1
remains finite. Let ψn,k for k ∈ {Kn−1 + 1, . . . , Kn } index these new locations.
Let θn,k be the weight of Θpost,n at ψn,k for k ∈ {Kn−1 + 1, . . . , Kn }. And let
xn,k be the value of X at ψn,k .
31
We check that the inductive hypothesis holds. By repeated application of
Theorem 3.1, the ordinary component of Θ|X1 , . . . , Xn−1 has rate measure
ν(dθ)h(0|θ)n−1 .
So, again by Theorem 3.1, we have that
indep
θn,k ∼ Fn.k (dθ) ∝ ν(dθ)h(0|θ)n−1 h(xn,k |θ).
Since Xm has value 0 at ψn,k for m ∈ {1, . . . , n − 1} by construction, we have
that the inductive hypothesis holds.
As in the case of size-biased representations (Section 5 and Corollary 5.2),
we can find a more detailed recipe when the prior is in a conjugate exponential
CRM to the likelihood.
Corollary 6.2 (Exponential CRM marginal representations). Let Θ be an exponential CRM with no fixed-location atoms (thereby trivially satisfying Assumption A0) such that Assumption A1 holds.
Let X be generated conditional on Θ according to an exponential CRM with
fixed-location atoms at {ψk }∞
k=1 and no ordinary component. Let the distribution
of the weight xn,k at ψk have probability mass function
h(x|θk ) = κ(x) exp {hη(θk ), φ(x)i − A(θk )} .
Suppose that Θ and X jointly satisfy Assumption A2. And let Θ be conjugate
to X as in Theorem 4.2. Then the marginal distribution of X1:N is the same as
that provided by the following construction.
For each n = 1, 2, . . . , N ,
K
n−1
be the union of atom locations in X1 , . . . , Xn−1 . Let xm,k :=
1. Let {ψk }k=1
Xm ({ψk }). Let xn,k denote the weight of Xn |X1 , . . . , Xn−1 at ψk . Then
xn,k has distribution described by the following probability mass function:
hcond xn,k = x x1:(n−1),k
(
)
n−1
n−1
X
X
= κ(x) exp −B(ξ +
xm , λ + n − 1) + B(ξ +
xm + x, λ + n) .
m=1
m=1
2. For each x = 1, 2, . . .
ρ
n,x
,
• Xn has ρn,x new atoms. That is, Xn has atoms at locations {ψn,x,j }j=1
where
ρn,x
Kn−1
= ∅ a.s.
∩ {ψk }k=1
{ψn,x,j }j=1
Moreover,
Mn,x := γ · κ(0)n−1 κ(x) · exp {B(ξ + (n − 1)φ(0) + φ(x), λ + n)}
across n, x
indep
ρn,x ∼ Poisson (ρ |Mn,x ) across n, x
iid
ψn,x,j ∼ G(dψ) across n, x, j.
32
Proof. The corollary follows from Theorem 6.1 by plugging in the forms for
ν(dθ) and h(x|θ).
In particular,
Z
ν(dθ)
θ∈R+
=
Z
n
Y
h(xm,k |θ)
m=1
γ exp {hξ, η(θ)i + λ [−A(θ)]} ·
θ∈R+
=γ
"
n
Y
"
n
Y
κ(xm,k ) exp {hη(θ), φ(xm,k )i − A(θ)}
m=1
#
κ(xm,k ) B
m=1
ξ+
n
X
#
!
φ(xm,k ), λ + n .
m=1
So
hcond xn,k = x x1:(n−1),k
R
Qn−1
ν(dθ)h(x|θ) m=1
h(xm,k |θ)
θ∈R+
=
R
Qn−1
m=1 h(xm,k |θ)
θ∈R+ ν(dθ)
(
)
n−1
n−1
X
X
= κ(x) exp −B(ξ +
xm , λ + n − 1) + B(ξ +
xm + x, λ + n) .
m=1
m=1
In Example C.1 in Appendix C we show that Corollary 6.2 can be used to
recover the Indian buffet process marginal from a beta process prior together
with a Bernoulli process likelihood. In the following example, we discover a new
marginal for the Poisson likelihood process with gamma process prior.
Example 6.3. Let Θ be a gamma process, and let Xn be iid Poisson likelihood
processes conditioned on Θ for each n as in Example 4.3. That is, we have
ν(dθ) = γθξ e−λθ dθ
and
h(x|θk ) =
1 x −θk
θ e
x! k
with
γ > 0,
ξ ∈ (−2, −1],
λ > 0;
ξf ix,k > −1 and λf ix,k > 0
for all k ∈ [Kprior,f ix ]
by Example 4.3.
We can pick out the following components of h:
κ(x) =
And we calculate
Z
exp {B(ξ, λ)} =
1
,
x!
φ(x) = x,
η(θ) = log(θ),
exp {hξ, η(θ)i + λ[−A(θ)]} dθ =
A(θ) = θ.
Z
θ∈R+
θ∈R+
33
θξ e−λθ = Γ(ξ + 1)λ−(ξ+1) .
So, for k ∈ Z∗ , we have
(
P(xn = x) = κ(x) exp −B(ξ +
n−1
X
xm , λ + n − 1) + B(ξ +
m=1
Pn−1
n−1
X
)
xm + x, λ + n)
m=1
Pn−1
1 (λ + n − 1)ξ+ m=1 xm +1 Γ(ξ + m=1 xm + x + 1)
·
·
Pn−1
Pn−1
x!
Γ(ξ + m=1 xm + 1)
(λ + n)ξ+ m=1 xm +x+1
Pn−1
ξ+Pnm=1 xm +1
x
Γ(ξ + m=1 xm + x + 1)
1
λ+n−1
=
·
Pn−1
λ+n
λ+n
Γ(x + 1)Γ(ξ + m=1 xm + 1)
!
n−1
X
= NegBin x ξ +
xm + 1, (λ + n)−1 .
=
m=1
And
Mn,x := γ · κ(0)n−1 κ(x) · exp {B(ξ + (n − 1)φ(0) + φ(x), λ + n)}
1
=γ·
· Γ(ξ + x + 1)(λ + n)−(ξ+x+1) .
x!
We summarize the marginal distribution representation of X1:N that follows
from Corollary 6.2 in the following result.
Corollary 6.4. Let Θ be a gamma process with fixed-location atom weight distributions as in Eq. (19) and ordinary component weight measure as in Eq. (20).
Let Xn be drawn, iid across n, conditional on Θ according to a Poisson likelihood process with fixed-location atom weight distributions as in Eq. (18). Then
X1:N has the same distribution as the following construction.
For each n = 1, 2, . . . , N ,
K
n−1
be the union of atom locations in X1 , . . . , Xn−1 . Let xm,k :=
1. Let {ψk }k=1
Xm ({ψk }). Let xn,k denote the weight of Xn |X1 , . . . , Xn−1 at ψk . Then
xn,k has distribution described by the following probability mass function:
!
n−1
X
xm,k + 1, (λ + n)−1 .
hcond xn,k = x x1:(n−1),k = NegBin x ξ +
m=1
2. For each x = 1, 2, . . .
ρ
n,x
,
• Xn has ρn,x new atoms. That is, Xn has atoms at locations {ψn,x,j }j=1
where
ρn,x
Kn−1
= ∅ a.s.
∩ {ψk }k=1
{ψn,x,j }j=1
Moreover,
1 Γ(ξ + x + 1)
·
x! (λ + n)ξ+x+1
across n, x
Mn,x := γ ·
34
indep
ρn,x ∼ Poisson (ρ |Mn,x ) independently across n, x
iid
ψn,x,j ∼ G(dψ) independently across n, x and iid across j.
7
Discussion
In the preceding sections, we have shown how to calculate posteriors for general
CRM-based priors and likelihoods for Bayesian nonparametric models. We have
also shown how to represent Bayesian nonparametric priors as a sequence of finite draws, and full Bayesian nonparametric models via finite marginals. We
have introduced a notion of exponential families for CRMs, which we call exponential CRMs, that has allowed us to specify automatic Bayesian nonparametric
conjugate priors for exponential CRM likelihoods. And we have demonstrated
that our exponential CRMs allow particularly straightforward recipes for sizebiased and marginal representations of Bayesian nonparametric models. Along
the way, we have proved that the gamma process is a conjugate prior for the
Poisson likelihood process and the beta prime process is a conjugate prior for
the odds Bernoulli process. We have discovered a size-biased representation of
the gamma process and a marginal representation of the gamma process coupled
with a Poisson likelihood process.
All of this work has relied heavily on the description of Bayesian nonparametric models in terms of completely random measures. As such, we have worked
very particularly with pairings of real values—the CRM atom weights, which we
have interpreted as trait frequencies or rates—together with trait descriptors—
the CRM atom locations. However, all of our proofs broke into essentially two
parts: the fixed-location atom part and the ordinary component part. The fixedlocation atom development essentially translated into the usual finite version of
Bayes Theorem and could easily be extended to full Bayesian models where the
prior describes a random element that need not be real-valued. Moreover, the
ordinary component development relied entirely on its generation as a Poisson
point process over a product space. It seems reasonable to expect that our development might carry through when the first element in this tuple need not be
real-valued. And thus we believe our results are suggestive of broader results
over more general spaces.
Acknowledgements
Support for this project was provided by ONR under the Multidisciplinary University Research Initiative (MURI) program (N00014-11-1-0688). T. Broderick
was supported by a Berkeley Fellowship. A. C. Wilson was supported by an
NSF Graduate Research Fellowship.
35
A
Further automatic conjugate priors
We use Theorem 4.2 to calculate automatic conjugate priors for further exponential CRMs.
Example A.1. Let X be generated according to a Bernoulli process as in
Example 2.1. That is, X has an exponential CRM distribution with Klike,f ix
fixed-location atoms, where Klike,f ix < ∞ in accordance with Assumption A0:
Klike,f ix
X=
X
xlike,k δψlike,k .
k=1
The weight of the kth atom, xlike,k , has support on {0, 1} and has a Bernoulli
density with parameter θk ∈ (0, 1]:
h(x|θk ) = θkx (1 − θk )1−x
= exp {x log(θk /(1 − θk )) + log(1 − θk )} .
The final line is rewritten to emphasize the exponential family form of this
density, with
κ(x) = 1
φ(x) = x
η(θ) = log
θ
1−θ
A(θ) = − log(1 − θ).
Then, by Theorem 4.2, X has a Bayesian nonparametric conjugate prior for
Klike,f ix
Θ :=
X
θk δ ψ k .
k=1
This conjugate prior has two parts.
First, Θ has a set of Kprior,f ix fixed-location atoms at some subset of the
Klike,f ix fixed locations of X. The kth such atom has random weight θf ix,k
with density
fprior,f ix,k (θ) = exp {hξf ix,k , η(θ)i + λf ix,k [−A(θ)] − B(ξf ix,k , λf ix,k )}
= θξf ix,k (1 − θ)λf ix,k −ξf ix,k exp {−B(ξf ix,k , λf ix,k )}
= Beta (θ |ξf ix,k + 1, λf ix,k − ξf ix,k + 1 ) ,
where Beta(θ|a, b) denotes the beta density with shape parameters a > 0 and
b > 0. So we must have fixed hyperparameters ξf ix,k > −1 and λf ix,k >
ξf ix,k − 1. Further,
exp {−B(ξf ix,k , λf ix,k )} =
Γ(λf ix,k + 2)
Γ(ξf ix,k + 1)Γ(λf ix,k − ξf ix,k + 1)
36
to ensure normalization.
Second, Θ has an ordinary component characterized by any proper distribution G and weight rate measure
ν(dθ) = γ exp {hξ, η(θ)i + λ [−A(θ)]} dθ
= γθξ (1 − θ)λ−ξ dθ.
Finally, we need to choose the allowable hyperparameter ranges for γ, ξ, and
λ. γ > 0 ensures ν is a measure. By Assumption A1, we must have ν(R+ ) = ∞,
so ν must represent an improper beta distribution. As such, we require either
ξ + 1 ≤ 0 or λ − ξ ≤ 0. By Assumption A2, we must have
∞ Z
X
ν(dθ) · h(x|θ)
θ∈R+
x=1
=
Z
ν(dθ)h(1|θ)
θ∈(0,1]
since the support of x is {0, 1} and the support of θ is (0, 1]
Z
=γ
θξ (1 − θ)λ−ξ dθ · θ
θ∈(0,1]
<∞
Since the integrand is the kernel of a beta distribution, the integral is finite if
and only if ξ + 2 > 0 and λ − ξ + 1 > 0.
Finally, then the hyperparameter restrictions can be summarized as:
γ>0
ξ ∈ (−2, −1]
λ>ξ−1
ξf ix,k > −1 and λf ix,k > ξf ix,k − 1 for all k ∈ [Kprior,f ix ]
By setting α = ξ+1, c = λ+2, ρf ix,k = ξf ix,k +1, and σf ix,k = λf ix,k −ξf ix,k +1,
we recover the hyperparameters of Eq. (11) in Example 2.1. Here, by contrast to
Example 2.1, we found the conjugate prior and its hyperparameter settings given
just the Bernoulli process likelihood. Henceforth, we use the parameterization
of the beta process above.
B
Further size-biased representations
Example B.1. Let Θ be a beta process, and let Xn be iid Bernoulli processes
conditioned on Θ for each n as in Example A.1. That is, we have
ν(dθ) = γθξ (1 − θ)λ−ξ dθ.
And
h(x|θk ) = θkx (1 − θk )1−x
37
with
γ>0
ξ ∈ (−2, −1]
λ>ξ−1
ξf ix,k > −1 and λf ix,k > ξf ix,k − 1 for all k ∈ [Kprior,f ix ]
by Example A.1.
We can pick out the following components of h:
κ(x) = 1
φ(x) = x
η(θ) = log
θ
1−θ
A(θ) = − log(1 − θ).
Thus, by Corollary 5.2,
Θ=
m,x
∞ X
∞ ρX
X
θm,x,j δψm,x,j
m=1 x=1 j=1
iid
ψm,x,j ∼ G iid across m, x, j
indep
θm,x,j ∼ fsize,m,x (θ) dθ
∝ θξ+x (1 − θ)λ+m−ξ−x dθ
∝ Beta (θ |ξ + x, λ − ξ + m − x ) dθ
iid across j and independently across m, x
Mm,x := γ ·
Γ(ξ + x + 1)Γ(λ − ξ + m − x + 1)
Γ(λ + m + 2)
indep
ρm,x ∼ Poisson (Mm,x )
across m, x
Broderick et al. [2012] and Paisley et al. [2012] have previously noted that
this size-biased representation of the beta process arises from the Poisson point
process.
C
Further marginals
Example C.1. Let Θ be a beta process, and let Xn be iid Bernoulli processes
conditioned on Θ for each n as in Examples A.1 and B.1.
We calculate the main components of Corollary 6.2 for this pair of processes.
In particular, we have
(
)
n−1
n−1
X
X
P(xn = 1) = κ(k) exp −B(ξ +
xm , λ + n − 1) + B(ξ +
xm + 1, λ + n)
m=1
m=1
38
Γ(λ + n − 1 + 2)
Pn−1
x
+
1)Γ(λ + n − 1 − ξ − m=1 xm + 1)
m=1 m
Pn−1
Pn−1
Γ(ξ + m=1 xm + 1 + 1)Γ(λ + n − ξ − m=1 xm − 1 + 1)
·
Γ(λ + n + 2)
Pn−1
ξ + m=1 xm + 1
=
λ+n+1
=
Γ(ξ +
Pn−1
And
Mn,1 := γ · κ(0)n−1 κ(1) · exp {B(ξ + (n − 1)φ(0) + φ(1), λ + n)}
=γ·
Γ(ξ + 1 + 1)Γ(λ + n − ξ − 1 + 1)
Γ(λ + n + 2)
Thus, the marginal distribution of X1:N is the same as that provided by the
following construction.
For each n = 1, 2, . . . , N ,
1. At any location ψ for which there is some atom in X1 , . . . , Xn−1 , let xm be
the weight of Xm at ψ for m ∈ [n−1]. Then we have that Xn |X1 , . . . , Xn−1
has weight xn at ψ, where
!
Pn−1
ξ + m=1 xm + 1
P(dxn ) = Bern xn
λ+n+1
2. Xn has ρn,1 atoms at locations {ψn,1,j } with j ∈ [ρn,1 ] where there have
not yet been atoms in any of X1 , . . . , Xn−1 . Moreover,
Γ(ξ + 1 + 1)Γ(λ + n − ξ − 1 + 1)
Γ(λ + n + 2)
across n
Mn,1 := γ ·
indep
ρn,1 ∼ Poisson (Mn,1 ) across n, x
iid
ψn,1,j ∼ G(dψ) across n, j
Here, we have recovered the three-parameter extension of the Indian buffet
process [Teh and Görür, 2009, Broderick et al., 2013].
39
References
E. M. Airoldi, D. Blei, E. A. Erosheva, and S. E. Fienberg. Handbook of Mixed
Membership Models and Their Applications. CRC Press, 2014.
D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent Dirichlet allocation. Journal of
Machine Learning Research, 3:993–1022, 2003.
T. Broderick, M. I. Jordan, and J. Pitman. Beta processes, stick-breaking, and
power laws. Bayesian Analysis, 7(2):439–476, 2012.
T. Broderick, M. I. Jordan, and J. Pitman. Cluster and feature modeling from
combinatorial stochastic processes. Statistical Science, 2013.
T. Broderick, L. Mackey, J. Paisley, and M. I. Jordan. Combinatorial clustering
and the beta negative binomial process. IEEE TPAMI, 2015.
P. Damien, J. Wakefield, and S. Walker. Gibbs sampling for Bayesian nonconjugate and hierarchical models by using auxiliary variables. Journal of the
Royal Statistical Society: Series B, 61(2):331–344, 1999.
M. H. DeGroot. Optimal Statistical Decisions. John Wiley & Sons, Inc, 1970.
P. Diaconis and D. Ylvisaker. Conjugate priors for exponential families. The
Annals of Statistics, 7(2):269–281, 1979.
K. Doksum. Tailfree and neutral random probabilities and their posterior distributions. The Annals of Probability, pages 183–201, 1974.
F. Doshi, K. T. Miller, J. Van Gael, and Y. W. Teh. Variational inference for
the Indian buffet process. In AISTATS, 2009.
M. D. Escobar. Estimating normal means with a Dirichlet process prior. Journal
of the American Statistical Association, 89(425):268–277, 1994.
M. D. Escobar and M. West. Bayesian density estimation and inference using
mixtures. Journal of the American Statistical Association, 90(430):577–588,
1995.
M. D. Escobar and M. West. Computing nonparametric hierarchical models. In
Practical nonparametric and semiparametric Bayesian statistics, pages 1–22.
Springer, 1998.
T. S. Ferguson. A Bayesian analysis of some nonparametric problems. The
Annals of Statistics, pages 209–230, 1973.
T. S. Ferguson. Prior distributions on spaces of probability measures. The
Annals of Statistics, pages 615–629, 1974.
T. Griffiths and Z. Ghahramani. Infinite latent feature models and the Indian
buffet process. In NIPS, 2006.
40
N. L. Hjort. Nonparametric Bayes estimators based on beta processes in models
for life history data. The Annals of Statistics, pages 1259–1294, 1990.
H. Ishwaran and L. F. James. Gibbs sampling methods for stick-breaking priors.
Journal of the American Statistical Association, 96(453), 2001.
L. F. James. Poisson latent feature calculus for generalized Indian buffet processes. arXiv preprint arXiv:1411.2936, 2014.
L. F. James, A. Lijoi, and I. Prünster. Posterior analysis for normalized random
measures with independent increments. Scandinavian Journal of Statistics,
36(1):76–97, 2009.
M. Kalli, J. E. Griffin, and S. G. Walker. Slice sampling mixture models. Statistics and Computing, 21(1):93–105, 2011.
Y. Kim. Nonparametric Bayesian estimators for counting processes. Annals of
Statistics, pages 562–588, 1999.
J. F. C. Kingman. Completely random measures. Pacific Journal of Mathematics, 21(1):59–78, 1967.
J. F. C. Kingman. Poisson Processes, volume 3. Oxford University Press, 1992.
A. Lijoi and I. Prünster. Models beyond the Dirichlet process. In N. L. Hjort,
C. Holmes, P. Müller, and S. G. Walker, editors, Bayesian Nonparametrics.
Cambridge Series in Statistical and Probabilistic Mathematics, 2010.
A. Y. Lo. Bayesian nonparametric statistical inference for Poisson point processes. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete, 59
(1):55–66, 1982.
A. Y. Lo. On a class of Bayesian nonparametric estimates: I. Density estimates.
Annals of Statistics, 12(1):351–357, 1984.
S. N. MacEachern. Estimating normal means with a conjugate style Dirichlet
process prior. Communications in Statistics-Simulation and Computation, 23
(3):727–741, 1994.
R. M. Neal. Markov chain sampling methods for Dirichlet process mixture
models. Journal of Computational and Graphical Statistics, 9(2):249–265,
2000.
R. M. Neal. Slice sampling. Annals of Statistics, pages 705–741, 2003.
P. Orbanz. Conjugate projective limits. arXiv preprint arXiv:1012.0363, 2010.
J. W. Paisley, A. K. Zaas, C. W. Woods, G. S. Ginsburg, and L. Carin. A
stick-breaking construction of the beta process. In ICML, pages 847–854,
2010.
41
J. W. Paisley, L. Carin, and D. M. Blei. Variational inference for stick-breaking
beta process priors. In ICML, pages 889–896, 2011.
J. W. Paisley, D. M. Blei, and M. I. Jordan. Stick-breaking beta processes and
the Poisson process. In AISTATS, pages 850–858, 2012.
M. Perman, J. Pitman, and M. Yor. Size-biased sampling of poisson point
processes and excursions. Probability Theory and Related Fields, 92(1):21–39,
1992.
J. Pitman. Random discrete distributions invariant under size-biased permutation. Advances in Applied Probability, pages 525–539, 1996a.
J. Pitman. Some developments of the Blackwell-MacQueen urn scheme. Lecture
Notes-Monograph Series, pages 245–267, 1996b.
J. Pitman. Poisson-Kingman partitions. Lecture Notes-Monograph Series, pages
1–34, 2003.
J. Sethuraman. A constructive definition of Dirichlet priors. Statistica Sinica,
4:639–650, 1994.
Y. W. Teh and D. Görür. Indian buffet processes with power-law behavior. In
NIPS, pages 1838–1846, 2009.
Y. W. Teh, M. I. Jordan, M. J. Beal, and D. M. Blei. Hierarchical Dirichlet
processes. Journal of the American Statistical Association, 101(476), 2006.
Y. W. Teh, D. Görür, and Z. Ghahramani. Stick-breaking construction for the
Indian buffet process. In AISTATS, pages 556–563, 2007.
R. Thibaux and M. I. Jordan. Hierarchical beta processes and the Indian buffet
process. In AISTATS, pages 564–571, 2007.
R. J. Thibaux. Nonparametric Bayesian Models for Machine Learning. PhD
thesis, UC Berkeley, 2008.
M. K. Titsias. The infinite gamma-Poisson feature model. In NIPS, pages
1513–1520, 2008.
S. G. Walker. Sampling the Dirichlet mixture model with slices. Communications in Statistics—Simulation and Computation, 36(1):45–54, 2007.
C. Wang and D. M. Blei. Variational inference in nonconjugate models. The
Journal of Machine Learning Research, 14(1):1005–1031, 2013.
M. West and M. D. Escobar. Hierarchical priors and mixture models, with application in regression and density estimation. In P. R. Freeman and A. F. M.
Smith, editors, Aspects of Uncertainty: A Tribute to D. V. Lindley. Institute
of Statistics and Decision Sciences, Duke University, 1994.
M. Zhou, L. Hannah, D. Dunson, and L. Carin. Beta-negative binomial process
and Poisson factor analysis. AISTATS, 2012.
42
| 10math.ST
|
AVA: A Video Dataset of Spatio-temporally Localized Atomic Visual Actions
Chunhui Gu∗
Yeqing Li∗
Chen Sun∗
David A. Ross∗
Sudheendra Vijayanarasimhan∗
arXiv:1705.08421v3 [cs.CV] 29 Nov 2017
Rahul Sukthankar∗
Carl Vondrick∗
George Toderici∗
Cordelia Schmid† ∗
Caroline Pantofaru∗
Susanna Ricco∗
Jitendra Malik‡ ∗
Abstract
This paper introduces a video dataset of spatiotemporally localized Atomic Visual Actions (AVA). The AVA
dataset densely annotates 80 atomic visual actions in 192
15-minute video clips, where actions are localized in space
and time, resulting in 740k action labels with multiple labels per person occurring frequently. The key characteristics of our dataset are: (1) the definition of atomic visual actions, rather than composite actions; (2) precise
spatio-temporal annotations with possibly multiple annotations for each person; (3) exhaustive annotation of these
atomic actions over 15-minute video clips; (4) people temporally linked across consecutive segments; and (5) using
movies to gather a varied set of action representations. This
departs from existing datasets for spatio-temporal action
recognition, which typically provide sparse annotations for
composite actions in short video clips. We will release the
dataset publicly.
AVA, with its realistic scene and action complexity, exposes the intrinsic difficulty of action recognition. To benchmark this, we present a novel approach for action localization that builds upon the current state-of-the-art methods, and demonstrates better performance on JHMDB and
UCF101-24 categories. While setting a new state of the art
on existing datasets, the overall results on AVA are low at
16.2% mAP, underscoring the need for developing new approaches for video understanding.
Figure 1. The bounding box and action annotations in sample
frames of the AVA dataset. Each bounding box is associated with
1 pose action (in orange), 0–3 interactions with objects (in red),
and 0–3 interactions with other people (in blue). Note that some
of these actions require temporal context to accurately label.
the actor’s pose (orange text) — standing, sitting, walking,
swimming etc. — and there may be additional actions corresponding to interactions with objects (red text) or interactions with other persons (blue text). Each person in a
frame containing multiple actors is labeled separately.
To label the actions performed by a person, a key choice
is the annotation vocabulary, which in turn is determined
by the “temporal grain” at which actions are classified. We
use short segments (±1.5 seconds centered on a keyframe)
to provide temporal context for labeling the actions in the
middle frame. This enables the annotator to use movement cues for disambiguating actions such as pick up or put
down that cannot be resolved in a static frame. We keep the
temporal context relatively brief because we are interested
in (temporally) fine-scale annotation of physical actions,
which motivates “Atomic Visual Actions” (AVA). The vocabulary consists of 80 different atomic visual actions. Our
dataset is sourced from the 15th to 30th minute time intervals of 192 different movies, which given the 1 Hz sampling
frequency gives us 900 keyframes for each movie. In each
keyframe, every person is labeled with (possibly multiple)
1. Introduction
We introduce a new annotated video dataset, AVA, to advance action recognition research (see Fig. 1). The annotation is person-centric at a sampling frequency of 1 Hz.
Every person is localized using a bounding box and the attached labels correspond to (possibly multiple) actions being performed by the actor: one action corresponding to
∗ Google
Inc., USA
Laboratoire Jean Kuntzmann, Grenoble, France
‡ University of California at Berkeley, USA
† Inria,
1
Figure 2. This figure illustrates the hierarchical nature of an activity. From Barker and Wright [3], pg. 247.
actions from the AVA vocabulary. Each person is linked
to the consecutive keyframes to provide short temporal sequences of action labels (Section 4.3). We now motivate the
main design choices of AVA.
Atomic action categories. Barker & Wright [3] noted the
hierarchical nature of activity (Fig. 2) in their classic study
of the ”behavior episodes” in the daily lives of the residents
of a small town in Kansas. At the finest level, the actions
consist of atomic body movements or object manipulation
but at coarser levels, the most natural descriptions are in
terms of intentionality and goal-directed behavior.
This hierarchy makes defining a vocabulary of action labels ill posed, contributing to the slower progress of our
field compared to object recognition; exhaustively listing
high-level behavioral episodes is impractical. However if
we limit ourselves to fine time scales, then the actions
are very physical in nature and have clear visual signatures. Here, we annotate keyframes at 1 Hz as this is sufficiently dense to capture the complete semantic content
of actions while enabling us to avoid requiring unrealistically precise temporal annotation of action boundaries. The
THUMOS challenge [18] observed that action boundaries
(unlike objects) are inherently fuzzy, leading to significant
inter-annotator disagreement. By contrast, annotators can
easily determine (using ±1.5s of context) whether a frame
contains a given action. Effectively, AVA localizes action
start and end points to an acceptable precision of ±0.5 s.
Person-centric action time series. While events such as
trees falling do not involve people, our focus is on the activities of people, treated as single agents. There could be
multiple people as in sports or two people hugging, but each
one is an agent with individual choices, so we treat each separately. The action labels assigned to a person over time is
a rich source of data for temporal modeling (Section 4.3).
Annotation of movies. Ideally we would want behavior “in
the wild”. We do not have that, but movies are a compelling
approximation, particularly when we consider the diversity
of genres and countries with flourishing film industries. We
do expect some bias in this process. Stories have to be interesting and there is a grammar of the film language [2]
that communicates through the juxtaposition of shots. That
said, in each shot we can expect an unfolding sequence of
human actions, somewhat representative of reality, as conveyed by competent actors. AVA complements the current
datasets sourced from user generated video because we expect movies to contain a greater range of activities as befits
the telling of diverse stories.
Exhaustive action labeling. We label all the actions of
all the people in all the keyframes. This will naturally result in a Zipf’s law type of imbalance across action categories. There will be many more examples of typical
actions (standing or sitting) than memorable ones (dancing), but this is how it should be! Recognition models
need to operate on realistic “long tailed” action distributions [15] rather than being scaffolded using artificially balanced datasets. Another consequence of our protocol is that
since we do not retrieve examples of action categories by
explicit querying of internet video resources, we avoid a
certain kind of bias: opening a door is a common event that
occurs frequently in movie clips; however a door opening
action that has been tagged as such on YouTube is likely
attention worthy in a way that makes it atypical.
We believe that AVA, with its realistic complexity, exposes the inherent difficulty of action recognition hidden by
many popular datasets in the field. A video clip of a single person performing a visually salient action like swimming in typical background is easy to discriminate from,
say, one of a person running. Compare with AVA where
we encounter multiple actors, small in image size, performing actions which are only subtly different such as touching vs. holding an object. To verify this intuition, we do
comparative bench-marking on JHMDB [20], UCF101-24
categories [31] and AVA. The approach we use for spatiotemporal action localization (see Section 5) builds upon
multi-frame approaches [16, 40], but classifies tubelets with
I3D convolutions [6]. We obtain state-of-the-art performance on JHMDB [20] and UCF101-24 categories [31]
(see Section 6) while the mAP on AVA is only 16.2%.
We have released a preliminary version of AVA (v1.0,
with annotation rate of 1/3 Hz and no person links) at
https://research.google.com/ava/. The full
version (v2.0), as described in this paper, will be made
available online soon.
2. Related work
Action recognition datasets. Most popular action classification datasets, such as KTH [34], Weizmann [4],
Hollywood-2 [26], HMDB [24], UCF101 [38] consist of
short clips, manually trimmed to capture a single action. These datasets are ideally suited for training fullysupervised, whole-clip, forced-choice video classifiers. Recently, datasets, such as TrecVid MED [28], Sports1M [21], YouTube-8M [1], Something-something [12] and
Kinetics [22] have focused on large-scale video classification, often with automatically generated – and hence poten-
tially noisy – annotations. They serve a valuable purpose
but address a different need than AVA.
Some recent work has moved towards temporal localization. ActivityNet [5], THUMOS [18], MultiTHUMOS [45]
and Charades [36] use large numbers of untrimmed videos,
each containing multiple actions, obtained either from
YouTube (ActivityNet, THUMOS, MultiTHUMOS) or
from crowdsourced actors (Charades). The datasets provide temporal (but not spatial) localization for each action
of interest. AVA differs from them, as we provide spatiotemporal annotations for each subject performing an action
and annotations are dense over 15-minute clips.
A few datasets, such as CMU [23], MSR Actions [46],
UCF Sports [31] and JHMDB [20] provide spatio-temporal
annotations in each frame for short videos. The main differences with our AVA dataset are: the small number of actions; the small number of video clips; and the fact that
clips are very short. Furthermore, actions are composite (e.g., pole-vaulting) and not atomic as in AVA. Recent
extensions, such as UCF101 [38], DALY [43] and Hollywood2Tubes [27] evaluate spatio-temporal localization
in untrimmed videos, which makes the task significantly
harder and results in a performance drop. However, the
action vocabulary is still restricted to a limited number of
composite actions. Moreover, they do not densely cover
the actions; a good example is BasketballDunk in UCF101,
where only the dunking player is annotated. However, realworld applications often require a continuous annotations of
atomic actions of all humans, which can then be composed
into higher-level events. This motivates AVA’s exhaustive
labeling over 15-minute clips.
AVA is also related to still image action recognition
datasets [7, 9, 13] that are limited in two ways. First, the
lack of motion can make action disambiguation difficult.
Second, modeling composite events as a sequence of atomic
actions is not possible in still images. This is arguably out
of scope here, but clearly required in many real-world applications, for which AVA does provide training data.
Methods for spatio-temporal action localization. Most
recent approaches [11, 29, 33, 42] rely on object detectors
trained to discriminate action classes at the frame level with
a two-stream variant, processing RGB and flow data separately. The resulting per-frame detections are then linked
using dynamic programming [11, 37] or tracking [42].
All these approaches rely on integrating frame-level detections. Very recently, multi-frame approaches have emerged:
Tubelets [40] jointly estimate localization and classification
over several frames, T-CNN [16] use 3D convolutions to
estimate short tubes, micro-tubes rely on two successive
frames [32] and pose-guided 3D convolutions add pose to
a two-stream approach [47]. We build upon the idea of
spatio-temporal tubes, but employ state-of-the-art I3D convolution [6] and Faster R-CNN [30] region proposals to out-
Figure 3. User interface for action annotation. Details in Sec 3.5.
perform the state of the art.
3. Data collection
Annotation of the AVA dataset consists of five stages:
action vocabulary generation, movie and segment selection,
person bounding box annotation, person linking and action
annotation.
3.1. Action vocabulary generation
We follow three principles to generate our action vocabulary. The first one is generality. We collect generic actions
in daily-life scenes, as opposed to specific activities in specific environments (e.g., playing basketball on a basketball
court). The second one is atomicity. Our action classes have
clear visual signatures, and are typically independent of interacted objects (e.g., hold without specifying what object
to hold). This keeps our list short yet complete. The last
one is exhaustivity. We initialized our list using knowledge
from previous datasets, and iterated the list in several rounds
until it covered ∼99% of actions in the AVA dataset labeled
by annotators. We end up with 14 pose classes, 49 personobject interaction classes and 17 person-person interaction
classes in the vocabulary.
3.2. Movie and segment selection
The raw video content of the AVA dataset comes from
YouTube. We begin by assembling a list of top actors of
many different nationalities. For each name we issue a
YouTube search query, retrieving up to 2000 results. We
only include videos with the “film” or “television” topic
annotation, a duration of over 30 minutes, at least 1 year
since upload, and at least 1000 views. We further exclude
black & white, low resolution, animated, cartoon, and gaming videos, as well as those containing mature content.
To create a representative dataset within constraints, our
selection criteria avoids filtering by action keywords, using
automated action classifiers, or forcing a uniform label distribution. We aim to create an international collection of
films by sampling from large film industries. However, the
depiction of action in film is biased, e.g. by gender [10], and
does not reflect the “true” distribution of human activity.
clink glass → drink
open → close
turn → open
look at phone → answer phone
grab (a person) → hug
fall down → lie/sleep
Figure 4. We show examples of how atomic actions change over time in AVA. The text shows pairs of atomic actions for the people in red
bounding boxes. Temporal information is key for recognizing many of the actions and appearance can substantially vary within an action
category, such as opening a door or bottle.
Each movie contributes equally to the dataset, as we only
label a sub-part ranging from the 15th to the 30th minute.
We skip the beginning of the movie to avoid annotating
titles or trailers. We choose a duration of 15 minutes so
we are able to include more movies under a fixed annotation budget, and thus increase the diversity of our dataset.
Each 15-min clip is then partitioned into 900 overlapping 3s
movie segments with a stride of 1 second.
3.3. Person bounding box annotation
We localize a person and his or her actions with a bounding box. When multiple subjects are present in a keyframe,
each subject is shown to the annotator separately for action
annotation, and thus their action labels can be different.
Since bounding box annotation is manually intensive,
we choose a hybrid approach. First, we generate an initial set of bounding boxes using the Faster-RCNN person
detector [30]. We set the operating point to ensure highprecision. Annotators then annotate the remaining bounding boxes missed by our detector. This hybrid approach ensures full bounding box recall which is essential for benchmarking, while minimizing the cost of manual annotation.
This manual annotation retrieves only 5% more bounding
boxes missed by our person detector, validating our design
choice. Any incorrect bounding boxes are marked and removed by annotators in the next stage of action annotation.
3.4. Person link annotation
We link the bounding boxes over short periods of time
to obtain ground-truth person tracklets. We calculate the
pairwise similarity between bounding boxes in adjacent
key frames using a person embedding [44] and solve for
the optimal matching with the Hungarian algorithm [25].
While automatic matching is generally strong, we further
remove false positives with human annotators who verify
each match. This procedure results in 81,000 tracklets ranging from a few seconds to a few minutes.
3.5. Action annotation
The action labels are generated by crowd-sourced annotators using the interface shown in Figure 3. The left panel
shows both the middle frame of the target segment (top)
and the segment as a looping embedded video (bottom).
The bounding box overlaid on the middle frame specifies
the person whose action needs to be labeled. On the right
are text boxes for entering up to 7 action labels, including 1 pose action (required), 3 person-object interactions
(optional), and 3 person-person interactions (optional). If
none of the listed actions is descriptive, annotators can flag
a check box called “other action”. In addition, they could
flag segments containing blocked or inappropriate content,
or incorrect bounding boxes.
In practice, we observe that it is inevitable for annotators to miss correct actions when they are instructed to find
all correct ones from a large vocabulary of 80 classes. Inspired by [35], we split the action annotation pipeline into
two stages: action proposal and verification. We first ask
multiple annotators to propose action candidates for each
question, so the joint set possesses a higher recall than individual proposals. Annotators then verify these proposed
candidates in the second stage. Results show significant recall improvement using this two-stage approach, especially
on actions with fewer examples. See detailed analysis in
the supplemental material. On average, annotators take 22
seconds to annotate a given video segment at the propose
stage, and 19.7 seconds at the verify stage.
Each video clip is annotated by three independent annotators and we only regard an action label as ground truth if it
is verified by at least two annotators. Annotators are shown
segments in randomized order.
3.6. Training and test sets
Our training/test sets are split at the video level, so that
all segments of one video appear only in one split. The
192 videos are split into 154 training and 38 test videos,
resulting in 138k training segments and 34k test segments,
roughly a 80:20 split.
Figure 5. Sizes of each action class in the AVA dataset sorted by descending order, with colors indicating action types. A full list of counts
are in supplemental material.
4. Characteristics of the AVA dataset
We first build intuition on the diversity and difficulty of
our AVA dataset through visual examples. Then, we characterize the annotations of our dataset quantitatively. Finally,
we explore action and temporal structure.
4.1. Diversity and difficulty
Figure 4 shows examples of atomic actions as they
change over consecutive segments. Besides variations in
bounding box size and cinematography, many of the categories will require discriminating fine-grained differences,
such as “clinking glass” versus “drinking” or leveraging
temporal context, such as “opening” versus “closing”.
Figure 4 also shows two examples for the action “open”.
Even within an action class the appearance varies with
vastly different contexts: the object being opened may even
change. The wide intra-class variety will allow us to learn
features that identify the critical spatio-temporal parts of an
action — such as the breaking of a seal for “opening”. Additional examples are in the supplemental material.
4.2. Annotation Statistics
Figure 5 shows the distribution of action annotations in
AVA. The distribution roughly follows Zipf’s law. Figure 6
illustrates bounding box size distribution. A large portion
of people take up the full height of the frame. However,
there are still many boxes with smaller sizes. The variability can be explained by both zoom level as well as pose.
For example, boxes with the label “enter” show the typical
pedestrian aspect ratio of 1:2 with average widths of 30%
of the image width, and an average heights of 72%. On the
other hand, boxes labeled “lie/sleep” are close to square,
with average widths of 58% and heights of 67%. The box
widths are widely distributed, showing the variety of poses
people undertake to execute the labeled actions.
There are multiple labels for the majority of person
bounding boxes. All bounding boxes have one pose label,
28% of bounding boxes have at least 1 person-object interaction label, and 67% of them have at least 1 person-person
Figure 6. Size and aspect ratio variations of annotated bounding
boxes in the AVA dataset. Note that our bounding boxes consist
of a large variation of sizes, many of which are small and hard to
detect. Large variation also applies to the aspect ratios of bounding
boxes, with mode at 2:1 ratio (e.g., sitting pose).
interaction label.
4.3. Temporal Structure
A key characteristic of AVA is the rich temporal structure that evolves from segment to segment. Since we have
linked people between segments, we can discover common
consecutive actions by looking at pairs of actions performed
by the same person. We sort pairs by Normalized Pointwise
Mutual Information (NPMI) [8], which is commonly used
in linguistics to represent
between two
the co-occurrence
p(x,y)
words: NPMI(x, y) = ln p(x)p(y)
/ (− ln p(x, y)). Values intuitively fall in the range (−1, 1], with −1 for pairs
of words that never co-occur, 0 for independent pairs, and 1
for pairs that always co-occur.
Table 1 shows pairs of actions with top NPMI in consecutive one-second segments for the same person. After removing identity transitions, some interesting common
sense temporal patterns arise. Frequently, there are transitions from “click glass” → “drink”, “fall down” → “crawl”,
or “answer phone” → “put down”. We also analyze interperson action pairs. Table 2 shows top pairs of actions performed at the same time, but by different people. Several
meaningful pairs emerge, such as “take” ↔ “give”, “play
music” ↔ “hand clap”, or “ride” ↔ “drive”. The transitions between atomic actions, despite the relatively coarse
temporal sampling, provide excellent data for building more
complex models of actions and activities with longer tem-
First Action
Second Action
NPMI
paint
hit (object)
0.54
open (window/car door) close (door/box)
0.48
drink
0.44
clink glass
fall down
crawl
0.41
text/look at cellphone
answer phone
0.40
answer phone
text on/look cellphone
0.40
row boat
0.34
ride (eg bike/car/horse)
ride (eg bike/car/horse)
sail boat
0.34
put down
0.34
answer phone
turn (eg screwdriver)
open (window/door)
0.33
Table 1. We show top pairs of consecutive actions that are likely
to happen before/after for the same person. We sort by NPMI.
Person 1 Action
take from (person)
play musical instrument
ride (eg bike/car/horse)
talk to (self/person)
stand
sing to (self/person/group)
play musical instrument
lie/sleep
hug (person)
hand clap
Table 2. We show top pairs of
people. We sort by NPMI.
poral structure.
work. The output feature map at Mixed 4e has a stride of 16,
which is equivalent to the conv4 block of ResNet [14]. Second, for action proposal generation, we use a 2D ResNet-50
model on the keyframe as the input for the region proposal
network, avoiding the impact of I3D with different input
lengths on the quality of generated action proposals. Finally, we extend ROI Pooling to 3D by applying the 2D ROI
Pooling at the same spatial location over all time steps. To
understand the impact of optical flow for action detection,
we fuse the RGB stream and the optical flow stream at the
feature map level using average pooling.
Baseline. To compare to a frame-based two-stream approach on AVA, we have implemented a variant of [29]. We
use Faster RCNN [30] with ResNet-50 [14] to jointly learn
action proposals and action labels. Region proposals are
obtained with the RGB stream only. The region classifier
takes as input RGB along with optical flow features stacked
over 5 consecutive frames. As for our I3D approach, we
jointly train the RGB and the optical flow streams by fusing
the conv4 feature maps with average pooling.
Implementation details.
We implemented FlowNet
v2 [19] to extract optical flow features. We train FasterRCNN with asynchronous SGD. For all training tasks, we
use a validation set to determine the number of training
steps, which ranges from 600K to 1M iterations. We fix
the input resolution to be 320 by 400 pixels. All the
other model parameters are set based on the recommended
values from [17], which were tuned for object detection.
5. Action Localization Model
Performance numbers on popular action recognition
datasets such as UCF101 or JHMDB have gone up considerably in recent years, but we believe that this may present
an artificially rosy picture of the state of the art. When the
video clip involves only a single person performing something visually characteristic like swimming in an equally
characteristic background scene, it is easy to classify accurately. Difficulties come in when actors are multiple, or
small in image size, or performing actions which are only
subtly different, and when the background scenes are not
enough to tell us what is going on. AVA has these aspects
galore, and we will find that performance at AVA is much
poorer as a result. Indeed this finding was foreshadowed by
the poor performance at the Charades dataset [36].
To prove our point, we develop a state of the art action localization approach, which is inspired by recent approaches for spatio-temporal action localization that operate on multi-frame temporal information [16, 40]. Here,
we rely on the impact of larger temporal context based on
I3D [6] for action detection. See Fig. 7 for an overview of
our approach.
Following Peng and Schmid [29], we apply the Faster
RCNN algorithm [30] for end-to-end localization and classification of actions. However, in their approach, the temporal information is lost at the first layer where input channels from multiple frames are concatenated over time. We
propose to use the Inception 3D (I3D) architecture by Carreira and Zisserman [6] to model temporal context. The
I3D architecture is designed based on the Inception architecture [39], but replaces 2D convolutions with 3D convolutions. Temporal information is kept throughout the network. I3D achieves state-of-the-art performance on a wide
range of video classification benchmarks.
To use I3D with Faster RCNN, we make the following changes to the model: first, we feed input frames of
length T to the I3D model, and extract 3D feature maps of
size T 0 × W 0 × H 0 × C at the Mixed 4e layer of the net-
TxHxWx3
RGB frames
Person 2 Action
give/serve to (person)
listen (music)
drive (car/truck)
listen to (person)
sit
listen (music)
hand clap
crouch/kneel
grab (person)
dance
simultaneous actions by
ROI
Pooling
RGB
I3D
Region Proposal
Network
RGB ResNet-50
conv4
TxHxWx2
Flow frames
Flow
I3D
ROI
Pooling
Mixed 4e
H’ x W’ x C
T’ x H’ x W’ x C
Mixed 4e
Key
frame
Avg
Pooling
NPMI
0.46
0.43
0.40
0.37
0.27
0.24
0.23
0.21
0.20
0.19
different
+
Classification
Box
Refinement
Avg
Pooling
T’ x H’ x W’ x C
H’ x W’ x C
Figure 7. Illustration of our approach for spatio-temporal action
localization. Region proposals are detected and regressed with
Faster-RCNN on RGB keyframes. Spatio-temporal tubes are classified with two-stream I3D convolutions.
The ResNet-50 networks are initialized with ImageNet pretrained models. For the optical flow stream, we duplicate
the conv1 filters to input 5 frames. The I3D networks are
initialized with Kinetics [22] pre-trained models, for both
the RGB and optical flow streams. Note that although I3D
were pre-trained on 64-frame inputs, the network is fully
convolutional over time and can take any number of frames
as input. All feature layers are jointly updated during training. The output frame-level detections are post-processed
with non-maximum suppression with threshold 0.6.
One key difference between AVA and existing action detection datasets is that the action labels of AVA are not mutually exclusive. To address this, we replace the standard
softmax loss function by a sum of binary Sigmoid losses,
one for each class. We use Sigmoid loss for AVA and softmax loss for all other datasets.
Linking. Once we have per frame-level detections, we
link them to construct action tubes. We report video-level
performance based on average scores over the obtained
tubes. We use the same linking algorithm as described in
[37], except that we do not apply temporal labeling. We
choose this approach because it is robust to missing detections. It is also used to determine temporal detections, as the
initialization and termination steps result in tubes with different temporal extents. Since AVA is annotated at 1 Hz and
each tube may have multiple labels, we modify the videolevel evaluation protocol to estimate an upper bound. We
use ground truth links to infer detection links, and when
computing IoU score of a class between a ground truth tube
and a detection tube, we only take tube segments that are
labeled by that class into account.
6. Experiments and Analysis
We now experimentally analyze key characteristics of
AVA and motivate challenges for action understanding.
6.1. Datasets and Metrics
AVA benchmark. Since the label distribution in AVA
roughly follows Zipf’s law (Figure 5) and evaluation on a
very small number of examples could be unreliable, we use
classes that have at least 25 test instances to benchmark performance. Our resulting benchmark consists of 63 action
classes. We randomly select 10% of the training data as validation set and use them to set the model parameters. Our
benchmark consists of a total of 122,800 training, 13,696
validation and 33,226 test examples.
Datasets. Besides AVA, we also analyze standard video
datasets in order to compare difficulty. JHMDB [20] consists of 928 trimmed clips over 21 classes. We report results
for split one in our ablation study, but results are averaged
over three splits for comparison to the state of the art. For
UCF101, we use spatio-temporal annotations for a 24-class
Frame-mAP
Actionness [41]
Peng w/o MR [29]
Peng w/ MR [29]
ACT [40]
Our approach
JHMDB
39.9%
56.9%
58.5%
65.7%
73.3%
UCF101-24
64.8%
65.7%
69.5%
76.3%
Video-mAP
Peng w/ MR [29]
Singh et al. [37]
ACT [40]
TCNN [16]
Our approach
JHMDB
73.1%
72.0%
73.7%
76.9%
78.6%
UCF101-24
35.9%
46.3%
51.4%
59.9%
Table 3. Frame-mAP (top) and video-mAP (bottom) @ IoU 0.5
for JHMDB and UCF101-24. For JHMDB, we report averaged
performance over three splits. Our approach outperforms previous
state-of-the-art on both metrics by a considerable margin.
subset with 3207 videos, provided by Singh et al. [37]. We
conduct experiments on the official split1 as is standard.
Metrics. For evaluation, we follow standard practice when
possible. We report intersection-over-union (IoU) performance on frame level and video level. For frame-level IoU,
we follow the standard protocol used by the PASCAL VOC
challenge [9] and report the average precision (AP) using
an IoU threshold of 0.5. For each class, we compute the average precision and report the average over all classes. For
video-level IoU, we compute 3D IoUs between ground truth
tubes and linked detection tubes at the threshold of 0.5. The
mean AP is computed by averaging over all classes.
6.2. Comparison to the state-of-the-art
Table 3 shows our model performance on two standard
video datasets. Our 3D two-stream model obtains stateof-the-art performance on UCF101 and JHMDB, outperforming well-established baselines for both frame-mAP and
video-mAP metrics.
However, the picture is less auspicious when recognizing
atomic actions. Table 4 shows that the same model obtains
relatively low performance on AVA (frame-mAP of 16.2%,
video-mAP of 10.3% at 0.5 IoU and 16.0% at 0.2 IoU). We
attribute this to the design principles behind AVA: we collected a vocabulary where context and object cues are not
as discriminative for action recognition. Instead, recognizing fine-grained details and rich temporal models may be
needed to succeed at AVA, posing a new challenge for visual action recognition. In the remainder of this paper, we
analyze what makes AVA challenging and discuss how to
move forward.
6.3. Ablation study
How important is temporal information for recognizing
AVA categories? Table 4 shows the impact of the temporal
length and the type of model. All 3D models outperform the
Figure 8. Top: We plot the performance of models for each action category, sorting by the number of training examples. Bottom: We plot
the number of training examples per category. While more data is better, the outliers suggest that not all categories are of equal complexity.
For example, one of the smallest categories “swim” has one of the highest performances because the associated scenes make it relatively
easy. Instead, the challenging categories are ones with large diversity, such as “touch,” where context is not as discriminative.
Model
2D
3D
3D
3D
3D
3D
3D
3D
Temp.+ Mode
1 RGB + 5 Flow
5 RGB + 5 Flow
10 RGB + 10 Flow
20 RGB + 20 Flow
40 RGB + 40 Flow
50 RGB + 50 Flow
20 RGB
20 Flow
JHMDB
52.1%
67.9%
73.4%
76.4%
76.7%
73.2%
67.0%
UCF101-24
60.1%
76.1%
78.0%
78.3%
76.0%
73.2%
77.0%
71.3%
AVA
12.8%
13.4%
13.9%
14.9%
16.2%
15.8%
14.1%
10.9%
Table 4. Frame-mAP @ IoU 0.5 for action detection on JHMDB
(split1), UCF101 (split1) and AVA. Note that JHMDB has up to
40 frames per clip. For UCF101-24, we randomly sample 20,000
frame subset for evaluation. Although our model obtains state-ofthe-art performance on JHMDB and UCF101-24, the fine-grained
nature of AVA makes it a challenge.
2D baseline. We can also see that increasing the length of
the temporal window helps for the 3D two-stream models
across all datasets. As expected, combining RGB and optical flow features improves the performance over a single
input modality. Moreover, AVA benefits more from larger
temporal context than JHMDB and UCF101, whose performances saturate at 20 frames. This gain and the consecutive
actions in Table 1 suggests that one may obtain further gains
by leveraging the rich temporal context in AVA.
How challenging is localization versus recognition? Table 5 compares the performance of end-to-end action localization and recognition versus class agnostic action localization. We can see that although action localization is more
challenging on AVA than on JHMDB, the gap between localization and end-to-end detection performance is nearly
60% on AVA, while less than 15% on JHMDB and UCF101.
This suggests that the main difficulty of AVA lies in action
classification rather than localization. Figure 9 shows examples of high-scoring false alarms, suggesting that the difficulty in recognition lies in the fine-grained details.
Which categories are challenging? How important is
number of training examples? Figure 8 breaks down
performance by categories and the number of training ex-
Action detection
Actor detection
JHMDB
76.7%
92.8%
UCF101-24
78.3%
84.8%
AVA
16.2%
75.8%
Table 5. Frame-mAP @ IoU 0.5 for action detection and actor detection performance on JHMDB (split1), UCF101-24 (split1) and
AVA benchmarks. Since human annotators are consistent, our results suggest there is significant headroom to improve on recongizing atomic visual actions.
Figure 9. Red boxes show high-scoring false alarms for smoking.
The model often struggles to discriminate fine-grained details.
amples. While more data generally yields better performance, the outliers reveals that not all categories are of
equal complexity. Categories correlated with scenes and
objects (such as swimming) or categories with low diversity
(such as jumping) obtain high performance despite having
fewer training examples. In contrast, categories with lots
of data, such as touching, obtain low performance possibly
because they have large visual variations and require fine
grained discrimination, motivating work on person-object
interaction [7, 12]. We hypothesize that the gains on recognizing atomic actions will need not only large datasets, such
as AVA, but also rich models of motion and interactions.
7. Conclusion
This paper introduces the AVA dataset with spatiotemporal annotations of atomic actions at 1 Hz over diverse
15-min. movie segments. In addition we propose a method
that outperforms the current state of the art on standard
benchmarks to serve as a baseline. This method highlights
the difficulty of the AVA dataset as its performance is significantly lower than on UCF101 or JHMDB, underscoring
the need for developing new action recognition approaches.
Future work includes modeling more complex activities
based on our atomic actions. Our present day visual classification technology may enable us to classify events such
as “eating in a restaurant” at the coarse scene/video level,
but models based on AVA’s fine spatio-temporal granularity
facilitate understanding at the level of an individual agents
actions. These are essential steps towards imbuing computers with “social visual intelligence” – understanding what
humans are doing, what they might do next, and what they
are trying to achieve.
8. Acknowledgement
We thank Ahbinav Gupta, Ahbinav Shrivastava, Andrew
Gallagher, Irfan Essa, and Vicky Kalogeiton for discussion
and comments about this work.
References
[1] S. Abu-El-Haija, N. Kothari, J. Lee, P. Natsev, G. Toderici,
B. Varadarajan, and S. Vijayanarasimhan.
YouTube8M: A large-scale video classification benchmark.
arXiv:1609.08675, 2016. 2
[2] D. Arijon. Grammar of the film language. Silman-James
Press, 1991. 2
[3] R. Barker and H. Wright. Midwest and its children: The
psychological ecology of an American town. Row, Peterson
and Company, 1954. 2
[4] M. Blank, L. Gorelick, E. Shechtman, M. Irani, and R. Basri.
Actions as space-time shapes. In ICCV, 2005. 2
[5] F. Caba Heilbron, V. Escorcia, B. Ghanem, and J. C. Niebles.
ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 3
[6] J. Carreira and A. Zisserman. Quo vadis, action recognition?
A new model and the Kinetics dataset. In CVPR, 2017. 2, 3,
6
[7] Y.-W. Chao, Z. Wang, Y. He, J. Wang, and J. Deng. HICO:
A benchmark for recognizing human-object interactions in
images. In ICCV, 2015. 3, 8
[8] K.-W. Church and P. Hanks. Word association norms, mutual information, and lexicoraphy. Computational Linguistics, 16(1), 1990. 5
[9] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I.
Williams, J. Winn, and A. Zisserman. The PASCAL Visual
Object Classes Challenge: A retrospective. IJCV, 2015. 3, 7
[10] Geena Davis Institute on Gender in Media. The Reel Truth:
Women Aren’t Seen or Heard. https://seejane.
org/research-informs-empowers/data/, 2016.
3
[11] G. Gkioxari and J. Malik. Finding action tubes. In CVPR,
2015. 3
[12] R. Goyal, S. E. Kahou, V. Michalski, J. Materzynska,
S. Westphal, H. Kim, V. Haenel, I. Fründ, P. Yianilos,
M. Mueller-Freitag, F. Hoppe, C. Thurau, I. Bax, and
R. Memisevic. The ”something something” video database
for learning and evaluating visual common sense. In ICCV,
2017. 2, 8
[13] S. Gupta and J. Malik. Visual semantic role labeling. CoRR,
abs/1505.04474, 2015. 3
[14] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In CVPR, 2016. 6
[15] G. V. Horn and P. Perona. The devil is in the tails: Finegrained classification in the wild. arXiv:1709.01450, 2017.
2
[16] R. Hou, C. Chen, and M. Shah. Tube convolutional neural network (t-cnn) for action detection in videos. In ICCV,
2017. 2, 3, 6, 7
[17] J. Huang, V. Rathod, C. Sun, M. Zhu, A. Korattikara,
A. Fathi, I. Fischer, Z. Wojna, Y. Song, S. Guadarrama, and
K. Murphy. Speed/accuracy trade-offs for modern convolutional object detectors. In CVPR, 2017. 6
[18] H. Idrees, A. R. Zamir, Y. Jiang, A. Gorban, I. Laptev,
R. Sukthankar, and M. Shah. The THUMOS challenge on
action recognition for videos “in the wild”. CVIU, 2017. 2,
3
[19] E. Ilg, N. Mayer, T. Saikia, M. Keuper, A. Dosovitskiy, and
T. Brox. FlowNet 2.0: Evolution of optical flow estimation
with deep networks. In CVPR, 2017. 6
[20] H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. Black. Towards understanding action recognition. In ICCV, 2013. 2,
3, 7
[21] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar,
and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 2
[22] W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier,
S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev,
M. Suleyman, and A. Zisserman. The Kinetics human action
video dataset. arXiv:1705.06950, 2017. 2, 7
[23] Y. Ke, R. Sukthankar, and M. Hebert. Efficient visual event
detection using volumetric features. In ICCV, 2005. 3
[24] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre.
HMDB: A large video database for human motion recognition. In ICCV, 2011. 2
[25] H. W. Kuhn. The Hungarian method for the assignment problem. Naval Research Logistics (NRL), 2(1-2):83–97, 1955.
4
[26] M. Marszalek, I. Laptev, and C. Schmid. Actions in context.
In CVPR, 2009. 2
[27] P. Mettes, J. van Gemert, and C. Snoek. Spot On: Action
localization from pointly-supervised proposals. In ECCV,
2016. 3
[28] P. Over, G. Awad, M. Michel, J. Fiscus, G. Sanders,
W. Kraaij, A. Smeaton, and G. Quénot. TRECVID 2014 –
an overview of the goals, tasks, data, evaluation mechanisms
and metrics, 2014. 2
[29] X. Peng and C. Schmid. Multi-region two-stream R-CNN
for action detection. In ECCV, 2016. 3, 6, 7
[30] S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NIPS, 2015. 3, 4, 6
[31] M. Rodriguez, J. Ahmed, and M. Shah. Action MACH: a
spatio-temporal maximum average correlation height filter
for action recognition. In CVPR, 2008. 2, 3
[32] S. Saha, G.Sing, and F. Cuzzolin. AMTnet: Action-microtube regression by end-to-end trainable deep architecture. In
ICCV, 2017. 3
[33] S. Saha, G. Singh, M. Sapienza, P. Torr, and F. Cuzzolin.
Deep learning for detecting multiple space-time action tubes
in videos. In BMVC, 2016. 3
[34] C. Schuldt, I. Laptev, and B. Caputo. Recognizing human
actions: a local SVM approach. In ICPR, 2004. 2
[35] G. Sigurdsson, O. Russakovsky, A. Farhadi, I. Laptev, and
A. Gupta. Much ado about time: Exhaustive annotation of
temporal data. In Conference on Human Computation and
Crowdsourcing, 2016. 4
[36] G. Sigurdsson, G. Varol, X. Wang, A. Farhadi, I. Laptev, and
A. Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016. 3, 6
[37] G. Singh, S. Saha, M. Sapienza, P. Torr, and F. Cuzzolin.
Online real-time multiple spatiotemporal action localisation
and prediction. In ICCV, 2017. 3, 7
[38] K. Soomro, A. Zamir, and M. Shah. UCF101: A dataset of
101 human actions classes from videos in the wild. Technical Report CRCV-TR-12-01, University of Central Florida,
2012. 2, 3
[39] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna.
Rethinking the inception architecture for computer vision. In
CVPR, 2016. 6
[40] V.Kalogeiton, P. Weinzaepfel, V. Ferrari, and C. Schmid. Action tubelet detector for spatio-temporal action localization.
In ICCV, 2017. 2, 3, 6, 7
[41] L. Wang, Y. Qiao, X. Tang, and L. Van Gool. Actionness estimation using hybrid fully convolutional networks. In CVPR,
2016. 7
[42] P. Weinzaepfel, Z. Harchaoui, and C. Schmid. Learning to
track for spatio-temporal action localization. In ICCV, 2015.
3
[43] P. Weinzaepfel, X. Martin, and C. Schmid. Towards weaklysupervised action localization. arXiv:1605.05197, 2016. 3
[44] L. Wu, C. Shen, and A. van den Hengel. PersonNet: Person
re-identification with deep convolutional neural networks.
arXiv preprint arXiv:1601.07255, 2016. 4
[45] S. Yeung, O. Russakovsky, N. Jin, M. Andriluka, G. Mori,
and L. Fei-Fei. Every moment counts: Dense detailed labeling of actions in complex videos. IJCV, 2017. 3
[46] J. Yuan, Z. Liu, and Y. Wu. Discriminative subvolume search
for efficient action detection. In CVPR, 2009. 3
[47] M. Zolfaghari, G. Oliveira, N. Sedaghat, and T. Brox.
Chained multi-stream networks exploiting pose, motion, and
appearance for action classification and detection. In ICCV,
2017. 3
Appendix
In the following, we present additional quantitative information and examples for our AVA dataset as well as for our
action detection approach on AVA.
9. Additional details on the annotation
Figure 10 shows the user interface for bounding box annotation. As described in Section 3.3, we employ a hybrid approach to tradeoff accuracy with annotation cost. We
show annotators frames overlaid by detected person bounding boxes, so they can add boxes to include more persons
missed by the detector.
Figure 10. User interface for bounding box annotation. The purple
box was generated by the person detector. The orange box (missed
by the detector) was manually added by an annotator.
In Section 3.5 of our paper submission, we explain why
our two-stage action annotation design is crucial for preserving high recall of action classes. Here we show quantitative analysis. Figure 11 shows the proportion of labels per
action class generated from each stage. (Blue ones are generated from the first (propose) stage and red ones from the
second (verify) stage). As we can see, for more than half of
our action labels, the majority labels are derived from the
verification stage. Furthermore, the smaller the action class
size, the more likely that they are missed by the first stage
(e.g., kick, exit, extract), and require the second stage to
boost recall. The second stage helps us to build more robust
models for long tail classes that are more sensitive to the
sizes of the training data.
10. Additional details on the dataset
Table 6 and 7 present the number of instances for each
class of the AVA dataset. We observe a significant class imbalance to be expected in real-world data [c.f. Zipf’s Law].
As stated in the paper, we select a subset of these classes
Figure 11. Action class recall improvement due to the two-stage process. For each class, the blue bar shows the proportion of labels
annotated without verification (majority voted results over raters’ selections from 80 classes.), and the red bar shows the proportion of
labels revived from the verification stage. More than half of the action classes doubles their recalls thanks to the additional verification.
(without asterisks) for our benchmarking experiment, in order to have a sufficient number of test examples. Note that
we consider the presence of the “rare” classes as an opportunity for approaches to learn from a few training examples.
Figure 12 shows more examples of common consecutive
atomic actions in AVA.
11. Examples of our action detection
Figure 13 and Figure 14 show the top true positives and
false alarms returned by our best Faster-RCNN with I3D
model.
Pose
stand
sit
walk
bend/bow (at the waist)
lie/sleep
dance
run/jog
crouch/kneel
martial art
get up
jump/leap
fall down
crawl*
swim
#
150482
91131
41004
8949
5351
4155
3404
2520
2007
1347
463
415
228
223
Person-Person Interaction
#
watch (a person)
120804
talk to (e.g. self/person)
99239
listen to (a person)
86080
grab (a person)
3689
fight/hit (a person)
2210
sing to (e.g., self, a person, a group)
1986
hug (a person)
1508
hand clap
1495
give/serve (an object) to (a person)
1195
kiss (a person)
834
take (an object) from (a person)
770
hand shake
584
lift (a person)
540
hand wave
537
426
push (another person)
play with kids
175
kick (a person)*
92
Table 6. Number of instances for pose (left) and person-person (right) interaction labels in the AVA dataset, sorted in decreasing order.
Labels marked by asterisks are not included in the benchmark dataset.
Person-Object Interaction
Person-Object Interaction
#
#
carry/hold (an object)
71339
work on a computer
262
touch (an object)
6697
hit (an object)
224
ride (e.g., a bike, a car, a horse)
4657
play with pets
220
3149
215
answer phone
take a photo
watch (e.g., TV)
2851
point to (an object)
190
2689
183
smoke
climb (e.g., a mountain)*
eat
turn (e.g., a screwdriver)* 178
2257
read
play board game*
1916
157
1610
146
play musical instrument
cut
1486
138
open (e.g., a window, a car door)
press
drive (e.g., a car, a truck)
1352
row boat*
132
drink
shoot*
1225
129
listen (e.g., to music)
1178
exit*
120
lift/pick up
920
clink glass
102
write
892
dig*
95
847
79
close (e.g., a door, a box)
fishing*
put down
733
paint*
69
pull (an object)
591
stir
66
catch (an object)
586
cook*
65
push (an object)
507
shovel*
62
dress/put on clothing
498
chop
56
sail boat*
351
extract*
45
text on/look at a cellphone
brush teeth*
341
41
enter
320
kick (an object)*
39
throw
318
Table 7. Number of instances for person-object interactions in the AVA dataset, sorted in decreasing order. Labels marked by asterisks are
not included in the benchmark.
answer (eg phone) → look at (eg phone)
answer (eg phone) → put down
clink glass → drink
crouch/kneel → crawl
grab → handshake
grab → hug
open → close
Figure 12. We show more examples of how atomic actions change over time in AVA. The text shows pairs of atomic actions for the people
in red bounding boxes.
cut
throw
hand clap
work on a computer
Figure 13. Most confident action detections on AVA. True positives are in green, false alarms in red.
open (e.g. window, door)
get up
smoke
take (something) from (someone)
Figure 14. Most confident action detections on AVA. True positives are in green, false alarms in red.
| 1cs.CV
|
THOMPSON GROUPS FOR SYSTEMS OF GROUPS,
AND THEIR FINITENESS PROPERTIES
arXiv:1405.5491v2 [math.GR] 30 Mar 2016
STEFAN WITZEL AND MATTHEW C. B. ZAREMSKY
Abstract. We describe a procedure for constructing a generalized Thompson group
out of a family of groups that is equipped with what we call a cloning system. The
previously known Thompson groups F , V , Vbr and Fbr arise from this procedure using,
respectively, the systems of trivial groups, symmetric groups, braid groups and pure
braid groups.
We give new examples of families of groups that admit a cloning system and study how
the finiteness properties of the resulting generalized Thompson group depend on those of
the original groups. The main new examples here include upper triangular matrix groups,
mock reflection groups, and loop braid groups. For generalized Thompson groups of
upper triangular matrix groups over rings of S-integers of global function fields, we
develop new methods for (dis-)proving finiteness properties, and show that the finiteness
length of the generalized Thompson group is exactly the limit inferior of the finiteness
lengths of the groups in the family.
Introduction
In 1965 Richard Thompson introduced three groups that today are usually denoted F ,
T , and V . These have received a lot of recent attention for their interesting and often
surprising properties. Most prominently, T and V are finitely presented, infinite, simple
groups, and F is torsion-free with infinite cohomological dimension and of type F∞ .
Numerous generalizations of Thompson’s groups have been introduced in the literature;
see for example [Hig74, Ste92, GS97, Röv99, Bri04, Hug09, MPN13, BF]. Most of these
constructions either generalize the way in which branching can occur, or mimic the selfsimilarity in some way. Here we describe a more algebraic construction of Thompson-like
groups, which combines the usual branching of the group F with a chosen family of
groups. The construction is based on Brin’s description on the braided Thompson group
Vbr [Bri07], which utilizes the family of braid groups. Another example is the pure braided
Thompson group Fbr introduced by Brady, Burillo, Cleary and Stein in [BBCS08], using
the pure braid groups. Classical examples include F , using the trivial group, and V , using
the symmetric groups.
The input to our construction is a directed system of groups (Gn )n∈N together with a
cloning system, which essentially determines how a group element is moved past a split.
A cloning system consists of morphisms Gn → Sn (where Sn is the symmetric group on
n symbols), and cloning maps κnk : Gn → Gn+1 , 1 ≤ k ≤ n, subject to certain conditions
(see Definition 2.18). The output is a group T (G∗ ):
Proposition 2.24. Let (Gn )n∈N be an injective directed system of groups equipped with a
cloning system. Then there is a generalized Thompson group T (G∗ ) that contains all of
the Gn . There are homomorphisms F ֒→ T (G∗ ) → V whose composition is the inclusion
F ֒→ V .
Date: March 31, 2016.
2010 Mathematics Subject Classification. Primary 20F65; Secondary 57M07, 20G30.
Key words and phrases. Thompson’s group, finiteness properties, upper triangular matrix, mock reflection group, loop braid group.
2
S. WITZEL AND M. C. B. ZAREMSKY
The groups F , V , Fbr and Vbr are all examples of groups of the form T (G∗ ).
One of our main motivations for constructing these new Thompson-like groups is the
analysis of their finiteness properties. Recall that a group G is of type Fn if there is a
K(G, 1) with finite n-skeleton. For example, F1 means finitely generated and F2 means
finitely presented. We are in particular interested in understanding how the finiteness
properties of T (G∗ ) depend on the finiteness properties of the groups Gn . Our main
results are:
Theorem 8.28. Let k be a global function field, let S be a set of places of k, and let
OS be the ring of S-integers in k. Let Bn denote the algebraic group of invertible upper
triangular n-by-n matrices. There is a generalized Thompson group T (B∗ (OS )) and it is
of type F|S|−1 but not of type F|S| .
To put this into context it is important to know that the groups Bn (OS ) are themselves
of type F|S|−1 but not of type F|S| by [Bux04]. In particular, for every n ∈ N, we get an
example of a generalized Thompson group of type Fn−1 but not of type Fn .
Theorem 8.10. Let Ab n (Z[1/p]) be the nth Abels group (see Section 7). There is a
generalized Thompson group T (Ab ∗ (Z[1/p])) and it is of type F∞ .
The groups Ab n (Z[1/p]) are known to be of type Fn−1 but not of type Fn by [AB87, Bro87].
To be of type F∞ for a generalization of Thompson’s groups is a relatively common
phenomenon, but what is interesting about this example is that it organizes the groups
Ab n (Z[1/p]), none of which is individually of type F∞ , into a group of type F∞ .
To formulate the above statements in a unified way, it is helpful to introduce the finiteness
length φ(G) of a group G, which is just the supremum over all n for which G is of type Fn .
Now Theorems 8.28 and 8.10 can be formulated to say that
φ(T (G∗ )) = lim inf φ(Gn )
n
(1)
for the respective groups.
This relation is not coincidental but is suggested by the structure of the groups. In fact,
we give a general construction which reduces proving the inequality ≥ for (1) to showing
that certain complexes Ln (G∗ ) are asymptotically highly connected. This construction is
an abstraction of the well developed methods from [Bro92, Ste92, Bro06, Far03, FMWZ13,
BFM+ 14] (which were all used to prove that the respective groups are of type F∞ ). For
this reason, the proof of the inequality ≥ in Theorem 8.28 works without change for the
groups Bn (R) where R is an arbitrary ring. This evidence leads us to ask:
Question 5.1. For which generalized Thompson groups T (G∗ ) does (1) hold?
The group T (G∗ ) may be thought of as a limit of the groups Gn , for example since it
contains all of them. From this point of view, it is rather remarkable that (1) holds in
such generality. For example compare this to an ascending direct limit of groups with
good finiteness properties, which will not even be finitely generated.
Another reason why (1) is interesting is that it describes how finiteness properties of groups
change when they are subject to a certain operation (here Thompsonifying). A different
such operation is braiding: when V is “braided,” we get Vbr , and similarly F yields
Fbr . The question of the finiteness properties of Fbr and Vbr was answered in [BFM+ 14];
they are still of type F∞ , just like F and V . When reinterpreting F , V , Fbr and Vbr as
Thompsonifications (of the trivial group, the symmetric groups, the pure braid groups,
and the braid groups, respectively), they provide more examples where (1) holds: in all
of these cases all the groups Gn are of type F∞ and so are the corresponding Thompson
groups. This is in some cases related to a similar program carried out in [BdCK15] for
wreath products, see Remark 3.4.
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
3
In addition to the groups discussed so far, we also construct generalized Thompson groups
for more families of groups. All of them are relatives of the family of symmetric groups
in some way and it is very natural to put them into a generalized Thompson group. The
first is a family of mock reflection groups that were studied by Davis, Januszkiewicz and
Scott [DJS03]. The groups naturally arise as blowups of symmetric groups and we call
them mock symmetric groups. Constructing a generalized Thompson group for the mock
symmetric groups was suggested to us by Januszkiewicz. The second family consists of
loop braid groups, which are a melding of symmetric groups and braid groups.
Theorem 9.2, 10.2. There exist generalized Thompson groups Vmock , Vloop and Floop
built from (and thus containing) all mock symmetric groups, all loop braid groups, and
all pure loop braid groups. The groups Vmock and Vloop surject onto V and Floop surjects
onto F .
We expect that all of these groups belong to the list of groups that answer Question 5.1
positively, and thus:
Conjecture 9.3, 10.3. Vmock , Vloop and Floop are of type F∞ .
To investigate the finiteness properties of a generalized Thompson group T (G∗ ) we let it
act on a contractible cube complex X (G∗ ) which we call the Stein–Farley complex. This
space exists for arbitrary cloning systems and in many cases has been used previously.
When the cloning system is properly graded (Definition 2.16), the action has certain
desirable properties: the cell stabilizers are subgroups of the groups Gn and there is a
natural cocompact filtration. To show that the generalized Thompson group is of type Fn ,
assuming that all the Gn are, (which gives one half of (1)) thus amounts to showing that
the descending links Ln (G∗ ) in this filtration are eventually (n − 1)-connected. This is
the only part of the proof that needs to be done for every properly graded cloning system
individually and depends on the nature of the concrete example. This treats the positive
case, which so far has been sufficient for most existing Thompson groups since they have
been of type F∞ .
For the negative finiteness properties we have to develop new methods. For example we
give a condition on a group homomorphism G → H that ensures that if the morphism
factors through a group K then K cannot be of type FPn , see Theorem 5.14 (type FPn
is a homological, and slightly weaker, version of type Fn ). This is a similar idea to
that of [KM97] and may be of independent use. Unlike the proof that T (B∗ (OS )) is
of type F|S|−1 , the proof that it is not of type FP|S| borrows large parts from the proof
in [Bux04] of the same fact for Bn (OS ). For example, the space for T (B∗ (OS )) is built
out of the space for B2 (OS ) (which is a Bruhat–Tits tree).
The paper is organized as follows. In Section 1 we recall some background on monoids and
the Zappa–Szép product. In Section 2 we introduce cloning systems (Definition 2.18) and
explain how they give rise to generalized Thompson groups. Section 3 collects some group
theoretic consequences that follow directly from the construction. To study finiteness
properties, the Stein–Farley complex is introduced in Section 4. The filtration and its
descending links are described in Section 5, and we discuss some background on Morse
theory and other related techniques for proving high connectivity, including a new method
in Section 5.5.
Up to this point everything is mostly generic. The following sections discuss examples.
Section 6 gives an elementary example where Gn = H n for some group H. Section 7
discusses cloning systems for groups of upper triangular matrices. In Section 8 we study
their finiteness properties. The last two sections 9 and 10 introduce the groups Vmock and
Vloop and Floop .
4
S. WITZEL AND M. C. B. ZAREMSKY
splits
a
c
=
=
group element
d
b
merges
=
=
Figure 1. On the left, an element of Vbr in its standard form consisting of
splitting, braiding and merging. On the right, some relations: (a) splitting
and then merging is trivial; (b) merging and then splitting is trivial; (c)
splits and merges on different strands commute. The main relation, (d),
which is encoded by the Zappa–Szép product, is how splits and group
elements interact.
Acknowledgments. We are grateful to Matt Brin and Kai-Uwe Bux for helpful discussions, to Tadeusz Januszkiewicz for proposing to us the group Vmock , and to Werner
Thumann and an anonymous referee for many helpful comments. Both authors were supported by the SFB 878 in Münster. The second author was also supported directly by
the DFG through project WI 4079/2 and by the SFB 701 in Bielefeld. All of this support
is gratefully acknowledged.
0. Motivation
Starting with the first section we will spend some ten pages introducing notions and
technical results from the theory of monoids. Before we dive into these preparations, we
want to explain why they are precisely the ones needed to describe generalized Thompson
groups. We illustrate this on the example of Vbr .
We want to think of an element of a Thompson group as consisting of a tree of splittings,
followed by a group element from a chosen group (a braid in the example), and finally
an inverse tree of merges. An element of Vbr is illustrated in Figure 1. Two elements are
multiplied by stacking them on top of each other and reducing, as in Figure 2. Among the
relations available to reduce an element are the fact that splitting and then merging again
is a trivial operation, as well as merging and then splitting (Figure 1(a),(b)). Another
relation that is implicit in the pictures is that a group element followed by another group
element is the same as the product. However, these relations are not typically sufficient
to bring a diagram into the form that we want: splits, group element, merges. To move
all the splits to the top (and all the merges to the bottom), we eventually will have to
move a split λ past a group element g. In Figure 2 this point is reached in the third step.
Expressed algebraically, we need to rewrite gλ = λ′ g′ for some group element g′ and some
split λ′ (Figure 1(d)). The algebraic operation that defines how a split is moved past a
group element is the Zappa–Szép product.
The trees of splittings will be elements of the forest monoid F . We will then form the
Zappa-Szép product F ⊲⊳ G with the chosen group G. To also obtain merges, we will pass
to the group of fractions — a merge is just the inverse of a split. For technical reasons,
we will have started with infinitely many strands and in a final step have to reduce to
elements that start and end with one strand. With this outline in mind, we hope the
reader will find the following technical pages more illuminating.
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
1
2
3
4
5
5
Figure 2. Computing the product of two elements of Vbr . First, both
elements are stacked onto each other. Second, pairs of merges and splits
are resolved. Third, merges and splits are moved past each other. In the
fourth and fifth step a merge and a split are moved past a group element
(here a braid).
1. Preliminaries
Much of the material in this section is taken from [Bri07].
1.1. Monoids. A monoid is an associative binary structure with a two-sided identity. A
monoid M is called left cancellative if for all x, y, z ∈ M , we have that xy = xz implies
y = z. Elements x, y ∈ M have a common left multiple m if there exist z, w ∈ M such
that zx = wy = m. This is the least common left multiple if for all p, q ∈ M such that
px = qy, we have that px is a left multiple of m. There are the obvious definitions of right
cancellative, common right multiples and least common right multiples. We say that M
has common right/left multiples if any two elements have a common right/left multiple.
It is said to have least common right/left multiples if any two elements that have some
common right/left multiple have a least common right/left multiple. Finally, we say M
is cancellative if it is both left and right cancellative. The importance of these notions
lies in the following classical theorem (see [CP61, Theorems 1.23, 1.25]):
Theorem 1.1 (Ore). A cancellative monoid with common right multiples has a unique
group of right fractions.
Recall that for every monoid M there exists a group GM and a monoid morphism ω : M →
GM such that every monoid morphism from M to a group factors through ω (namely the
group generated by all the elements of M subject to all the relations that hold in M ). This
is the group of fractions of M . The morphism ω will be injective if and only if M embeds
into a group. A group G is called a group of right fractions of M if it contains M and every
element of G can be written as m · n−1 with m, n ∈ M . A group of right fractions exists
precisely in the situation of Ore’s theorem and is unique up to isomorphism; see [CP61,
Section 1.10] for details. We call a monoid satisfying the hypotheses of Theorem 1.1 an
Ore monoid. The group of right fractions of an Ore monoid is its group of fractions (see
for example [KS06, Theorem 7.1.16]):
Lemma 1.2. Let M be an Ore monoid, let G be its group of right fractions and let H be
any group. Let ϕ : M → H be a monoid morphism. Then the map ϕ̃ : G → H defined by
ϕ̃(mn−1 ) = ϕ(m) · ϕ(n)−1 is a group homomorphism and ϕ = ϕ̃|M .
Proof. That inverses map to inverses is clear. Let m1 , m2 , n1 , n2 ∈ M and let n1 ·x = m2 ·y
−1
−1 −1
be a common right multiple so that m1 n−1
1 m2 n2 = m1 xy n2 . We have to check that
ϕ(m1 )ϕ(n1 )−1 ϕ(m2 )ϕ(n2 )−1 = ϕ(m1 x)ϕ(n2 y)−1 .
(1.1)
6
S. WITZEL AND M. C. B. ZAREMSKY
The fact that ϕ is a monoid morphism means that ϕ(n1 )ϕ(x) = ϕ(m2 )ϕ(y) which entails
ϕ(n1 )−1 ϕ(m2 ) = ϕ(x)ϕ(y)−1 . Extending by ϕ(m1 ) from the left and by ϕ(n2 )−1 from
the right gives (1.1).
1.2. Posets from monoids. Throughout this section let M be an Ore monoid and let G
be its group of right fractions. The notions of left/right multiple/factor are uninteresting
for G as a monoid because it is a group. Instead we introduce these notions relative to
the monoid M . Concretely, assume that elements a, b, c ∈ G satisfy
ab = c.
If a ∈ M then we call b a right factor of c and c a left multiple of b. If b ∈ M then
we call a a left factor of c and c a right multiple of a. If g is a left factor (respectively
right multiple) of both h and h′ then we say that it is a common left factor (respectively
common right multiple). If g is a common left factor of h and h′ and any other left factor
of h and h′ is also a left factor of g then g is called a greatest common left factor. If g is a
common right multiple of h and h′ and every other right multiple is also a right multiple
of g then g is called a least common right multiple of h and h′ . Thus we obtain notions of
when G has (least) common right/left multiples and (greatest) common right/left factors.
We say that two elements have no common right factor if they have greatest common
right factor 1.
Under a moderate additional assumption, having least common right multiples is inherited
by G from M :
Lemma 1.3. Let M have least common right multiples. Let n, n′ , m, m′ ∈ M be such
that n and m have no common right factor and neither do n′ and m′ . Let nv = n′ u be a
least common right multiple of n and n′ . Then nv = n′ u is a least common right multiple
of nm−1 and n′ m′ −1 .
We call a monoid homomorphism len : M → N0 a length function if every element of the
kernel is a unit. It induces a length function len : G → Z. Note that if M admits a length
function then every element of G can be written as mn−1 where m and n are elements of
M with no common right factor.
The following is an extension of [Bri07, Lemma 2.3] to G.
Lemma 1.4. Assume that M admits a length function. Then G has least common right
multiples if and only if it has greatest common left factors.
One reason for our interest in least common right multiples and greatest common left
factors is order theoretic. Define a relation on G by declaring g ≤ h if g is a left factor
of h. This relation is reflexive and transitive but fails to satisfy antisymmetry if M has
non-trivial units. We denote the relation induced on G/M × also by ≤. It is an order
relation so G/M × becomes a partially ordered set (poset). Spelled out, the relation is
given by gM × ≤ hM × if g−1 h ∈ M .
The algebraic properties discussed before immediately translate into order theoretic properties: recall that a poset P is a join-semilattice if any two elements of P have a supremum
(their join). We say that P has conditional meets if any two elements that have a lower
bound have an infimum.
Observation 1.5. If M has common right multiples, least common right multiples, and
greatest common left factors then M/M × is a join-semilattice with conditional meets.
Similarly, if G has common right multiples, least common right multiples and greatest
common left factors then G/M × is a join-semilattice with conditional meets.
Putting everything together, we find:
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
...
7
...
=
...
Figure 3. Multiplication of forests.
Corollary 1.6. Let M be a cancellative monoid with common right multiples, least common right multiples and length function. Let G be its group of right fractions. Then
G/M × is a join-semilattice with conditional meets.
1.3. The monoid of forests. Since we are interested in Thompson’s groups, an important monoid in all that follows will be the monoid of forests, which we define in this
section.
For us, a tree is always a finite rooted full binary tree. In other words, every vertex has
either no outgoing edges or a left and right outgoing edge, and every vertex other than
the root has an incoming edge. The vertices without outgoing edges are called leaves.
The distinction between left and right induces a natural order on the leaves. If a tree has
only one leaf, then the leaf is also its root and the tree is the trivial tree.
By a forest we mean a sequence of trees E = (Ti )i∈N such that all but finitely many Ti
are trivial. The roots are numbered in the obvious way, i.e., the ith root of E is the root
of Ti . If all the Ti are trivial we call E trivial. If the Ti are trivial for i > 1 then the forest
is called semisimple (here we deviate from Brin’s notation; what we call “semisimple” is
called “simple” in [Bri07], and what we will later call “simple”, Brin calls “simple and
balanced”). The rank of E is the least index i such that Tj is trivial for j > i. So E is
semisimple if it has rank at most 1. The leaves of all the Ti are called the leaves of E. The
order on the leaves of the trees induces an order on the leaves of the forest by declaring
that any leaf of Ti comes before any leaf of Tj , whenever i < j. We may equivalently
think of the leaves as numbered by natural numbers. The number of feet of a semisimple
forest (Ti )i∈N is the number of leaves of T1 .
Let F be the set of forests. Define a multiplication on F as follows. Let E = (Tk ) and
E ′ = (Tk′ ) be forests, and set EE ′ to be the forest obtained by identifying the ith leaf of
E with the ith root of E ′ , for each i. This product is associative, and the trivial forest
is a left and right identity, so F is a monoid. Some more details on F can be found in
Section 3 of [Bri07]. Figure 3 illustrates the multiplication of two elements.
There is an obvious set of generators of F , namely the set of single-caret forests. Such a
forest can be characterized by the property that there exists k ∈ N such that for i < k,
the ith root is also the ith leaf, and for i > k, the ith root is also the (i + 1)st leaf. Denote
this forest by λk . Every tree in λk is trivial except for the kth tree, which is a single caret.
Proposition 1.7 (Presentation of the forest monoid). [Bri07, Proposition 3.3] F is
generated by the λk , and defining relations are given by
λj λi = λi λj+1
for
i < j.
(1.2)
Every element of F can be uniquely expressed as a word of the form λk1 λk2 · · · λkr for
some k1 ≤ · · · ≤ kr .
8
S. WITZEL AND M. C. B. ZAREMSKY
A consequence is that the number of carets is an invariant of a forest, and is exactly
the length of the word in the λk representing the forest. The following is part of [Bri07,
Lemma 3.4].
Lemma 1.8. The monoid F has the following properties.
(1) It is cancellative.
(2) It has common right multiples.
(3) It has no non-trivial units.
(4) There is a monoid homomorphism len : F → N0 sending each generator to 1.
(5) It has greatest common right factors and least common left multiples.
(6) It has greatest common left factors and least common right multiples.
In view of Theorem 1.1, properties (1) and (2) imply that F has a unique group of right
fractions, which we denote Fb.
1.4. Zappa–Szép products. In this section we recall the background on Zappa–Szép
products of monoids. Our main reference is [Bri07, Section 2.4], and also see [Bri05].
When the monoids are groups, Zappa–Szép products generalize semidirect products by
dropping the assumption that one of the groups be normal.
The internal Zappa–Szép product is straightforward to define. Let M be a monoid with
submonoids U and A such that every m ∈ M can be written in a unique way as m = uα
for u ∈ U and α ∈ A. In particular, for α ∈ A and u ∈ U there exist u′ ∈ U and α′ ∈ A
such that αu = u′ α′ , and the u′ and α′ are uniquely determined by α and u, so we denote
them u′ = α · u and α′ = αu , following [Bri07]. The maps (α, u) 7→ α · u and (α, u) 7→ αu
should be thought of as mutual actions of U and A on each other. Then we can define a
multiplication on U × A via
(u, α)(v, β) := (u(α · v), αv β),
(1.3)
for u, v ∈ U and α, β ∈ A, and the map (u, α) 7→ uα is a monoid isomorphism from U × A
(with this multiplication) to M ; see [Bri07, Lemma 2.7]. We say that M is the (internal)
Zappa–Szép product of U and A, and write M = U ⊲⊳ A.
Example 1.9 (Semidirect product). Suppose G is a group that is a semidirect product
G = U ⋉ A for U, A ≤ G. Then for u ∈ U and α ∈ A we have αu = u(u−1 αu), and
u−1 αu ∈ A, so the actions defined above are just α · u = u and αu = u−1 αu.
We actually need to use the external Zappa–Szép product. This is discussed in detail in
[Bri07, Section 2.4] (and in even more detail in [Bri05]).
Definition 1.10 (External Zappa–Szép product). Let U and A be monoids with maps
(α, u) 7→ α · u ∈ U and (α, u) 7→ αu ∈ A satisfying the following eight properties for all
u, v ∈ U and α, β ∈ A:
1)
2)
3)
4)
5)
6)
7)
8)
1A · u = u
(αβ) · u = α · (β · u)
α1U = α
α(uv) = (αu )v
(1A )u = 1A
(αβ)u = α(β·u) β u
α · 1U = 1U
α · (uv) = (α · u)(αu · v).
(Identity acting on U )
(Product acting on U )
(Identity acting on A)
(Product acting on A)
(U acting on identity)
(U acting on product)
(A acting on identity)
(A acting on product)
Then the maps are called a Zappa–Szép action. The set U × A together with the multiplication defined by (1.3) is called the (external) Zappa–Szép product of U and A, denoted
U ⊲⊳ A.
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
9
It is shown in Lemma 2.9 in [Bri07] that the external Zappa–Szép product turns U ⊲⊳ A
into a monoid and coincides with the internal Zappa–Szép product of U and A with
respect to the embeddings u 7→ (u, 1A ) and α 7→ (1U , α).
Some pedantry about the use of the word “action” might now be advisable. The action of
U on A is a right action described by a homomorphism of monoids U → Symm(A), where
Symm(A) is the symmetric group on A (and is not the group of monoid automorphisms).
The action of A on U is a left action described by a homomorphism of monoids A →
Symm(U ), again not to Aut(U ). In a phrase, both actions are actions of monoids as
monoids, but on monoids as sets.
Brin [Bri07] regards the action (α, u) 7→ αu of U on A as a family of maps from A to itself
parametrized by U and defines properties of this family. For brevity we apply the same
adjectives to the action itself but one should think of the family of maps. The action is
called injective if for all u ∈ U , αu = β u implies α = β. It is surjective if for every α ∈ A
and u ∈ U there exists a β ∈ A with β u = α. The action is strongly confluent if the
following holds: if u, v ∈ U have a least common left multiple ru = sv and α = β u = γ v
for some β, γ ∈ A then there is a θ ∈ A such that θ r = β and θ s = γ. Note that if the
action is injective then for this to happen it is sufficient that θ ru = α. The notions for
the action of A on U are defined by analogy.
The following lemma can be found as Lemma 2.12 in [Bri07], or as Lemma 3.15 in [Bri05].
Lemma 1.11. Let U be a cancellative monoid with least common left multiples and let
A be a group. Let U and A act on each other via Zappa–Szép actions. Assume that the
action (α, u) 7→ αu of U on A is strongly confluent. Then M = U ⊲⊳ A has least common
left multiples.
A least common left multiple (r, α)(u, θ) = (s, β)(v, φ) of (u, θ) and (v, φ) in M can be
constructed so that r(α · u) = s(β · v) is the least common left multiple of (α · u) and (β · v)
in U . If M is cancellative, every least common left multiple will have that property.
Being actions of monoids, Zappa–Szép actions are already determined by the actions of
generating sets. It is not obvious, but also true, that they are often also determined by
the actions of generating sets on generating sets. This means that, in order to define
the actions, we need only define α · u and uα where both α and u come from generating
sets. Brin [Bri07, pp. 768–769] gives a sufficient condition for such partial actions to
extend to well defined Zappa–Szép actions, which we restate here. Given sets X and Y ,
let X ∗ and Y ∗ denote the free monoids generated respectively by them. Suppose maps
Y × X → Y ∗ , (α, u) 7→ αu and Y × X → X, (α, u) 7→ α · u are given (so α · u should be
a single generator, but αu may be a string of generators). Let W be the set of relations
(αu, (α · u)(αu )) with α ∈ Y, u ∈ X. Then
hX ∪ Y | W i
is a Zappa–Szép product of X ∗ and Y ∗ . In particular, the above maps extend to Zappa–
Szép actions Y ∗ × X ∗ → Y ∗ and Y ∗ × X ∗ → X ∗ .
Lemma 1.12 ([Bri07, Lemma 2.14]). Let U = hX | Ri and A = hY | T i be presentations
of monoids (with X ∩ Y = ∅). Assume that functions Y × X → Y ∗ , (α, u) 7→ αu and
Y × X → X, (α, u) 7→ α · u are given. Let ∼R and ∼T denote the equivalence relations on
X ∗ and Y ∗ imposed by the relation sets R and T .
Extend the above maps to Y ∗ × X ∗ as above. Assume that the following are satisfied. If
(u, v) ∈ R then for every α ∈ Y we have (α · u, α · v) ∈ R or (α · v, α · u) ∈ R, and also
αu ∼T αv . If (α, β) ∈ T then for all u ∈ X we have α · u = β · u and αu ∼T β u .
Then the lifted maps induce well defined Zappa–Szép actions and the restriction of the
map A × U → U to A × X has its image in X. A presentation for U ⊲⊳ A is
hX ∪ Y | R ∪ T ∪ W i
10
S. WITZEL AND M. C. B. ZAREMSKY
where W consists of all pairs (αu, (α · u)(αu )) for (α, u) ∈ Y × X.
2. Cloning systems and generalized Thompson groups
2.1. Brin–Zappa–Szép products and cloning systems. To construct Thompson-like
groups we now consider Zappa–Szép products F ⊲⊳ G of the forest monoid F with a group
G.
Definition 2.1 (BZS products). Suppose we have Zappa–Szép actions (g, E) 7→ g · E and
(g, E) 7→ g E on G × F , for G a group. For each standard generator λk of F the map
κk = κλk : G → G given by g 7→ gλk is called the kth cloning map. If every such cloning
map is injective, we call the actions Brin–Zappa–Szép (BZS) actions and call the monoid
F ⊲⊳ G the Brin–Zappa–Szép (BZS) product.
Since the action of F on G is a right action we will also write the cloning maps κk on
the right.
The monoid F is cancellative and has common right multiples, and the same is true of
G, being a group. Since G is a group these properties are inherited by F ⊲⊳ G:
Observation 2.2. A BZS product F ⊲⊳ G is cancellative and has (least) common right
multiples. In particular it has a group of right fractions.
Proof. This follows easily from the statements about F using the unique factorization in
Zappa–Szép products and that E is a right multiple and left factor of (E, g).
In Definition 2.1 we have already simplified the data needed to describe BZS products by
using the fact that F is generated by the λk . In a similar fashion the following lemma
reduces the data needed to describe the action of G on F . We denote by Sω the group
Symm(N) of permutations of N and by S∞ ≤ Sω the subgroup of permutations that fix
almost all elements of N.
Lemma 2.3 (Carets to carets). Let F ⊲⊳ G be a BZS product. The action of G on F
preserves the set Λ = {λk }k∈N and so induces a homomorphism ρ : G → Sω . Conversely,
the action of G on F is completely determined by ρ and (κk )k∈N .
Proof. For g ∈ G and E, F ∈ F , we know that g ·(EF ) = (g ·E)(g E ·F ) by Definition 1.10.
We show that the action of G preserves Λ. If g · λk = EF then g −1 · (EF ) = λk , so one of
g−1 · E or (g −1 )E · F equals 1F . Again by Definition 1.10, we see that either E = 1F or
F = 1F . We conclude that g · λk equals λℓ for some ℓ depending on k and g. The map ρ
then is defined via ρ(g)k = ℓ.
To see that the action of G on F is determined by ρ and (κk ), we use repeated applications
of the equation g · (λk E) = λρ(g)k ((g)κk · E).
As a consequence we see that the action of G on F preserves the length of an element:
Corollary 2.4. There is a monoid homomorphism len : F ⊲⊳ G → N0 taking (E, g) to
the length of E in the standard generators. The kernel of len is G = (F ⊲⊳ G)× .
In particular, len is a length function in the sense of Section 1.2. The induced morphism
from the group of right fractions to Z (Lemma 1.2) is also denoted len.
The next result is a technical lemma that tells us that ρ and the cloning maps always
behave well together, in any BZS product.
Lemma 2.5 (Compatibility). Let F ⊲⊳ G be a BZS product. The homomorphism ρ : G →
Sω and the maps (κk )k∈N satisfy the following compatibility condition for k < ℓ:
If ρ(g)k < ρ(g)ℓ then ρ((g)κℓ )k = ρ(g)k and ρ((g)κk )(ℓ + 1) = ρ(g)ℓ + 1.
If ρ(g)k > ρ(g)ℓ then ρ((g)κℓ )k = ρ(g)k + 1 and ρ((g)κk )(ℓ + 1) = ρ(g)ℓ.
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
11
Proof. For k < ℓ we know that
g · (λℓ λk ) = g · (λk λℓ+1 ).
Writing this out using the axioms for Zappa–Szép products we obtain that
(g · λℓ )(gλℓ · λk ) = (g · λk )(gλk · λℓ+1 )
which can be rewritten using the action morphism ρ as
λρ(g)ℓ λρ(gλℓ )k = λρ(g)k λρ(gλk )(ℓ+1) .
Using the normal form for F (see Proposition 1.7) we can distinguish cases for how this
could occur. The first case is that both pairs of indices
(ρ(g)ℓ, ρ(g λℓ )k) and (ρ(g)k, ρ(g λk )(ℓ + 1))
are ordered increasingly and coincide. But this is impossible because ρ(g)ℓ 6= ρ(g)k. The
second case is that both pairs are ordered strictly decreasingly and coincide, which is
impossible for the same reason. The remaining two cases have that one pair is ordered
increasingly and the other strictly decreasingly. In either case the monoid relation now
yields a relationship among the indices, namely either
ρ(g λk )(ℓ + 1) − 1 = ρ(g)ℓ > ρ(g λℓ )k = ρ(g)k
or
ρ(g)ℓ = ρ(g λk )(ℓ + 1) < ρ(g)k = ρ(g λℓ )k − 1.
Finally, replacing the action of λk by the map κk yields the result.
The compatibility condition
ρ((g)κℓ )(k) =
can also be rewritten as
ρ(g)(k)
ρ(g)(k) + 1
ρ(g)(k − 1)
ρ(g)(k − 1) + 1
k < ℓ, ρ(g)k < ρ(g)ℓ,
k < ℓ, ρ(g)k > ρ(g)ℓ,
k − 1 > ℓ, ρ(g)(k − 1) < ρ(g)ℓ,
k − 1 > ℓ, ρ(g)(k − 1) > ρ(g)ℓ.
(2.1)
Lemma 2.3 said that the action of G on F is uniquely determined by ρ and the cloning
maps. The action of F on G is also uniquely determined by the cloning maps, simply
because F is generated by the λk . Our findings can be summarized as:
Proposition 2.6 (Uniqueness). A BZS product F ⊲⊳ G induces a homomorphism ρ : G →
Sω and injective maps κk : G → G, k ∈ N satisfying the following conditions for k, ℓ ∈ N
with k < ℓ and g, h ∈ G:
(CS1) (gh)κk = (g)κρ(h)k (h)κk .
(CS2) κℓ ◦ κk = κk ◦ κℓ+1 .
(CS3) If ρ(g)k < ρ(g)ℓ then ρ((g)κℓ )k = ρ(g)k and
ρ((g)κk )(ℓ + 1) = ρ(g)ℓ + 1.
If ρ(g)k > ρ(g)ℓ then ρ((g)κℓ )k = ρ(g)k + 1 and
ρ((g)κk )(ℓ + 1) = ρ(g)ℓ.
The BZS product is uniquely determined by these data.
(Cloning a product)
(Product of clonings)
(Compatibility)
The converse is also true:
Proposition 2.7 (Existence). Let G be a group, ρ : G → Sω a homomorphism and
(κk )k∈N a family of injective maps from G to itself. Assume that for k < ℓ and g, h ∈ G
the conditions (CS1), (CS2) and (CS3) in Proposition 2.6 are satisfied.
Then there is a well defined BZS product F ⊲⊳ G corresponding to these data.
12
S. WITZEL AND M. C. B. ZAREMSKY
Proof. We will verify the assumptions of Lemma 1.12. This will produce a Zappa–Szép
action, which will be a Brin–Zappa–Szép action by construction. We take U to be F
with the presentation
hλk for k ∈ N | (λℓ λk , λk λℓ+1 ) for k < li.
Let R denote the set of relations used here and let Rsym be the symmetrization. We take
A to be G with the trivial presentation
hg for g ∈ G | (gh, g ′ ) for gh = g′ i.
The maps on generators are defined as g λk := (g)κk and g · λk := λρ(g)k .
First, for k < ℓ and g ∈ G we need to verify that
(g · (λℓ λk ), g · (λk λℓ+1 )) ∈ Rsym
and gλℓ λk = gλk λℓ+1 .
The latter of these is just condition (CS2). The former condition means that
(λρ(g)ℓ λρ((g)κℓ )k , λρ(g)k λρ((g)κk )(ℓ+1) )
should lie in
Rsym .
If ρ(g)k > ρ(g)ℓ we can use condition (CS3) to rewrite this as
(λρ(g)ℓ λρ(g)k+1 , λρ(g)k λρ(g)ℓ )
which is in
Rsym .
If ρ(g)k < ρ(g)ℓ then the tuple is
(λρ(g)ℓ λρ(g)k , λρ(g)k λρ(g)ℓ+1 )
which already lies in R.
Second, for every relation (gh, g′ ) of G and every k ∈ N we have to verify that
(gh) · λk = g ′ · λk
λk
and (gh)λk = (g′ )
for k ∈ N. The former is not really a condition because the partial action was already
defined using G (rather than the free monoid spanned by G). The latter means that we
need
λ
(g ′ ) k = g λρ(h)k hλk
which is just condition (CS1).
Definition 2.8. Let G be a group, ρ : G → Sω a homomorphism and (κk )k∈N : G → G a
family of maps, also denoted κ∗ for brevity. The triple (G, ρ, κ∗ ) is called a cloning system
if the data satisfy conditions (CS1), (CS2) and (CS3) above. We may also refer to ρ and
κ∗ as a forming a cloning system on G.
We now discuss an extended example, of the infinite symmetric group, and show that we
have a cloning system. It is exactly the cloning system that gives rise to Thompson’s
group V .
Example 2.9 (Symmetric groups). Let G = S∞ . Let ρ : S∞ → Sω just be inclusion. The
action of G on F is thus given by g · λk = λρ(g)k = λgk .
Since we will use the specific cloning maps in this example even in the future general
setting, we will give them their own name, ςℓ . They are defined by the formula
gm
m ≤ k, gm ≤ gk,
gm + 1
m < k, gm > gk,
((g)ςk )(m) =
(2.2)
g(m
−
1)
m > k, g(m − 1) < gk,
g(m − 1) + 1 m > k, g(m − 1) ≥ gk.
If we draw permutations as strands crossing each other, the word “cloning” becomes
more or less literal: applying the kth cloning map creates a parallel copy of the kth
strand, where we count the strands at the bottom. See Figure 4 for an example.
We will prove that this defines a cloning system by verifying (CS1), (CS2) and (CS3). For
this example we will just verify them directly, and not use any specific presentation for S∞ .
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
13
ς
2
−→
Figure 4. An example of cloning in symmetric groups. Here we see that
(1 2)ς2 = (1 3 2).
It is immediate from (2.2) that the compatibility condition (CS3) in the formulation (2.1)
is satisfied.
To aid in checking condition (CS1), we define two families of maps, πk : N → N and
τk : N → N, for k ∈ N:
m
m ≤ k,
m
m ≤ k,
(2.3)
and τk (m) =
πk (m) =
m + 1 m > k.
m−1 m>k
Note that πk ◦ τk = id and τk ◦ πk (m) = m, unless m = k + 1 in which case it equals m − 1.
In the m = k + 1 case, we see that
(gh)ςk (k + 1) = gh(k) + 1 = (g)ςhk (hk + 1) = (g)ςhk (h)ςk (k + 1),
by repeated use of the last case in the definition. It remains to check condition (CS1) in
the m 6= k + 1 case. According to the definitions, we have
((g)ςk )(m) = τgk (gπk (m))
whenever m 6= k + 1. Using this we see that
((g)ςhk ) ◦ ((h)ςk )(m) = τghk gπhk ◦ τhk hπk (m)
= τghk ghπk (m)
= ((gh)ςk )(m)
for m 6= k + 1.
To check condition (CS2), we consider k < ℓ. We first verify, from the definition, the
special cases
((g)ςℓ ◦ ςk )(k + 1) = gk + 1 = ((g)ςk ◦ ςℓ+1 )(k + 1) and
((g)ςℓ ◦ ςk )(ℓ + 2) = gℓ + 2 = ((g)ςk ◦ ςℓ+1 )(ℓ + 2).
For the remaining case, when m 6= k + 1, ℓ + 2, we have
((g)ςℓ ◦ ςk )(m) = τk τℓ gπℓ πk (m) and
((g)ςk ◦ ςℓ+1 )(m) = τℓ+1 τk gπk πℓ+1 (m)
and it is straightforward to check that
πℓ πk = πk πℓ+1 and τk τℓ = τℓ+1 τk .
(2.4)
We conclude that (S∞ , ρ, (ςk )k ) is a cloning system.
Remark 2.10. Besides the example of symmetric groups there are two more examples
of cloning systems previously existing in the literature (though of course not using this
language): they are for the families of braid groups and pure braid groups and were used
in [Bri07, BBCS08] to construct Vbr and Fbr .
Observation 2.11 (Simplified compatibility). Condition (CS3) in Proposition 2.6 can
equivalently be rewritten as
ρ((g)κk )(i) = (ρ(g))ςk (i) for all i 6= k, k + 1.
14
S. WITZEL AND M. C. B. ZAREMSKY
All the examples in the later sections satisfy the condition in Observation 2.11 even when
i = k, k + 1.
Remark 2.12. Proposition 2.7 is an application of Lemma 1.12 to the trivial presentation.
As this example demonstrates, it can be rather involved to verify the conditions for a
cloning system. If the group in question comes equipped with a presentation involving
only short relations, it may be easier to re-run the proof of Proposition 2.7 with that
presentation by applying Lemma 1.12. In this case one has to check (CS2) and (CS3)
only on generators, but also has to check a variant of (CS1) for every relation.
We finish by discussing the case when we have least common left multiples. Let κ∗ be the
cloning maps of a cloning system. For E = λk1 · · · λkr define κE := κk1 ◦ · · · ◦ κkr . Note
that this is well defined by condition (CS2) and is just the map g 7→ g E .
Observation 2.13. Let G be a group and let (ρ, κ∗ ) be a cloning system on G. The action
of F on G defines a strongly confluent family if and only if im(κE1 ) ∩ im(κE2 ) = im(κF )
whenever E1 and E2 have least common left multiple F .
In particular the BZS product F ⊲⊳ G has least common left multiples in that case.
Proof. This is just unraveling the definition and using the remark before Lemma 1.11.
Assume that the above condition holds. Write F = F1 E1 = F2 E2 . Assume that g =
g1E1 = g2E2 , that is, g ∈ im(κE1 ) ∩ im(κE2 ). By assumption there is an h ∈ G such that
g = (h)κF . That is g = hF = hF1 E1 = g1E1 . Injectivity of the action of F on G now
implies hF1 = g1 . A similar argument shows hF2 = g2 .
Conversely assume that the action of F on G is strongly confluent and write F as before.
Let g ∈ im(κE1 ) ∩ im(κE2 ). Write g = (g1 )κE1 and g = (g2 )κE2 , that is g = g1E1 and
g = g2E2 . By strong confluence there is an h ∈ G such that hF1 = g1 and hF2 = g2 . Then
g = hF = (h)κF as desired.
To check this global confluence condition one either needs a good understanding of the
action of F on G (as was the case for Vbr [Bri07, Section 5.3]) or one has to reduce it to
local confluence statements.
2.2. Interlude: hedges. In the above example of the symmetric group, the action of F
on S∞ factors through an action of a proper quotient. This amounts to a further relation
being satisfied in addition to the product of clonings relation (CS2). The quotient turns
out to be what Brin [Bri07] called the monoid of hedges. Without going into much detail
we want to explain the action of the hedge monoid on S∞ .
...
...
Figure 5. A forest and the corresponding hedge.
The hedge monoid H is the monoid of monotone surjective maps N → N. Multiplication
is given by composition: f · h = f ◦ h. There is an action of S∞ on H given by the
property that, for g ∈ S∞ and f ∈ H , the cardinality of (g · f )−1 (i) is that of f −1 (g−1 i).
There is an obvious equivariant morphism c : F → H (see Figure 5) given by c(λk ) = ηk
where
m
m ≤ k,
ηk (m) =
m − 1 m > k.
This morphism is surjective but not injective, in fact (see [Bri07, Proposition 4.4]):
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
15
Lemma 2.14. The monoid H has the presentation
hηk , k ∈ N | ηℓ ηk = ηk ηℓ+1 , ℓ ≥ ki.
Observe that the only difference between this and the presentation of F is that the
relation also holds for ℓ = k, rather than only for ℓ > k. It turns out that the action of
F on S∞ defined in Example 2.9 factors through c:
Observation 2.15. The maps ςk defined in (2.2) satisfy ςk ςk = ςk ςk+1 . Thus they define
an action of H on S∞ .
Proof. The verification of (CS2) above extends to the case k = ℓ.
2.3. Filtered cloning systems. Typically one will want to think of Thompson’s group
V not as built from S∞ but rather from the family (Sn )n∈N . We will now describe this
approach. We regard S∞ as the direct limit lim Sn where the maps ιm,n : Sm → Sn are
−→
induced by the inclusions {1, . . . , m} ֒→ {1, . . . , n}.
Let (Gn )n∈N be a family of groups with monomorphisms ιm,n : Gm → Gn for each m ≤ n.
For convenience we will sometimes write G∗ for (Gn )n∈N ; note that in this case the index
set is always N. The maps ιm,n will be written on the right, e.g., (g)ιm,n for g ∈ Gm .
Suppose that ιm,m = id and ιm,n ◦ιn,ℓ = ιm,ℓ for all m ≤ n ≤ ℓ. Then ((Gn )n∈N , (ιm,n )m≤n )
is a directed system of groups with a direct limit G:=lim Gn . Since all the ιm,n are injective,
−→
we may equivalently think of a group G filtered by subgroups Gn .
Consider injective maps κnk : Gn → Gn+1 for k, n ∈ N, k ≤ n. We call such maps a family
of cloning maps for the directed system (Gn )n∈N if for m, k ≤ n they satisfy
m
κk ◦ ιm+1,n+1 if k ≤ m
n
(2.5)
ιm,n ◦ κk =
ιm,n+1
if m < k.
This amounts to setting κnk := ιn,n+1 for k > n and requiring that
ιm,n ◦ κnk = κm
k ◦ ιm+1,n+1 ,
i.e., that the family (κnk )n∈N defines a morphism of directed systems of sets. From that it
is clear that a family of cloning maps induces a family of injective maps κk : G → G by
setting
(g)ιn ◦ κk = (g)κnk ◦ ιn+1
for g ∈ Gn . Here ιn : Gn → G denotes the map given by the universal property of G.
Definition 2.16 (Properly graded). We say that the cloning maps are properly graded
if the following strong confluence condition holds: if g ∈ Gn+1 can be written as (h)κnk =
g = (ḡ)ιn,n+1 then there is an h̄ ∈ Gn−1 with (h̄)κkn−1 = ḡ and (h̄)ιn−1,n = h.
In view of the injectivity of all maps involved this is equivalent to saying that
im κnk ∩ im ιn,n+1 ⊆ im(ιn−1,n ◦ κnk )
(2.6)
(where the converse inclusion is automatic) or to saying that the diagram
Gn−1
κkn−1
ιn−1,n✲
Gn
κnk
❄
❄ ιn,n+1
✲ Gn+1
Gn
is a pullback diagram of sets. A formulation in terms of the direct limit G is that if
(h)κk ∈ Gn for k ≤ n then h ∈ Gn−1 . Note that a filtered cloning system satisfying the
confluence condition of Observation 2.13 is automatically properly graded.
16
S. WITZEL AND M. C. B. ZAREMSKY
Example 2.17. Take Gn = Sn as in Example 2.9. A family of cloning maps ςkn is obtained
by restriction of the maps from Example 2.9:
S
.
ςkn := ςk |Sn+1
n
(2.7)
This family of cloning maps is properly graded: if g ∈ im ιn,n+1 then g fixes n + 1; if
moreover g = (h)ςk then it follows from (2.2) that h fixes n so h ∈ im ιn−1,n .
Now suppose further that we have a family of homomorphisms ρn : Gn → Sn for each
n ∈ N that are compatible with the directed systems, i.e., ρn ((g)ιm,n ) = (ρm (g))ιm,n for
m < n and g ∈ Gm . Let ρ : G → S∞ be the induced homomorphism. We are of course
interested in the case when ρ and the family (κk )k∈N define a cloning system on G. The
corresponding defining formulas are obtained by adding decorations to the formulas from
Section 2.1:
Definition 2.18 (Cloning system). Let ((Gn )n∈N , (ιm,n )m≤n ) be an injective directed
system of groups. Let (ρn )n∈N : Gn → Sn be a homomorphism of directed systems of
groups and let (κnk )k≤n : Gn → Gn+1 be a family of cloning maps. The quadruple
((Gn )n∈N , (ιm,n )m≤n , (ρn )n∈N , (κnk )k≤n )
is called a cloning system if the following hold for all k ≤ n, k < ℓ, and g, h ∈ Gn :
(FCS1) (gh)κnk = (g)κnρ(h)k (h)κnk .
(Cloning a product)
(Product of clonings)
= κnk ◦ κn+1
(FCS2) κnℓ ◦ κn+1
ℓ+1 .
k
n
n
(FCS3) ρn+1 ((g)κk )(i) = (ρn (g))ςk (i) for all i 6= k, k + 1
(Compatibility)
We may also refer to ρ∗ and (κnk )k≤n as forming a cloning system on the directed system
G∗ . The cloning system is properly graded if the cloning maps are properly graded.
Note that condition (FCS3) is phrased more concisely than (CS3), but this is just in light
of Observation 2.11. Again, condition (FCS3) will in practice often be satisfied even when
i = k, k + 1.
Observation 2.19. Let (Gn )n∈N be an injective directed system of groups. A cloning
system on (Gn )n∈N gives rise to a cloning system on G := lim Gn . Conversely a cloning
−→
system on G gives rise to a cloning system on (Gn )n∈N provided (Gn )κnk ⊆ Gn+1 and
ρn (Gn ) ⊆ Sn .
We will usually not distinguish explicitly between a cloning system on G∗ and a cloning
system on lim G∗ that preserves the filtration. In particular, given a cloning system on a
−→
directed system of groups we will implicitly define ρ := lim ρn and κk := lim κnk .
−→
−→
2.4. Thompson groups from cloning systems. Let (G, ρ, (κk )k∈N ) be a cloning sysc(G) for the
tem and let F ⊲⊳ G be the associated BZS product. We now define a group T
cloning system. This is a supergroup of the actual group T (G∗ ) that we construct later
in the case when G arises as a limit of a family (Gn )n (Definition 2.25).
Definition 2.20 (Thompson group of a cloning system). The group of right fractions of
c(G) and is called the large generalized Thompson group of G.
F ⊲⊳ G is denoted by T
c(G, ρ, (κk )k ) and call it the large generalized
If more context is required we denote it T
Thompson group of the cloning system (G, ρ, (κk )k ).
c(G) can be written as
By Observation 2.2 and Theorem 1.1 every element t of of T
−1
t = (E− , g)(E+ , h)
for some E− , E+ ∈ F and g, h ∈ G. If it can also be written
′
′
−1
t = (E− , g )(E+ , h ) then gh−1 = g′ h′ −1 . It therefore makes sense to represent it by just
the triple (E− , gh−1 , E+ ). Of course, this representation is still not unique, for example
(E, 1G , E) represents the identity element for every E ∈ F . We will denote the element
represented by (E− , g, E+ ) by [E− , g, E+ ]. Note that [E− , g, E+ ]−1 = [E+ , g−1 , E− ]. We
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
17
will call (E− (g · F ), g F , E+ F ) an expansion of (E− , g, E+ ), and the latter a reduction of
the former, so any reduction or expansion of a triple (E− , g, E+ ) represents the same
c(G) as (E− , g, E+ ).
element of T
Now assume that G = lim Gn is an injective direct limit of groups (Gn )n∈N and that the
−→
cloning system is a cloning system on (Gn )n∈N . Recall from Section 1.3 that a forest E
is called semisimple if all but its first tree are trivial and in that case its number of feet
is the number of leaves of the first tree.
We collect some facts about semisimple elements of F .
Observation 2.21. Let E, E1 , E2 , F ∈ F .
(1) The number of feet of a non-trivial semisimple element of F is its length plus
one.
(2) Any two semisimple elements of F have a semisimple common right multiple.
More generally, any two elements of rank at most m have a common right multiple
of rank at most m.
(3) If E is semisimple with n feet then EF is semisimple if and only if F has rank at
most n.
More generally, if E is non-trivial of rank m and length n − m then EF has rank
m if and only if F has rank at most n.
(4) If E1 , E2 are semisimple with n feet then E1 E is semisimple if and only if E2 E
is.
Now we upgrade these facts to F ⊲⊳ G. We say that an element (E, g) ∈ F ⊲⊳ G is
semisimple if E is semisimple with n feet (for some n) and g ∈ Gn . In this case we also
say (E, g) has n feet.
Lemma 2.22. Let E, E1 , E2 , F ∈ F and g, h ∈ G.
(1) The number of feet of a semisimple element of F ⊲⊳ G is its length plus one.
(2) Any two semisimple elements of F ⊲⊳ G have a semisimple common right multiple.
(3) If (E, g) is semisimple then (E, g)F = (E(g · F ), g F ) is semisimple if and only if
E(g · F ) is semisimple.
(4) If (E, g) is semisimple with n feet then (E, g)F is semisimple if and only if F has
rank at most n.
(5) If (E1 , g) and (E2 , h) are semisimple with same number of feet then (E1 , g)E is
semisimple if and only if (E2 , g)E is semisimple.
Proof. The first statement is clear by definition. The second statement can be reduced
to the corresponding statement in F because E is a right multiple of (E, g).
In the third statement only the implication from right to left needs justification, namely
that gF ∈ Gn where n is the number of feet of E(g · F ). This is because if g ∈ Gm and
len E = k then g E ∈ Gm+k as can be seen by induction on len E using κk (Gn ) ⊆ Gn+1 .
For (4) note that g ∈ Gn . But ρ(Gn ) ⊆ Sn so having rank at most n is preserved under
the action of Gn , i.e., rk(g · F ) ≤ n ⇔ rk F ≤ n. Thus the statement follows from the one
for F . The last statement is immediate from (4).
Definition 2.23 (Simple). A triple (E− , g, E+ ) (and the element [E− , g, E+ ] represented
by it) is said to be simple if E− and E+ are semisimple, both of them with n feet and
g ∈ Gn . This is the case if it can be written as (E− , g)(E+ , h)−1 with both factors
semisimple with the same number of feet.
c(G) is a subgroup.
Proposition 2.24. The set of simple elements in T
Proof. The proof closely follows [Bri07, Section 7].
18
S. WITZEL AND M. C. B. ZAREMSKY
Consider two simple elements s = [E− , g, E+ ], t = [F− , h, F+ ]. Let
E+ E = F− F
(2.8)
be a semisimple common right multiple of E+ and F− (Observation 2.21 (2)). Then
st = E− gEF −1 hF+−1
= (E− (g · E), g E )(F+ (h−1 · F ), (h−1 )F )−1
= [E− (g · E), g E hh
−1 ·F
(2.9)
, F+ (h−1 · F )].
−1
In the last line we used that (hF )−1 = (h−1 )h·F so that ((h−1 )F )−1 = hh ·F .
We claim that the last expression of (2.9) is simple. Indeed, (E− , g) and E+ are semisimple with the same number of feet and E+ E is semisimple so (E− , g)E = (E− (g · E), gE ) is
F
semisimple by Lemma 2.22 (5). Similar reasoning applies to (F+ (h−1 · F ), h−1 ). Moreover, we can use Corollary 2.4 to compute
len(E− , g) + len E
s simple
=
(2.8)
len E+ + len E = len F− + len F
t simple
=
len(F+ , (h−1 )F ) + len F .
By Lemma 2.22 (1) this shows that the last expression of (2.9) is simple.
Definition 2.25 (Thompson group of a filtered cloning system). The group of simple
c(G) is denoted T (G∗ ) and called the generalized Thompson group of G∗ .
elements in T
c(G), we can include other data from the cloning
If we need to be more precise, as with T
system in the notation as in T (G∗ , ρ∗ , (κ∗k )k ).
Notationally, when we talk about a generalized Thompson group, the asterisk will always
take the position of the index of the family. For instance, the generalized Thompson
group for the family (Gn )n∈N of direct powers in Section 6 will be denoted T (G∗ ); and
the generalized Thompson group for the family of matrix groups (Bn (R))n∈N in Section 7
will be denoted T (B∗ (R)).
c(G) →
Recall from the discussion after Corollary 2.4 that there is a length morphism len : T
Z which takes an element [E, g, F ] to len(E) − len(F ). The group T (G∗ ) lies in the kernel
of that morphism, that is, simple elements have length 0.
Given a simple element [E, g, F ] with E = (Ti )i∈N and F = (Ui )i∈N , since all the Ti and
Ui are trivial for i > 1, we will often write our element as [T1 , g, U1 ] instead. In other
words, we view an element of T (G∗ ) as being a tree with n leaves, followed by an element
of Gn , followed by another tree with n leaves.
c(G) is
Remark 2.26. Constructing T (G∗ ) as the subgroup of simple elements of T
somewhat artificial as can be seen in some of the proofs above. The more natural approach
would be to have each element of F “know” on which level it can be applied. This
amounts to considering the category of forests P that has objects the natural numbers
and morphisms λnk : n → n + 1, 1 ≤ k ≤ n subject to the forest relations (1.2), cf. [Bel04,
Section 7]. Let G be another category that also has objects the natural numbers and
morphisms from n to n that form a group Gn . So while P has only “vertical” arrows,
G has only “horizontal” arrows. One would then want to form the Zappa–Sźep product
P ⊲⊳ G which would be specified by commutative squares of the form γλnk = λnρ(γ)k γ λk
with γ ∈ Gn and γ λk ∈ Gn+1 . Localizing everywhere one would obtain a groupoid of
fractions Q and T (G∗ ) should be just HomQ (1, 1).
The reason that we have not chosen that description is simply that Zappa–Sźep products
for categories are not well-developed to our knowledge, while for monoids all the needed
statements were already available thanks to Brin’s work [Bri05, Bri07].
Artifacts of this approach, which should be overcome by the general approach above,
include the maps ιn,n+1 and the property of being properly graded. Not having to collect
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
19
all the groups Gn in a common group G would also make it possible to construct, for
example, the Thompson groups T and Tbr .
H
H
2.5. Morphisms. Let (G, ρG , (κG
k )k∈N ) and (H, ρ , (κk )k∈N ) be cloning systems. A homomorphism ϕ : G → H is a morphism of cloning systems if
G
(1) (ϕ(g))κH
k = ϕ((g)κk ) for all k ∈ N and g ∈ G, and
H
G
(2) ρ ◦ ϕ = ρ .
Observation 2.27. Let ϕ : G → H be a morphism of cloning systems. There is an
c(ϕ) : T
c(G) → T
c(H). If ϕ is injective or surjective then so is
induced homomorphism T
c(ϕ). In particular, there is always a homomorphism T
c(G) → T
c(Sω ).
T
Proof. We show that a morphism of cloning systems induces a morphism F ⊲⊳ G →
c(ϕ) is defined by
F ⊲⊳ H. The statement then follows from Lemma 1.2. Naturally, T
c(ϕ)(Eg) = Eϕ(g). Well definedness amounts to T
c(ϕ)((g · E)gE ) = (ϕ(g) · E)(ϕ(g)E )
T
which follows from (1) and (2) above by writing E as a product of λk s and inducting on
the length.
The injectivity and surjectivity statements are clear.
Similarly let (Gn )n∈N and (Hn )n∈N be injective direct systems equipped with cloning
systems. A morphism of directed systems of groups ϕ∗ : G∗ → H∗ is a morphism of
cloning systems if
(1) (ϕn (g))κH,n
= ϕn+1 ((g)κG,n
k
k ) for all 1 ≤ k ≤ n and g ∈ Gn , and
G for all n ∈ N.
◦
ϕ
=
ρ
(2) ρH
n
n
n
Observation 2.28. Let ϕ∗ : G∗ → H∗ be a morphism of cloning systems. There is an
induced homomorphism T (ϕ) : T (G∗ ) → T (H∗ ). If ϕ is injective or surjective then so is
T (ϕ). In particular, there is always a homomorphism T (G∗ ) → T (S∗ ), the latter being
Thompson’s group V .
Proof. We have to show that if Eg ∈ F ⊲⊳ G is semisimple with n feet then so is
c(ϕ)(Eg) = Eϕ(g). But this follows since E is semisimple with n feet and g ∈ Gn , so
T
ϕ(g) ∈ Hn .
Functoriality is straightforward:
Observation 2.29. If ϕ : G∗ → H∗ and ψ : H∗ → K∗ are morphisms of cloning systems
c(ψϕ) = T
c(ψ)Tc(ϕ) : T
c(G) → T
c(K). If ϕ and ψ are morphisms of filtered cloning
then T
systems then T (ψϕ) = T (ψ)T (ϕ) : T (G∗ ) → T (K∗ ).
3. Basic properties
Throughout this section let T (G∗ ) be the generalized Thompson group of a cloning system
on an injective directed system of groups (Gn )n∈N and let G = lim Gn . We collect some
−→
properties of T (G∗ ) that follow directly from the construction.
3.1. A short exact sequence.
Observation 3.1. Let T ∈ F be semisimple with n feet. The map g 7→ [T, g, T ] is an
injective homomorphism Gn → T (G∗ ).
c(G) are all injective. The element [T, g, T ] is
Proof. The maps Gn → G → F ⊲⊳ G → T
simple, so the image lies in T (G∗ ). The map is visibly a homomorphism.
In fact, this can be explained more globally. For a semisimple forest T with n feet let
GT denote the subgroup (isomorphic to Gn ) of T (G∗ ) that consists of elements [T, g, T ].
The cloning map κk induces an embedding GT ֒→ GU where U is obtained from T by
20
S. WITZEL AND M. C. B. ZAREMSKY
adding a split to the kth foot (so U = T λk ). Finite binary trees form a directed set and
the condition (FCS2) (product of clonings) ensures that that the groups (GT )T form a
directed system of groups.
Lemma 3.2. Consider a cloning system satisfies condition (FCS3) even for i = k, k + 1
(this is the case in particular if ρ = 0). There is a directed subsystem (KT )T of (GT )T
and a short exact sequence
1 → lim KT → T (G∗ ) → W → 1
−→
T
where the quotient morphism is the morphism T (ρ∗ ) from Observation 2.28 and W is its
image.
Note that W contains Thompson’s group F .
Proof. For each T , say with n feet, let KT be the kernel of ρn : GT → Sn . The assumption
on the cloning system implies that if ρ(g) = 1 then ρ((g)κk ) = 1, showing that (KT )T is
indeed a subsystem of (GT )T . It remains to see that the direct limit is isomorphic to the
kernel of T (ρ∗ ). This is clear once one realizes that it consists of all elements that can
be written in the form [T, g, T ], for some T and g ∈ KT .
In what follows we will concentrate on the case where ρ = 0 is the trivial morphism
ρ(g) = 1, so KT = GT for all T . Examples are F and Fbr but not V and Vbr .
Observation 3.3. Suppose ρ = 0. Then T (G∗ ) = K (G∗ ) ⋊ F .
Proof. Since each ρn = 0, we have W = F , which is T ({1}). Then the splitting map
F → T (G∗ ) is T (ι∗ ) where ι∗ : {1} → G∗ is the trivial homomorphism.
Remark 3.4. Bartholdi, Cornulier, and Kochloukova [BdCK15] studied finiteness properties of wreath products. Observation 3.3 shows how this relates to our groups. A wreath
product is built by taking a direct product of copies of a group H, indexed by a set X, and
combining this with another group G acting on X. The generalized Thompson groups in
Observation 3.3 can be viewed as the result of taking a direct limit (instead of product)
of groups from a family (GT )T , indexed by a set of trees T on which there is a partial
(instead of full) action of F , and combining these data into a group T (G∗ ).
The question of whether F is amenable or not is probably the most famous question about
Thompson’s groups. The following observation does not purport to be deep, but it seems
worth recording nonetheless.
Observation 3.5 (Amenability). Suppose ρ = 0. Then T (G∗ ) is amenable if and only
if F and every Gn is amenable.
Proof. We have seen that K (G∗ ) is a direct limit of copies of Gn . Since amenability is
preserved under taking subgroups and direct limits, this tells us that K (G∗ ) is amenable
if and only if every Gn is. Then since T (G∗ ) = K (G∗ ) ⋊ F , the conclusion follows since
amenability is also closed under group extensions.
Observation 3.6 (Free group-free). Suppose ρ = 0. If none of the Gn contains a nonabelian free group then neither does T (G∗ ).
Proof. Suppose H ≤ T (G∗ ) is free. If H ∩ K (G∗ ) = {1} then H embeds into F , and so
H must be cyclic, since F does not contain a non-abelian free group. Now suppose there
is some 1 6= x ∈ H ∩ K (G∗ ). For any y ∈ H, the conjugate xy is in H ∩ K (G∗ ). Since
K (G∗ ) is a direct limit of copies of the Gn , it does not contain a non-abelian free group
by assumption, and so hx, xy i is abelian. But y ∈ H was arbitrary, so H must already be
abelian.
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
21
The next result does not require ρ = 0. It fits into the context of this section but to prove
it we need some of the tools of Section 4.3.
Lemma 3.7 (Torsion-free). Assume that the cloning system is properly graded. If all the
Gn are torsion-free then so is T (G∗ ).
3.2. Truncation. For g ∈ Gn and k ≤ n we have the equation gλk = (g · λk )gλk in
c(G) this implies
F ⊲⊳ G where gλk ∈ Gn+1 . In T
g = (g · λk )g λk λ−1
k .
(3.1)
This elementary observation has an interesting consequence. Let N ∈ N be arbitrary
and define a directed system of groups (G′n )n∈N by G′n := {1} for n ≤ N and G′n := Gn
for n > N . Define a cloning system on G′∗ by letting (κ′ )nk : G′n → G′n+1 be the trivial
homomorphism when n ≤ N , and (κ′ )nk = κnk and ρ′n = ρn when n > N . We call G′∗ the
truncation of G∗ at N and ((ρ′n )n , ((κ′ )nk )k≤n ) the truncation of ((ρn )n , (κnk )k≤n ) at N .
Proposition 3.8 (Truncation isomorphism). Let G′∗ be the truncation of G∗ at N . The
morphism T (G′∗ ) → T (G∗ ) induced by the obvious homomorphism G′∗ → G∗ is an
isomorphism.
Proof. The morphism G′∗ → G∗ is injective hence so is T (G′∗ ) → T (G∗ ). To show that
it is surjective let [T, g, U ] ∈ T (G∗ ) be such that T and U have n leaves. If n > N there
is nothing to show. Otherwise use (3.1) to write
[T, g, U ] = [T (g · λk ), gλk , U λk ]
for some k ≤ n. The trees in the right hand side expression have n + 1 leaves. Proceeding
inductively, we obtain an element whose trees have N + 1 leaves and therefore the element
is in T (G′∗ ).
This proposition is in line with treating T (G∗ ) as a sort of limit of G∗ since it does not
depend on an initial segment of data.
4. Spaces for generalized Thompson groups
The goal of this section is to produce for each generalized Thompson group T (G∗ ) a
space on which it acts. The space will be contractible and have stabilizers isomorphic to
the groups Gn , assuming the cloning system on G∗ is properly graded. The ideas used in
the construction were used before in [Ste92, Bro92, Far03, Bro06, FMWZ13, BFM+ 14].
Throughout let G∗ be an injective directed system of groups equipped with a cloning
system and let G = lim G∗ .
−→
As a starting point we note that Corollary 1.6, Observation 2.2 and Corollary 2.4 imply
c(G)/G is a join-semilattice with conditional meets, under the relation xG ≤ yG if
that T
−1
x y ∈ F ⊲⊳ G. Later on it will be convenient to have a symbol for the quotient relation
so we let x ∼G y if x−1 y ∈ G.
4.1. Semisimple group elements. We generalize some of the notions that were introduced in Sections 1.3 and 2.4. We say that an arbitrary (not necessarily semisimple)
element E of F has n feet if it has rank m and length n − m. Visually this means that
the last leaf that is not a root is numbered n. An element (E, g) of F ⊲⊳ G has n feet if E
c(G) semisimple
has at most n feet and g ∈ Gn . Finally, we call an element [E, g, F ] of T
if (E, g) is semisimple with n feet and F has at most n feet (note F need not be semisimple). This is consistent with the previous definition of “semisimple”: If an element of the
c(G) is semisimple in this sense, and is an element of the monoid F ⊲⊳ G, then it
group T
e1 denote the set of all semisimple elements
must be semisimple in the monoid. We let P
c(G).
of T
22
S. WITZEL AND M. C. B. ZAREMSKY
Lemma 4.1. If [E1 , g1 , F1 ] is simple and [E2 , g2 , F2 ] is semisimple then [E1 , g, F1 ][E2 , g, F2 ]
is semisimple. As a consequence, T (G∗ ) acts on Pe1 .
Proof. This is shown analogously to Proposition 2.24.
If [E, g, F ] is semisimple we say that it has len([E, g, F ]) + 1 = len(E) − len(F ) + 1 feet,
which is well defined by Corollary 2.4. This can be visualized as the number of roots
e1,n denote the set of all
of F that can be “reached” from the first root of E. We let P
semisimple elements with at most n feet. We define P1,n to be the quotient Pe1,n /∼G and
c(G)/G. We
call the passage from Pe1,n to P1,n dangling. Note that P1,n is a subposet of T
also denote Pe1 /∼G by P1 .
For context, the term “dangling” comes from the case when G∗ is the system of braid
groups B∗ , and the elements of P1,n can be pictured as “dangling braided strand diagrams”
[BFM+ 14], originating on one strand and ending on n strands.
The next lemma is the reason for having introduced the notion of a cloning system being
properly graded.
Lemma 4.2. Assume that the cloning system is properly graded. If x, y ∈ Pe1,n are
semisimple then x ∼G y if and only if x−1 y ∈ Gn .
Proof. What needs to be shown is that if x−1 y ∈ G then x−1 y ∈ Gn . Write x =
[E1 , g −1 , F1 ] and y = [E2 , h−1 , F2 ]. Let E = E1 E1′ = E2 E2′ be a common right multi′
′
ple so that x−1 y = [F1 (g · E1′ ), gE1 (hE2 )−1 , F2 (h · E2′ )] =: [A, b, C]. For this to equal some
d ∈ G it is necessary that Ab = dC in F ⊲⊳ G, that is, A = d · C and b = dC .
Say that E has length m. Then we compute that len(A) = len(C) ≥ m − n + 1. Since
the cloning system is properly graded, the fact that b = dC implies that d has to lie in
Gm+1−len(C) ⊆ Gn .
4.2. Poset structure. Consider the geometric realization |P1 |. This is the simplicial
complex with a k-simplex for each chain x0 ≤ · · · ≤ xk of elements of P1 , and face
relation given by subchains.
Lemma 4.3. The poset P1 is a join-semilattice with conditional meets, in particular |P1 |
is contractible.
c(G)/G is a join-semilattice with conditional meets so it
Proof. We already know that T
suffices to show that P1 is closed under taking suprema and infima. In other words, it
suffices to show that least common right multiples of semisimple elements are semisimple
and that left factors of semisimple elements are semisimple. The first is similar to the
proof of Proposition 2.24 and the second is easy.
In |P1 | every vertex is contained in a simplex of arbitrarily large dimension, which makes
it too big for practical purposes. It has proven helpful to consider a subspace called the
Stein–Farley complex, which we introduce next.
e1 was defined by declaring that
4.3. The Stein–Farley complex. The preorder on P
x ≤ y if y = x(E, g) for some (E, g) ∈ F ⊲⊳ G. The basic idea in constructing the
Stein–Farley complex is to regard this relation as a transitive hull of a finer relation
and to use this finer relation in constructing the space. It is defined by declaring x y if
y = x(E, g) for some (E, g) ∈ F ⊲⊳ G with the additional assumption that E is elementary.
An elementary forest is one in which every tree has at most two leaves. That is, a forest
is elementary if it can be written as λk1 · · · λkr with ki+1 > ki + 1 for i < r. Note that if
x ∈ Pe1,n , in order for x(E, g) to be in Pe1 as well, it is necessary that E has rank at most
n and that g ∈ Gn+len(E) . Note also that if E is elementary then so is g · E for any g ∈ G
because the action of G (via ρ : G → Sω ) just permutes the trees of E.
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
23
As a consequence is invariant under dangling and we also write for the relation
induced on P1 . Note that is not transitive, but it is true that if x z and x ≤ y ≤ z
then x y z. Given a simplex x0 ≤ · · · ≤ xk in |P1 |, call the simplex elementary if
x0 xk . The property of being elementary is preserved under passing to subchains, so
the elementary simplices form a subcomplex.
Definition 4.4 (Stein–Farley complex). The subcomplex of elementary simplices of |P1 |
is denoted by X (G∗ ) and called the Stein–Farley complex of T (G∗ ).
The Stein–Farley complex has the structure of a cubical complex, which we now describe.
The key point is:
Observation 4.5. If E is elementary then the set of right factors of E forms a boolean
lattice under .
For x y in P1 we consider the closed interval [x, y] := {z ∈ P1 | x ≤ z ≤ y} as well as
the open and half open intervals (x, y), [x, y) and (x, y] that are defined analogously. As a
consequence of Observation 4.5 we obtain that the interval [x, y] := {z ∈ P1 | x ≤ z ≤ y}
is a boolean lattice and so |[x, y]| has the structure of a cube. The intersection of two such
cubes |[x, y]| and |[z, w]| is empty if y and w do not have a common lower bound and is
|[sup(x, z), inf (y, w)]| (which may be empty if the supremum is larger than the infimum)
otherwise. In particular the intersection of cubes is either empty or is again a cube. Hence
X (G∗ ) is a cubical complex in the sense of Definition 7.32 of [BH99].
Observation 4.6. For any vertex x in X (G∗ ), there are only finitely many vertices y in
X (G∗ ) with x y.
Proof. If x̃ ∈ Pe1 is a vertex representative (modulo dangling) for x, it is clear using
dangling that every vertex y with x y has a representative ỹ with ỹ = x̃(E, 1) for some
some elementary forest E. In order for ỹ to be semisimple, E can have rank at most
len(x̃) − 1, and there are only finitely many elementary forests of a given rank, so the
result follows.
The next step is to show that X (G∗ ) is itself contractible. The argument is similar to
that given in Section 4 of [Bro92]. We follow the exposition in [BFM+ 14].
Lemma 4.7. For x < y with x 6≺ y, |(x, y)| is contractible.
Proof. For any z ∈ (x, y] let z0 be the unique largest element of [x, z] such that x z0 .
By hypothesis z0 ∈ [x, y), and by the definition of it is clear that z0 ∈ (x, y], so in fact
z0 ∈ (x, y). Also, z0 ≤ y0 for any z ∈ (x, y). The inequalities z ≥ z0 ≤ y0 then imply that
|(x, y)| is contractible, by Section 1.5 of [Qui78].
Proposition 4.8. X (G∗ ) is contractible.
Proof. We know that |P1 | is contractible by Lemma 4.3. We can build up from X (G∗ ) to
|P1 | by attaching new subcomplexes, and we claim that this never changes the homotopy
type, so X (G∗ ) is contractible. Given a closed interval [x, y], define r([x, y]) := len(y) −
len(x). As a remark, if x y then r([x, y]) is the dimension of the cube given by [x, y].
We attach the contractible subcomplexes |[x, y]| for x 6 y to X (G∗ ) in increasing order
of r-value. When we attach |[x, y]| then, we attach it along |[x, y)| ∪ |(x, y]|. But this is
the suspension of |(x, y)|, and so is contractible by the previous lemma. We conclude that
attaching |[x, y]| does not change the homotopy type, and since |P1 | is contractible, so is
X (G∗ ).
Lemma 4.9 (Stabilizers). Assume that the cloning system is properly graded. The stabilizer in T (G∗ ) of a vertex in X (G∗ ) with n feet is isomorphic to Gn . The stabilizer in
T (G∗ ) of an arbitrary cell is isomorphic to a finite index subgroup of some Gn .
24
S. WITZEL AND M. C. B. ZAREMSKY
Proof. First consider the stabilizer of a vertex x with n feet. We claim that StabT (G∗ ) (x) ∼
=
e
Gn . Choose x̃ ∈ P1 representing x and let g ∈ StabT (G∗ ) (x). By the definition of dangling, and by Lemma 4.2, there is a (unique) h ∈ Gn such that gx̃ = x̃h. Then the map
g 7→ h = x̃−1 gx̃ is a group isomorphism.
Now let σ = |[x, y]|, x y be a an arbitrary cube. Since the action of T (G∗ ) preserves
the number of feet, the stabilizer Gσ of σ fixes x and y. Hence Gσ is contained in Gx
and contains the kernel of the map Gx → Symm({w | x w y}), the image of which is
finite by Observation 4.6.
We close this section by providing the proof of Lemma 3.7, left out in the last section,
which says that T (G∗ ) is torsion-free as soon as all the Gn are.
Proof of Lemma 3.7. The vertices in X (G∗ ) coincide with the vertices of |P1 |, and, as
we just proved, any vertex has some Gn as a stabilizer. Hence it suffices to prove that
if g ∈ T (G∗ ) has finite order then it fixes an element of the directed poset P1 . By
Lemma 4.3, P1 is a join-semilattice, so any finite collection of elements has a unique least
upper bound. But then if g has finite order, for any x ∈ P1 the unique least upper bound
of the finite set hgi.x is necessarily fixed by g.
5. Finiteness properties
One of our main motivations for defining the functor T (−) is to study how it behaves
with respect to finiteness properties. Recall that a group G if said to be of type Fn if
there is a K(G, 1) whose n-skeleton is compact. Most of the known Thompson’s groups
are of type F∞ , that is, of type Fn for all n. To efficiently speak about groups that are
not of type F∞ recall that the finiteness length of G, denoted φ(G), is the supremum over
all n ∈ N such that G is of type Fn .
We will see below that proofs of the finiteness properties of T (G∗ ) depend on the finiteness
properties of the individual groups Gn as well as on the asymptotic connectivity of certain
descending links, which is infinite in many cases. Since finite initial intervals of G∗ can
always be ignored by Proposition 3.8 we ask:
Question 5.1. For which directed systems of groups G∗ equipped with properly graded
cloning systems do we have
φ(T (G∗ )) = lim inf φ(G∗ )?
Note that for any directed system of groups G∗ one can take all ρk to be trivial and all κnk
to be ιn,n+1 . In this case T (G∗ ) = (limn Gn ) × F , which would seem to give a negative
answer to Question 5.1. However, in order to be properly graded in this example we
would need im ιn,n+1 ⊆ im ιn−1,n+1 , and this implies that the ιn,n+1 are all isomorphisms.
Thus, in fact this does provide a positive answer to the question.
5.1. Morse theory. One of the main tools to study connectivity properties of spaces,
and thus to study finiteness properties of groups, is combinatorial Morse theory. We
collect here the main ingredients that will be needed later on.
Let X be a Euclidean cell complex. A map h : X (0) → N0 is called a Morse function if the
maximum of h over the vertices of a cell of X is attained in a unique vertex. We typically
think of h as assigning a height to each vertex. If h is a Morse function and r ∈ R, the
sublevel set Xr = X ≤r consists of all cells of X whose vertices have height at most r. For
a vertex x ∈ X (0) of height r, the descending link lk↓(x) of x is the subcomplex of lk(x)
spanned by all vertices of strictly lower height. The main observation that makes Morse
theory work is that keeping track of the connectivity of descending links allows one to
deduce global (relative) connectivity statements:
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
25
Lemma 5.2 (Morse Lemma). Let X be a Euclidean cell complex and let h : X (0) → N0
be a Morse function on X. Let s, t ∈ R ∪ {∞} with s < t. If lk↓(x) is (k − 1)-connected
for every vertex in Xt \ Xs then the pair (Xt , Xs ) is k-connected.
The connection between connectivity of spaces and finiteness properties of groups is most
directly made using Brown’s criterion. A Morse function on X gives rise to a filtration
(Xr )r∈N0 by subcomplexes. We say that the filtration is essentially k-connected if for
every i ∈ N0 there exists a j ≥ i such that πℓ (Xi → Xj ) is trivial for all ℓ ≤ k.
Now assume that a group G acts on X. If h is G-invariant then so is the filtration (Xr )r .
We say that the filtration is cocompact if the quotient G\Xr is compact for all r. This is
the setup for Brown’s criterion, see [Bro87, Theorems 2.2, 3.2].
Theorem 5.3 (Brown’s criterion). Let n ∈ N and assume a group G acts on an (n − 1)connected CW complex X. Assume that the stabilizer of every p-cell of X is of type Fn−p .
Let (Xr )r∈N0 be a G-cocompact filtration of X. Then G is of type Fn if and only if (Xr )r
is essentially (n − 1)-connected.
Putting both statements together we obtain the version that we will mostly use.
Corollary 5.4. Let G act on a contractible Euclidean cell complex X and let h : X (0) →
N0 be a G-invariant Morse function. Assume that the stabilizer of every p-cell of X is
of type Fn−p and that the sublevel sets Xr are cocompact. If there is an s ∈ R such that
lk↓(x) is (n − 1)-connected for all vertices x ∈ X (0) \ Xs then G is of type Fn .
If G∗ is a system of groups equipped with a properly graded cloning system then T (G∗ )
acts on the Stein–Farley complex X (G∗ ), which is contractible (Proposition 4.8) with
stabilizers from G∗ (Lemma 4.9). Our next goal is to define an invariant, cocompact
Morse function and to describe the descending links.
5.2. The Morse function. Recall that the vertices of X (G∗ ) are classes [E, g, F ] of
semisimple elements modulo dangling. The height function we will be using assigns to
such a vertex its number of feet (see Section 4.1). That is, X (G∗ )n = |P1,n |∩X (G∗ ). This
c(G) → Z
height function is T (G∗ )-invariant because it is induced by the morphism len : T
and every element of T (G∗ ) has length 0.
Lemma 5.5 (Cocompactness). The action of T (G∗ ) is transitive on vertices of X (G∗ )
with a fixed number of feet. Consequently the action of T (G∗ ) on X (G∗ )n is cocompact
for every n.
Proof. Let x̃ = [E− , g, E+ ] and ỹ = [F− , h, F+ ] be semisimple with n feet. We know
that x̃ỹ −1 takes ỹ to x̃, so it suffices to show that x̃ỹ −1 is simple. Note that E+ and
F+ have rank at most n. By Observation 2.21 (2) they admit a common right multiple
E+ E = F+ F of rank at most n. Let the length of this multiple be m, so it has at most
m + n feet. Then
x̃ỹ −1 = [E− (g · E), g E (hF )−1 , F− (h · F )]
and both E− (g · E) and F− (h · F ) are semisimple by Observation 2.21 (3). They have
m + n feet and both gE and hF lie in Gn+m . Thus x̃ỹ −1 is simple.
The second statement now follows from Observation 4.6.
5.3. Descending links. Let x be a vertex in X (G∗ ), with n feet. We want to describe
the descending link of x. A vertex y is in the link of x if either x y or y x. In
the first case y is ascending so the descending link is spanned by vertices y with y x.
These are by definition of the form x(E, g)−1 for E an elementary forest and g ∈ Gn . In
particular, for a fixed n, the descending links of any vertices of height n look the same,
and are all isomorphic to the simplicial complex of products gE −1 where g ∈ Gn and E
is an elementary forest with at most n feet, modulo the relation ∼G .
26
S. WITZEL AND M. C. B. ZAREMSKY
It is helpful to describe this complex somewhat more explicitly. In doing so we slightly
shift notation by making use of the fact that elementary forests can be parametrized by
subgraphs of linear graphs.
Let Ln be the graph with n vertices, labeled 1 through n, and a single edge connecting
i to i + 1, for each 1 ≤ i ≤ n − 1. This is the linear graph with n vertices. Denote the
edge from i to i + 1 by ei . We will exclusively consider spanning subgraphs of Ln , that is,
subgraphs whose vertex set is {1, . . . , n}. We call the spanning subgraph without edges
trivial. A matching on a graph is a spanning subgraph in which no two edges share a
vertex. For an elementary forest E with at most n feet, define Γ(E) to be the spanning
subgraph of Ln that has an edge from i to i + 1 if and only if the ith and (i + 1)st leaves
of E are leaves of a common caret. Note that this is a matching. Conversely, given a
matching Γ of Ln , there is an elementary forest E(Γ) = λik · · · λi1 where Γ has edges
ei1 , . . . , eik . Both operations are inverse to each other so we conclude:
Observation 5.6. There is a one-to-one correspondence between matchings of Ln and
elementary forests with at most n feet.
In particular, if Γ is a matching with m edges and n vertices we obtain a cloning map
κΓ : Gn−m → Gn which is just the cloning map of E(Γ) as defined before Observation 2.13.
We also get an action of Gn−m on Γ which is given by the action of ρ(Gn−m ) on connected
components. For future reference we also note:
Observation 5.7. There is a one-to-one correspondence between spanning subgraphs of
Ln and hedges with at most n feet.
Now define a simplicial complex Ln (G∗ ) as follows. A simplex in Ln (G∗ ) is represented
by a pair (g, Γ), where g ∈ Gn and Γ is a non-trivial matching of Ln . Two such pairs
(g1 , Γ1 ), (g2 , Γ2 ) are equivalent (under dangling) if the following conditions hold:
(1) Γ1 and Γ2 both have m edges for some 1 ≤ m ≤ n/2,
(2) g2−1 g1 lies in the image of κΓ1 , and
(3) Γ2 = (g2−1 g1 )κ−1
Γ1 · Γ1 .
We make Ln (G∗ ) into a simplicial complex with face relation given by passing to subgraphs of the second term in the pair. Denote the equivalence class of (g, Γ) under dangling
by [g, Γ]. In summary,
Ln (G∗ ) has simplex set {[g, Γ] | Γ is a matching of Ln and g ∈ Gn }.
Observation 5.8. If x has n feet, the correspondence (g, Γ) 7→ xgE(Γ)−1 induces an
isomorphism Ln (G∗ ) → lk↓(x).
In particular, the Ln (G∗ ) are indeed simplicial complexes as claimed, since X (G∗ ) is a
cubical complex.
We now have all the pieces together to apply Brown’s criterion to our setting.
Proposition 5.9. Let G∗ be equipped with a properly graded cloning system. If Gk is
eventually of type Fn and Lk (G∗ ) is eventually (n−1)-connected then T (G∗ ) is of type Fn .
Proof. Suppose first that all Gk are of type Fn . Let X = X (G∗ ), which is contractible
by Proposition 4.8. Our Morse function “number of feet” has cocompact sublevel sets by
Lemma 5.5. The stabilizer of any cell is a finite index subgroup of some Gk by Lemma 4.9.
Since finiteness properties are inherited by finite index subgroups, our assumption implies
that all stabilizers are of type Fn . By the second assumption there is an s such that Lk (G∗ )
is (n − 1)-connected for k > s, which by Observation 5.8 means that descending links are
(n−1)-connected from s on. Applying Corollary 5.4 we conclude that T (G∗ ) is of type Fn .
If the Gk are of type Fn only from t on, we use Proposition 3.8 to replace T (G∗ ) by the
isomorphic group T (G′∗ ) where G′k = Gk for k ≥ t and Gk = {1} for k < t. In particular,
all of the G′k are of type Fn .
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
27
Of course X (G′∗ ) is not isomorphic to X (G∗ ) and neither are the Lm (G′∗ ) isomorphic
to the Lm (G∗ ). However, the k-skeleton of Lm (G′∗ ) is isomorphic to the k-skeleton of
Lm (G∗ ) once m > k + t. Since (n − 1)-connectivity only depends on the n-skeleton, if
the Lm (G∗ ) are eventually (n − 1)-connected then so are the Lm (G′∗ ).
For a negative counterpart to this statement, this is, to show that T (G∗ ) is not of type Fn ,
we would need stabilizers with good finiteness properties and a filtration that is not
essentially (n − 1)-connected – at least as long as we are trying to apply Brown’s criterion.
Hence if we have groups Gn whose finiteness lengths do not have a limit inferior of ∞,
we would need an action on a different space to show that T (G∗ ) answers Question 5.1
affirmatively.
Returning to the positive statement, we remark that inspecting the homotopy type of
Ln (G∗ ) does not seem possible uniformly. Instead, in what follows we will focus on
examples and in particular find some instances of Ln (G∗ ) being highly connected. In the
case where the Gn are braid groups, these complexes were modeled by arc complexes in
[BFM+ 14]. In Section 7 below, where the Gn are matrix groups, we will directly work
with the combinatorial description. General tools that have turned out to be helpful will
be collected in Sections 5.4 and 5.5.
We can make one positive statement about finiteness properties without knowing much
at all about G∗ . Before stating this as a lemma, we need to define the matching complex
of Ln . This is a simplicial complex, denoted M(Ln ), whose simplices are matchings on
Ln and with face relation given by passing to subgraphs. It is well-known and not hard
to see that M(Ln ) is (⌊ n−2
3 ⌋ − 1)-connected. A precise description of the homotopy type
is given in [Koz08, Proposition 11.16] where M(Ln ) arises as the independence complex
Ind(Ln−1 ).
Lemma 5.10 (Finite generation). Let G∗ be a family of groups equipped with a properly
graded cloning system, with cloning maps κnk . Suppose that for n sufficiently large, all
Gn are finitely generated and also are generated by the images of the cloning maps with
codomain Gn . Then T (G∗ ) is finitely generated.
Proof. By the above discussion, we need only show that the Ln (G∗ ) are connected, for
large enough n. Suppose n is large enough that: (a) Gn is generated by images of
cloning maps, and (b) n ≥ 5 so M(Ln ) is connected. We will show that every vertex
can be connected by an edge path to the vertex [1, J1 ], where Ji denotes the spanning
graph whose only edge connects the ith vertex to the (i + 1)st. So let [g, Γ] be a vertex
of Ln (G∗ ) and write g = s1 · · · sr , where the si are generators coming from images of
cloning maps si ∈ im(κki ) for some ki . Since M(Ln ) is connected, there is a path in
Ln (G∗ ) from [s1 · · · sr , Γ] to [s1 · · · sr , Jkr ] = [s1 · · · sr−1 , ((sr )κ−1
kr ) · Jkr ]. Repeating this r
times, we connect to [1, Jk ] for some k, and then to [1, J1 ].
5.4. Proving high connectivity. As we have seen, Morse theory is a tool that allows
one to show that a pair (X, X0 ) is highly connected. We will eventually want to inductively
apply this to the situation where X = Ln (G∗ ) and X0 = Ln−k (G∗ ) for some k ∈ N. This
is insufficient to conclude that the connectivity tends to infinity though, because we would
be trying to get X to be more highly connected than X0 . The following lemma expresses
the degree of insufficiency. The lemma is straightforward to prove but can be seen as a
roadmap for the argument that follows.
Lemma 5.11. Let (X, X0 ) be a k-connected CW-pair. Assume that X0 is (k − 1)connected. Then X is k-connected if and only if πk (X0 → X) is trivial.
Proof. Consider the part of the homotopy long exact sequence associated to (X, X0 ):
ιj
πj+1 (X, X0 ) → πj (X0 ) → πj (X) → πj (X, X0 ).
28
S. WITZEL AND M. C. B. ZAREMSKY
For j < k the map ιj is an isomorphism and πj (X0 ) trivial. For j = k it is an epimorphism,
so indeed πk (X) is trivial if and only if ιk is.
In our applications we will know X0 to be (k − 1)-connected by induction and (X, X0 )
will be seen to be k-connected using Morse theory. To show that πk (X0 → X) is trivial
we will use a relative variant of the Hatcher flow for arc complexes that was shown to
us by Andrew Putman (Proposition 5.13 below). Before we can prove it we need some
technical preliminaries.
A combinatorial k-sphere (respectively k-disk) is a simplicial complex that can be subdivided to be isomorphic to a subdivision of the boundary of a (k + 1)-simplex (respectively to a subdivision of a k-simplex). An m-dimensional combinatorial manifold is
an m-dimensional simplicial complex in which the link of every simplex σ of dimension
k is a combinatorial (m − k − 1)-sphere. In an m-dimensional combinatorial manifold
with boundary the link of a k-simplex σ is allowed to be homeomorphic to a combinatorial (m − k − 1)-disk; its boundary consists of all the simplices whose link is indeed a
disk.
A simplicial map is called simplexwise injective if its restriction to any simplex is injective.
The following is Lemma 3.8 of [BFM+ 14], cf. also the proof of Proposition 5.2 in [Put].
Lemma 5.12. Let Y be a k-dimensional combinatorial manifold. Let X be a simplicial
complex and assume that the link of every d-simplex in X is (k − 2d − 2)-connected for
d ≥ 0. Let ψ : Y → X be a simplicial map whose restriction to ∂Y is simplexwise injective.
Upon changing the simplicial structure of Y , ψ is homotopic relative ∂Y to a simplexwise
injective map.
In practice Y will be a sphere, so the lemma allows us to restrict attention to simplexwise
injective combinatorial maps when collapsing spheres.
For the proposition, we need one more technical definition. Let X be a simplicial complex
and w a vertex. We say that X is conical at w if for any simplex σ, as soon as every
vertex of σ lies in the closed star st(w) then so does σ (that is, the star of w is the cone
over the link of w). In particular, if X is a flag complex then it is conical at every vertex.
Proposition 5.13. Let X0 ⊆ X1 ⊆ X be simplicial complexes. Assume that (X, X0 ) is
k-connected, that X0 is (k−1)-connected and that the link of every d-simplex is (k−2d−2)connected for d ≥ 0. Further assume the following “exchange condition”:
(EXC) There is a vertex w ∈ X at which X is conical, such that for every vertex v ∈ X0
that is not in stw there is a vertex v ′ ∈ stX1 w such that lkX1 v ⊆ lkX1 v ′ and lkX1 v
is (k − 1)-connected.
Then X is k-connected.
Proof. Let ι : X0 → X denote the inclusion. In view of Lemma 5.11, all that needs to be
shown is that if ϕ : S k → X0 is a map from a k-sphere then ϕ̄ := ι ◦ ϕ is homotopically
trivial.
By simplicial approximation [Spa66, Theorem 3.4.8] we may assume ϕ (and thus ϕ̄) to
be a simplicial map Y → X0 and by our assumptions and Lemma 5.12 we may assume it
to be simplexwise injective. Our goal is to homotope ϕ̄ to a map to stw. Once we have
achieved that, we are done since stw is contractible.
The simplicial sphere Y contains finitely many vertices x whose image v = ϕ̄(x) does not
lie in stw. Pick one and define ϕ̄′ : Y → X to be the map that coincides with ϕ̄ outside
the open star of x and takes x to the vertex v ′ from the statement. We claim that ϕ̄ is
homotopic to ϕ̄′ . Inductively replacing vertices then finishes the proof, since X is conical
at w.
It remains to show that ϕ̄|stx and ϕ̄′ |stx are homotopic relative to lk x. Note that ϕ̄(lk x) ⊆
lk v by simplexwise injectivity. Furthermore the complex spanned by v, v ′ and lk v is the
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
29
suspension Σ(lk v) of lk v (unless v and v ′ are adjacent in which case there is nothing to
show). So both ϕ̄|stx and ϕ̄′ |stx are maps (D k , S k−1 ) ∼
= (stx, lk x) → (Σ(lk v), lk v). But
lk v is (k − 1)-connected by assumption so (Σ(lk v), lk v) is k-connected and both maps
are homotopic.
5.5. Proving negative finiteness properties. We have already seen that if the G∗ are
not eventually of type Fn , then Brown’s criterion applied to the Stein–Farley complex
cannot be used to show that T (G∗ ) is not of type Fn . In Section 8.2, when the Gn are
matrix groups, we will instead use a different action, together with the following result.
It is formulated in terms of the homological finiteness properties FPn . The relationship
is explained for example in [Geo08, Chapter 8], but we mostly just need to know the fact
that a group of type Fn is also of type FPn . Note that for Λ = Γ the following theorem
is essentially one half of Brown’s criterion.
Theorem 5.14. Let Λ be a group and let Γ be a subgroup. Let Y be a CW complex on
which Λ acts. Assume that Y is (n − 1)-acyclic and that the stabilizer of every p-cell in
Y (in Λ as well as in Γ) is of type FPn−p . Let Z be a Γ-cocompact subspace of Y . Let
(Yα )α∈I be a Λ-cocompact filtration of Y . Assume that there is no α with Z ⊆ Yα such
that the map H̃n−1 (Z ֒→ Yα ) is trivial. Then no group ∆ through which the inclusion
Γ ֒→ Λ factors is of type FPn .
The application is similar in spirit to that of [KM97], where a morphism Γ → Λ is
constructed that cannot factor through a finitely presented group. The proof should be
compared to [Bro87, Theorem 2.2].
Proof. For n = 1 suppose that Γ is contained in a finitely generated subgroup hSi of Λ.
Let K be a compact subspace such that Γ.K = Z. Since Y is connected, we can add
finitely many edges to K and take Z to be its Γ-orbit, so without loss of generality K is
connected.
For every s ∈ S we may pick an edge path ps that connects K to s.K. Let
S
P := {ps | s ∈ S}. Now any two points in Z can be connected in Γ.(K ∪ P ). In other
words, the map H̃0 (Z → Γ.(K ∪ P )) is trivial. But K ∪ P and thus Γ.(K ∪ P ) is contained
in some Yα , contradicting the assumption.
From now on Q
we assume that n >Q1. Our goal is to find an index set J such that the
map Hn−1 (Γ, J ZΓ) → Hn−1 (Λ, J ZΛ) is non-trivial. The result then follows from
the Bieri–Eckmann
criterion [BE74, Proposition 1.2], because if this map factors through
Q
Hn−1 (∆, J Z∆) then the latter module cannot be zero. Note that Z is contained in a
subfiltration of (Yα )α so we may assume without loss of generality that Z is contained in
all Yα , α ∈ I.
Let J be a cofinal set in I (for instance all of I) and for α ∈ J let cα ∈ Hn−1 (Z) be
such that the image in Hn−1 (Yα ) is non-trivial. By the arguments in the proof of [Bro87,
30
S. WITZEL AND M. C. B. ZAREMSKY
Theorem 2.2] we have the isomorphisms in the diagram
Y
Y
✲ Hn−1 (Λ,
Hn−1 (Γ,
ZΓ)
ZΛ)
J
J
✻
∼
=
✻
∼
=
Γ
Hn−1
(Y,
Y
ZΓ)
✲ H Λ (Y,
n−1
✻
Y
J
∼
=
Y
ZΛ)
J
J
Γ
Hn−1
(Z,
Y
✻
∼
=
Λ
ZΓ) ✲ lim Hn−1
(Yα ,
−→
❄
Hn−1 (Z)
∼
=
Y
ZΛ)
J
Y ❄
✲ lim
Hn−1 (Yα ).
−→
J
J
(Essentially the two vertical arrows at the top are isomorphisms because Y is (n − 1)acyclic and the two vertical arrows at the bottom are isomorphisms by cocompactness of
the actions and the assumptions on the finiteness properties Q
of the stabilizers.)
Assuming that
Q the diagram commutes, the chain (cα )α∈J ∈ J Hn−1 (Z) has non-trivial
image in lim J Hn−1 (Yα ) and we are done.
−→
The rest of the proof will be concerned with the commutativity of the diagram. The only
square whose commutativity is not clear is the bottom one. In what follows, all products
are taken over J which we suppress from notation.
Let C∗ , C∗α , and D∗ be the cellular chain complexes of Y , Yα , and Z (respectively).
Let P∗ → Z be a resolution by projective ZΛ-modules (which are also projective
ZΓQ
modules). The
third
horizontal
map
is
induced
by
the
maps
P
⊗
(D
⊗
ZΓ)
→
q
Γ
pQ
Q
Q
Pq ⊗Λ (Cpα ⊗ ZΛ). (Or equivalently (Pq ⊗ Dp ) ⊗Γ ZΓ → (Pq ⊗ Cpα ) ⊗Λ ZΛ), which
is the same since the tensor product is associative and, using the notation from [Bro82,
p. 55], also commutative.) The bottom horizontal map is just induced by D∗ → C∗α . The
lower vertical maps come from spectral sequences
Y
Y
1
Γ
Epq
= TorΓq (Dp ,
ZΓ) ⇒ Hp+q
(Z,
ZΓ) and
(5.1)
Y
Y
α
Λ
1
(Yα ,
ZΛ).
(5.2)
= TorΛ
ZΛ) ⇒ Hp+q
Epq
q (Cp ,
The finiteness and cocompactness assumptions guarantee that Dp is of type FPn−p over
Q
ZΓ and Cpα is of type FPn−p over ZΛ so that the natural maps TorΓq (Dp , ZΓ) →
Q
Q
Q
α
α
TorΓq (Dp , ZΓ) and TorΛ
ZΛ) → TorΛ
q (Cp ,
q (Cp , ZΛ) are isomorphisms and the spectral sequences collapse on the second page. We have the commutative diagram of chain
complexes
Y
Y
α
ZΛ)
TorΓ0 (D∗ ,
ZΓ) ✲ TorΛ
0 (C∗ ,
∼
=
Y
∼
=
❄
❄
Y
Y
Y
Λ
Γ
✲
Tor0 (C∗α , ZΛ)=
C∗α
D∗ = Tor0 (D∗ , ZΓ)
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
31
and taking homology in degree n − 1 gives the commutative diagram
Y
Y
Λ
Γ
(Yα ,
ZΛ)
Hn−1
(Z,
ZΓ) ✲ Hn−1
∼
=
that we were looking for.
Y
∼
=
❄
Hn−1 (Z)
✲
Y
❄
Hn−1 (Yα )
6. A Thompson group for direct products of a group
The examples in this section were constructed independently by Slobodan Tanusevski in
his PhD thesis [Tan14], using entirely different techniques, and in discussions with him
we have determined that his groups are identical to those discussed here.
Fix a group G. Let Gn be the direct power Gn . We declare that ρn is trivial for all n,
and define cloning maps via (g1 , . . . , gk , . . . , gn )κnk := (g1 , . . . , gk , gk , . . . , gn ). This makes
rather literal the word “cloning.” To verify that this defines a cloning system, observe that
since the ρn are trivial, we need only check that the cloning maps are homomorphisms
(which they are) and that κnℓ ◦ κn+1
= κnk ◦ κn+1
k
ℓ+1 for 1 ≤ k < ℓ ≤ n (which is visibly
true). These respectively handle conditions (FCS1) and (FCS2) of Definition 2.18, and
condition (FCS3) is trivial. Lastly, the cloning system is visibly properly graded.
It turns out that this cloning system is an example answering Question 5.1 positively,
that is, the finiteness length of T (G∗ ) is exactly that of G (notationally, the asterisk
is a superscript now because we are considering the family of direct powers (Gn )n∈N ).
The proof is due to Tanusevski and we sketch a version of it here, using our setup and
language. For the positive finiteness properties, we just need that the complexes Ln (G∗ )
become increasingly highly connected. This follows by noting that every simplex fiber
of the projection Ln (G∗ ) → M(Ln ) is the join of its vertex fibers, and applying [Qui78,
Theorem 9.1]. For the negative finiteness properties, we claim that there is a sequence of
homomorphisms G → T (G∗ ) → G that composes to the identity. This is sufficient by
the Bieri–Eckmann criterion [BE74, Proposition 1.2]; see [Bux04, Proposition 4.1]. The
first map in the claim is g 7→ [1, g, 1], and the second is [T− , (g1 , . . . , gn ), T+ ] 7→ g1 . One
must check that this second map is well defined on equivalence classes under reduction
and expansion, and is a homomorphism, but this is not hard to see.
A variation of these groups was recently studied using cloning systems, by Berns-Zieve,
Fry, Gillings, and Mathews [BZFGM14]. With the above setup, they consider cloning
maps of the form (g1 , . . . , gk , . . . , gn )κnk := (g1 , . . . , gk , φ(gk ), . . . , gn ) where φ ∈ Aut(G).
They prove that for G finite, the resulting Thompson group is coCF . If these groups
turn out to not embed into V , which seems believable when φ 6= id, then they would be
counterexamples to the conjecture that V is universal coCF .
7. Thompson groups for matrix groups
Let R be a unital ring and consider the algebra of n-by-n matrices Mn (R). We will define
a family of injective functions Mn (R) → Mn+1 (R), which will become cloning maps after
we restrict to the subgroups of upper triangular matrices Bn (R). Consider the map κk
defined by
A<,< A<,k A<,k A<,>
A<,< A<,k A<,>
0
0
Ak,< Ak,k Ak,> κk = Ak,< Ak,k
0
0
Ak,k Ak,>
A>,< A>,k A>,>
A>,< A>,k A>,k A>,>
32
S. WITZEL AND M. C. B. ZAREMSKY
where the matrix has a block structure under which the middle column and row are the
kth column and row of the full matrix respectively. Given the block structure it is not
hard to see that κk is a morphism of monoids, but it generally fails to map invertible
elements to invertible elements. We therefore restrict to the groups Bn (R) of invertible
upper triangular matrices. Let B∞ (R) = lim Bn (R).
−→
Lemma 7.1. The trivial morphisms ρn and the maps κnk defined above describe a properly
graded cloning system on B∗ (R).
It may be noted that the action of F on B∞ (R) factors through H , that is κℓ κk = κk κℓ+1
even for ℓ = k.
Proof. Since ρ∗ is trivial, condition (FCS1) asks that the cloning maps be group homomorphisms. That κk is multiplicative and takes 1 to 1 is straightforward to check. Also,
A is invertible if and only if all the Ai,i are units, in which case (A)κk is also invertible.
To check condition (FCS2) it is helpful to note that for any A ∈ Mn (R), ((A)κk )i,j =
Aπk (i),πk (j) unless i = k or i > j (here πk is as in Example 2.9). One can now distinguish
cases similar to Example 2.9. The compatibility condition (FCS3) is vacuous for trivial
ρ∗ .
To see that the cloning system is properly graded note that g ∈ im ιn,n+1 if and only if the
last column of g is the vector en+1 . If at the same time g = (h)κk then by the definition
of κk the last column of h has to be en . Hence h ∈ im ιn−1,n .
Having equipped B∗ (R) with a cloning system, we get a generalized Thompson group
T (B∗ (R)). Elements are represented by triples (T− , A, T+ ) for trees T± with n leaves
and matrices A ∈ Bn (R), up to reduction and expansion. Figure 6 gives an example of
an element of T (B∗ (R)), represented as a triple and an expansion of that triple.
"
1 2 3
0 4 5
, 0 0 6 ,
#
=
"
1
0
, 0
0
2
4
0
0
2
0
4
0
3
0
5,
6
#
Figure 6. An example of expansion in T (B∗ (Q)).
We are interested in finiteness properties of T (B∗ (R)) because of the following examples
where the groups B∗ (R) themselves have interesting finiteness properties, see [Bux04,
Theorem A, Remarks 3.6, 3.7].
Theorem 7.2. Let k be a global function field, let S be a finite nonempty set of places
and OS the ring of S-integers. Then Bn (OS ) is of type F|S|−1 but not of type F|S| for any
n ≥ 2.
For instance, when R = Fp [t, t−1 ] then Bn (Fp [t, t−1 ]) is finitely generated but not finitely
presented, for n ≥ 2. What is particularly interesting about Theorem 7.2 is that the
finiteness properties of Bn (OS ) depend on |S| but not on n.
A class of examples where the finiteness properties do depend on n arises as subgroups of
groups of the form Bn (R). Let Ab n ≤ Bn+1 be the group of invertible upper triangular
n + 1-by-n + 1 matrices whose upper left and lower right entries are 1. The groups
Ab n (Z[1/p]) were studied by Abels and others and we call them the Abels groups. Their
finiteness length tends to infinity with n [AB87, Bro87]:
Theorem 7.3. For any prime p the group Ab n (Z[1/p]) is of type Fn−1 but not of type Fn
for n ≥ 1.
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
33
For any ring R, the cloning system described above for Bn (R) preserves the Abels groups
Ab n−1 (R). By restriction we obtain a generalized Thompson group T (Ab ∗−1 (R)) which
we will just denote by T (Ab ∗ (R)).
8. Finiteness properties of Thompson groups for matrix groups
We will prove below that the finiteness length of T (B∗ (OS )) is the same as that of all
the Bn (OS ). For consistency, we can state this as
φ(T (B∗ (OS ))) = lim inf φ(Bn (OS )).
n
The inequality ≥, i.e., that T (B∗ (OS )) is of type F|S|−1 , is proved in Section 8.1, and
follows the general strategy outlined in Sections 4 and 5. In fact, it applies to arbitrary
rings. To show the inequality ≤, i.e., that T (B∗ (OS )) is not of type FP|S| , we develop
some new tools in Section 8.2, and make use of the criterion established in Theorem 5.14.
The proof showing the inequality ≥ above also applies to T (Ab ∗ (R)). Since the right
hand side is infinite this time, this directly gives the full equation
φ(T (Ab ∗ (Z[1/p]))) = lim inf φ(Ab n (Z[1/p])).
n
8.1. Positive finiteness properties. The first main result of this section is that the
group T (B∗ (R)) has all the finiteness properties that the individual groups B∗ (R) eventually have:
Theorem 8.1. φ(T (B∗ (R))) ≥ lim inf (φ(Bn (R))).
n
In particular, together with Theorem 7.2 this implies:
Corollary 8.2. T (B∗ (OS )) is of type F|S|−1 .
In view of Proposition 5.9, to prove Theorem 8.1 it suffices to show that the connectivity
of Ln (B∗ (R)) goes to infinity with n. In fact, we will induct, so we need to consider a
slightly larger class of complexes.
For a spanning subgraph ∆ of the linear graph Ln , define Ln (B∗ (R); ∆) to be the subcomplex of Ln (B∗ (R)) whose elements only use graphs that are subgraphs of ∆. Define
e(∆) to be the number of edges of ∆. Define η(m):=⌊ m−1
4 ⌋. Taking ∆ = Ln , Theorem 8.1
will follow from:
Proposition 8.3. Ln (B∗ (R); ∆) is (η(e(∆)) − 1)-connected.
The base case is that Ln (B∗ (R); ∆) is non-empty provided e(∆) ≥ 1, which is clearly
true.
We need to do a bit of preparation before we can prove the proposition. To work with
simplices of Ln (B∗ (R)) it will be helpful to have simple representatives for dangling classes.
To define them we have to recall some of the origins of Ln (B∗ (R)): by Observation 5.6
matchings Γ of Ln correspond to elementary forests. Using this correspondence, it makes
sense to denote the corresponding cloning map by κΓ . In fact, since our cloning maps
factor through the hedge monoid, we even get a cloning map κΓ for any spanning subgraph
Γ of Ln using Observation 5.7. For the sake of readability, we describe this map explicitly.
Let Dk (λ) be the k-by-k matrix with all diagonal entries λ and all other entries 0. Let
Fk,ℓ (λ) be the k-by-ℓ matrix whose bottom row has all entries λ and all other entries are
0 and let Ck,ℓ (λ) be defined analogously for the top row. Assume that Γ has m connected
components which we think of as numbered from left to right. Then
κΓ : Mm (R) → Mn (R)
can be described as follows. The image κΓ (A) has a block structure where columns and
rows are grouped together if their indices lie in a common component of Γ. More precisely,
34
S. WITZEL AND M. C. B. ZAREMSKY
the (i, j)-block has k rows and ℓ columns if the ith (respectively jth) component of Γ has
k (respectively ℓ) vertices. The block is Dk (Ai,i ), Fk,ℓ (Ai,j ) or Ck,ℓ (Ai,j ) depending on
whether i = j, i < j, or i > j (see Figure 7).
a1,1 a1,2 a1,3
D2 (a1,1 ) F2,4 (a1,2 ) F2,3 (a1,3 )
κΓ
a2,1 a2,2 a2,3 7→
C4,2 (a2,1 ) D4 (a2,2 ) F4,3 (a2,3 )
a3,1 a3,2 a3,3
C3,2 (a3,1 ) C3,4 (a3,2 ) D3 (a3,3 )
a1,1
a1,1 a1,2 a1,2 a1,2 a1,2 a1,3 a1,3 a1,3
a
a
2,1 2,1 a2,2
a2,2
a2,2
a2,2 a2,3 a2,3 a2,3
a
a
a
a
a
a
3,1 3,1 3,2 3,2 3,2 3,2 a3,3
a3,3
a3,3
=
Figure 7. Visualization of the cloning map of a graph. The graph Γ is
drawn on top and to the left of the last matrix.
Recall that we denote by ek the kth edge of Ln . We denote by Jk the matching of Ln
whose only edge is ek (as we did in Lemma 5.10). For a spanning subgraph Γ of Ln we say
that an index i is fragile if ei ∈ Γ and we say that i is stable otherwise. In other words, i
is stable if it is the rightmost vertex of its component in Γ. A matrix A ∈ Mn (R) is said
to be modeled on Γ if Ai,j = 0 whenever both i and j are stable in Γ (see Figure 8).
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
0
∗
∗
∗
0
∗
∗
0
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
0
∗
∗
∗
0
∗
∗
0
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
0
∗
∗
∗
0
∗
∗
0
∗ ∗ ∗ ∗
1 ∗ ∗
∗ ∗
∗
∗
∗
∗
∗
∗
∗
0
∗
∗
∗
1
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
0
∗
∗
∗
0
∗
∗
1
Figure 8. A matrix that is modeled on a graph (left) and an upper triangular matrix that is reduced relative to a graph (right).
Lemma 8.4. Let Γ be a spanning subgraph of Ln with m components and let A ∈ Bn (R).
There is a representative B in the coset A(Bm (R))κΓ such that B − In is modeled on Γ.
Moreover, rows of zeroes in A (off the diagonal) can be preserved in B.
Proof. We inductively multiply A on the right by matrices in (Bm (R))κΓ to eventually
obtain B. Let Ei,j (λ) denote the matrix that coincides with the identity matrix in all
entries but (i, j) and is λ there.
We begin by clearing the diagonal. Let i be the (stable) rightmost vertex of the kth
component of Γ and let λ = A−1
i,i . Then A(Ek,k (λ))κΓ has (i, i)-entry one and no other
diagonal entry with stable indices was affected.
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
35
Now we clear the region above the diagonal. We proceed inductively by rows and columns.
Let (i, j) be the (lexicographically) minimal pair of stable indices of Γ such that 0 6=
Ai,j =:−λ. Let i and j lie in the kth respectively ℓth component of Γ. Then A(Ek,ℓ (λ))κm
has (i, j)-entry zero and no other entry with stable indices was affected.
For the last statement assume that the ith row of A was zero off the diagonal. Then none
of the matrices by which we multiplied had a nonzero off-diagonal entry in the ith row.
If i is fragile no such matrix even lies in (Bm (R))κΓ . If i is stable then the only matrices
we might have used of this form were meant to clear the ith row, but since the entries
there were zero, nothing happened in these steps.
Corollary 8.5 (Reduced form). Every simplex in Ln (B∗ (R)) has a representative (A, Γ)
such that the matrix A − In is modeled on Γ.
We will refer to a matrix A ∈ Bn (R) as being reduced relative Γ if it satisfies the conclusion
of Corollary 8.5.
The next sequence of lemmas is a gradual checking of the hypotheses of Proposition 5.13,
still in the context of an induction proof, ultimately leading to a proof of Proposition 8.3.
Lemma 8.6 (Flag complex). Ln (B∗ (R); ∆) is a flag complex.
Proof. We need to show that any collection of vertices {v1 , . . . , vr } that are pairwise
connected by edges spans a simplex. We induct on r (with the trivial base case of r ≤ 2).
Each vertex vi in our collection is of the form [Ai , Jki ] for Jki some single-edge subgraphs
of ∆. Assume without loss of generality that k1 < ki for all 1 < i ≤ r, so v1 is the vertex
whose lone merge occurs farthest to the left among all the vi . By induction, v2 , . . . , vr
span a simplex, σ. Thanks to the action of Bn (R), without loss of generality v1 is the
vertex [In , Jk ], where we have set k := k1 for brevity.
Represent σ = [A, Γ] with A reduced relative to Γ. Since k is less than the index of any
edge of Γ, we know that the kth column of A − In is all zeros. Since v1 shares an edge
with every vertex of σ, we know that in fact k is even less than the index of any edge of
Γ, minus one. Hence the (k + 1)st column of A − In is similarly all zeros. Our goal is
to show that A ∈ im κk , since then σ and v1 will share a simplex. Thanks to the setup,
it suffices to show that the kth row of A − In is all zeros. Since A is reduced relative Γ,
non-zero entries of A − In may only possibly occur in columns indexed by k2 , . . . , kr .
For each vertex [A, Jki ], 2 ≤ i ≤ r, of σ, let Ai be such that [Ai , Jki ] = [A, Jki ] and Ai is
reduced relative Jki . Let ℓ ∈ {k2 , . . . , kr }. Observe that Aℓ is obtained from A by right
multiplication by an element D of im(κℓ ). For 1 ≤ i ≤ n denote by M(i,∗) the ith row
of an n-by-n matrix M , and by M(∗,i) the ith column. When we multiply by D to get
AD = Aℓ , the (k, ℓ)-entry of Aℓ is A(k,∗) D(∗,ℓ) and the (k, ℓ + 1)-entry is A(k,∗) D(∗,ℓ+1) .
Since Aℓ is reduced relative Jℓ , we know that its (k, ℓ + 1)-entry must be 0. Also since
D ∈ im(κJℓ ), we have D(∗,ℓ) = D(∗,ℓ+1) + d(eℓ − eℓ+1 ) for some d ∈ R× . Let a denote
the (k, ℓ)-entry of A, and note that the (k, ℓ + 1)-entry of A is 0. We calculate that the
(k, ℓ)-entry of Aℓ is
A(k,∗) D(∗,ℓ) = A(k,∗) (D(∗,ℓ+1) + d(eℓ − eℓ+1 ))
= A(k,∗) (d(eℓ − eℓ+1 ))
= da.
Since d is a unit, this shows that the (k, ℓ)-entry of Aℓ is zero if and only if the (k, ℓ)entry of A is zero. By the same argument just given, this statement remains true with
Aℓ replaced by Aℓ D for any D ∈ im(κℓ ). But by assumption v1 shares an edge with vℓ ,
and so some such Aℓ D must have (k, ℓ)-entry zero. We conclude that A has (k, ℓ)-entry
zero. Since ℓ was arbitrary, the kth row of A − In is all zeros and so v1 and σ share a
simplex.
36
S. WITZEL AND M. C. B. ZAREMSKY
Let ∆0 := ∆ \ {e1 ∪ e2 }, and consider Ln (B∗ (R); ∆0 ) as a subcomplex of Ln (B∗ (R); ∆).
For a vertex [A, Jk ] ∈ Ln (B∗ (R); ∆0 ) we write lk0 ([A, Jk ]) for the link in Ln (B∗ (R); ∆0 ),
to differentiate from the link in Ln (B∗ (R); ∆) which is just denoted lk([A, Jk ]). To prove
Proposition 8.3 we follow the strategy outlined by Proposition 5.13: we want to show
that Ln (B∗ (R); ∆0 ) is (η(e(∆)) − 2)-connected, that (Ln (B∗ (R); ∆), Ln (B∗ (R); ∆0 )) is
(η(e(∆))−1)-connected and that there is a vertex satisfying condition (EXC). That vertex
is w := [In , J1 ] in our case. The following statements (up to the proof of Proposition 8.3)
are part of an induction, so we assume that Proposition 8.3 has been proven for graphs
∆′ with e(∆′ ) < e(∆) and intend to prove it for ∆.
Lemma 8.7 (Links are lower rank complexes). Let σ be a simplex of dimension d ≥ 0 in
Ln (B∗ (R); ∆). Then lk(σ) is isomorphic to a complex of the form Ln−(d+1) (B∗ (R); ∆′ )
where ∆′ is a spanning subgraph of Ln−(d+1) with at least e(∆)−3d−3 edges. In particular,
it is (η(e(∆) − 3d − 3) − 1)-connected by induction.
Proof. The simplex σ is of the form [g, Γ] with g ∈ Bn (R) and Γ ⊆ ∆. If it has dimension
d then Γ has d + 1 edges, say ei1 , . . . , eid+1 . Using the left action of Bn (R) we may assume
that g = 1. Then lk(σ) is Ln ((B∗ (R))κΓ ; ∆♯ ), where ∆♯ is ∆ with the edges eij −1 , eij , eij +1
removed for each 1 ≤ j ≤ d + 1. In particular ∆♯ has at least e(∆) − 3d − 3 edges. Now
consider the map bΓ : Ln → Ln−(d+1) given by blowing down the edges of Γ. The image
of ∆♯ under bΓ is what we will call ∆′ . Note that ∆′ still has at least e(∆) − 3d − 3 edges.
♯
Since κΓ is injective, we may now apply κ−1
Γ paired with bΓ to Ln ((B∗ (R))κΓ ; ∆ ) and
′
get an isomorphism to Ln−(d+1) (B∗ (R); ∆ ).
Lemma 8.8. The pair (Ln (B∗ (R); ∆), Ln (B∗ (R); ∆0 )) is (η(e(∆)) − 1)-connected.
Proof. Note that for any vertex of Ln (B∗ (R); ∆) \ Ln (B∗ (R); ∆0 ), the entire link of the
vertex lies in Ln (B∗ (R); ∆0 ). Hence the function sending vertices of the former to 1 and
vertices of the latter to 0 yields a Morse function in the sense of Section 5, and to prove
the statement we need only show that links of vertices in Ln (B∗ (R); ∆) \ Ln (B∗ (R); ∆0 )
are (η(e(∆)) − 2)-connected. By Lemma 8.7, each descending link is isomorphic to a
complex of the form Ln−1 (B∗ (R); ∆′ ) for ∆′ a graph with at least e(∆) − 3 edges. By
induction, these are (η(e(∆)) − 2)-connected as desired.
In addition to the subcomplex Ln (B∗ (R); ∆0 ) we will now need to consider Ln (B∗ (R); ∆1 )
where ∆1 := ∆ \ {e1 }. We will write links in this complex using the symbol lk1 .
Lemma 8.9 (Shared links). Let k > 2 and let A be reduced relative Jk . Let A′ be
obtained from A by setting the (1, k)-entry to 0. Then lk1 ([A, Jk ]) ⊆ lk1 ([A′ , Jk ]) and
[A′ , Jk ] ∈ lk w.
Proof. As a first observation, note that since A is reduced relative Jk and k > 2, the
(1, 1)-entry and (2, 2)-entry of A are both 1, and the entries of the top row of A past the
first entry is all 0’s except possibly in the kth column. Let −λ be the (1, k)-entry of A,
and note that A′ = AE1k (λ). The first row of A′ is now (1, 0, . . . , 0) and the (2, 2)-entry
is 1, which tells us that A′ ∈ (Bn−1 (R))κ1 . Hence [A′ , Jk ] ∈ lk0 w.
To see that lk1 ([A, Jk ]) ⊆ lk1 ([A′ , Jk ]) we first multiply by A−1 from the left and are reduced to showing that lk1 ([In , Jk ]) ⊆ lk1 ([E1k (λ), Jk ]). An arbitrary simplex of lk1 ([In , Jk ])
is of the form [B, Γ], with B ∈ im(κk ) and Γ not containing any of e1 , ek−1 , ek , or ek+1 .
Note that the kth row of B is zero off the diagonal. By Lemma 8.4 there is a B ′ ∈ B im(κΓ )
that is reduced relative Γ and has kth row zero off the diagonal. We have [B ′ , Γ] = [B, Γ].
Since e1 6∈ Γ and B ′ is reduced relative Γ, the first column of B ′ is e1 .
We now claim that B ′ commutes with E1k (λ). Indeed, left multiplication by E1k (λ) is the
row operation r1 7→ r1 + λrk , and right multiplication by E1k (λ) is the column operation
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
37
ck 7→ ck + λc1 . For our B ′ , both of these operations change the (1, k)-entry by adding λ
to it, and change no other entries. This proves the claim.
Now we have
[B, Γ] = [B ′ , Γ] = [E1k (λ)B ′ E1k (−λ), Γ] = [E1k (λ)B ′ , Γ] = [E1k (λ)B, Γ].
The second to last step works since E1k (−λ) ∈ im(κΓ ) by virtue of ek−1 , ek 6∈ Γ. This
shows that our arbitrary simplex of lk1 ([In , Jk ]) is also in lk1 ([E1k (λ), Jk ]).
Proof of Proposition 8.3. We want to apply Proposition 5.13. The complexes are X =
Ln (B∗ (R); ∆), X1 = Ln (B∗ (R); ∆1 ) and X0 = Ln (B∗ (R); ∆0 ), and k = η(e(∆)) − 1.
We check the assumptions. The pair (Ln (B∗ (R); ∆), Ln (B∗ (R); ∆0 )) is k-connected by
Lemma 8.8. Since X is a flag complex (Lemma 8.6), it is conical at every vertex, in
particular at our vertex w = [In , J1 ]. The complex Ln (B∗ (R); ∆0 ) is (η(e(∆0 )) − 1)connected by induction. This is sufficient because η(e(∆0 )) − 1 ≥ η(e(∆) − 2) − 1 ≥
η(e(∆)) − 2 = k − 1. The link of a d-simplex is (η(e(∆) − 3d − 3) − 1)-connected by
Lemma 8.7. This is sufficient because η(e(∆) − 3d − 3) − 1 ≥ η(e(∆)) − d − 2 = k − d − 1.
Finally condition (EXC) is satisfied by Lemma 8.9 where lk1 ([A, Jk ]) is at least (η(e(∆) −
4) − 1)-connected and η(e(∆) − 4) − 1 = η(e(∆)) − 2 = k − 1 as desired.
Shifting focus to the Abels groups, thanks to the flexibility of Lemma 8.4, the above
arguments also show high connectivity of Ln (Ab ∗ (Z[1/p])), and using Proposition 5.9
and Theorem 7.3 we conclude:
Theorem 8.10. T (Ab ∗ (Z[1/p])) is of type F∞ .
This, despite none of the Ab n (Z[1/p]) individually being F∞ .
The remaining question is whether φ(T (B∗ (R))) = lim inf n (φ(Bn (R))), that is whether
negative finiteness properties of the Bn (R) can impose negative finiteness properties on
T (B∗ (R)). For R the ring of S-integers of a global function field, we will answer this
question affirmatively in the next section.
Before we do that, we need to treat one more relative of the family Bn (R): Let Bn2 be the
normal subgroup of Bn consisting of matrices that differ from the identity only from the
second off-diagonal on (the second term of the lower central series), and let B̄n := Bn /Bn2
be the quotient group. Set ν(n) = ⌊ n−2
3 ⌋. One could check that the above proof for B∗
goes through for the family B̄∗ as well, but instead we will prove directly:
Proposition 8.11. The descending link Ln (B̄∗ (R)) is (ν(n)−1)-connected. Thus φ(T (B̄∗ (R))) ≥
lim inf φ(B̄∗ (R)).
n
Proof. Using reductions as in Lemma 8.4 one can see the following: every simplex in
Ln (B̄∗ (R)) has a representative [A, Γ] where the matrix A has a diagonal block of the
form
∗ ∗
1
above every edge of Γ and otherwise equals the identity matrix (here the representative
is modulo dangling as well as modulo Bn2 (R)). What makes this case particularly easy is
that this representative is unique. That is, we may think of Ln (B̄∗ (R)) as consisting of
pairs (A, Γ) where A is as above and the face relation is given by removing an edge of Γ
and turning the diagonal block above it into an identity block.
Let sLn denote the linear graph with vertices {1, . . . , n} and with every pair of adjacent
vertices i and i + 1 connected by s distinct edges. By what we just said, Ln (B̄∗ (R)) is
isomorphic to the matching complex M(sLn ) where s = |R∗ × R|. There is an obvious
map M(sLn ) → M(Ln ). The fiber of this map over a k-simplex is a (k + 1)-fold join
of s-element sets, thus k-spherical. Moreover M(Ln ) is (ν(n) − 1)-connected by [Koz08,
38
S. WITZEL AND M. C. B. ZAREMSKY
Proposition 11.16] (and links in M(Ln ) are highly connected as well, being joins of lowerrank copies of the complex). Thus we can apply [Qui78, Theorem 9.1] to conclude that
M(sLn ) is (ν(n) − 1)-connected.
As a remark, this simple approach for B̄∗ (R) would not have worked for B∗ (R), since the
analogous fibers are not joins of vertex fibers.
8.2. Negative finiteness properties. In the last section we saw that for any R, the
generalized Thompson group T (B∗ (R)) is of type Fn if all but finitely many Bk (R) are.
In this section we prove the converse in the case we are most interested in (cf. [Bux04]):
Let k be a global function field and let S be a non-empty set of places. Denote by OS
the ring of S-integers in k.
Theorem 8.12. The group T (B∗ (OS )) not of type FP|S| .
Remark 8.13. Unlike the positive statement from the previous section, for the proof of
Theorem 8.28 we cannot just use the results from [Bux04] but have to use parts of the
proof. By using the more substantial parts of the proof, it is quite possible that the setup
of this section could be used to prove the positive finiteness properties as well, but we
will not do so.
We will actually prove first that T (B̄∗ (OS )) is not of type FP|S| . We then use the result
from Section 5.5 to deduce Theorem 8.28. Instead of the Stein–Farley complex on which
T (B̄∗ (OS )) acts with stabilizers isomorphic to the B̄∗ (OS ) we will construct a new space
Y for which the stabilizers are themselves generalized Thompson groups of smaller cloning
systems. In particular the stabilizers on Y will have good finiteness properties and the
negative finiteness properties of the B̄∗ (OS ) are reflected in bad connectivity properties.
For any place s ∈ S denote by ks the completion of k at s, and by Os the ring of integers
of ks . As before we let Bn be the linear algebraic group of invertible upper triangular
n-by-n matrices, let Bn2 be the normal subgroup of matrices that differ from the identity
only from the second off-diagonal on, and let B̄n := Bn /Bn2 be the quotient group. Let
Zn ≤ Bn be the group of homotheties, i.e., scalar multiples of the identity matrix, and
let PB 2 = B2 /Z2 .
All of this is relevant to us for the following reason: For any of the local fields ks the group
PGL2 (ks ) admits
a Bruhat–Tits tree Vs on which it acts properly. Since OS is discrete as
Q
a subset of s∈S ks when embedded diagonally, we get a properly discontinuous action of
PGL2 (OS ) on
Y
Vs .
V :=
s∈S
Our goal is to use this action to understand finiteness properties of T (B̄∗ (OS )). Note
that the group PGL2 (Os ) is the stabilizer of a vertex in Vs , call it zs . Define z := (zs )s∈S ,
so z is a vertex in V .
Denote the quotient morphism from Bn to B̄n by ¯: Bn → B̄n , g 7→ ḡ. For 1 ≤ i ≤ n − 1
let πi denote the homomorphism B̄n → PB 2 , [A] 7→ [Ai ] where Ai is the ith diagonal
2-by-2 block of A. For brevity we denote the composition πi ◦ ¯ by π̄i . Now for any i,
1 ≤ i ≤ n − 1 consider the composition
Q
Y
πi (ks ) Y
B̄n (ks ) →
PB 2 (ks )
αi : B̄n (OS ) →
s∈S
s∈S
where the first morphism is induced by the diagonal inclusion OS →
!
\
Y
−1
Kn :=
αi
PB 2 (Os ) .
1≤i≤n−1
s∈S
Q
s∈S
ks . Define
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
39
Lemma 8.14. The group Kn is of type F∞ .
Remark 8.15. The importance of the Lemma lies in the fact that the groups Kn will
appear in stabilizers of an action of T (B̄∗ (OS )). It is worth noting that the statement
does not remain true if B̄n (OS ) is replaced by Bn (OS ) so that the strategy does not
immediately carry over to T (B∗ (OS )). Instead we will have to apply Theorem 5.14 in
the end to conclude that T (B∗ (OS )) is not of type FP|S| .
Proof of Lemma 8.14. We first study the map πi (ks ) : B̄n (ks ) → PB 2 (ks ). The kernel
Ni (ks ) is determined by the conditions that the (i, i + 1)-entry of a matrix is 0 and that
the (i, i) and the (i + 1, i + 1)-entry coincide. The inverse image of PB 2 (Os ) under πi (ks )
is
by Ni (ks ) and a copy of B2 (Os ). Intersecting over all i, we find that
T thus generated
−1
i πi (ks ) (PB 2 (Os )) = Zn (ks )B̄n (Os ).
The intersection of this group with B̄n (OS ) is Zn (OS )B̄n (OS\{s} ). Intersecting over all
s ∈ S we find that Kn = Zn (OS )B̄n (ℓ) where ℓ := O∅ is the coefficient field of k, which
is finite. In particular B̄n (ℓ) is finite and of type F∞ . By the Dirichlet Unit Theorem, as
extended to S-units by Hasse and Chevalley, Zn (OS ) is finitely generated abelian and so
of type F∞ . Since Kn is a central product of these groups, this finishes the proof.
Now consider the action of B̄n (OS ) on V n−1 via the maps αi .
Corollary 8.16. The stabilizers for the action of B̄n (OS ) on V n−1 are all of type F∞ .
Proof. The group Kn is precisely the stabilizer of (z, . . . , z) ∈ V n−1 . Since the product of
trees V n−1 is locally finite and the action is proper, every stabilizer is commensurable to
Kn and therefore of type F∞ as well.
We are about to define a space Y for T (B̄∗ (OS )) to act on. The advantage over the
Stein–Farley complex will be that the stabilizers have better finiteness properties. Let
D = Z[1/2] ∩ (0, 1) be the set of dyadic points in (0, 1). Let V D be the set of all maps
D → V . We will usually regard these elements as tuples; that is, we write xq for the value
of x ∈ V D at q ∈ D and sometimes we write x as (xq )q∈D . Let
Y := V (D)
be the subset consisting of those maps that evaluate to z at all but finitely many points. An
alternative description is as a direct limit limI⊆D finite V I . Note that this set is naturally
−→
equipped with a (unique) topology: the topology induced from the product topology and
the CW topology coincide.
Note that Thompson’s group F acts on D from the right, via q.f = f −1 (q) for f ∈ F and
q ∈ D. To describe this action in terms of paired tree diagrams, note that every point
in D corresponds to a caret in the leafless rooted binary tree. Thus every finite rooted
binary tree T determines a finite subset D(T ) of D, namely that consisting of points that
correspond to its carets. An element [T, U ] of F takes D(T ) to D(U ) (preserving the
order) and is linear between these break points.
As a consequence, F acts from the left on the set V D via (f.x)q = xq.f where x ∈ V D ,
q ∈ D and f ∈ F . Clearly this induces an action of F on Y . Explicitly, the action of F
on Y satisfies
([T, U ].x)ti = xui
where D(T ) = {t1 < . . . < tn−1 } and D(U ) = {u1 < . . . < un−1 }. Away from the break
points, the values are interpolated linearly: [T, U ].xsti +(1−s)ti+1 = xsui +(1−s)ui+1 .
There is also an action of K (B̄∗ (OS )) on Y which is given as follows: if T is a finite
rooted binary tree and D(T ) = {q1 < . . . < qn−1 } then
αi (g).xqi if q = qi
([T, g, T ].x)q =
xq
else.
40
S. WITZEL AND M. C. B. ZAREMSKY
This is just the action obtained by taking the direct limit over the actions of B̄T (OS ) on
V D(T ) .
These actions are compatible and so give an action of T (B̄∗ (OS )) on Y , which is given
by
([T, g, U ].x)ti = αi (g).xui
and ([T, g, U ].x)t = ([T, U ].x)t for t 6∈ D(T ); see Figure 9.
1 2
4 5
6 =
a b
a b̄
c̄
c
1 2
45
6
a
b
=
a b̄
c̄
c
Figure 9. Two points of view on the action of T (B̄∗ (OS )) on Y . On
the left the action is described in terms of tree diagrams, on the right in
terms of piecewise linear homeomorphisms. In both pictures b̄ = ( 1 42 ) b
and c̄ = ( 4 56 ) c. All unspecified values are z.
Next we want to understand stabilizers of this action. First observe that the action has
a nontrivial kernel, namely the center of T (B̄∗ (OS )), which is isomorphic to OS× .
Observation 8.17. Let (G∗ , (κk )k ) be a cloning system and let H be a group. Define a
new cloning system (H ×G∗ , (κ̂k )k ) by taking κ̂k :=id ×κk . Then T (H ×G∗ ) = H ×T (G∗ ).
Proof. The isomorphism is given by (h, [T, g, U ]) 7→ [T, hg, U ].
We now turn to one particular stabilizer.
Observation 8.18. The cloning system on B̄∗ (OS ) induces a cloning system on K∗ . The
stabilizer in T (B̄∗ (OS )) of the point (z)q is T (K∗ ).
Proof. For the first statement it suffices to show that (Zn (OS ))κk ⊆ Zn+1 (OS ) and that
(B̄n (ℓ))κk ⊆ B̄n+1 (ℓ) which is easy to see. The second statement is clear.
Corollary 8.19. The group T (K∗ ) is of type F∞ .
Proof. By Observation 8.17 T (K∗ ) is isomorphic to a central product OS× T (B̄∗ (ℓ)). The
second factor is of type F∞ by Proposition 8.11.
We now turn to general stabilizers. For a point x ∈ V write [x] for its PB 2 (OS )-orbit.
We call a point (xq )q ∈ Y reduced if xq = z whenever [xq ] = [z].
Lemma 8.20. Every point of Y has a reduced point in its T (B̄∗ (OS ))-orbit.
Proof. If (xq )q ∈ Y is arbitrary, let T be a tree such that D(T ) contains all of the finitely
many indices q ∈ D for which xq 6= z. Write D(T ) = {q1 , . . . , qn−1 } where the indices are
in increasing order. For each i pick gi ∈ PB 2 (OS ) such that gi .xqi = z whenever possible
(i.e., when [xqi ] = [z]) and arbitrarily otherwise. Take g ∈ B̄n (OS ) such that αi (g) = gi
for all i. Then [T, g, T ].(xq )q is reduced.
Lemma 8.21. The stabilizer in T (B̄∗ (OS )) of any reduced point is of type F∞ .
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
41
Proof. Let (xq )q∈D be a reduced point and let I = {q ∈ D | xq 6= z} = {q1 , . . . , qn−1 }.
Let H be the stabilizer of (xq )q in T (B̄∗ (OS )) and let K be the kernel of the action of
H. Since [xq ] 6= [z] for q ∈ I, we see that the stabilizer has to fix I (when acting on D
via the canonical homomorphism to F ). Thus the action of the stabilizer H on Y = V (D)
decomposes into an action on V I and on V (D\I) .
Modulo K we find that H is a direct product of the pointwise stabilizer (in H) of V I
and the pointwise stabilizer of V (D\I) . The action of the pointwise stabilizer of V (D\I)
in T (B̄∗ (OS )) is isomorphic to B̄n (OS ) acting on V I . Thus its intersection with H is
isomorphic to a point stabilizer in B̄n (OS ), and hence is of type F∞ by Corollary 8.16.
The pointwise stabilizer of V I decomposes further. Let D1 := D ∩ (0, q1 ), D2 :=D ∩ (q1 , q2 ),
. . . , Dn := D ∩ (qn−1 , 1). The pointwise stabilizer of V (D\Dj ) is itself isomorphic to a copy
of T (B̄∗ (OS )) and therefore the stabilizer of (z)q∈DJ in this stabilizer is isomorphic to a
copy of T (K∗ ), which is of type F∞ by Corollary 8.19.
Putting everything together we find that H/K is a product of groups of type F∞ , and K
is of type F∞ as well, so H is of type F∞ .
In summary we have:
Proposition 8.22. The group T (B̄∗ (OS )) acts on Y with stabilizers of type F∞ .
Proof. Every point is in the orbit of a reduced point by Lemma 8.20 so every stabilizer is
isomorphic to that of a reduced point. Those are of type F∞ by Lemma 8.21.
It remains to provide a cocompact filtration and determine its essential connectivity. For
this purpose we will use the key result from [Bux04] used to show that PB 2 (OS ) is not
of type FP|S| :
Theorem 8.23 ([Bux04]). There is a filtration (Vr )r∈N of V that is PB 2 (OS )-invariant
and -cocompact and is essentially (|S| − 2)-connected but not essentially (|S| − 1)-acyclic.
In fact, by Brown’s criterion any cocompact filtration of V has that property just because
B2 (OS ) is of type F|S|−1 but not of type FP|S| . We use this filtration to construct a
cocompact filtration of Y as follows. For r ∈ N let Y (r) be the set of all points (xq )q∈D for
which {q ∈ D | [xq ] 6= [z]} has at most r elements. Note that Y (r) is T (B̄∗ (OS ))-invariant.
The filtration we want to consider is
Yr := Y (r) ∩ Vr(D) .
The last piece that is missing to conclude that T (B̄∗ (OS )) is not of type FP|S| is the
following:
Proposition 8.24. The filtration (Yr )r∈N is T (B̄∗ (OS ))-invariant and -cocompact. It is
not essentially (|S| − 1)-acyclic.
Before we can prove the second part, we have to state a technical lemma which says that
taking products does not help to kill cycles:
Lemma 8.25. Let (X1 , A1 ) and (X2 , A2 ) be pairs of CW complexes and assume that the
map H̃n (A1 → X1 ) is non-trivial and that A2 is non-empty. Then the map H̃n (A1 ×A2 →
X1 × X2 ) is non-trivial as well.
Proof. The case n = 0 is clear so assume n > 0 from now on.
42
S. WITZEL AND M. C. B. ZAREMSKY
Let c be an n-cycle in A1 that is mapped non-trivially into X1 and let d be a non-trivial
0-cycle in A2 . Consider the diagram
Hn (A1 ) ⊗ H0 (A2 ) ⊂ ✲ Hn (A1 × A2 )
❄
❄
Hn (X1 ) ⊗ H0 (X2 ) ⊂✲ Hn (X1 × X2 ).
where the rows are parts of the Künneth formula (see [Hat01, Theorem 3B.6]) and the
columns are the maps induced from the inclusions. The diagram commutes by naturality
of the Künneth formula. The cycle c ⊗ d in the upper left maps non-trivially into the
lower left which injects into the lower right. Hence it has non-trivial image in the lower
right. Since the diagram commutes, it follows that its image in the upper right also has
non-trivial image in the lower right, which is what we want.
Proof of Proposition 8.24. For cocompactness let Cr ⊆ V be compact such that its PB 2 (OS )translates cover Vr . Let Ĉr ⊆ Y be the product of r copies of Cr (say at positions q1 , . . . , qr )
and {z} otherwise. We claim that the translates of Ĉr cover Yr .
Indeed, let (xq )q ∈ Yr be arbitrary. Since it lies in Y (r) there are at most r positions
where [xq ] 6= [z]. Using the action of F we can achieve that these positions are (some of)
q1 to qr . Now, since each xqi lies in Vr , we can move it into Cr , using an element of the
form [T, g, T ], without moving any of the other xq . At all other coordinates q, i.e., where
[xq ] = [z], we can move xq to z using the same method. Since all but finitely many xq
were z to begin with, we have moved (xq )q into Ĉr in finitely many steps.
For the second statement let N = |S| − 1, so we want to show that (H̃N (Yr ))r is not
essentially trivial. Let k be such that the map H|S|−1 (V
Qk → Vm ) is non-trivial for every
m ≥ k. For arbitrary m ≥ k take A1 = Vk , A2 =
q∈D {z}, X1 = Vm , and X2 =
q6
=1/2
Q
q∈D Vm . Then A1 × A2 ⊆ Yk and Ym ⊆ X1 × X2 (on the infinite products we take
q6=1/2
the CW topology, not the product topology). By Lemma 8.25 the map HN (A1 × A2 →
X1 ×X2 ) is non-trivial. But this factors through the map HN (Yk → Ym ) which is therefore
non-trivial as well. This shows that (HN (Yr ))r is not essentially trivial.
Theorem 8.26. The group T (B̄∗ (OS )) is not of type FP|S| .
Proof. The group acts on Y , which is contractible, with stabilizers of type F∞ (Proposition 8.22). There is an invariant, cocompact filtration (Yr )r which is not essentially
(|S| − 1)-acyclic (Proposition 8.24). We conclude using Brown’s criterion.
Remark 8.27. As far as we can tell, none of the established methods in the literature can
be used now to show that T (B∗ (OS )) is not of type FP|S| . The kernel of the morphism
T (B∗ (OS )) → T (B̄∗ (OS )) is very unlikely to be even finitely generated (or else one could
apply [Bie76, Proposition 2.7] or the following exercise, see also [Geo08, Theorem 7.2.21]).
Also the projection does not split (or else one could apply the retraction argument [Bux04,
Proposition 4.1]). For this reason we will now use the new methods established in Section 5.5, which can be regarded as a generalization of the retraction argument. We should
mention that one can deduce that T (B∗ (OS )) is not finitely generated if |S| = 1, without
using this new machinery.
Proof of Theorem 8.12. We apply Theorem 5.14 to the inclusion homomorphism B2 (OS ) ֒→
T (B̄∗ (OS )) that takes g to [λ1 , g, λ1 ] where λ1 is a single caret. We take Z to be the subspace of Y consisting of points (xq )q with xq = z for q 6= 1/2, so Z is B2 (OS )-cocompact.
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
43
Our T (B̄∗ (OS ))-cocompact filtration of Y is (Yr )r∈N . Observe that H|S|−1 (Z → Yr ) is
not eventually trivial by the proof of Proposition 8.24. Thus we can apply Theorem 5.14.
Since the inclusion B2 (OS ) ֒→ T (B̄∗ (OS )) clearly factors through T (B∗ (OS )) we conclude that this group is not of type FP|S|−1 .
Combining Theorem 8.1 and Theorem 8.12, we obtain:
Theorem 8.28. The group T (B∗ (OS )) is of type F|S|−1 but not of type FP|S| .
9. Thompson groups for mock-symmetric groups
The groups discussed in this section are instances of what Davis, Januszkiewicz and Scott
call “mock reflection groups” [DJS03]. These are groups generated by involutions, and
act on associated cell complexes very much like Coxeter groups, with the only difference
being that some of the generators may be “mock reflections” that do not fix their reflection
mirror pointwise. Here we will only be concerned with one family of groups consisting of
the minimal blow up of Coxeter groups of type An . These Coxeter groups are symmetric
groups and so we call their blow ups mock symmetric groups. For n ∈ N the mock
symmetric group Snmock is given by the presentation
Snmock = hsi,j , 1 ≤ i < j ≤ n | s2i,j = 1 for all i, j
si,j sk,ℓ = sk,ℓ si,j for i < j < k < ℓ
(9.1)
sk,ℓsi,j = sk+ℓ−j,k+ℓ−isk,ℓ for k ≤ i < j ≤ ℓi.
mock
S∞
lim S mock .
−→ n
We also set
=
See Figure 10 for a visualization of elements of Snmock , and
a visualization of the last relation.
=
Figure 10. The relation si,j sk,ℓ = sk,ℓ sk+ℓ−j,k+ℓ−i of Snmock in the case
i = 3, j = 4, k = 1, ℓ = 5, n = 5.
i+j
Let s̄i,j ∈ Sn be the involution (i j)((i + 1) (j − 1)) · · · (⌊ i+j
2 ⌋ ⌈ 2 ⌉) (this is the longest
element in the Coxeter group generated by (i i+1), . . . , (j −1 j)). Taking si,j to s̄i,j defines
mock
a surjective homomorphism ρn : Snmock → Sn . We define cloning maps κnk : Snmock → Sn+1
by first defining them on the generators:
for j < k
si,j
si,j+1 sk,k+1 for i ≤ k ≤ j
(si,j )κnk =
(9.2)
si+1,j+1
for k < i.
mock as in the paragraph leading up to Lemma 1.12.
Now we extend κnk to a map Snmock → Sn+1
See Figure 11 for an example of cloning.
=
mock .
Figure 11. The relation s1,4 λ3 = λ2 s1,5 s3,4 of F ⊲⊳ S∞
44
S. WITZEL AND M. C. B. ZAREMSKY
Proposition 9.1. The above data define a cloning system on S∗mock .
Proof. Note first that (9.1) is a presentation for Snmock as a monoid because all the generators are involutions by the first relation. Following the advice from Remark 2.12, we
will apply Lemma 1.12 with this presentation rather than the trivial presentation used in
Proposition 2.7.
We have to verify conditions coming from relations of F and conditions coming from
relations of Snmock , after which the proof proceeds as that of Proposition 2.7. For the
relations of F we must verify the conditions (FCS2) (product of clonings) and (FCS3)
(compatibility)
(si,j )κℓ κk = (si,j )κk κℓ+1
ρ((si,j )κk ) = (ρ(si,j ))ςk
for k < ℓ and i < j
(9.3)
for i < j.
(9.4)
(Note that we verified (FCS3) for all i, which is not technically necessary; see the remark
after Observation 2.11). For the relations of Snmock we have to check that ρ is a well defined
homomorphism, and check that the following equations, standing in for (CS1) (cloning a
product), are satisfied:
(si,j )κρ(sk,ℓ )p (sk,ℓ )κp = (sk,ℓ )κρ(si,j )p (si,j )κp
for i < j < k < ℓ
(9.5)
(sk+ℓ−j,k+ℓ−i)κρ(sk,ℓ )p (sk,ℓ )κp = (sk,ℓ )κρ(si,j )p (si,j )κp
for k ≤ i < j ≤ ℓ.
(9.6)
Note that the conditions coming from the relations s2i,j = 1 are vacuous.
Condition (9.3) is easy to check if k < i or ℓ > j so we consider the situation where
i ≤ k < ℓ ≤ j. In this case we have
(si,j )κℓ κk = (si,j+1 sℓ,ℓ+1 )κk = (si,j+1 )κk (sℓ,ℓ+1 )κk = si,j+2 sk,k+1 sℓ+1,ℓ+2
= si,j+2 sℓ+1,ℓ+2 sk,k+1 = (si,j+1 )κℓ+1 (sk,k+1 )κℓ+1 = (si,j+1 sk,k+1)κℓ+1 = (si,j )κk κℓ+1
since ρ(sk,k+1 )(ℓ + 1) = (ℓ + 1), ρ(sℓ,ℓ+1 )k = k and
Condition (9.4) amounts to showing that
s̄i+1,j+1
s̄i,j+1 s̄k,k+1
(s̄i,j )ςk =
s̄i,j
sk,k+1 and sℓ+1,ℓ+2 commute.
k<i
i≤k≤j
k > j.
The cases k < i and k > j are clear. For the remaining case we first note that
s̄i,j+1 s̄k,k+1 (m) = τi+j−k s̄i,j πk (m) = ((s̄i,j )ςk )(m)
for m 6= k, k+1 (which is also the same as s̄i,j+1 (m)). Here τk and πk are as in Example 2.9.
Finally one checks that
s̄i,j+1 s̄k,k+1 (k) = i + j − k = (s̄i,j )ςk (k)
and that
s̄i,j+1 s̄k,k+1(k + 1) = i + j − k + 1 = (s̄i,j )ςk (k + 1).
That ρ is a well defined homomorphism amounts to saying that the defining relations of
Snmock hold in Sn with si,j replaced by s̄i,j , which they do.
Condition (9.5) is also easy to check unless i ≤ p ≤ j or k ≤ p ≤ ℓ. We treat the case
i ≤ p ≤ j, the other remaining case being similar. We have
(si,j )κρ(sk,ℓ )p (sk,ℓ )κp = si,j+1 sp,p+1 sk+1,ℓ+1 = sk+1,ℓ+1 si,j+1 sp,p+1 = (sk,ℓ )κρ(si,j )p (si,j )κp .
Finally, the interesting case of condition (9.6) is when i ≤ p ≤ j. We have
(sk,ℓ )κρ(si,j )p (si,j )κp = sk,ℓ+1si+j−p,i+j−p+1si,j+1 sp,p+1 = sk,ℓ+1 si,j+1
= sk+ℓ−j,k+ℓ−i+1sk+ℓ−p,k+ℓ−p+1sk,ℓ+1 sp,p+1 = (sk+ℓ−j,k+ℓ−i)κρ(sk,ℓ )p (sk,ℓ )κp
using the defining relations of Snmock several times.
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
45
As a consequence we get
Theorem 9.2. There is a generalized Thompson group T (S∗mock ) which contains all the
Snmock and canonically surjects onto V . We denote it Vmock .
Conjecture 9.3. Vmock is of type F∞ .
Since each Snmock is of type F∞ [DJS03, Section 4.7, Corollary 3.5.4], to prove the conjecture it suffices to show that the the cloning system is properly graded and that the
connectivity of the complexes Ln (S∗mock ) goes to infinity as n goes to infinity.
10. Thompson groups for loop braid groups
Our next example of a cloning system comes from the family of loop braid groups LB n ,
also known as groups ΣAutn of symmetric automorphisms of free groups, or as braidpermutation groups. This will produce a generalized Thompson group Vloop that contains
both Vbr and V as subgroups. There is also a pure version of this cloning system, using
the pure loop braid groups, which we will discuss as well, yielding a group Floop .
We first describe the family of groups in terms of free group automorphisms. Fix a set of
generators {x1 , . . . , xn } for Fn , and call an automorphism φ ∈ Aut(Fn ) symmetric if for
every 1 ≤ i ≤ n there exists 1 ≤ j ≤ n such that φ(xi ) is conjugate to xj . If every φ(xi )
is even conjugate to xi , call φ pure symmetric. The group of symmetric automorphisms
of Fn is denoted ΣAutn , and the group of pure symmetric automorphisms is denoted
PΣAutn . The latter is also denoted by P LBn , for pure loop braid group. The reader is
cautioned that in the literature “symmetric” sometimes allows for generators to map to
conjugates of inverses of generators, but we do not allow this.
The LB n fit into a directed system. The map ιn,n+1 : LB n ֒→ LB n+1 is given by sending
the automorphism φ of Fn to the automorphism of Fn+1 that does nothing to the new
generator and otherwise acts like φ. This restricts to PLB n as well, and so we have
directed systems LB ∗ and PLB ∗ .
Our presentation for LB n = ΣAutn will be taken from [FRR97]. The generators are as
follows, for 1 ≤ i ≤ n.
xi
βi :
x
i+1
xj
xi
xi+1
σi :
xj
7→ xi+1
7→ x−1
i+1 xi xi+1
7→ xj
(j 6= i, i + 1)
7 xi+1
→
7→ xi
7→ xj
(j 6= i, i + 1)
The βi together with the σi generate ΣAutn . The βi by themselves generate a copy of Bn
in ΣAutn , and the σi generate a copy of Sn . As seen in [FRR97], defining relations for
ΣAutn are as follows (with 1 ≤ i ≤ n − 1):
46
S. WITZEL AND M. C. B. ZAREMSKY
βi βj = βj βi (|i − j| > 1)
βi βi+1 βi = βi+1 βi βi+1
σi2 = 1
σi σj = σj σi (|i − j| > 1)
σi σi+1 σi = σi+1 σi σi+1
βi σj = σj βi (|i − j| > 1)
σi σi+1 βi = βi+1 σi σi+1
βi βi+1 σi = σi+1 βi βi+1 .
This is a group presentation, and it becomes a monoid presentation after adding generators
βi−1 with relations βi βi−1 = βi−1 βi = 1.
Since we already have cloning systems on S∗ (from Example 2.9) as well as on B∗ (from
[Bri07]), we already know how the cloning system on LB ∗ = ΣAut∗ should be defined.
The only thing to check is that it is actually well defined.
The homomorphism ρn : LB n → Sn just takes βi as well as σi to σi ∈ Sn . This is easily
seen to be well defined.
The cloning maps are defined as they are defined for the symmetric groups and braid
groups respectively: for ε ∈ {±1} this means that
ε
β
if k < i
i+1
εβε
β
if k = i
i i+1
(βiε )κk :=
(10.1)
ε
ε
β βi if k = i + 1
i+1
βε
if k > i + 1
i
σ
if k < i
i+1
σi σi+1 if k = i
(σi )κk :=
(10.2)
σ
if k = i + 1
i+1 σi
σi
if k > i + 1
Lemma 10.1. The above data ρ∗ and κ∗k define cloning systems on LB ∗ and on PLB ∗ .
Proof. We already noted that ρ is a well defined group homomorphism. We have to
check (CS2) (product of clonings) and (CS3) (compatibility) on generators of LB n . But
since every generator is a generator of either Sn or of Bn , each verification needed has
been performed in establishing the cloning systems on either S∗ or B∗ .
It remains to check that cloning a relation is well defined, standing in for (CS1) (cloning
a product). Again, the relations involving only elements of Sn or Bn are already verified.
This leaves the last three kinds of relations.
For the first relation we have to check that
(βi )κρ(σj )k (σj )κk = (βi )κσj k (σj )κk = (σj )κσi k (βi )κk = (σj )κρ(βi )k (βi )κk
which is easy to do case by case. For the other two relations we must show that
(σi )κ(i
i+2 i+1)k (σi+1 )κ(i i+1)k (βi )κk
= (βi+1 )κ(i
i+1 i+2)k (σi )κ(i+1 i+2)k (σi+1 )κk
(βi )κ(i
i+2 i+1)k (βi+1 )κ(i i+1)k (σi )κk
= (σi+1 )κ(i
i+1 i+2)k (βi )κ(i+1 i+2)k (βi+1 )κk
which can be treated formally equivalently as long as we do not use either of the relations
σi2 = 1 or βi βi−1 = βi−1 βi = 1. The cases k < i and k > i + 2 are easy. For k = i we apply
only mixed relations to find
(σi )κi+2 (σi+1 )κi+1 (βi )κi = σi σi+1 σi+2 βi βi+1
= βi+1 βi+2 σi σi+1 σi+2 = (βi+1 )κi+1 (σi )κi (σi+1 )κi .
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
47
Similarly for k = i + 1 we get
(σi )κi (σi+1 )κi (βi )κi+1 = σi σi+1 σi+2 βi+1 βi
= βi+2 βi+1 σi σi+1 σi+2 = (βi+1 )κi+2 (σi )κi+2 (σi+1 )κi+1 .
Lastly for k = i + 2 we first apply a braid relation and then use the mixed relations to get
(σi )κi+1 (σi+1 )κi+2 (βi )κi+2 = σi+1 σi σi+2 σi+1 βi = σi+1 σi+2 σi σi+1 βi
= βi+2 σi+1 σi σi+2 σi+1 = (βi+1 )κi (σi )κi+1 (σi+1 )κi+2 .
Finally, that the cloning system on LB ∗ restricts to one on PLB ∗ is straightforward.
Theorem 10.2. There are generalized Thompson groups
Vloop := T (LB ∗ )
and
Floop := T (PLB ∗ )
containing the loop braid groups and the pure loop braid groups, respectively. The group
Vloop canonically surjects onto V , and the group Floop canonically surjects onto F .
The group LB n is known to be of type F∞ , for instance it acts properly cocompactly on
the contractible space of marked cactus graphs [Col89]. For this reason understanding
the finiteness properties of Vloop and Floop amounts to showing that the cloning systems
are properly graded, and understanding the connectivity of Ln (LB ∗ ) and Ln (PLB ∗ ). We
expect that these should be increasingly highly connected and thus:
Conjecture 10.3. Vloop and Floop are of type F∞ .
We do not attempt to prove this conjecture here. However, we end by sketching a more
geometric viewpoint of these cloning systems, which could be useful in the future. To
do so, we will view LB n as a group of motions of loops (which is where the name comes
from); see [BWC07], [BH13] and [Wil12]. Let R3 be Euclidean 3-space, and define a loop
1
3
γ to be a smooth, unknotted, oriented embedded copy of the circle
` S in R . Now fix a
3
γ. A motion of Cn
set L of n pairwise disjoint, unlinked loops in R , and let Cn :=
γ∈L
is a path of diffeomorphisms ft ∈ Diff(R3 ) for t ∈ [0, 1] such that f0 is the identity and
f1 stabilizes Cn set-wise, preserving orientations of the loops. Two motions ft,0 and ft,1
are considered equivalent if they are smoothly isotopic via an isotopy ft,s with f0,s and
f1,s setwise stabilizing Cn . If f1 also stabilizes each γ ∈ L then the motion ft is a pure
motion. These constructions and the above ones yield isomorphic groups, that is to say
LB n is the group of motions, and PLB n is the group of pure motions. This is explained,
e.g., in [Gol81] and [Wil12, Section 3]. One should picture σi as the motion in which the
ith and (i + 1)st loops move around each other and take each other’s old spots. Then βi
is similar, except that during the motion the (i + 1)st loop passes through the ith instead
of around. See Figure 12 for an idea.
i+1
i
i+1
i
σi
βi
Figure 12. Generators of LB n .
There is a bit of inconsistency in the literature: all that we have described here is as in,
e.g., [FRR97], but in, e.g., [BH13], instead of the generators βi their inverses are used
(called ρi there), and then the relevant relations look slightly different.
48
S. WITZEL AND M. C. B. ZAREMSKY
In [BWC07] there are some helpful diagrams, analogous to strand diagrams for braids,
illustrating elements of LB n . The pictures are four-dimensional, and show one loop
passing through another in a sort of movie. Using a bit of artistic license, we can draw
similar diagrams to demonstrate cloning; see Figure 13.
=
Figure 13. An example of cloning, namely (β1 )κ22 = β2 β1 . The picture
shows β1 λ2 = λ1 β2 β1 . The vertical direction is time, while the missing
spatial direction is indicated by breaking the surfaces; see [BWC07, p. 717]
for a detailed explanation.
Alternatively we can draw cloning using the welded braid diagrams from [FRR97]. See
Figure 14.
=
Figure 14. Another example of cloning, now using welded braid diagrams.
We see that σ1 β2 λ3 = λ1 σ2 σ1 β3 β2 .
One might expect the descending links to be modeled on disjoint “tubes” in 3-space with
prescribed boundaries, or “welded arcs” of some sort. This is in analogy to the disjoint
arcs in 2-space with prescribed boundaries for descending links in the braid group case.
References
[AB87]
[BBCS08]
[BdCK15]
[BE74]
[Bel04]
[BF]
H. Abels and K. S. Brown. Finiteness properties of solvable S-arithmetic groups: an example.
J. Pure Appl. Algebra, 44(1-3):77–83, 1987.
T. Brady, J. Burillo, S. Cleary, and M. Stein. Pure braid subgroups of braided Thompson’s
groups. Publ. Mat., 52(1):57–89, 2008.
L. Bartholdi, Y. de Cornulier, and D. H. Kochloukova. Homological finiteness properties of
wreath products. Q. J. Math., 66(2):437–457, 2015.
R. Bieri and B. Eckmann. Finiteness properties of duality groups. Comment. Math. Helv.,
49:74–83, 1974.
J. Belk. Thompson’s Group F . PhD thesis, Cornell University, 2004.
J. M. Belk and B. Forrest. A Thompson Group for the Basilica. arXiv:1201.4225.
THOMPSON GROUPS FOR SYSTEMS OF GROUPS, AND THEIR FINITENESS PROPERTIES
[BFM+ 14]
49
K.-U. Bux, M. Fluch, M. Marschler, S. Witzel, and M. C. B. Zaremsky. The braided Thompson’s groups are of type F∞ . J. Reine Angew. Math., 2014. To appear.
[BH99]
M. R. Bridson and A. Haefliger. Metric spaces of non-positive curvature, volume 319 of
Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, Berlin, 1999.
[BH13]
T. Brendle and A. Hatcher. Configuration spaces of rings and wickets. Comment. Math. Helv.,
88(1):131–162, 2013.
[Bie76]
R. Bieri. Homological dimension of discrete groups. Mathematics Department, Queen Mary
College, London, 1976.
[Bri04]
M. G. Brin. Higher dimensional Thompson groups. Geom. Dedicata, 108:163–192, 2004.
[Bri05]
M. G. Brin. On the Zappa-Szép product. Comm. Algebra, 33(2):393–424, 2005.
[Bri07]
M. G. Brin. The algebra of strand splitting. I. A braided version of Thompson’s group V . J.
Group Theory, 10(6):757–788, 2007.
[Bro82]
K. S. Brown. Cohomology of groups, volume 87 of Graduate Texts in Mathematics. SpringerVerlag, New York, 1982.
[Bro87]
K. S. Brown. Finiteness properties of groups. J. Pure Appl. Algebra, 44(1-3):45–75, 1987.
[Bro92]
K. S. Brown. The geometry of finitely presented infinite simple groups. In Algorithms and
classification in combinatorial group theory (Berkeley, CA, 1989), volume 23 of Math. Sci.
Res. Inst. Publ., pages 121–136. Springer, New York, 1992.
[Bro06]
K. S. Brown. The homology of Richard Thompson’s group F . In Topological and asymptotic
aspects of group theory, volume 394 of Contemp. Math., pages 47–59. Amer. Math. Soc.,
Providence, RI, 2006.
[Bux04]
K.-U. Bux. Finiteness properties of soluble arithmetic groups over global function fields.
Geom. Topol., 8:611–644, 2004.
[BWC07]
J. C. Baez, D. K. Wise, and A. S. Crans. Exotic statistics for strings in 4D BF theory. Adv.
Theor. Math. Phys., 11(5):707–749, 2007.
[BZFGM14] R. Berns-Zieve, D. Fry, J. Gillings, and H. Mathews. Groups with context-free co-word problem and embeddings into Thompson’s group V. The Journal of the Summer Undergraduate
Mathematical Science Research Institute (SUMSRI), 2014. arXiv:1407.7745.
[Col89]
D. J. Collins. Cohomological dimension and symmetric automorphisms of a free group. Comment. Math. Helv., 64(1):44–61, 1989.
[CP61]
A. H. Clifford and G. B. Preston. The algebraic theory of semigroups. Vol. I. Mathematical
Surveys, No. 7. American Mathematical Society, Providence, R.I., 1961.
[DJS03]
M. Davis, T. Januszkiewicz, and R. Scott. Fundamental groups of blow-ups. Adv. Math.,
177(1):115–179, 2003.
[Far03]
D. S. Farley. Finiteness and CAT(0) properties of diagram groups. Topology, 42(5):1065–1082,
2003.
[FMWZ13] M. G. Fluch, M. Marschler, S. Witzel, and M. C. B. Zaremsky. The Brin–Thompson groups
sV are of type F∞ . Pacific J. Math., 266(2):283–295, 2013.
[FRR97]
R. Fenn, R. Rimányi, and C. Rourke. The braid-permutation group. Topology, 36(1):123–135,
1997.
[Geo08]
R. Geoghegan. Topological Methods in Group Theory, volume 243 of Graduate Texts in Mathematics. Springer, 2008.
[Gol81]
D. L. Goldsmith. The theory of motion groups. Michigan Math. J., 28(1):3–17, 1981.
[GS97]
V. Guba and M. Sapir. Diagram groups. Mem. Amer. Math. Soc., 130(620), 1997.
[Hat01]
A. Hatcher. Algebraic Topology. Cambridge University Press, 2001.
[Hig74]
G. Higman. Finitely presented infinite simple groups. Department of Pure Mathematics, Department of Mathematics, I.A.S. Australian National University, Canberra, 1974. Notes on
Pure Mathematics, No. 8 (1974).
[Hug09]
B. Hughes. Local similarities and the Haagerup property. Groups Geom. Dyn., 3(2):299–315,
2009. With an appendix by D. S. Farley.
[KM97]
S. Krstić and J. McCool. The non-finite presentability of IA(F3 ) and GL2 (Z[t, t−1 ]). Invent.
Math., 129(3):595–606, 1997.
[Koz08]
D. Kozlov. Combinatorial Algebraic Topology. Springer, 2008.
[KS06]
M. Kashiwara and P. Schapira. Categories and sheaves, volume 332 of Grundlehren der
Mathematischen Wissenschaften. Springer, 2006.
[MPN13]
C. Martı́nez-Pérez and B. E. A. Nucinkis. Bredon cohomological finiteness conditions for
generalisations of Thompson groups. Groups Geom. Dyn., 7(4):931–959, 2013.
[Put]
A. Putman. Representation stability, congruence subgroups, and mapping class groups.
arXiv:1201.4876v2.
[Qui78]
D. Quillen. Homotopy properties of the poset of nontrivial p-subgroups of a group. Adv. in
Math., 28(2):101–128, 1978.
50
[Röv99]
[Spa66]
[Ste92]
[Tan14]
[Wil12]
S. WITZEL AND M. C. B. ZAREMSKY
C. E. Röver. Constructing finitely presented simple groups that contain Grigorchuk groups.
J. Algebra, 220(1):284–313, 1999.
E. H. Spanier. Algebraic Topology. Springer, 1966.
M. Stein. Groups of piecewise linear homeomorphisms. Trans. Amer. Math. Soc., 332(2):477–
514, 1992.
S. Tanusevski. Generalized Thompson groups. PhD thesis, Binghamton University, 2014.
J. C. H. Wilson. Representation stability for the cohomology of the pure string motion groups.
Algebr. Geom. Topol., 12(2):909–931, 2012.
Mathematical Institute, University of Münster, Einsteinstraße 62, 48149 Münster, Germany
E-mail address: [email protected]
Department of Mathematical Sciences, Binghamton University, Binghamton, NY 13902
E-mail address: [email protected]
| 4math.GR
|
Logical Methods in Computer Science
Vol. 1 (2:5) 2005, pp. 1–39
www.lmcs-online.org
Submitted
Published
Apr. 1, 2005
Nov. 8, 2005
AN OPERATIONAL FOUNDATION FOR DELIMITED
CONTINUATIONS IN THE CPS HIERARCHY
MALGORZATA BIERNACKA, DARIUSZ BIERNACKI, AND OLIVIER DANVY
BRICS, Department of Computer Science, University of Aarhus
IT-parken, Aabogade 34, DK-8200 Aarhus N, Denmark
e-mail address: {mbiernac,dabi,danvy}@brics.dk
Abstract. We present an abstract machine and a reduction semantics for the lambdacalculus extended with control operators that give access to delimited continuations in
the CPS hierarchy. The abstract machine is derived from an evaluator in continuationpassing style (CPS); the reduction semantics (i.e., a small-step operational semantics with
an explicit representation of evaluation contexts) is constructed from the abstract machine;
and the control operators are the shift and reset family.
We also present new applications of delimited continuations in the CPS hierarchy:
finding list prefixes and normalization by evaluation for a hierarchical language of units
and products.
1. Introduction
The studies of delimited continuations can be classified in two groups: those that use
continuation-passing style (CPS) and those that rely on operational intuitions about control instead. Of the latter, there is a large number proposing a variety of control operators [5, 37, 40, 41, 49, 52, 53, 65, 70, 74, 80] which have found applications in models of control,
concurrency, and type-directed partial evaluation [8,52,75]. Of the former, there is the work
revolving around the family of control operators shift and reset [27–29,32,42,43,55,56,66,80]
which have found applications in non-deterministic programming, code generation, partial evaluation, normalization by evaluation, computational monads, and mobile computing [6, 7, 9, 17, 22, 23, 33, 34, 44, 46, 48, 51, 57, 59, 61, 72, 77–79].
The original motivation for shift and reset was a continuation-based programming pattern involving several layers of continuations. The original specification of these operators
relied both on a repeated CPS transformation and on an evaluator with several layers
of continuations (as is obtained by repeatedly transforming a direct-style evaluator into
continuation-passing style). Only subsequently have shift and reset been specified operationally, by developing operational analogues of a continuation semantics and of the CPS
transformation [32].
2000 ACM Subject Classification: D.1.1; F.3.2.
Key words and phrases: Delimited continuations, abstract machines, reduction semantics.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-1 (2:5) 2005
CC
M. Biernacka, D. Biernacki, and O. Danvy
Creative Commons
2
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
The goal of our work here is to establish a new operational foundation for delimited
continuations, using CPS as a guideline. To this end, we start with the original evaluator
for shift1 and reset1 . This evaluator uses two layers of continuations: a continuation and a
meta-continuation. We then defunctionalize it into an abstract machine [1] and we construct
the corresponding reduction semantics [36], as pioneered by Felleisen and Friedman [39].
The development scales to shiftn and resetn . It is reusable for any control operators that
are compatible with CPS, i.e., that can be characterized with a (possibly iterated) CPS
translation or with a continuation-based evaluator. It also pinpoints where operational
intuitions go beyond CPS.
This article is structured as follows. In Section 2, we review the enabling technology of our work: Reynolds’s defunctionalization, the observation that a defunctionalized
CPS program implements an abstract machine, and the observation that Felleisen’s evaluation contexts are the defunctionalized continuations of a continuation-passing evaluator;
we demonstrate this enabling technology on a simple example, arithmetic expressions. In
Section 3, we illustrate the use of shift and reset with the classic example of finding list
prefixes, using an ML-like programming language. In Section 4, we then present our main
result: starting from the original evaluator for shift and reset, we defunctionalize it into
an abstract machine; we analyze this abstract machine and construct the corresponding
reduction semantics. In Section 5, we extend this result to the CPS hierarchy. In Section 6,
we illustrate the CPS hierarchy with a class of normalization functions for a hierarchical
language of units and products.
2. From evaluator to reduction semantics for arithmetic expressions
We demonstrate the derivation from an evaluator to a reduction semantics. The derivation consists of the following steps:
(1) we start from an evaluator for a given language; if it is in direct style, we CPStransform it;
(2) we defunctionalize the CPS evaluator, obtaining a value-based abstract machine;
(3) we modify the abstract machine to make it term-based instead of value-based; in
particular, if the evaluator uses an environment, then so does the corresponding
value-based abstract machine, and in that case, making the machine term-based
leads us to use substitutions rather than an environment;
(4) we analyze the transitions of the term-based abstract machine to identify the evaluation strategy it implements and the set of reductions it performs; the result is a
reduction semantics.
The first two steps are based on previous work on a functional correspondence between
evaluators and abstract machines [1–3, 17, 26], which itself is based on Reynolds’s seminal
work on definitional interpreters [71]. The last two steps follow the lines of Felleisen and
Friedman’s original work on a reduction semantics for the call-by-value λ-calculus extended
with control operators [39]. The last step has been studied further by Hardin, Maranget,
and Pagano [50] in the context of explicit substitutions and by Biernacka, Danvy, and
Nielsen [15, 16, 31].
In the rest of this section, our running example is the language of arithmetic expressions,
formed using natural numbers (the values) and additions (the computations):
exp ∋ e ::= pmq | e1 + e2
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY
3
• Values: val ∋ v ::= m
• Evaluation function: eval : exp → val
eval (pmq) = m
eval (e1 + e2 ) = eval (e1 ) + eval (e2 )
• Main function:
evaluate : exp → val
evaluate (e) = eval (e)
Figure 1: A direct-style evaluator for arithmetic expressions
• Values: val ∋ v ::= m
• Continuations: cont = val → val
• Evaluation function: eval : exp × cont → val
eval (pmq, k) = k m
eval (e1 + e2 , k) = eval (e1 , λm1 . eval (e2 , λm2 . k (m1 + m2 )))
• Main function:
evaluate : exp → val
evaluate (e) = eval (e, λv. v)
Figure 2: A continuation-passing evaluator for arithmetic expressions
2.1. The starting point: an evaluator in direct style. We define an evaluation function for arithmetic expressions by structural induction on their syntax. The resulting directstyle evaluator is displayed in Figure 1.
2.2. CPS transformation. We CPS-transform the evaluator by naming intermediate results, sequentializing their computation, and introducing an extra functional parameter,
the continuation [29, 68, 76]. The resulting continuation-passing evaluator is displayed in
Figure 2.
2.3. Defunctionalization. The generalization of closure conversion [60] to defunctionalization is due to Reynolds [71]. The goal is to represent a functional value with a first-order
data structure. The means is to partition the function space into a first-order sum where
each summand corresponds to a lambda-abstraction in the program. In a defunctionalized
program, function introduction is thus represented as an injection, and function elimination
as a call to a first-order apply function implementing a case dispatch. In an ML-like functional language, sums are represented as data types, injections as data-type constructors,
and apply functions are defined by case over the corresponding data types [30].
Here, we defunctionalize the continuation of the continuation-passing evaluator in Figure 2. We thus need to define a first-order algebraic data type and its apply function. To
this end, we enumerate the lambda-abstractions that give rise to the inhabitants of this
function space; there are three: the initial continuation in evaluate and the two continuations in eval. The initial continuation is closed, and therefore the corresponding algebraic
constructor is nullary. The two other continuations have two free variables, and therefore
4
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
• Values: val ∋ v ::= m
• Defunctionalized continuations: cont ∋ k ::= [ ] | ADD2 (e, k) | ADD1 (v, k)
• Functions eval : exp × cont → val and apply cont : cont × val → val:
eval (pmq, k) = apply cont (k, m)
eval (e1 + e2 , k) = eval (e1 , ADD2 (e2 , k))
apply cont ([ ], v) = v
apply cont (ADD2 (e2 , k), v1 ) = eval (e2 , ADD1 (v1 , k))
apply cont (ADD1 (m1 , k), m2 ) = apply cont (k, m1 + m2 )
• Main function: evaluate : exp → val
evaluate (e) = eval (e, [ ])
Figure 3: A defunctionalized continuation-passing evaluator for arithmetic expressions
the corresponding constructors are binary. As for the apply function, it interprets the
algebraic constructors. The resulting defunctionalized evaluator is displayed in Figure 3.
2.4. Abstract machines as defunctionalized continuation-passing programs. Elsewhere [1, 26], we have observed that a defunctionalized continuation-passing program implements an abstract machine: each configuration is the name of a function together with
its arguments, and each function clause represents a transition. (As a corollary, we have
also observed that the defunctionalized continuation of an evaluator forms what is known
as an ‘evaluation context’ [25, 30, 39].)
Indeed Plotkin’s Indifference Theorem [68] states that continuation-passing programs
are independent of their evaluation order. In Reynolds’s words [71], all the subterms in
applications are ‘trivial’; and in Moggi’s words [64], these subterms are values and not
computations. Furthermore, continuation-passing programs are tail recursive [76]. Therefore, since in a continuation-passing program all calls are tail calls and all subcomputations
are elementary, a defunctionalized continuation-passing program implements a transition
system [69], i.e., an abstract machine.
We thus reformat Figure 3 into Figure 4. The correctness of the abstract machine with
respect to the initial evaluator follows from the correctness of CPS transformation and of
defunctionalization.
2.5. From value-based abstract machine to term-based abstract machine. We
observe that the domain of expressible values in Figure 4 can be embedded in the syntactic
domain of expressions. We therefore adapt the abstract machine to work on terms rather
than on values. The result is displayed in Figure 5; it is a syntactic theory [36].
2.6. From term-based abstract machine to reduction semantics. The method of
deriving a reduction semantics from an abstract machine was introduced by Felleisen and
Friedman [39] to give a reduction semantics for control operators. Let us demonstrate it.
We analyze the transitions of the abstract machine in Figure 5. The second component
of eval -transitions—the stack representing “the rest of the computation”—has already been
identified as the evaluation context of the currently processed expression. We thus read a
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY
5
• Values: v ::= m
• Evaluation contexts: C ::= [ ] | ADD2 (e, C) | ADD1 (v, C)
• Initial transition, transition rules, and final transition:
e ⇒ he, [ ]ieval
hpmq, Cieval ⇒ hC, micont
he1 + e2 , Cieval
hADD2 (e2 , C), v1 icont
⇒ he1 , ADD2 (e2 , C)ieval
⇒ he2 , ADD1 (v1 , C)ieval
hADD1 (m1 , C), m2 icont
h[ ], vicont
⇒ hC, m1 + m2 icont
⇒ v
Figure 4: A value-based abstract machine for evaluating arithmetic expressions
• Expressions and values: e ::= v | e1 + e2
v ::= pmq
• Evaluation contexts: C ::= [ ] | ADD2 (e, C) | ADD1 (v, C)
• Initial transition, transition rules, and final transition:
e ⇒ he, [ ]ieval
hpmq, Cieval
he1 + e2 , Cieval
⇒ hC, pmqicont
⇒ he1 , ADD2 (e2 , C)ieval
hADD2 (e2 , C), v1 icont
hADD1 (pm1q, C), pm2qicont
⇒ he2 , ADD1 (v1 , C)ieval
⇒ hC, pm1 + m2qicont
h[ ], vicont
⇒ v
Figure 5: A term-based abstract machine for processing arithmetic expressions
configuration he, Cieval as a decomposition of some expression into a sub-expression e and
an evaluation context C.
Next, we identify the reduction and decomposition rules in the transitions of the machine. Since a configuration can be read as a decomposition, we compare the left-hand side
and the right-hand side of each transition. If they represent the same expression, then the
given transition defines a decomposition (i.e., it searches for the next redex according to
some evaluation strategy); otherwise we have found a redex. Moreover, reading the decomposition rules from right to left defines a ‘plug’ function that reconstructs an expression
from its decomposition.
Here the decomposition function as read off the abstract machine is total. In general,
however, it may be undefined for stuck terms; one can then extend it straightforwardly into
a total function that decomposes a term into a context and a potential redex, i.e., an actual
redex (as read off the machine), or a stuck redex.
6
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
In this simple example there is only one reduction rule. This rule performs the addition
of natural numbers:
(add) C [pm1q + pm2q] → C [pm1 + m2q]
The remaining transitions decompose an expression according to the left-to-right strategy.
2.7. From reduction semantics to term-based abstract machine. In Section 2.6,
we have constructed the reduction semantics corresponding to the abstract machine of
Figure 5, as pioneered by Felleisen and Friedman [38, 39]. Over the last few years [15,
16, 24, 31], Biernacka, Danvy, and Nielsen have studied the converse transformation and
systematized the construction of an abstract machine from a reduction semantics. The
main idea is to short-cut the decompose-contract-plug loop, in the definition of evaluation
as the transitive closure of one-step reduction, into a refocus-contract loop. The refocus
function is constructed as an efficient (i.e., deforested) composition of plug and decompose
that maps a term and a context either to a value or to a redex and a context. The result
is a ‘pre-abstract machine’ computing the transitive closure of the refocus function. This
pre-abstract machine can then be simplified into an eval/apply abstract machine.
It is simple to verify that using refocusing, one can go from the reduction semantics of
Section 2.6 to the eval/apply abstract machine of Figure 5.
2.8. Summary and conclusion. We have demonstrated how to derive an abstract machine out of an evaluator, and how to construct the corresponding reduction semantics out
of this abstract machine. In Section 4, we apply this derivation and this construction to the
first level of the CPS hierarchy, and in Section 5, we apply them to an arbitrary level of the
CPS hierarchy. But first, let us illustrate how to program with delimited continuations.
3. Programming with delimited continuations
We present two examples of programming with delimited continuations. Given a list
xs and a predicate p, we want
(1) to find the first prefix of xs whose last element satisfies p, and
(2) to find all such prefixes of xs.
For example, given the predicate λm.m > 2 and the list [0, 3, 1, 4, 2, 5], the first prefix is
[0, 3] and the list of all the prefixes is [[0, 3], [0, 3, 1, 4], [0, 3, 1, 4, 2, 5]].
In Section 3.1, we start with a simple solution that uses a first-order accumulator.
This simple solution is in defunctionalized form. In Section 3.2, we present its higher-order
counterpart, which uses a functional accumulator. This functional accumulator acts as a
delimited continuation. In Section 3.3, we present its direct-style counterpart (which uses
shift and reset) and in Section 3.4, we present its continuation-passing counterpart (which
uses two layers of continuations). In Section 3.5, we introduce the CPS hierarchy informally.
We then mention a typing issue in Section 3.6 and review related work in Section 3.7.
3.1. Finding prefixes by accumulating lists. A simple solution is to accumulate the
prefix of the given list in reverse order while traversing this list and testing each of its
elements:
• if no element satisfies the predicate, there is no prefix and the result is the empty
list;
• otherwise, the prefix is the reverse of the accumulator.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY
find first prefix a (p, xs)
def
=
letrec visit (nil , a)
= nil
| visit (x :: xs, a)
= let a ′ = x :: a
in if p x
then reverse (a ′ , nil )
else visit (xs, a ′ )
and reverse (nil , xs)
= xs
| reverse (x :: a, xs)
= reverse (a, x :: xs)
in visit (xs, nil)
find all prefixes a (p, xs)
def
letrec visit (nil , a)
= nil
| visit (x :: xs, a)
= let a ′ = x :: a
in if p x
then (reverse (a ′ , nil)) :: (visit (xs, a ′ ))
else visit (xs, a ′ )
and reverse (nil , xs)
= xs
| reverse (x :: a, xs)
= reverse (a, x :: xs)
in visit (xs, nil)
=
7
To find the first prefix, one stops as soon as a satisfactory list element is found. To list
all the prefixes, one continues the traversal, adding the current prefix to the list of the
remaining prefixes.
We observe that the two solutions are in defunctionalized form [30,71]: the accumulator
has the data type of a defunctionalized function and reverse is its apply function. We present
its higher-order counterpart next [54].
3.2. Finding prefixes by accumulating list constructors. Instead of accumulating the
prefix in reverse order while traversing the given list, we accumulate a function constructing
the prefix:
• if no element satisfies the predicate, the result is the empty list;
• otherwise, we apply the functional accumulator to construct the prefix.
8
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
find first prefix c 1 (p, xs)
def
=
letrec visit (nil , k )
= nil
| visit (x :: xs, k )
= let k ′ = λvs.k (x :: vs)
in if p x
then k ′ nil
else visit (xs, k ′ )
in visit (xs, λvs.vs)
find all prefixes c 1 (p, xs)
def
letrec visit (nil , k )
= nil
| visit (x :: xs, k )
= let k ′ = λvs.k (x :: vs)
in if p x
then (k ′ nil) :: (visit (xs, k ′ ))
else visit (xs, k ′ )
in visit (xs, λvs.vs)
=
To find the first prefix, one applies the functional accumulator as soon as a satisfactory list
element is found. To list all such prefixes, one continues the traversal, adding the current
prefix to the list of the remaining prefixes.
Defunctionalizing these two definitions yields the two definitions of Section 3.1.
The functional accumulator is a delimited continuation:
• In find first prefix c 1 , visit is written in CPS since all calls are tail calls and all
sub-computations are elementary. The continuation is initialized in the initial call
to visit, discarded in the base case, extended in the induction case, and used if a
satisfactory prefix is found.
• In find all prefixes c 1 , visit is almost written in CPS except that the continuation is
composed if a satisfactory prefix is found: it is used twice—once where it is applied
to the empty list to construct a prefix, and once in the visit of the rest of the list to
construct a list of prefixes; this prefix is then prepended to this list of prefixes.
These continuation-based programming patterns (initializing a continuation, not using it,
or using it more than once as if it were a composable function) have motivated the control
operators shift and reset [28, 29]. Using them, in the next section, we write visit in direct
style.
3.3. Finding prefixes in direct style. The two following local functions are the directstyle counterpart of the two local functions in Section 3.2:
find first prefix c 0 (p, xs)
def
=
letrec visit nil
= Sk.nil
| visit (x :: xs)
= x :: (if p x then nil else visit xs)
in h visit xsii
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY
find all prefixes c 0 (p, xs)
def
=
9
letrec visit nil
= Sk.nil
| visit (x :: xs)
= x :: if p x
then Sk′ .hhk ′ nilii :: h k′ (visit xs)ii
else visit xs
in h visit xsii
In both cases, visit is in direct style, i.e., it is not passed any continuation. The initial calls
to visit are enclosed in the control delimiter reset (noted h ·ii for conciseness). In the base
cases, the current (delimited) continuation is captured with the control operator shift (noted
S), which has the effect of emptying the (delimited) context; this captured continuation is
bound to an identifier k, which is not used; nil is then returned in the emptied context.
In the induction case of find all prefixes c 0 , if the predicate is satisfied, visit captures the
current continuation and applies it twice—once to the empty list to construct a prefix, and
once to the result of visiting the rest of the list to construct a list of prefixes; this prefix is
then prepended to the list of prefixes.
CPS-transforming these two local functions yields the two definitions of Section 3.2 [29].
3.4. Finding prefixes in continuation-passing style. The two following local functions
are the continuation-passing counterpart of the two local functions in Section 3.2:
def
find first prefix c 2 (p, xs) = letrec visit (nil , k1 , k2 )
= k2 nil
| visit (x :: xs, k1 , k2 )
= let k1′ = λ(vs, k2′ ).k1 (x :: vs, k2′ )
in if p x
then k1′ (nil , k2 )
else visit (xs, k1′ , k2 )
in visit (xs, λ(vs, k2 ).k2 vs, λvs.vs)
def
find all prefixes c 2 (p, xs) = letrec visit (nil , k1 , k2 )
= k2 nil
| visit (x :: xs, k1 , k2 )
= let k1′ = λ(vs, k2′ ).k1 (x :: vs, k2′ )
in if p x
then k1′ (nil , λvs.visit (xs, k1′ , λvss.k2 (vs :: vss)))
else visit (xs, k1′ , k2 )
in visit (xs, λ(vs, k2 ).k2 vs, λvss.vss)
CPS-transforming the two local functions of Section 3.2 adds another layer of continuations and restores the syntactic characterization of all calls being tail calls and all subcomputations being elementary.
3.5. The CPS hierarchy. If k2 were used non-tail recursively in a variant of the examples
of Section 3.4, we could CPS-transform the definitions one more time, adding one more layer
of continuations and restoring the syntactic characterization of all calls being tail calls and
all sub-computations being elementary. We could also map this definition back to direct
style, eliminating k2 but accessing it with shift. If the result were mapped back to direct
10
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
style one more time, k2 would then be accessed with a new control operator, shift2 , and k1
would be accessed with shift (or more precisely with shift1 ).
All in all, successive CPS-transformations induce a CPS hierarchy [28,32], and abstracting control up to each successive layer is achieved with successive pairs of control operators
shift and reset—reset to initialize the continuation up to a level, and shift to capture a
delimited continuation up to this level. Each pair of control operators is indexed by the
corresponding level in the hierarchy. Applying a captured continuation packages all the
current layers on the next layer and restores the captured layers. When a captured continuation completes, the packaged layers are put back into place and the computation proceeds.
(This informal description is made precise in Section 4.)
3.6. A note about typing. The type of find all prefixes c 1 , in Section 3.2, is
(α → bool ) × α list → α list list
and the type of its local function visit is
α list × (α list → α list) → α list list.
In this example, the co-domain of the continuation is not the same as the co-domain of
visit.
Thus find all prefixes c 0 provides a simple and meaningful example where Filinski’s
typing of shift [42] does not fit, since it must be used at type
((β → ans) → ans) → β
for a given type ans, i.e., the answer type of the continuation and the type of the computation must be the same. In other words, control effects are not allowed to change the types
of the contexts. Due to a similar restriction on the type of shift, the example does not fit
either in Murthy’s pseudo-classical type system for the CPS hierarchy [66] and in Wadler’s
most general monadic type system [80, Section 3.4]. It however fits in Danvy and Filinski’s
original type system [27] which Ariola, Herbelin, and Sabry have recently embedded in
classical subtractive logic [5].
3.7. Related work. The example considered in this section builds on the simpler function
that unconditionally lists the successive prefixes of a given list. This simpler function is a
traditional example of delimited continuations [21, 73]:
• In the Lisp Pointers [21], Danvy presents three versions of this function: a typed
continuation-passing version (corresponding to Section 3.2), one with delimited control (corresponding to Section 3.3), and one in assembly language.
• In his PhD thesis [73, Section 6.3], Sitaram presents two versions of this function:
one with an accumulator (corresponding to Section 3.1) and one with delimited
control (corresponding to Section 3.3).
In Section 3.2, we have shown that the continuation-passing version mediates the version
with an accumulator and the version with delimited control since defunctionalizing the
continuation-passing version yields one and mapping it back to direct style yields the other.
3.8. Summary and conclusion. We have illustrated delimited continuations with the
classic example of finding list prefixes, using CPS as a guideline. Direct-style programs using
shift and reset can be CPS-transformed into continuation-passing programs where some
calls may not be tail calls and some sub-computations may not be elementary. One more
CPS transformation establishes this syntactic property with a second layer of continuations.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 11
• Terms: term ∋ t ::= pmq | x | λx .t | t0 t1 | succ t | h tii | Sk .t
• Values: val ∋ v ::= m | f
• Answers, meta-continuations, continuations and functions:
ans = val
k2 ∈ cont2 = val → ans
k1 ∈ cont1 = val × cont2 → ans
f ∈ fun = val × cont1 × cont2 → ans
• Initial continuation and meta-continuation: θ1 = λ(v, k2 ). k2 v
θ2 = λv. v
• Environments: env ∋ e ::= eempty | e[x 7→ v]
• Evaluation function: eval : term × env × cont1 × cont2 → ans
eval (pmq,
eval (x ,
eval (λx .t,
eval (t0 t1 ,
eval (succ t,
eval (hhtii ,
eval (Sk.t,
e,
e,
e,
e,
e,
e,
e,
k1 ,
k1 ,
k1 ,
k1 ,
k1 ,
k1 ,
k1 ,
k2 )
k2 )
k2 )
k2 )
k2 )
k2 )
k2 )
=
=
=
=
=
=
=
k1 (m, k2 )
k1 (e(x ), k2 )
k1 (λ(v, k1′ , k2′ ). eval (t, e[x 7→ v], k1′ , k2′ ), k2 )
eval (t0 , e, λ(f, k2′ ). eval (t1 , e, λ(v, k2′′ ). f (v, k1 , k2′′ ), k2′ ), k2 )
eval (t, e, λ(m, k2′ ). k1 (m + 1, k2′ ), k2 )
eval (t, e, θ1 , λv. k1 (v, k2 ))
eval (t, e[k 7→ c], θ1 , k2 )
where c = λ(v, k1′ , k2′ ). k1 (v, λv ′ . k1′ (v ′ , k2′ ))
• Main function: evaluate : term → val
evaluate (t) = eval (t, eempty , θ1 , θ2 )
Figure 6: An environment-based evaluator for the first level of the CPS hierarchy
Further CPS transformations provide the extra layers of continuation that are characteristic
of the CPS hierarchy.
In the next section, we specify the λ-calculus extended with shift and reset.
4. From evaluator to reduction semantics for delimited continuations
We derive a reduction semantics for the call-by-value λ-calculus extended with shift and
reset, using the method demonstrated in Section 2. First, we transform an evaluator into
an environment-based abstract machine. Then we eliminate the environment from this
abstract machine, making it substitution-based. Finally, we read all the components of a
reduction semantics off the substitution-based abstract machine.
Terms consist of integer literals, variables, λ-abstractions, function applications, applications of the successor function, reset expressions, and shift expressions:
t ::= pmq | x | λx .t | t0 t1 | succ t | h tii | Sk .t
Programs are closed terms.
12
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
This source language is a subset of the language used in the examples of Section 3.
Adding the remaining constructs is a straightforward exercise and does not contribute to
our point here.
4.1. An environment-based evaluator. Figure 6 displays an evaluator for the language
of the first level of the CPS hierarchy. This evaluation function represents the original
call-by-value semantics of the λ-calculus with shift and reset [28], augmented with integer
literals and applications of the successor function. It is defined by structural induction
over the syntax of terms, and it makes use of an environment e, a continuation k1 , and a
meta-continuation k2 .
The evaluation of a terminating program that does not get stuck (i.e., a program where
no ill-formed applications occur in the course of evaluation) yields either an integer, a
function representing a λ-abstraction, or a captured continuation. Both evaluate and eval
are partial functions to account for non-terminating or stuck programs. The environment
stores previously computed values of the free variables of the term under evaluation.
The meta-continuation intervenes to interpret reset expressions and to apply captured
continuations. Otherwise, it is passively threaded through the evaluation of literals, variables, λ-abstractions, function applications, and applications of the successor function. (If
it were not for shift and reset, and if eval were curried, k2 could be eta-reduced and the
evaluator would be in ordinary continuation-passing style.)
The reset control operator is used to delimit control. A reset expression h tii is interpreted by evaluating t with the initial continuation and a meta-continuation on which the
current continuation has been “pushed.” (Indeed, and as will be shown in Section 4.2,
defunctionalizing the meta-continuation yields the data type of a stack [30].)
The shift control operator is used to abstract (delimited) control. A shift expression
Sk.t is interpreted by capturing the current continuation, binding it to k, and evaluating t
in an environment extended with k and with a continuation reset to the initial continuation.
Applying a captured continuation is achieved by “pushing” the current continuation on the
meta-continuation and applying the captured continuation to the new meta-continuation.
Resuming a continuation is achieved by reactivating the “pushed” continuation with the
corresponding meta-continuation.
4.2. An environment-based abstract machine. The evaluator displayed in Figure 6
is already in continuation-passing style. Therefore, we only need to defunctionalize its expressible values and its continuations to obtain an abstract machine. This abstract machine
is displayed in Figure 7.
The abstract machine consists of three sets of transitions: eval for interpreting terms,
cont 1 for interpreting the defunctionalized continuations (i.e., the evaluation contexts),1 and
cont 2 for interpreting the defunctionalized meta-continuations (i.e., the meta-contexts).2
The set of possible values includes integers, closures and captured contexts. In the original
1The grammar of evaluation contexts in Figure 7 is isomorphic to the grammar of evaluation contexts in
the standard inside-out notation:
C1
::=
[ ] | C1 [[ ] (t, e)] | C1 [succ [ ]] | C1 [v [ ]]
2To build on Peyton Jones’s terminology [62], this abstract machine is therefore in ‘eval/apply/metaapply’ form.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 13
evaluator, the latter two were represented as higher-order functions, but defunctionalizing
expressible values of the evaluator has led them to be distinguished.
This eval/apply/meta-apply abstract machine is an extension of the CEK machine [39],
which is an eval/apply machine, with the meta-context C2 and its two transitions, and
the two transitions for shift and reset. C2 intervenes to process reset expressions and to
apply captured continuations. Otherwise, it is passively threaded through the processing of
literals, variables, λ-abstractions, function applications, and applications of the successor
function. (If it were not for shift and reset, C2 and its transitions could be omitted and the
abstract machine would reduce to the CEK machine.)
Given an environment e, a context C1 , and a meta-context C2 , a reset expression h tii
is processed by evaluating t with the same environment e, the empty context •, and a
meta-context where C1 has been pushed on C2 .
Given an environment e, a context C1 , and a meta-context C2 , a shift expression Sk.t is
processed by evaluating t with an extension of e where k denotes C1 , the empty context [ ],
and a meta-context C2 . Applying a captured context C1′ is achieved by pushing the current
context C1 on the current meta-context C2 and continuing with C1′ . Resuming a context
C1 is achieved by popping it off the meta-context C2 · C1 and continuing with C1 .
The correctness of the abstract machine with respect to the evaluator is a consequence
of the correctness of defunctionalization. In order to express it formally, we define a partial
function evale mapping a term t to a value v whenever the environment-based machine,
started with t, stops with v. The following theorem states this correctness by relating
observable results:
Theorem 1. For any program t and any integer value m, evaluate (t) = m if and only if
evale (t) = m.
Proof. The theorem follows directly from the correctness of defunctionalization [10, 67].
The environment-based abstract machine can serve both as a foundation for implementing functional languages with control operators for delimited continuations and as a
stepping stone in theoretical studies of shift and reset. In the rest of this section, we use it
to construct a reduction semantics of shift and reset.
4.3. A substitution-based abstract machine. The environment-based abstract machine of Figure 7, on which we want to base our development, makes a distinction between
terms and values. Since a reduction semantics is specified by purely syntactic operations
(it gives meaning to terms by specifying their rewriting strategy and an appropriate notion
of reduction, and is indeed also referred to as ‘syntactic theory’), we need to embed the
domain of values back into the syntax. To this end we transform the environment-based
abstract machine into the substitution-based abstract machine displayed in Figure 8. The
transformation is standard, except that we also need to embed evaluation contexts in the
syntax; hence the substitution-based machine operates on terms where “quoted” (in the
sense of Lisp) contexts can occur. (If it were not for shift and reset, C2 and its transitions
could be omitted and the abstract machine would reduce to the CK machine [39].)
We write t{v/x } to denote the result of the usual capture-avoiding substitution of the
value v for x in t.
Formally, the relationship between the two machines is expressed with the following
simulation theorem, where evaluation with the substitution-based abstract machine is captured by the partial function evals , defined analogously to evale .
14
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
• Terms: t ::= pmq | x | λx .t | t0 t1 | succ t | h tii | Sk .t
• Values (integers, closures, and captured continuations): v ::= m | [x , t, e] | C1
• Environments: e ::= eempty | e[x 7→ v]
• Evaluation contexts: C1 ::= [ ] | ARG((t, e), C1 ) | SUCC(C1 ) | FUN(v, C1 )
• Meta-contexts: C2 ::= • | C2 · C1
• Initial transition, transition rules, and final transition:
t ⇒ ht, eempty , [ ], •ieval
hpmq,
hx ,
hλx .t,
ht0 t1 ,
hsucc t,
hhhtii ,
hSk .t,
e,
e,
e,
e,
e,
e,
e,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
h[ ], v,
hARG((t, e), C1 ), v,
hSUCC(C1 ), m,
hFUN([x , t, e], C1 ), v,
hFUN(C1′ , C1 ), v,
C2 ieval
C2 ieval
C2 ieval
C2 ieval
C2 ieval
C2 ieval
C2 ieval
⇒
⇒
⇒
⇒
⇒
⇒
⇒
hC1 , m, C2 icont1
hC1 , e (x ), C2 icont1
hC1 , [x , t, e], C2 icont1
ht0 , e, ARG((t1 , e), C1 ), C2 ieval
ht, e, SUCC(C1 ), C2 ieval
ht, e, [ ], C2 · C1 ieval
ht, e[k 7→ C1 ], [ ], C2 ieval
C2 icont1
C2 icont1
C2 icont1
C2 icont1
C2 icont1
⇒
⇒
⇒
⇒
⇒
hC2 , vicont2
ht, e, FUN(v, C1 ), C2 ieval
hC1 , m + 1, C2 icont1
ht, e[x 7→ v], C1 , C2 ieval
hC1′ , v, C2 · C1 icont1
hC2 · C1 , vicont2
h•, vicont2
⇒ hC1 , v, C2 icont1
⇒ v
Figure 7: An environment-based abstract machine for the first level of the CPS hierarchy
Theorem 2. For any program t, either both evals (t) and evale (t) are undefined, or there
exist values v, v ′ such that evals (t) = v, evale (t) = v ′ and T (v ′ ) = v. The function T relates
a semantic value with its syntactic representation and is defined as follows:3
T (m) = pmq
T ([x , t, e]) = λx .t{T (e(x1 ))/x1 } . . . {T (e(xn ))/xn },
where F V (λx . t) = {x1 , . . . , xn }
T ([ ]) = [ ]
T (ARG((t, e), C1 )) = ARG(t{T (e(x1 ))/x1 } . . . {T (e(xn ))/xn }, T (C1 )),
where F V (t) = {x1 , . . . , xn }
T (FUN(v, C1 )) = FUN(T (v), T (C1 ))
T (SUCC(C1 )) = SUCC(T (C1 ))
3T is a generalization of Plotkin’s function Real [68].
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 15
• Terms and values:
t ::= v | x | t0 t1 | succ t | h tii | Sk .t
v ::= pmq | λx .t | C1
• Evaluation contexts: C1 ::= [ ] | ARG(t, C1 ) | SUCC(C1 ) | FUN(v, C1 )
• Meta-contexts: C2 ::= • | C2 · C1
• Initial transition, transition rules, and final transition:
t ⇒ ht, [ ], •ieval
hpmq,
hλx .t,
hC1′ ,
ht0 t1 ,
hsucc t,
hhhtii ,
hSk .t,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
h[ ], v,
hARG(t, C1 ), v,
hSUCC(C1 ), pmq,
hFUN(λx .t, C1 ), v,
hFUN(C1′ , C1 ), v,
C2 ieval
C2 ieval
C2 ieval
C2 ieval
C2 ieval
C2 ieval
C2 ieval
⇒
⇒
⇒
⇒
⇒
⇒
⇒
hC1 , pmq, C2 icont1
hC1 , λx .t, C2 icont1
hC1 , C1′ , C2 icont1
ht0 , ARG(t1 , C1 ), C2 ieval
ht, SUCC(C1 ), C2 ieval
ht, [ ], C2 · C1 ieval
ht{C1 /k }, [ ], C2 ieval
C2 icont1
C2 icont1
C2 icont1
C2 icont1
C2 icont1
⇒
⇒
⇒
⇒
⇒
hC2 , vicont2
ht, FUN(v, C1 ), C2 ieval
hC1 , pm + 1q, C2 icont1
ht{v/x }, C1 , C2 ieval
hC1′ , v, C2 · C1 icont1
hC2 · C1 , vicont2
h•, vicont2
⇒ hC1 , v, C2 icont1
⇒ v
Figure 8: A substitution-based abstract machine for the first level of the CPS hierarchy
Proof. We extend the translation function T to meta-contexts and configurations, in the
expected way, e.g.,
T (ht, e, C1 , C2 ieval ) = ht{T (e(x1 ))/x1 } . . . {T (e(xn ))/xn }, T (C1 ), T (C2 )ieval
where F V (t) = {x1 , . . . , xn }
Then it is straightforward to show that the two abstract machines operate in lock step with
respect to the translation. Hence, for any program t, both machines diverge or they both
stop (after the same number of transitions) with the values v and T (v), respectively.
We now proceed to analyze the transitions of the machine displayed in Figure 8. We
can think of a configuration ht, C1 , C2 ieval as the following decomposition of the initial term
into a meta-context C2 , a context C1 , and an intermediate term t:
C2 # C1 [t]
where # separates the context and the meta-context. Each transition performs either a
reduction, or a decomposition in search of the next redex. Let us recall that a decomposition
is performed when both sides of a transition are partitions of the same term; in that case,
depending on the structure of the decomposition C2 # C1 [t], a subpart of the term is chosen
16
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
to be evaluated next, and the contexts are updated accordingly. We also observe that eval transitions follow the structure of t, cont 1 -transitions follow the structure of C1 when the
term has been reduced to a value, and cont 2 -transitions follow the structure of C2 when a
value in the empty context has been reached.
Next we specify all the components of the reduction semantics based on the analysis of
the abstract machine.
4.4. A reduction semantics. A reduction semantics provides a reduction relation on
expressions by defining values, evaluation contexts, and redexes [36, 38, 39, 82]. In the
present case,
• the values are already specified in the (substitution-based) abstract machine:
v ::= pmq | λx .t | C1
• the evaluation contexts and meta-contexts are already specified in the abstract machine, as the data-type part of defunctionalized continuations;
C1 ::= [ ] | ARG(t, C1 ) | FUN(v, C1 ) | SUCC(C1 )
C2 ::= • | C2 · C1
• we can read the redexes off the transitions of the abstract machine:
r ::= succ pmq | (λx .t) v | Sk .t | C1′ v | h vii
Based on the distinction between decomposition and reduction, we single out the following reduction rules from the transitions of the machine:
(δ)
(βλ )
(Sλ )
(βctx )
(Reset)
C2 # C1 [succ pmq]
C2 # C1 [(λx .t) v]
C2 # C1 [Sk .t]
C2 # C1 [C1′ v]
C2 # C1 [hhvii ]
→
→
→
→
→
C2 # C1 [pm + 1q]
C2 # C1 [t{v/x }]
C2 # [t{C1 /k }]
C2 · C1 # C1′ [v]
C2 # C1 [v]
(βλ ) is the usual call-by-value β-reduction; we have renamed it to indicate that the applied
term is a λ-abstraction, since we can also apply a captured context, as in (βctx ). (Sλ ) is
plausibly symmetric to (βλ ) — it can be seen as an application of the abstraction λk .t to
the current context. Moreover, (βctx ) can be seen as performing both a reduction and a
decomposition: it is a reduction because an application of a context with a hole to a value
is reduced to the value plugged into the hole; and it is a decomposition because it changes
the meta-context, as if the application were enclosed in a reset. Finally, (Reset) makes
it possible to pass the boundary of a context when the term inside this context has been
reduced to a value.
The βctx -rule and the Sλ -rule give a justification for representing a captured context C1
as a term λx .hhC1 [x ]ii, as found in other studies of shift and reset [55, 56, 66]. In particular,
the need for delimiting the captured context is a consequence of the βctx -rule.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 17
Finally, we can read the decomposition
machine:
decompose (t) =
decompose ′ (t0 t1 , C1 , C2 ) =
decompose ′ (succ t, C1 , C2 ) =
decompose ′ (hhtii , C1 , C2 ) =
decompose ′ (v, ARG(t, C1 ), C2 ) =
function off the transitions of the abstract
decompose ′ (t, [ ], •)
decompose ′ (t0 , ARG(t1 , C1 ), C2 )
decompose ′ (t, SUCC(C1 ), C2 )
decompose ′ (t, [ ], C2 · C1 )
decompose ′ (t, FUN(v, C1 ), C2 )
In the remaining cases either a value or a redex has been found:
decompose ′ (v, [ ], •)
decompose ′ (v, [ ], C2 · C1 )
decompose ′ (Sk .t, C1 , C2 )
decompose ′ (v, FUN((λx .t), C1 ), C2 )
decompose ′ (v, FUN(C1′ , C1 ), C2 )
decompose ′ (pmq, SUCC(C1 ), C2 )
=
=
=
=
=
=
• # [v]
C2 # C1 [hhvii ]
C2 # C1 [Sk .t]
C2 # C1 [(λx .t) v]
C2 # C1 [C1′ v]
C2 # C1 [succ pmq]
An inverse of the decompose function, traditionally called plug , reconstructs a term
from its decomposition:
plug (• # [t])
plug (C2 · C1 # [t])
plug (C2 # (ARG(t′ , C1 ))[t])
plug (C2 # (FUN(v, C1 ))[t])
plug (C2 # (SUCC(C1 ))[t])
=
=
=
=
=
t
plug (C2
plug (C2
plug (C2
plug (C2
#
#
#
#
C1 [hhtii ])
C1 [t t′ ])
C1 [v t])
C1 [succ t])
In order to talk about unique decomposition, we need to define the set of potential
redexes (i.e., the disjoint union of actual redexes and stuck redexes). The grammar of
potential redexes reads as follows:
p ::= succ v | v0 v1 | Sk .t | h vii
Lemma 1 (Unique decomposition). A program t is either a value v or there exist a
unique context C1 , a unique meta-context C2 and a potential redex p such that t =
plug (C2 # C1 [p]). In the former case decompose (t) = • # [v] and in the latter case either
decompose (t) = C2 # C1 [p] if p is an actual redex, or decompose (t) is undefined.
Proof. The first part follows by induction on the structure of t. The second part follows
from the equation decompose (plug (C2 # C1 [r])) = C2 # C1 [r] which holds for all C2 , C1 and
r.
It is evident that evaluating a program either using the derived reduction rules or using
the substitution-based abstract machine yields the same result.
Theorem 3. For any program t and any value v, evals (t) = v if and only if t →∗ v, where
→∗ is the reflexive, transitive closure of the one-step reduction defined by the relation →.
Proof. When evaluating with the abstract machine, each contraction is followed by decomposing the contractum in the current context and meta-context. When evaluating with
the reduction rules, however, each contraction is followed by plugging the contractum and
decomposing the resulting term. Therefore, the theorem follows from the equation
decompose ′ (t, C1 , C2 ) = decompose (plug (C2 # C1 [t]))
which holds for any C2 , C1 and t.
18
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
We have verified that using refocusing [16,31], one can go from this reduction semantics
to the abstract machine of Figure 8.
4.5. Beyond CPS. Alternatively to using the meta-context to compose delimited continuations, as in Figure 7, we could compose them by concatenating their representation [41].
Such a concatenation function is defined as follows:
[ ] ⋆ C1′
(ARG((t, e), C1 )) ⋆ C1′
(SUCC(C1 )) ⋆ C1′
(FUN(v, C1 )) ⋆ C1′
=
=
=
=
C1′
ARG((t, e), C1 ⋆ C1′ )
SUCC(C1 ⋆ C1′ )
FUN(v, C1 ⋆ C1′ )
(The second clause would read (ARG(t, C1 )) ⋆ C1′ = ARG(t, C1 ⋆ C1′ ) for the contexts of
Figure 8.)
Then, in Figures 7 and 8, we could replace the transition
hFUN(C1′ , C1 ), v, C2 icont1
⇒ hC1′ , v, C2 · C1 icont1
hFUN(C1′ , C1 ), v, C2 icont1
⇒ hC1′ ⋆ C1 , v, C2 icont1
by the following one:
This replacement changes the control effect of shift to that of Felleisen et al.’s F operator [37]. Furthermore, the modified abstract machine is in structural correspondence with
Felleisen et al.’s abstract machine for F and # [37, 41].
This representation of control (as a list of ‘stack frames’) and this implementation of
composing delimited continuations (by concatenating these lists) are at the heart of virtually
all non-CPS-based accounts of delimited control. However, the modified environment-based
abstract machine does not correspond to a defunctionalized continuation-passing evaluator
because it is not in the range of defunctionalization [30] since the first-order representation
of functions should have a single point of consumption. Here, the constructors of contexts
are not solely consumed by the cont1 transitions of the abstract machine as in Figures 7
and 8, but also by ⋆. Therefore, the abstract machine that uses ⋆ is not in the range of
Reynolds’s defunctionalization and it thus does not immediately correspond to a higherorder, continuation-passing evaluator. In that sense, control operators using ⋆ go beyond
CPS.
Elsewhere [18], we have rephrased the modified abstract machine to put it in defunctionalized form, and we have exhibited the corresponding higher-order evaluator and the
corresponding ‘dynamic’ continuation-passing style. This dynamic CPS is not just plain
CPS but is a form of continuation+state-passing style where the threaded state is a list of
intermediate delimited continuations. Unexpectedly, it is also in structural correspondence
with the architecture for delimited control recently proposed by Dybvig, Peyton Jones, and
Sabry on other operational grounds [35].
4.6. Static vs. dynamic delimited continuations. Irrespectively of any new dynamic
CPS and any new architecture for delimited control, there seems to be remarkably few
examples that actually illustrate the expressive power of dynamic delimited continuations.
We have recently presented one, breadth-first traversal [19], and we present another one
below.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 19
The two following functions traverse a given list and return another list. The recursive
call to visit is abstracted into a delimited continuation, which is applied to the tail of the
list:
foo xs
def
=
letrec visit nil
= nil
| visit (x :: xs)
= visit (Sk .x :: (k xs))
in h visit xsii
bar xs
def
=
letrec visit nil
= nil
| visit (x :: xs)
= visit (Fk .x :: (k xs))
in h visit xsii
On the left, foo uses S and on the right, bar uses F; for the rest, the two definitions are
identical. Given an input list, foo copies it and bar reverses it.
To explain this difference and to account for the extended source language, we need to
expand the grammar of evaluation contexts, e.g., with a production to account for calls to
the list constructor:
C1 ::= [ ] | ARG(t, C1 ) | SUCC(C1 ) | FUN(v, C1 ) | CONS(v, C1 ) | . . .
Similarly, we need to expand the definition of concatenation as follows:
(CONS(v, C1 )) ⋆ C1′ = CONS(v, C1 ⋆ C1′ )
Here is a trace of the two computations in the form of the calls to and returns from
visit for the input list 1 :: 2 :: nil :
foo: Every time the captured continuation is resumed, its representation is kept separate from the current context. The meta-context therefore grows whereas the
captured context solely consists of FUN(visit, [ ]) throughout (writing visit in the
context for simplicity):
C2
C2 · C1
C2 · C1 · (CONS(1, [ ]))
C2 · C1 · (CONS(1, [ ])) · (CONS(2, [ ]))
C2 · C1 · (CONS(1, [ ])) · (CONS(2, [ ]))
C2 · C1 · (CONS(1, [ ]))
C2 · C1
C2
#
#
#
#
#
#
#
#
C1 [hhvisit (1 :: 2 :: nil )ii]
[visit (1 :: 2 :: nil)]
[visit (2 :: nil )]
[visit nil ]
[nil ]
[2 :: nil ]
[1 :: 2 :: nil]
C1 [1 :: 2 :: nil ]
bar : Every time the captured continuation is resumed, its representation is concatenated to the current context. The meta-context therefore remains the same whereas
the context changes dynamically. The first captured context is FUN(visit, [ ]); concatenating it to CONS(1, [ ]) yields CONS(1, FUN(visit, [ ])), which is the second
captured context:
C2
C2 · C1
C2 · C1
C2 · C1
C2 · C1
C2 · C1
C2 · C1
C2
#
#
#
#
#
#
#
#
C1 [hhvisit (1 :: 2 :: nil )ii]
[visit (1 :: 2 :: nil)]
(CONS(1, [ ]))[visit (2 :: nil )]
(CONS(2, CONS(1, [ ])))[visit nil]
(CONS(2, CONS(1, [ ])))[nil ]
(CONS(2, [ ]))[1 :: nil ]
[2 :: 1 :: nil]
C1 [2 :: 1 :: nil ]
20
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
4.7. Summary and conclusion. We have presented the original evaluator for the λcalculus with shift and reset; this evaluator uses two layers of continuations. From this
call-by-value evaluator we have derived two abstract machines, an environment-based one
and a substitution-based one; each of these machines uses two layers of evaluation contexts.
Based on the substitution-based machine we have constructed a reduction semantics for
the λ-calculus with shift and reset; this reduction semantics, by construction, is sound with
respect to CPS. Finally, we have pointed out the difference between the static and dynamic
delimited control operators at the level of the abstract machine and we have presented a
simple but meaningful example illustrating their differing behavior.
5. From evaluator to reduction semantics for the CPS hierarchy
We construct a reduction semantics for the call-by-value λ-calculus extended with shiftn
and resetn . As in Section 4, we go from an evaluator to an environment-based abstract
machine, and from a substitution-based abstract machine to a reduction semantics. Because of the regularity of CPS, the results can be generalized from level 1 to higher levels
without repeating the actual construction, based only on the original specification of the
hierarchy [28]. In particular, the proofs of the theorems generalize straightforwardly from
level 1.
5.1. An environment-based evaluator. At the nth level of the hierarchy, the language
is extended with operators shifti and reseti for all i such that 1 ≤ i ≤ n. The evaluator
for this language is shown in Figures 9 and 10. If n = 1, it coincides with the evaluator
displayed in Figure 6.
The evaluator uses n+1 layers of continuations. In the five first clauses (literal, variable,
λ-abstraction, function application, and application of the successor function), the continuations k2 , . . . , kn+1 are passive: if the evaluator were curried, they could be eta-reduced.
In the clauses defining shifti and reseti , the continuations ki+2 , . . . , kn+1 are also passive.
Each pair of control operators is indexed by the corresponding level in the hierarchy: reseti
is used to “push” each successive continuation up to level i onto level i+ 1 and to reinitialize
• Terms (1 ≤ i ≤ n): term ∋ t ::= pmq | x | λx .t | t0 t1 | succ t | h tiii | Si k .t
• Values: val ∋ v ::= m | f
• Answers, continuations and functions (1 ≤ i ≤ n):
ans
kn+1 ∈ contn+1
ki ∈ conti
f ∈ fun
=
=
=
=
val
val → ans
val × conti+1 × . . . × contn+1 → ans
val × cont1 × . . . × contn+1 → ans
• Initial continuations (1 ≤ i ≤ n):
θi = λ(v, ki+1 , ki+2 , . . ., kn+1 ). ki+1 (v, ki+2 , . . ., kn+1 )
θn+1 = λv. v
• Environments: env ∋ e ::= eempty | e[x 7→ v]
• Evaluation function: see Figure 10
Figure 9: An environment-based evaluator for the CPS hierarchy at level n
evaln (pmq, e, k1 , k2 , ..., kn+1 ) = k1 (m, k2 , ..., kn+1 )
evaln (x , e, k1 , k2 , ..., kn+1 ) = k1 (e(x ), k2 , ..., kn+1 )
′
′
), k2 , ..., kn+1 )
). evaln (t, e[x 7→ v], k1′ , k2′ , ..., kn+1
evaln (λx .t, e, k1 , k2 , ..., kn+1 ) = k1 (λ(v, k1′ , k2′ , ..., kn+1
evaln (t0 t1 , e, k1 , k2 , ..., kn+1 ) = evaln (t0 , e,
′
). evaln (t1 , e,
λ(f, k2′ , ..., kn+1
′′ ). f (v, k , k ′′ , ..., k ′′ ),
λ(v, k2′′ , ..., kn+1
1
n+1
2
′
′
k2 , ..., kn+1 ),
k2 , ..., kn+1 )
′
′
evaln (succ t, e, k1 , k2 , ..., kn+1 ) = evaln (t, e, λ(m, k2′ , ..., kn+1
). k1 (m + 1, k2′ , ..., kn+1
), k2 , ..., kn+1 )
′ , ..., k ′
′
′
evaln (hhtiii , e, k1 , k2 , ..., kn+1 ) = evaln (t, e, θ1 , . . ., θi , λ(v, ki+2 , ..., kn+1 ). k1 (v, k2 , ..., ki+1 , ki+2
n+1 ), ki+2 , ..., kn+1 )
evaln (Si k.t, e, k1 , k2 , ..., kn+1 ) = evaln (t, e[k 7→ ci ], θ1 , ..., θi , ki+1 , ..., kn+1 )
′
′′
′′
′
′′
′′
′
′
where ci = λ(v, k1′ , ..., kn+1
). k1 (v, k2 , ..., ki , λ(v ′ , ki+2
, ..., kn+1
). k1′ (v ′ , k2′ , ..., ki+1
, ki+2
, ..., kn+1
), ki+2
, ..., kn+1
)
• Main function: evaluaten : term → val
evaluaten (t) = evaln (t, eempty , θ1 , ..., θn , θn+1 )
Figure 10: An environment-based evaluator for the CPS hierarchy at level n, ctd.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 21
• Evaluation function (1 ≤ i ≤ n): evaln : term × env × cont1 × . . . × contn+1 → ans
22
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
them with θ1 , . . . , θi , which are the successive CPS counterparts of the identity function;
shifti is used to abstract control up to level i into a delimited continuation and to reinitialize
the successive continuations up to level i with θ1 , . . . , θi .
Applying a delimited continuation that was abstracted up to level i “pushes” each
successive continuation up to level i onto level i + 1 and restores the successive continuations that were captured in a delimited continuation. When such a delimited continuation
completes, and when an expression delimited by reseti completes, the successive continuations that were pushed onto level i + 1 are “popped” back into place and the computation
proceeds.
5.2. An environment-based abstract machine. Defunctionalizing the evaluator of Figures 9 and 10 yields the environment-based abstract machine displayed in Figures 11 and
12. If n = 1, it coincides with the abstract machine displayed in Figure 7.
The abstract machine consists of n + 2 sets of transitions: eval for interpreting terms
and cont 1 , . . . , cont n+1 for interpreting the successive defunctionalized continuations. The
set of possible values includes integers, closures and captured contexts.
This abstract machine is an extension of the abstract machine displayed in Figure 7
with n + 1 contexts instead of 2 and the corresponding transitions for shifti and reseti .
Each metai+1 -context intervenes to process reseti expressions and to apply captured continuations. Otherwise, the successive contexts are passively threaded to process literals,
variables, λ-abstractions, function applications, and applications of the successor function.
Given an environment e and a series of successive contexts, a reseti expression h tiii is
processed by evaluating t with the same environment e, i empty contexts, and a metai+1 context over which all the intermediate contexts have been pushed on.
Given an environment e and a series of successive contexts, a shift expression Si k.t
is processed by evaluating t with an extension of e where k denotes a composition of the
i surrounding contexts, i empty contexts, and the remaining outer contexts. Applying a
captured context is achieved by pushing all the current contexts on the next outer context,
restoring the composition of the captured contexts, and continuing with them. Resuming
a composition of captured contexts is achieved by popping them off the next outer context
and continuing with them.
In order to relate the resulting abstract machine to the evaluator, we define a partial
function evalen mapping a term t to a value v whenever the machine for level n, started with
• Terms (1 ≤ i ≤ n): t ::= pmq | x | λx .t | t0 t1 | succ t | h tiii | Si k .t
• Values (1 ≤ i ≤ n): v ::= m | [x , t, e] | Ci
• Evaluation contexts (2 ≤ i ≤ n + 1):
C1 ::= [ ] | ARG((t, e), C1 ) | SUCC(C1 ) | FUN(v, C1 )
Ci ::= [ ] | Ci · Ci−1
• Environments: e ::= eempty | e[x 7→ v]
• Initial transition, transition rules, and final transition: see Figure 12
Figure 11: An environment-based abstract machine for the CPS hierarchy at level n
t ⇒ ht, eempty , [ ], [ ], ..., [ ]ieval
hpmq,
hx ,
hλx .t,
ht0 t1 ,
hsucc t,
hhhtiii ,
hSi k .t,
e,
e,
e,
e,
e,
e,
e,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
h[ ], v,
hARG((t, e), C1 ), v,
hSUCC(C1 ), m,
hFUN([x , t, e], C1 ), v,
hFUN(Ci′ · (...(C2′ · C1′ )...), C1 ), v,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
...,
...,
...,
...,
...,
...,
...,
...,
...,
...,
...,
...,
Cn+1 ieval
Cn+1 ieval
Cn+1 ieval
Cn+1 ieval
Cn+1 ieval
Cn+1 ieval
Cn+1 ieval
⇒
⇒
⇒
⇒
⇒
⇒
⇒
hC1 , m, C2 , ..., Cn+1 icont1
hC1 , e (x ), C2 , ..., Cn+1 icont1
hC1 , [x , t, e], C2 , ..., Cn+1 icont1
ht0 , e, ARG((t1 , e), C1 ), C2 , ..., Cn+1 ieval
ht, e, SUCC(C1 ), C2 , ..., Cn+1 ieval
ht, e, [ ], ..., [ ], Ci+1 · (...(C2 · C1 )...), Ci+2 , ..., Cn+1 ieval
ht, e[k 7→ Ci · (...(C2 · C1 )...)], [ ], ..., [ ], Ci+1 , ..., Cn+1 ieval
Cn+1 icont1
Cn+1 icont1
Cn+1 icont1
Cn+1 icont1
Cn+1 icont1
⇒
⇒
⇒
⇒
⇒
hC2 , v, C3 , ..., Cn+1 icont2
ht, e, FUN(v, C1 ), C2 , ..., Cn+1 ieval
hC1 , m + 1, C2 , ..., Cn+1 icont1
ht, e[x 7→ v], C1 , C2 , ..., Cn+1 ieval
hC1′ , v, C2′ , ..., Ci′ , Ci+1 · (...(C2 · C1 )...), Ci+2 , ..., Cn+1 icont1
h[ ], v, Cj+1 , ..., Cn+1 icontj
hCj · (...(C2 · C1 )...), v, Cj+1 , ..., Cn+1 icontj
hCn+1 · (...(C2 · C1 )...), vicontn+1
h[ ], vicontn+1
⇒ hCj+1 , v, Cj+2 , ..., Cn+1 icontj +1
⇒ hC1 , v, C2 , ..., Cn+1 icont1
⇒ hC1 , v, C2 , ..., Cn+1 icont1
⇒ v
Figure 12: An environment-based abstract machine for the CPS hierarchy at level n, ctd.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 23
• Initial transition, transition rules, and final transition (1 ≤ i ≤ n, 2 ≤ j ≤ n):
24
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
t, stops with v. The correctness of the machine with respect to the evaluator is ensured by
the following theorem:
Theorem 4. For any program t and any integer value m, evaluaten (t) = m if and only if
evalen (t) = m.
5.3. A substitution-based abstract machine. In the same fashion as in Section 4.3, we
construct the substitution-based abstract machine corresponding to the environment-based
abstract machine of Section 5.2. The result is displayed in Figures 13 and 14. If n = 1, it
coincides with the abstract machine displayed in Figure 8.
The nth level contains n + 1 evaluation contexts and each context Ci can be viewed as
a stack of non-empty contexts Ci−1 . Terms are decomposed as
Cn+1 #n Cn #n−1 Cn−1 #n−2 · · · #2 C2 #1 C1 [t],
where each #i represents a context delimiter of level i. All the control operators that occur
at the jth level (with j < n) of the hierarchy do not use the contexts j + 2, . . . , n + 1. The
functions decompose and its inverse plug can be read off the machine, as for level 1.
The transitions of the machine for level j are “embedded” in the machine for level j + 1;
the extra components are threaded but not used.
We define a partial function evalsn capturing the evaluation by the substitution-based
abstract machine for an arbitrary level n, analogously to the definition of evalen . Now we can
relate evaluation with the environment-based and the substitution-based abstract machines
for level n.
Theorem 5. For any program t, either both evalsn (t) and evalen (t) are undefined, or there
exist values v, v ′ such that evalsn (t) = v, evalen (t) = v ′ and Tn (v ′ ) = v.
The definition of Tn extends that of T from Theorem 2 in such a way that it is homomorphic for all the contexts Ci , with 2 ≤ i ≤ n.
5.4. A reduction semantics. Along the same lines as in Section 4.4, we construct the
reduction semantics for the CPS hierarchy based on the abstract machine of Figures 13
and 14. For an arbitrary level n we obtain the following set of reduction rules, for all
• Terms and values (1 ≤ i ≤ n):
t ::= v | x | t0 t1 | succ t | h tiii | Si k .t
v ::= pmq | λx .t | Ci
• Evaluation contexts (2 ≤ i ≤ n + 1):
C1 ::= [ ] | ARG(t, C1 ) | SUCC(C1 ) | FUN(v, C1 )
Ci ::= [ ] | Ci · Ci−1
• Initial transition, transition rules, and final transition: see Figure 14
Figure 13: A substitution-based abstract machine for the CPS hierarchy at level n
t ⇒ ht, [ ], [ ], ..., [ ]ieval
hpmq,
hλx .t,
hCi′ ,
ht0 t1 ,
hsucc t,
hhhtiii ,
hSi k .t,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
C1 ,
h[ ], v,
hARG(t, C1 ), v,
hSUCC(C1 ), pmq,
hFUN((λx .t), C1 ), v,
hFUN(Ci′ · (...(C2′ · C1′ )...), C1 ), v,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
C2 ,
...,
...,
...,
...,
...,
...,
...,
...,
...,
...,
...,
...,
Cn+1 ieval
Cn+1 ieval
Cn+1 ieval
Cn+1 ieval
Cn+1 ieval
Cn+1 ieval
Cn+1 ieval
⇒
⇒
⇒
⇒
⇒
⇒
⇒
hC1 , pmq, C2 , ..., Cn+1 icont1
hC1 , λx .t, C2 , ..., Cn+1 icont1
hC1 , Ci′ , C2 , ..., Cn+1 icont1
ht0 , ARG((t1 , e), C1 ), C2 , ..., Cn+1 ieval
ht, SUCC(C1 ), C2 , ..., Cn+1 ieval
ht, [ ], ..., [ ], Ci+1 · (...(C2 · C1 )...), Ci+2 , ..., Cn+1 ieval
ht{Ci · (...(C2 · C1 )...)/k }, [ ], ..., [ ], Ci+1 , ..., Cn+1 ieval
Cn+1 icont1
Cn+1 icont1
Cn+1 icont1
Cn+1 icont1
Cn+1 icont1
⇒
⇒
⇒
⇒
⇒
hC2 , v, C3 , ..., Cn+1 icont2
ht, FUN(v, C1 ), C2 , ..., Cn+1 ieval
hC1 , pm + 1q, C2 , ..., Cn+1 icont1
ht{v/x }, C1 , C2 , ..., Cn+1 ieval
hC1′ , v, C2′ , ..., Ci′ , Ci+1 · (...(C2 · C1 )...), Ci+2 , ..., Cn+1 icont1
h[ ], v, Cj+1 , ..., Cn+1 icontj
hCj · (...(C2 · C1 )...), v, Cj+1 , ..., Cn+1 icontj
hCn+1 · (...(C2 · C1 )...), vicontn+1
h[ ], vicontn+1
⇒ hCj+1 , v, Cj+2 , ..., Cn+1 icontj +1
⇒ hC1 , v, C2 , ..., Cn+1 icont1
⇒ hC1 , v, C2 , ..., Cn+1 icont1
⇒ v
Figure 14: A substitution-based abstract machine for the the CPS hierarchy at level n, ctd.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 25
• Initial transition, transition rules, and final transition (1 ≤ i ≤ n, 2 ≤ j ≤ n):
26
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
1 ≤ i ≤ n; they define the actual redexes:
(δ)
Cn+1 #n · · · #1 C1 [succ pmq] →n Cn+1 #n · · · #1 C1 [pm + 1q]
(βλ )
Cn+1 #n · · · #1 C1 [(λx .t) v] →n Cn+1 #n · · · #1 C1 [t{v/x }]
(Sλi )
Cn+1 #n · · · #1 C1 [Si k .t] →n
Cn+1 #n · · · #i+1 Ci+1 #i [ ] . . . #1 [t{Ci · (. . . (C2 · C1 ) . . .)/k }]
i )
(βctx
Cn+1 #n · · · #1 C1 [Ci′ · (. . . (C2′ · C1′ ) . . .) v] →n
Cn+1 #n · · · #i+1 Ci+1 · (. . . (C2 · C1 ) . . .) #i Ci′ #i−1 · · · #1 C1′ [v]
(Reseti )
Cn+1 #n · · · #1 C1 [hhviii ] →n Cn+1 #n · · · #1 C1 [v]
Each level contains all the reductions from lower levels, and these reductions are compatible with additional layers of evaluation contexts. In particular, at level 0 there are only
δ- and βλ -reductions.
The values and evaluation contexts are already specified in the abstract machine. Moreover, the potential redexes are defined according to the following grammar:
pn ::= succ v | v0 v1 | Si k .t | h viii
(1 ≤ i ≤ n)
Lemma 2 (Unique decomposition for level n). A program t is either a value or there
exists a unique sequence of contexts C1 , . . . , Cn+1 and a potential redex pn such that
t = plug (Cn+1 #n · · · #1 C1 [pn ]).
Evaluating a term using either the derived reduction rules or the substitution-based
abstract machine from Section 5.3 yields the same result:
Theorem 6. For any program t and any value v, evalsn (t) = v if and only if t →∗n v, where
→∗n is the reflexive, transitive closure of →n .
As in Section 4.4, using refocusing, one can go from a given reduction semantics of
Section 5.4 into a pre-abstract machine and the corresponding eval/apply abstract machine
of Figures 13 and 14.
5.5. Beyond CPS. As in Section 4.5, one can define a family of concatenation functions
over contexts and use it to implement composable continuations in the CPS hierarchy,
giving rise to a family of control operators Fn and #n . Again the modified environmentbased abstract machine does not immediately correspond to a defunctionalized continuationpassing evaluator. Such control operators go beyond traditional CPS.
5.6. Static vs. dynamic delimited continuations. As in Section 4.6, one can illustrate
the difference between static and dynamic delimited continuations in the CPS hierarchy.
For example, replacing shift2 and reset2 respectively by F2 and #2 in Danvy and Filinski’s
version of Abelson and Sussman’s generator of triples [28, Section 3] yields a list in reverse
order.4
5.7. Summary and conclusion. We have generalized the results presented in Section 4
from level 1 to the whole CPS hierarchy of control operators shiftn and resetn . Starting
from the original evaluator for the λ-calculus with shiftn and resetn that uses n + 1 layers
of continuations, we have derived two abstract machines, an environment-based one and
a substitution-based one; each of these machines use n + 1 layers of evaluation contexts.
4Thanks are due to an anonymous reviewer for pointing this out.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 27
Based on the substitution-based machine we have obtained a reduction semantics for the
λ-calculus extended with shiftn and resetn which, by construction, is sound with respect to
CPS.
6. Programming in the CPS hierarchy
To finish, we present new examples of programming in the CPS hierarchy. The examples
are normalization functions. In Sections 6.1 and 6.2, we first describe normalization by
evaluation and we present the simple example of the free monoid. In Section 6.3, we present a
function mapping a proposition into its disjunctive normal form; this normalization function
uses delimited continuations. In Section 6.4, we generalize the normalization functions of
Sections 6.2 and 6.3 to a hierarchical language of units and products, and we express the
corresponding normalization function in the CPS hierarchy.
6.1. Normalization by evaluation. Normalization by evaluation is a ‘reduction-free’ approach to normalizing terms. Instead of reducing a term to its normal form, one evaluates
this term into a non-standard model and reifies its denotation into its normal form [34]:
eval
reify
normalize
normalize
:
:
:
=
term → value
value → term nf
term → term nf
reify ◦ eval
Normalization by evaluation has been developed in intuitionistic type theory [20, 63],
proof theory [12,13], category theory [4], and partial evaluation [22,23], where it has emerged
as a new field of application for delimited continuations [9, 23, 34, 44, 48, 51, 78].
6.2. The free monoid. A source term in the free monoid is either a variable, the unit
element, or the product of two terms:
term ∋ t ::= x | ε | t ⋆ t′
The product is associative and the unit element is neutral. These properties justify the
following conversion rules:
t ⋆ (t′ ⋆ t′′ ) ↔ (t ⋆ t′ ) ⋆ t′′
t⋆ε ↔ t
ε⋆t ↔ t
We aim (for example) for list-like flat normal forms:
t
term nf ∋ b
t ::= εnf | x ⋆nf b
In a reduction-based approach to normalization, one would orient the conversion rules
into reduction rules and one would apply these reduction rules until a normal form is
obtained:
t ⋆ (t′ ⋆ t′′ ) ← (t ⋆ t′ ) ⋆ t′′
ε⋆t → t
In a reduction-free approach to normalization, one defines a normalization function as
the composition of a non-standard evaluation function and a reification function. Let us
state such a normalization function.
28
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
The non-standard domain of values is the transformer
value = term nf → term nf .
The evaluation function is defined by induction over the syntax of source terms, and
the reification function inverts it:
eval x = λt.x ⋆nf t
eval ε = λt.t
eval (t ⋆ t′ ) = (eval t) ◦ (eval t′ )
reify v = v εnf
normalize t = reify (eval t)
In effect, eval is a homomorphism from the source monoid to the monoid of transformers (unit is mapped to unit and products are mapped to products) and the normalization
function hinges on the built-in associativity of function composition. Beylin, Dybjer, Coquand, and Kinoshita have studied its theoretical content [14, 20, 58]. From a (functional)
programming standpoint, the reduction-based approach amounts to flattening a tree iteratively by reordering it, and the reduction-free approach amounts to flattening a tree with
an accumulator.
6.3. A language of propositions. A source term, i.e., a proposition, is either a variable,
a literal (true or false), a conjunction, or a disjunction:
term ∋ t ::= x | true | t ∧ t′ | false | t ∨ t′
Conjunction and disjunction are associative and distribute over each other; true is neutral
for conjunction and absorbant for disjunction; and false is neutral for disjunction and
absorbant for conjunction.
We aim (for example) for list-like disjunctive normal forms:
term nf ∋ b
t ::= d
nf
term d ∋ d ::= false nf | c ∨nf d
nf | x ∧nf c
term nf
c ∋ c ::= true
Our normalization function is the result of composing a non-standard evaluation function
and a reification function. We state them below without proof.
Given the domains of transformers
nf
F1 = term nf
c → term c
nf
nf
F2 = term d → term d
the non-standard domain of values is ans 1 , where
ans 2 = F2
ans 1 = (F1 → ans 2 ) → ans 2 .
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 29
The evaluation function is defined by induction over the syntax of source terms, and the
reification function inverts it:
eval 0 x k d
eval 0 true k d
eval 0 (t ∧ t′ ) k d
eval 0 false k d
eval 0 (t ∨ t′ ) k d
=
=
=
=
=
k (λc.x ∧nf c) d
k (λc.c) d
eval 0 t (λf1 .eval 0 t′ (λf1′ .k (f1 ◦ f1′ ))) d
d
eval 0 t k (eval 0 t′ k d)
reify 0 v = v (λf1 .λd.(f1 true nf ) ∨nf d) false nf
normalize t = reify 0 (eval 0 t)
This normalization function uses a continuation k, an accumulator d to flatten disjunctions,
and another one c to flatten conjunctions. The continuation is delimited: the three first
clauses of eval 0 are in CPS; in the fourth, k is discarded (accounting for the fact that false
is absorbant for conjunction); and in the last, k is duplicated and used in non-tail position
(achieving the distribution of conjunctions over disjunctions). The continuation and the
accumulators are initialized in the definition of reify 0 .
Uncurrying the continuation and mapping eval 0 and reify 0 back to direct style yield
the following definition, which lives at level 1 of the CPS hierarchy:
eval 1 x d = (λc.x ∧nf c, d)
eval 1 true d = (λc.c, d)
eval 1 (t ∧ t′ ) d = let (f1 , d) = eval 1 t d
in let (f1′ , d) = eval 1 t′ d
in (f1 ◦ f1′ , d)
eval 1 false d = Sk.d
eval 1 (t ∨ t′ ) d = Sk.k (eval 1 t h k (eval 1 t′ d)ii)
reify 1 v = h let (f1 , d) = v false nf
in (f1 true nf ) ∨nf dii
normalize t = reify 1 (eval 1 t)
The three first clauses of eval 1 are in direct style; the two others abstract control with
shift. In the fourth clause, the context is discarded; and in the last clause, the context is
duplicated and composed. The context and the accumulators are initialized in the definition
of reify 1 .
This direct-style version makes it even more clear than the CPS version that the accumulator for the disjunctions in normal form is a threaded state. A continuation-based,
state-based version (or better, a monad-based one) can therefore be written—but it is out
of scope here.
6.4. A hierarchical language of units and products. We consider a generalization of
propositional logic where a source term is either a variable, a unit in a hierarchy of units,
or a product in a hierarchy of products:
term ∋ t ::= x | εi | t ⋆i t′
where 1 ≤ i ≤ n.
All the products are associative. All units are neutral for products with the same index.
30
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
The free monoid: The language corresponds to that of the free monoid if n = 1, as
in Section 6.2.
Boolean logic: The language corresponds to that of propositions if n = 2, as in
Section 6.3: ε1 is true, ⋆1 is ∧, ε2 is false, and ⋆2 is ∨.
Multi-valued logic: In general, for each n > 2 we can consider a suitable n-valued
logic [47]; for example, in case n = 4, the language corresponds to that of Belnap’s
bilattice FOU R [11]. It is also possible to modify the normalization function to
work for less regular logical structures (e.g., other bilattices).
Monads: In general, the language corresponds to that of layered monads [64]: each
unit element is the unit of the corresponding monad, and each product is the ‘bind’
of the corresponding monad. In practice, layered monads are collapsed into one for
programming consumption [43], but prior to this collapse, all the individual monad
operations coexist in the computational soup.
In the remainder of this section, we assume that all the products, besides being associative, distribute over each other, and that all units, besides being neutral for products
with the same index, are absorbant for products with other indices. We aim (for example)
for a generalization of disjunctive normal forms:
term nf ∋ b
t ::= tn
nf
nf
term nf
∋
t
n ::= εn | tn−1 ⋆n tn
n
..
.
nf
nf
term nf
1 ∋ t1 ::= ε1 | t0 ⋆1 t1
term nf
0 ∋ t0 ::= x
For presentational reasons, in the remainder of this section we arbitrarily fix n to be 5.
Our normalization function is the result of composing a non-standard evaluation function and a reification function. We state them below without proof. Given the domains of
transformers
nf
F1 = term nf
1 → term 1
nf
F2 = term nf
2 → term 2
nf
nf
F3 = term 3 → term 3
nf
F4 = term nf
4 → term 4
nf
F5 = term nf
5 → term 5
the non-standard domain of values is ans 1 , where
ans 5
ans 4
ans 3
ans 2
ans 1
=
=
=
=
=
F5
(F4
(F3
(F2
(F1
→
→
→
→
ans 5 )
ans 4 )
ans 3 )
ans 2 )
→
→
→
→
ans 5
ans 4
ans 3
ans 2 .
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 31
The evaluation function is defined by induction over the syntax of source terms, and the
reification function inverts it:
eval 0 x k1 k2 k3 k4 t5
eval 0 ε1 k1 k2 k3 k4 t5
eval 0 (t ⋆1 t′ ) k1 k2 k3 k4 t5
eval 0 ε2 k1 k2 k3 k4 t5
eval 0 (t ⋆2 t′ ) k1 k2 k3 k4 t5
eval 0 ε3 k1 k2 k3 k4 t5
eval 0 (t ⋆3 t′ ) k1 k2 k3 k4 t5
eval 0 ε4 k1 k2 k3 k4 t5
eval 0 (t ⋆4 t′ ) k1 k2 k3 k4 t5
eval 0 ε5 k1 k2 k3 k4 t5
eval 0 (t ⋆5 t′ ) k1 k2 k3 k4 t5
=
=
=
=
=
=
=
=
=
=
=
k1 (λt1 .x ⋆nf
1 t1 ) k2 k3 k4 t5
k1 (λt1 .t1 ) k2 k3 k4 t5
eval 0 t (λf1 .eval 0 t′ (λf1′ .k1 (f1 ◦ f1′ ))) k2 k3 k4 t5
k2 (λt2 .t2 ) k3 k4 t5
eval 0 t k1 (λf2 .eval 0 t′ k1 (λf2′ .k2 (f2 ◦ f2′ ))) k3 k4 t5
k3 (λt3 .t3 ) k4 t5
eval 0 t k1 k2 (λf3 .eval 0 t′ k1 k2 (λf3′ .k3 (f3 ◦ f3′ ))) k4 t5
k4 (λt4 .t4 ) t5
eval 0 t k1 k2 k3 (λf4 .eval 0 t′ k1 k2 k3 (λf4′ .k4 (f4 ◦ f4′ ))) t5
t5
eval 0 t k1 k2 k3 k4 (eval 0 t′ k1 k2 k3 k4 t5 )
nf
reify 0 v = v (λf1 .λk2 .k2 (λt2 .(f1 εnf
1 ) ⋆2 t2 ))
nf
(λf2 .λk3 .k3 (λt3 .(f2 εnf
2 ) ⋆3 t3 ))
nf
nf
(λf3 .λk4 .k4 (λt4 .(f3 ε3 ) ⋆4 t4 ))
nf
(λf4 .λt5 .(f4 εnf
4 ) ⋆5 t5 )
ε5
normalize t = reify 0 (eval 0 t)
This normalization function uses four delimited continuations k1 , k2 , k3 , k4 and five accumulators t1 , t2 , t3 , t4 , t5 to flatten each of the successive products. In the clause of each εi ,
the continuations k1 , . . . , ki−1 are discarded, accounting for the fact that εi is absorbant for
⋆1 , . . . , ⋆i−1 , and the identity function is passed to ki , accounting for the fact that εi is neutral for ⋆i . In the clause of each ⋆i+1 , the continuations k1 , . . . , ki are duplicated and used
in non-tail position, achieving the distribution of ⋆i+1 over ⋆1 , . . . , ⋆i . The continuations
and the accumulators are initialized in the definition of reify 0 .
This normalization function lives at level 0 of the CPS hierarchy, but we can express
it at a higher level using shift and reset. For example, uncurrying k3 and k4 and mapping
eval 0 and reify 0 back to direct style twice yield the following intermediate definition, which
lives at level 2:
eval 2 x k1 k2 t5 = k1 (λt1 .x ⋆nf
1 t1 ) k2 t5
eval 2 ε1 k1 k2 t5 = k1 (λt1 .t1 ) k2 t5
eval 2 (t ⋆1 t′ ) k1 k2 t5 = eval 2 t (λf1 .eval 2 t′ (λf1′ .k1 (f1 ◦ f1′ ))) k2 t5
eval 2 ε2 k1 k2 t5 = k2 (λt2 .t2 ) t5
eval 2 (t ⋆2 t′ ) k1 k2 t5 = eval 2 t k1 (λf2 .eval 2 t′ k1 (λf2′ .k2 (f2 ◦ f2′ ))) t5
eval 2 ε3 k1 k2 t5 = (λt3 .t3 , t5 )
eval 2 (t ⋆3 t′ ) k1 k2 t5 = let (f3 , t5 ) = eval 2 t k1 k2 t5
in let (f3′ , t5 ) = eval 2 t′ k1 k2 t5
in (f3 ◦ f3′ , t5 )
32
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
eval 2 ε4 k1 k2 t5 = S1 k3 .(λt4 .t4 , t5 )
eval 2 (t ⋆4 t′ ) k1 k2 t5 = S1 k3 .let (f4 , t5 ) = h k3 (eval 2 t k1 k2 t5 )ii1
in let (f4′ , t5 ) = h k3 (eval 2 t′ k1 k2 t5 )ii1
in (f4 ◦ f4′ , t5 )
eval 2 ε5 k1 k2 t5 = S2 k4 .t5
eval 2 (t ⋆5 t′ ) k1 k2 t5 = S1 k3 .S2 k4 .let t5 = h k4 h k3 (eval 2 t′ k1 k2 t5 )ii1 i 2
in h k4 h k3 (eval 2 t k1 k2 t5 )ii1 i 2
nf
reify 2 v = h let (f4 , t5 ) = h let (f3 , t5 ) = v (λf1 .λk2 .k2 (λt2 .(f1 εnf
1 ) ⋆2 t2 ))
nf
nf
(λf2 .λt3 .(f2 ε2 ) ⋆3 t3 )
ε5
nf
i1
in (λf4 .(f3 ε3 ) ⋆nf
4 t4 , t5 )i
nf
nf
in (f4 ε4 ) ⋆5 t5 i 2
normalize t = reify 2 (eval 2 t)
Whereas eval 0 had four layered continuations, eval 2 has only two layered continuations
since it has been mapped back to direct style twice. Where eval 0 accesses k3 as one of its
parameters, eval 2 abstracts the first layer of control with shift1 , and where eval 0 accesses
k4 as one of its parameters, eval 2 abstracts the first and the second layer of control with
shift2 .
Uncurrying k1 and k2 and mapping eval 2 and reify 2 back to direct style twice yield the
following direct-style definition, which lives at level 4 of the CPS hierarchy:
eval 4 x t5 = (λt1 .x ⋆nf
1 t1 , t5 )
eval 4 ε1 t5 = (λt1 .t1 , t5 )
eval 4 (t ⋆1 t′ ) t5 = let (f1 , t5 ) = eval 4 t t5
in let (f1′ , t5 ) = eval 4 t′ t5
in (f1 ◦ f1′ , t5 )
eval 4 ε2 t5 = S1 k1 .(λt2 .t2 , t5 )
eval 4 (t ⋆2 t′ ) t5 = S1 k1 .let (f2 , t5 ) = h k1 (eval 4 t t5 )ii1
in let (f2′ , t5 ) = h k1 (eval 4 t′ t5 )ii1
in (f2 ◦ f2′ , t5 )
eval 4 ε3 t5 = S2 k2 .(λt3 .t3 , t5 )
eval 4 (t ⋆3 t′ ) t5 = S1 k1 .S2 k2 .let (f3 , t5 ) = h k2 h k1 (eval 4 t t5 )ii1 i 2
in let (f3′ , t5 ) = h k2 h k1 (eval 4 t′ t5 )ii1 i 2
in (f3 ◦ f3′ , t5 )
eval 4 ε4 t5 = S3 k3 .(λt4 .t4 , t5 )
eval 4 (t ⋆4 t′ ) t5 = S1 k1 .S2 k2 .S3 k3 .let (f4 , t5 ) = h k3 h k2 h k1 (eval 4 t t5 )ii1 i 2 i 3
in let (f4′ , t5 ) = h k3 h k2 h k1 (eval 4 t′ t5 )ii1 i 2 i 3
in (f4 ◦ f4′ , t5 )
eval 4 ε5 t5 = S4 k4 .t5
eval 4 (t ⋆5 t′ ) t5 = S1 k1 .S2 k2 .S3 k3 .S4 k4 .let t5 = h k4 h k3 h k2 h k1 (eval 4 t′ t5 )ii1 i 2 i 3 i 4
in h k4 h k3 h k2 h k1 (eval 4 t t5 )ii1 i 2 i 3 i 4
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 33
reify 4 v = h let (f4 , t5 ) = h let (f3 , t5 ) = h let (f2 , t5 ) = h let (f1 , t5 ) = v ε5
nf
i1
in (λf2 .(f1 εnf
1 ) ⋆2 t2 , t5 )i
nf
nf
in (λf3 .(f2 ε2 ) ⋆3 t3 , t5 )ii2
nf
in (λf4 .(f3 εnf
)
i3
3 ⋆4 t4 , t5 )i
nf
nf
in (f4 ε4 ) ⋆5 t5 i 4
normalize t = reify 4 (eval 4 t)
Whereas eval 2 had two layered continuations, eval 4 has none since it has been mapped back
to direct style twice. Where eval 2 accesses k1 as one of its parameters, eval 4 abstracts the
first layer of control with shift1 , and where eval 2 accesses k2 as one of its parameters, eval 4
abstracts the first and the second layer of control with shift2 . Where eval 2 uses reset1 and
shift1 , eval 4 uses reset3 and shift3 , and where eval 2 uses reset2 and shift2 , eval 4 uses reset4
and shift4 .
6.5. A note about efficiency. We have implemented all the definitions of Section 6.4 as
well as the intermediate versions eval 1 and eval 3 in ML [32]. We have also implemented
hierarchical normalization functions for other values than 5.
For high products (i.e., in Section 6.4, for source terms using ⋆3 and ⋆4 ), the normalization function living at level 0 of the CPS hierarchy is the most efficient one. On the
other hand, for low products (i.e., in Section 6.4, for source terms using ⋆1 and ⋆2 ), the
normalization functions living at a higher level of the CPS hierarchy are the most efficient
ones. These relative efficiencies are explained in terms of resources:
• Accessing to a continuation as an explicit parameter is more efficient than accessing
to it through a control operator.
• On the other hand, the restriction of eval 4 to source terms that only use ε1 and ⋆1
is in direct style, whereas the corresponding restrictions of eval 2 and eval 0 pass a
number of extra parameters. These extra parameters penalize performance.
The better performance of programs in the CPS hierarchy has already been reported for
level 1 in the context of continuation-based partial evaluation [61], and it has been reported
for a similar “pay as you go” reason: a program that abstracts control relatively rarely is run
more efficiently in direct style with a control operator rather than in continuation-passing
style.
6.6. Summary and conclusion. We have illustrated the CPS hierarchy with an application of normalization by evaluation that naturally involves successive layers of continuations
and that demonstrates the expressive power of shiftn and resetn .
The application also suggests alternative control operators that would fit better its
continuation-based programming pattern. For example, instead of representing a delimited
continuation as a function and apply it as such, we could represent it as a continuation and
apply it with a ‘throw’ operator as in MacLisp and Standard ML of New Jersey. For another
example, instead of throwing a value to a continuation, we could specify the continuation of
a computation, e.g., with a reflect i special form. For a third example, instead of abstracting
control up to a layer n, we could give access to each of the successive layers up to n, e.g.,
with a Ln operator. Then instead of
eval 4 (t ⋆4 t′ ) t5 = S1 k1 .S2 k2 .S3 k3 .let (f4 , t5 ) = h k3 h k2 h k1 (eval 4 t t5 )ii1 i 2 i 3
in let (f4′ , t5 ) = h k3 h k2 h k1 (eval 4 t′ t5 )ii1 i 2 i 3
in (f4 ◦ f4′ , t5 )
34
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
one could write
eval 4 (t ⋆4 t′ ) t5 = L3 (k1 , k2 , k3 ).let (f4′ , t5 ) = reflect 3 (eval 4 t t5 , k1 , k2 , k3 )
in let (f4′ , t5 ) = reflect 3 (eval 4 t′ t5 , k1 , k2 , k3 )
in (f4 ◦ f4′ , t5 ).
Such alternative control operators can be more convenient to use, while being compatible
with CPS.
7. Conclusion and issues
We have used CPS as a guideline to establish an operational foundation for delimited
continuations. Starting from a call-by-value evaluator for λ-terms with shift and reset,
we have mechanically derived the corresponding abstract machine. From this abstract
machine, it is straightforward to obtain a reduction semantics of delimited control that,
by construction, is compatible with CPS—both for one-step reduction and for evaluation.
These results can also be established without the guideline of CPS, but less easily.
The whole approach generalizes straightforwardly to account for the shiftn and resetn
family of delimited-control operators and more generally for any control operators that are
compatible with CPS. These results would be non-trivial to establish without the guideline
of CPS.
Defunctionalization provides a key for connecting continuation-passing style and operational intuitions about control. Indeed most of the time, control stacks and evaluation
contexts are the defunctionalized continuations of an evaluator. Defunctionalization also
provides a key for identifying where operational intuitions about control go beyond CPS
(see Section 4.5).
We do not know whether CPS is the ultimate answer, but the present work shows yet
another example of its usefulness. It is like nothing can go wrong with CPS.
Acknowledgments
We are grateful to Mads Sig Ager, Julia Lawall, Jan Midtgaard, and the referees of
CW’04 and of LMCS for their comments. The third author would also like to thank Samuel
Lindley for our joint initial study of the normalization functions of Section 6.
This work is partially supported by the ESPRIT Working Group APPSEM II
(http://www.appsem.org), by the Danish Natural Science Research Council, Grant no. 2102-0474 (for the two first authors) and Grant no. 21-03-0545 (for the third author), and by
BRICS (Basic Research in Computer Science (http://www.brics.dk), funded by the Danish
National Research Foundation).
References
[1] Mads Sig Ager, Dariusz Biernacki, Olivier Danvy, and Jan Midtgaard. A functional correspondence between evaluators and abstract machines. In Dale Miller, editor, Proceedings of the Fifth ACM-SIGPLAN
International Conference on Principles and Practice of Declarative Programming (PPDP’03), pages 8–
19. ACM Press, August 2003.
[2] Mads Sig Ager, Olivier Danvy, and Jan Midtgaard. A functional correspondence between call-by-need
evaluators and lazy abstract machines. Information Processing Letters, 90(5):223–232, 2004. Extended
version available as the technical report BRICS-RS-04-3.
[3] Mads Sig Ager, Olivier Danvy, and Jan Midtgaard. A functional correspondence between monadic evaluators and abstract machines for languages with computational effects. Theoretical Computer Science,
342(1):149–172, 2005. Extended version available as the technical report BRICS RS-04-28.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 35
[4] Thorsten Altenkirch, Martin Hofmann, and Thomas Streicher. Categorical reconstruction of a reductionfree normalization proof. In David H. Pitt, David E. Rydeheard, and Peter Johnstone, editors, Category
Theory and Computer Science, number 953 in Lecture Notes in Computer Science, pages 182–199,
Cambridge, UK, August 1995. Springer-Verlag.
[5] Zena M. Ariola, Hugo Herbelin, and Amr Sabry. A type-theoretic foundation of continuations and
prompts. In Fisher [45], pages 40–53.
[6] Kenichi Asai. Online partial evaluation for shift and reset. In Peter Thiemann, editor, Proceedings of
the 2002 ACM SIGPLAN Workshop on Partial Evaluation and Semantics-Based Program Manipulation
(PEPM 2002), SIGPLAN Notices, Vol. 37, No 3, pages 19–30, Portland, Oregon, March 2002. ACM
Press.
[7] Kenichi Asai. Offline partial evaluation for shift and reset. In Nevin Heintze and Peter Sestoft, editors, Proceedings of the 2004 ACM SIGPLAN Symposium on Partial Evaluation and Semantics-Based
Program Manipulation (PEPM 2004), pages 3–14, Verona, Italy, August 2003. ACM Press.
[8] Vincent Balat, Roberto Di Cosmo, and Marcelo P. Fiore. Extensional normalisation and type-directed
partial evaluation for typed lambda calculus with sums. In Xavier Leroy, editor, Proceedings of the
Thirty-First Annual ACM Symposium on Principles of Programming Languages, pages 64–76, Venice,
Italy, January 2004. ACM Press.
[9] Vincent Balat and Olivier Danvy. Memoization in type-directed partial evaluation. In Don Batory,
Charles Consel, and Walid Taha, editors, Proceedings of the 2002 ACM SIGPLAN/SIGSOFT Conference on Generative Programming and Component Engineering, number 2487 in Lecture Notes in
Computer Science, pages 78–92, Pittsburgh, Pennsylvania, October 2002. Springer-Verlag.
[10] Anindya Banerjee, Nevin Heintze, and Jon G. Riecke. Design and correctness of program transformations based on control-flow analysis. In Naoki Kobayashi and Benjamin C. Pierce, editors, Theoretical
Aspects of Computer Software, 4th International Symposium, TACS 2001, number 2215 in Lecture
Notes in Computer Science, pages 420–447, Sendai, Japan, October 2001. Springer-Verlag.
[11] Nuel D. Belnap. How a computer should think. In Gilbert Ryle, editor, Proceedings of the Oxford
International Symposium on Contemporary Aspects of Philosophy, pages 30–56, Oxford, England, 1976.
Oriel Press.
[12] Ulrich Berger, Matthias Eberl, and Helmut Schwichtenberg. Normalization by evaluation. In Bernhard
Möller and John V. Tucker, editors, Prospects for hardware foundations (NADA), number 1546 in
Lecture Notes in Computer Science, pages 117–137, Berlin, Germany, 1998. Springer-Verlag.
[13] Ulrich Berger and Helmut Schwichtenberg. An inverse of the evaluation functional for typed λ-calculus.
In Gilles Kahn, editor, Proceedings of the Sixth Annual IEEE Symposium on Logic in Computer Science,
pages 203–211, Amsterdam, The Netherlands, July 1991. IEEE Computer Society Press.
[14] Ilya Beylin and Peter Dybjer. Extracting a proof of coherence for monoidal categories from a proof
of normalization for monoids. In Stefano Berardi and Mario Coppo, editors, Types for Proofs and
Programs, International Workshop TYPES’95, number 1158 in Lecture Notes in Computer Science,
pages 47–61, Torino, Italy, June 1995. Springer-Verlag.
[15] Malgorzata Biernacka and Olivier Danvy. A concrete framework for environment machines. Research
Report BRICS RS-05-15, DAIMI, Department of Computer Science, University of Aarhus, Aarhus,
Denmark, May 2005.
[16] Malgorzata Biernacka and Olivier Danvy. A syntactic correspondence between context-sensitive calculi
and abstract machines. Research Report BRICS RS-05-22, DAIMI, Department of Computer Science,
University of Aarhus, Aarhus, Denmark, July 2005.
[17] Dariusz Biernacki and Olivier Danvy. From interpreter to logic engine by defunctionalization. In Maurice Bruynooghe, editor, Logic Based Program Synthesis and Transformation, 13th International Symposium, LOPSTR 2003, number 3018 in Lecture Notes in Computer Science, pages 143–159, Uppsala,
Sweden, August 2003. Springer-Verlag.
[18] Dariusz Biernacki, Olivier Danvy, and Kevin Millikin. A dynamic continuation-passing style for dynamic
delimited continuations. Research Report BRICS RS-05-16, DAIMI, Department of Computer Science,
University of Aarhus, Aarhus, Denmark, May 2005.
[19] Dariusz Biernacki, Olivier Danvy, and Chung-chieh Shan. On the dynamic extent of delimited continuations. Information Processing Letters, 96(1):7–17, 2005. Extended version available as the technical
report BRICS RS-05-13.
36
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
[20] Thierry Coquand and Peter Dybjer. Intuitionistic model constructions and normalization proofs. Mathematical Structures in Computer Science, 7:75–94, 1997.
[21] Olivier Danvy. On listing list prefixes. LISP Pointers, 2(3-4):42–46, January 1989. ACM Press.
[22] Olivier Danvy. Type-directed partial evaluation. In Guy L. Steele Jr., editor, Proceedings of the TwentyThird Annual ACM Symposium on Principles of Programming Languages, pages 242–257, St. Petersburg
Beach, Florida, January 1996. ACM Press.
[23] Olivier Danvy. Type-directed partial evaluation. In John Hatcliff, Torben Æ. Mogensen, and Peter
Thiemann, editors, Partial Evaluation – Practice and Theory; Proceedings of the 1998 DIKU Summer
School, number 1706 in Lecture Notes in Computer Science, pages 367–411, Copenhagen, Denmark,
July 1998. Springer-Verlag.
[24] Olivier Danvy. From reduction-based to reduction-free normalization. In Sergio Antoy and Yoshihito
Toyama, editors, Proceedings of the Fourth International Workshop on Reduction Strategies in Rewriting
and Programming (WRS’04), number 124 in Electronic Notes in Theoretical Computer Science, pages
79–100, Aachen, Germany, May 2004. Elsevier Science. Invited talk.
[25] Olivier Danvy. On evaluation contexts, continuations, and the rest of the computation. In Hayo Thielecke, editor, Proceedings of the Fourth ACM SIGPLAN Workshop on Continuations, Technical report
CSR-04-1, Department of Computer Science, Queen Mary’s College, pages 13–23, Venice, Italy, January
2004. Invited talk.
[26] Olivier Danvy. A rational deconstruction of Landin’s SECD machine. In Clemens Grelck, Frank Huch,
Greg J. Michaelson, and Phil Trinder, editors, Implementation and Application of Functional Languages,
16th International Workshop, IFL’04, number 3474 in Lecture Notes in Computer Science, pages 52–71,
Lübeck, Germany, September 2004. Springer-Verlag. Recipient of the 2004 Peter Landin prize. Extended
version available as the technical report BRICS-RS-03-33.
[27] Olivier Danvy and Andrzej Filinski. A functional abstraction of typed contexts. DIKU Rapport 89/12,
DIKU, Computer Science Department, University of Copenhagen, Copenhagen, Denmark, July 1989.
[28] Olivier Danvy and Andrzej Filinski. Abstracting control. In Wand [81], pages 151–160.
[29] Olivier Danvy and Andrzej Filinski. Representing control, a study of the CPS transformation. Mathematical Structures in Computer Science, 2(4):361–391, 1992.
[30] Olivier Danvy and Lasse R. Nielsen. Defunctionalization at work. In Harald Søndergaard, editor, Proceedings of the Third International ACM SIGPLAN Conference on Principles and Practice of Declarative Programming (PPDP’01), pages 162–174, Firenze, Italy, September 2001. ACM Press. Extended
version available as the technical report BRICS RS-01-23.
[31] Olivier Danvy and Lasse R. Nielsen. Refocusing in reduction semantics. Research Report BRICS RS-0426, DAIMI, Department of Computer Science, University of Aarhus, Aarhus, Denmark, November 2004.
A preliminary version appears in the informal proceedings of the Second International Workshop on
Rule-Based Programming (RULE 2001), Electronic Notes in Theoretical Computer Science, Vol. 59.4.
[32] Olivier Danvy and Zhe Yang. An operational investigation of the CPS hierarchy. In S. Doaitse Swierstra,
editor, Proceedings of the Eighth European Symposium on Programming, number 1576 in Lecture Notes
in Computer Science, pages 224–242, Amsterdam, The Netherlands, March 1999. Springer-Verlag.
[33] Scott Draves. Implementing bit-addressing with specialization. In Mads Tofte, editor, Proceedings of
the 1997 ACM SIGPLAN International Conference on Functional Programming, pages 239–250, Amsterdam, The Netherlands, June 1997. ACM Press.
[34] Peter Dybjer and Andrzej Filinski. Normalization and partial evaluation. In Gilles Barthe, Peter Dybjer,
Luı́s Pinto, and João Saraiva, editors, Applied Semantics – Advanced Lectures, number 2395 in Lecture
Notes in Computer Science, pages 137–192, Caminha, Portugal, September 2000. Springer-Verlag.
[35] R. Kent Dybvig, Simon Peyton-Jones, and Amr Sabry. A monadic framework for subcontinuations.
Technical Report 615, Computer Science Department, Indiana University, Bloomington, Indiana, June
2005.
[36] Matthias Felleisen. The Calculi of λ-v-CS Conversion: A Syntactic Theory of Control and State in
Imperative Higher-Order Programming Languages. PhD thesis, Computer Science Department, Indiana
University, Bloomington, Indiana, August 1987.
[37] Matthias Felleisen. The theory and practice of first-class prompts. In Jeanne Ferrante and Peter Mager,
editors, Proceedings of the Fifteenth Annual ACM Symposium on Principles of Programming Languages,
pages 180–190, San Diego, California, January 1988. ACM Press.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 37
[38] Matthias Felleisen and Matthew Flatt. Programming languages and lambda calculi. Unpublished lecture
notes. http://www.ccs.neu.edu/home/matthias/3810-w02/readings.html, 1989-2003.
[39] Matthias Felleisen and Daniel P. Friedman. Control operators, the SECD machine, and the λ-calculus.
In Martin Wirsing, editor, Formal Description of Programming Concepts III, pages 193–217. Elsevier
Science Publishers B.V. (North-Holland), Amsterdam, 1986.
[40] Matthias Felleisen, Daniel P. Friedman, Bruce Duba, and John Merrill. Beyond continuations. Technical
Report 216, Computer Science Department, Indiana University, Bloomington, Indiana, February 1987.
[41] Matthias Felleisen, Mitchell Wand, Daniel P. Friedman, and Bruce F. Duba. Abstract continuations:
A mathematical semantics for handling full functional jumps. In Robert (Corky) Cartwright, editor,
Proceedings of the 1988 ACM Conference on Lisp and Functional Programming, pages 52–62, Snowbird,
Utah, July 1988. ACM Press.
[42] Andrzej Filinski. Representing monads. In Hans-J. Boehm, editor, Proceedings of the Twenty-First
Annual ACM Symposium on Principles of Programming Languages, pages 446–457, Portland, Oregon,
January 1994. ACM Press.
[43] Andrzej Filinski. Representing layered monads. In Alex Aiken, editor, Proceedings of the Twenty-Sixth
Annual ACM Symposium on Principles of Programming Languages, pages 175–188, San Antonio, Texas,
January 1999. ACM Press.
[44] Andrzej Filinski. Normalization by evaluation for the computational lambda-calculus. In Samson
Abramsky, editor, Typed Lambda Calculi and Applications, 5th International Conference, TLCA 2001,
number 2044 in Lecture Notes in Computer Science, pages 151–165, Kraków, Poland, May 2001.
Springer-Verlag.
[45] Kathleen Fisher, editor. Proceedings of the 2004 ACM SIGPLAN International Conference on Functional Programming, Snowbird, Utah, September 2004. ACM Press.
[46] Martin Gasbichler and Michael Sperber. Final shift for call/cc: direct implementation of shift and reset.
In Simon Peyton Jones, editor, Proceedings of the 2002 ACM SIGPLAN International Conference on
Functional Programming, SIGPLAN Notices, Vol. 37, No. 9, pages 271–282, Pittsburgh, Pennsylvania,
September 2002. ACM Press.
[47] Matthew L. Ginsberg. Multivalued logics: a uniform approach to reasoning in artificial intelligence.
Computational Intelligence, 4:265–316, 1988.
[48] Bernd Grobauer and Zhe Yang. The second Futamura projection for type-directed partial evaluation.
Higher-Order and Symbolic Computation, 14(2/3):173–219, 2001.
[49] Carl Gunter, Didier Rémy, and Jon G. Riecke. A generalization of exceptions and control in ML-like
languages. In Simon Peyton Jones, editor, Proceedings of the Seventh ACM Conference on Functional
Programming and Computer Architecture, pages 12–23, La Jolla, California, June 1995. ACM Press.
[50] Thérèse Hardin, Luc Maranget, and Bruno Pagano. Functional runtime systems within the lambdasigma calculus. Journal of Functional Programming, 8(2):131–172, 1998.
[51] Simon Helsen and Peter Thiemann. Two flavors of offline partial evaluation. In Jieh Hsiang and Atsushi
Ohori, editors, Advances in Computing Science - ASIAN’98, number 1538 in Lecture Notes in Computer
Science, pages 188–205, Manila, The Philippines, December 1998. Springer-Verlag.
[52] Robert Hieb and R. Kent Dybvig. Continuations and concurrency. In Proceedings of the Second ACM
SIGPLAN Symposium on Principles & Practice of Parallel Programming, SIGPLAN Notices, Vol. 25,
No. 3, pages 128–136, Seattle, Washington, March 1990. ACM Press.
[53] Robert Hieb, R. Kent Dybvig, and Claude W. Anderson, III. Subcontinuations. Lisp and Symbolic
Computation, 5(4):295–326, December 1993.
[54] John Hughes. A novel representation of lists and its application to the function “reverse”. Information
Processing Letters, 22(3):141–144, 1986.
[55] Yukiyoshi Kameyama. Axioms for delimited continuations in the CPS hierarchy. In Jerzy Marcinkowski
and Andrzej Tarlecki, editors, Computer Science Logic, 18th International Workshop, CSL 2004, 13th
Annual Conference of the EACSL, Proceedings, volume 3210 of Lecture Notes in Computer Science,
pages 442–457, Karpacz, Poland, September 2004. Springer.
[56] Yukiyoshi Kameyama and Masahito Hasegawa. A sound and complete axiomatization of delimited
continuations. In Olin Shivers, editor, Proceedings of the 2003 ACM SIGPLAN International Conference
on Functional Programming, pages 177–188, Uppsala, Sweden, August 2003. ACM Press.
[57] Richard Kelsey, William Clinger, and Jonathan Rees, editors. Revised5 report on the algorithmic language Scheme. Higher-Order and Symbolic Computation, 11(1):7–105, 1998.
38
M. BIERNACKA, D. BIERNACKI, AND O. DANVY
[58] Yoshiki Kinoshita. A bicategorical analysis of E-categories. Mathematica Japonica, 47(1):157–169, 1998.
[59] Oleg Kiselyov. How to remove a dynamic prompt: Static and dynamic delimited continuation operators are equally expressible. Technical Report 611, Computer Science Department, Indiana University,
Bloomington, Indiana, March 2005.
[60] Peter J. Landin. The mechanical evaluation of expressions. The Computer Journal, 6(4):308–320, 1964.
[61] Julia L. Lawall and Olivier Danvy. Continuation-based partial evaluation. In Carolyn L. Talcott, editor,
Proceedings of the 1994 ACM Conference on Lisp and Functional Programming, LISP Pointers, Vol. VII,
No. 3, pages 227–238, Orlando, Florida, June 1994. ACM Press.
[62] Simon Marlow and Simon L. Peyton Jones. Making a fast curry: push/enter vs. eval/apply for higherorder languages. In Fisher [45], pages 4–15.
[63] Per Martin-Löf. About models for intuitionistic type theories and the notion of definitional equality. In
Proceedings of the Third Scandinavian Logic Symposium (1972), volume 82 of Studies in Logic and the
Foundation of Mathematics, pages 81–109. North-Holland, 1975.
[64] Eugenio Moggi. Notions of computation and monads. Information and Computation, 93:55–92, 1991.
[65] Luc Moreau and Christian Queinnec. Partial continuations as the difference of continuations, a duumvirate of control operators. In Manuel Hermenegildo and Jaan Penjam, editors, Sixth International
Symposium on Programming Language Implementation and Logic Programming, number 844 in Lecture
Notes in Computer Science, pages 182–197, Madrid, Spain, September 1994. Springer-Verlag.
[66] Chethan R. Murthy. Control operators, hierarchies, and pseudo-classical type systems: A-translation
at work. In Olivier Danvy and Carolyn L. Talcott, editors, Proceedings of the First ACM SIGPLAN
Workshop on Continuations (CW 1992), Technical report STAN-CS-92-1426, Stanford University, pages
49–72, San Francisco, California, June 1992.
[67] Lasse R. Nielsen. A denotational investigation of defunctionalization. Research Report BRICS RS-00-47,
DAIMI, Department of Computer Science, University of Aarhus, Aarhus, Denmark, December 2000.
[68] Gordon D. Plotkin. Call-by-name, call-by-value and the λ-calculus. Theoretical Computer Science,
1:125–159, 1975.
[69] Gordon D. Plotkin. A structural approach to operational semantics. Technical Report FN-19, DAIMI,
Department of Computer Science, University of Aarhus, Aarhus, Denmark, September 1981.
[70] Christian Queinnec and Bernard Serpette. A dynamic extent control operator for partial continuations. In Robert (Corky) Cartwright, editor, Proceedings of the Eighteenth Annual ACM Symposium on
Principles of Programming Languages, pages 174–184, Orlando, Florida, January 1991. ACM Press.
[71] John C. Reynolds. Definitional interpreters for higher-order programming languages. Higher-Order and
Symbolic Computation, 11(4):363–397, 1998. Reprinted from the proceedings of the 25th ACM National
Conference (1972), with a foreword.
[72] Chung-chieh Shan. Shift to control. In Olin Shivers and Oscar Waddell, editors, Proceedings of the
2004 ACM SIGPLAN Workshop on Scheme and Functional Programming, Technical report TR600,
Computer Science Department, Indiana University, Snowbird, Utah, September 2004.
[73] Dorai Sitaram. Models of Control and their Implications for Programming Language Design. PhD thesis,
Computer Science Department, Rice University, Houston, Texas, April 1994.
[74] Dorai Sitaram and Matthias Felleisen. Control delimiters and their hierarchies. Lisp and Symbolic
Computation, 3(1):67–99, January 1990.
[75] Dorai Sitaram and Matthias Felleisen. Reasoning with continuations II: Full abstraction for models of
control. In Wand [81], pages 161–175.
[76] Guy L. Steele Jr. Rabbit: A compiler for Scheme. Master’s thesis, Artificial Intelligence Laboratory,
Massachusetts Institute of Technology, Cambridge, Massachusetts, May 1978. Technical report AI-TR474.
[77] Eijiro Sumii. An implementation of transparent migration on standard Scheme. In Matthias Felleisen,
editor, Proceedings of the Workshop on Scheme and Functional Programming, Technical Report 00-368,
Rice University, pages 61–64, Montréal, Canada, September 2000.
[78] Eijiro Sumii and Naoki Kobayashi. A hybrid approach to online and offline partial evaluation. HigherOrder and Symbolic Computation, 14(2/3):101–142, 2001.
[79] Peter Thiemann. Combinators for program generation. Journal of Functional Programming, 9(5):483–
525, 1999.
[80] Philip Wadler. Monads and composable continuations. LISP and Symbolic Computation, 7(1):39–55,
January 1994.
AN OPERATIONAL FOUNDATION FOR DELIMITED CONTINUATIONS IN THE CPS HIERARCHY 39
[81] Mitchell Wand, editor. Proceedings of the 1990 ACM Conference on Lisp and Functional Programming,
Nice, France, June 1990. ACM Press.
[82] Yong Xiao, Amr Sabry, and Zena M. Ariola. From syntactic theories to interpreters: Automating proofs
of unique decomposition. Higher-Order and Symbolic Computation, 14(4):387–409, 2001.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http://creativecommons.org/licenses/by-nd/2.0/ or send a
letter to Creative Commons, 559 Nathan Abbott Way, Stanford, California 94305, USA.
| 2cs.AI
|
arXiv:1711.05851v1 [cs.CL] 15 Nov 2017
G O FOR A WALK AND A RRIVE AT THE A NSWER :
R EASONING OVER PATHS IN K NOWLEDGE BASES
USING R EINFORCEMENT L EARNING
Rajarshi Das?,1 , Shehzaad Dhuliawala?,1 , Manzil Zaheer2 , Luke Vilnis1 , Ishan Durugkar3
Akshay Krishnamurthy1 , Alex Smola4 , Andrew McCallum1
{rajarshi, sdhuliawala, luke, akshay, mccallum}@cs.umass.edu
[email protected], [email protected], [email protected]
1 University of Massachusetts, Amherst, 2 Carnegie Mellon University
3 University of Texas at Austin, 4 Amazon Web Services
A BSTRACT
Knowledge bases (KB), both automatically and manually constructed, are often
incomplete — many valid facts can be inferred from the KB by synthesizing
existing information. A popular approach to KB completion is to infer new relations
by combinatory reasoning over the information found along other paths connecting
a pair of entities. Given the enormous size of KBs and the exponential number of
paths, previous path-based models have considered only the problem of predicting
a missing relation given two entities, or evaluating the truth of a proposed triple.
Additionally, these methods have traditionally used random paths between fixed
entity pairs or more recently learned to pick paths between them. We propose a
new algorithm, MINERVA1 , which addresses the much more difficult and practical
task of answering questions where the relation is known, but only one entity. Since
random walks are impractical in a setting with combinatorially many destinations
from a start node, we present a neural reinforcement learning approach which
learns how to navigate the graph conditioned on the input query to find predictive
paths. Empirically, this approach obtains state-of-the-art results on several datasets,
significantly outperforming prior methods.
1
I NTRODUCTION
Automated reasoning, the ability of computing systems to make new inferences from observed
evidence, has been a long standing goal of artificial intelligence. We are interested in automated
reasoning on large knowledge bases (KB) with rich and diverse semantics (Suchanek et al., 2007;
Bollacker et al., 2008; Carlson et al., 2010). KBs are highly incomplete (Min et al., 2013), and facts
not directly stored in a KB can often be inferred from those that are, creating exciting opportunities
and challenges for automated reasoning. For example, consider the small knowledge graph in figure 1.
We can infer the (unobserved fact) home stadium of Colin Kaepernick from the following reasoning
path: Colin Kaepernick → PlaysInTeam → 49ers → TeamHomeStadium → Levi’s Stadium. Our goal
is to automatically learn such reasoning paths in KBs. We frame the learning problem as one of query
answering, that is to say, answering questions of the form (Colin Kaepernick, PlaysInLeague, ?).
From its early days, the focus of automated reasoning approaches has been to build systems which can
learn crisp symbolic logical rules (McCarthy, 1960; Nilsson, 1991). Symbolic representations have
also been integrated with machine learning especially in statistical relational learning (Muggleton
et al., 1992; Getoor & Taskar, 2007; Kok & Domingos, 2007; Lao et al., 2011), but due to poor
generalization performance, these approaches have largely been superceded by distributed vector
representations. Learning embedding of entities and relations using tensor factorization or neural
methods has been a popular approach (Nickel et al., 2011; Bordes et al., 2013; Socher et al., 2013;
inter alia), but these methods cannot capture chains of reasoning expressed by KB paths. Neural
multi-hop models (Neelakantan et al., 2015; Guu et al., 2015; Toutanova et al., 2016) address the
aforementioned problems to some extent by operating on KB paths in vector space. However, these
1 https://github.com/shehzaadzd/MINERVA
1
LocatedIn
Levis
Stadium
HasState
CA
TeamHome
Stadium
StateIn
49ers
CityIn
Nationality?
e
et
hl In
At lays m
P ea
T
HasCity
USA
Player
Home
Stadium?
Football
AthletePlays
Sports
(AthletePlays
Sports)-1
Colin
Kaepernick
Milwaukee
BornInCity
Michael
Crabtree
PlaysinLeague
PlaysinLeague?
Figure 1: A small fragment of
a knowledge base represented
as a knowledge graph. Solid
edges are observed and dashed
edges are part of queries. Note
how each query (e.g. Nationality, PlaysInLeague, PLayerHomeStadium) can be answered
by traversing the graph via “logical” paths between entity ‘Colin
Kaepernick’ and the corresponding answer.
NFL
models take as input a set of paths which are gathered by performing random walks independent of
the query relation. Additionally, models such as Neelakantan et al. (2015); Das et al. (2017) use the
same set of initially collected paths to answer a diverse set of query types (e.g. MarriedTo, Nationality,
WorksIn etc.).
This paper presents a method for efficiently searching the graph for answer-providing paths using reinforcement learning (RL) conditioned on the input question, eliminating any need for precomputed paths. Given a massive knowledge graph, we learn a policy, which, given the query
(entity1 , relation, ?), starts from entity1 and learns to walk to the answer node by choosing to take a
labeled relation edge at each step, conditioning on the query relation and entire path history. This
formulates the query-answering task as a reinforcement learning (RL) problem where the goal is to
take an optimal sequence of decisions (choices of relation edges) to maximize the expected reward
(reaching the correct answer node). We call the RL agent MINERVA for ”Meandering In Networks of
Entities to Reach Verisimilar Answers.”
Our RL-based formulation has many desirable properties. First, MINERVA has the built-in flexibility
to take paths of variable length, which is important for answering harder questions that require
complex chains of reasoning (Shen et al., 2017). Secondly, MINERVA needs no pretraining and
trains on the knowledge graph from scratch with reinforcement learning; no other supervision or
fine-tuning is required representing a significant advance over prior applications of RL in NLP. Third,
our path-based approach is computationally efficient, since by searching in a small neighborhood
around the query entity it avoids ranking all entities in the KB as in prior work. Finally, the reasoning
paths found by our agent automatically form an interpretable provenance for its predictions.
The main contributions of the paper are: (a) We present agent MINERVA, which learns to do query
answering by walking on a knowledge graph conditioned on an input query, stopping when it reaches
the answer node. The agent is trained using reinforcement learning, specifically policy gradients (§ 2).
(b) We evaluate MINERVA on several benchmark datasets and compare favorably to Neural Theorem
Provers (NTP) (Rocktäschel & Riedel, 2017) and Neural LP (Yang et al., 2017), which do logical
rule learning in KBs, and also state-of-the-art embedding based methods such as DistMult (Yang
et al., 2015) and ComplEx (Trouillon et al., 2016). (c) We also extend MINERVA to handle partially
structured natural language queries and test it on the WikiMovies dataset (§ 4.3) (Miller et al., 2016).
We also compare to DeepPath (Xiong et al., 2017) which uses reinforcement learning to pick paths
between entity pairs. The main difference is that the state of their RL agent includes the answer entity
since it is designed for the simpler task of predicting if a fact is true or not. As such their method
cannot be applied directly to our more challenging query answering task where the second entity is
unknown and must be inferred. Nevertheless, MINERVA outperforms DeepPath on their benchmark
NELL -995 dataset when compared in their experimental setting (§ 4.1).
2
TASK AND M ODEL
We formally define the task of query answering in a KB. Let E denote the set of entities and R be the
set of binary relations. Then a KB is a collection of facts stored as triplets (e1 , r, e2 ) where e1 , e2 ∈ E
and r ∈ R. Query answering seeks to answer questions of the form (e1 , r, ?), e.g. Toronto, locatedIn,
?. We would also like to clearly point out the difference between query answering and the task of
fact prediction. Fact prediction involves predicting if a fact is true or not, e.g. (Toronto, locatedIn,
2
Canada)?. This task is easier than predicting the correct entity as the answer in query answering
since the latter require finding the answer entity among many possible entities.
Next we describe how we reduce the problem of query answering in a KB to a finite horizon sequential
decision making problem and solve it using reinforcement learning. We begin by representing the
environment as a deterministic Markov decision process on a knowledgegraph G derived from
the KB (§2.1). Our RL agent is given an input query of the form e1q , rq , ? . Starting from vertex
corresponding to e1q in the knowledge graph G, the agent learns to traverse the environment/graph to
mine the answer and stop when it determines the answer (§ 2.2). The agent is trained using policy
gradient more specifically by REINFORCE (Williams, 1992) with control variates (§ 2.3). Let us begin
by describing the environment.
2.1
E NVIRONMENT - S TATES , ACTIONS , T RANSITIONS AND R EWARDS
Our environment is a finite horizon, deterministic and partially observed Markov decision process
that lies on a knowledge graph derived from the KB. Recall that a KB is collection of facts stored as
triplets (e1 , r, e2 ) where e1 , e2 ∈ E and r ∈ R. From the KB, a knowledge graph G can be constructed
where the entities e1 , e2 are represented as the nodes and relation r as labeled edge between them.
Formally, a knowledge graph is a directed labeled multigraph G = (V, E, R), where V and E denote
the vertices and edges of the graph respectively. Note that V = E and E ⊆ V × R ×V . Also, following
previous approaches (Bordes et al., 2013; Neelakantan et al., 2015; Xiong et al., 2017), we add the
inverse relation of every edge, i.e. for an edge (e1 , r, e2 ) ∈ E, we add the edge (e2 , r−1 , e1 ) to the
graph. (If the set of binary relations R does not contain the inverse relation r−1 , it is added to R
as well.) On this graph we will now specify a deterministic partially observable Markov decision
process, which is a 5-tuple (S, O, A, δ, R), each of which we elaborate below.
States. The state space S consists of all possible query-answers cartesian product with the set of
entities. Intuitively, we want a state to encode the query (e1q , rq ), the answer (e2q ), and a location
of exploration et (current node of the entity). Thus overall a state S ∈ S is represented by S =
(et , e1q , rq , e2q ) and the state space consists of all valid combinations.
Observations. The complete state of the environment is not observable, but only its current location of
exploration and query can be observed but not the answer, i.e. only (et , e1q , rq ) is observed. Formally
the observation function O : S → V ×V × R is defined as O(s = (et , e1q , rq , e2q )) = (et , e1q , rq ).
Actions. The set of possible actions AS from a state S = (et , e1q , rq , e2q ) consists of all outgoing edges
of the vertex et in G. Formally AS = {(et , r, v) ∈ E : S = (et , e1q , rq , e2q ), r ∈ R, v ∈ V } ∪ {(s, ∅, s)}.
Basically, this means an agent at each state has option to select which outgoing edge it wishes to take
having the knowledge of the label of the edge r and destination vertex v.
During implementation, we unroll the computation graph up to a fixed number of time steps T. We
augment each node with a special action called ‘NO OP’ which goes from a node to itself. Some
questions are easier to answer and needs lesser steps of reasoning than others. This design decision
allows the agent to remain at a node for any number of time steps. This is especially helpful when
the agent has managed to reach a correct answer at a time step t < T and can continue to stay at the
‘answer node’ for the rest of the time steps. Alternatively, we could have allowed the agent to take a
special ‘STOP’ action, but we found the current setup to work sufficiently well. As mentioned before,
we also add the inverse relation of a triple, i.e. for the triple (e1 , r, e2 ), we add the triple (e2 , r−1 , e1 )
to the graph. We found this important because this actually equips our agent to undo a potentially
wrong decision as it can retract back to the current node in the next step.
Transition. The environment evolves deterministically by just updating the state to the new vertex
pointed by the edge selected by the agent through its action. The query and answer remains the
same. Formally, the transition function is δ : S × A → S defined by δ(S, A) = (v, e1q , rq , e2q ), where
S = (et , e1q , rq , e2q ) and A = (et , r, v)).
Rewards. We only have a terminal reward of +1 if the current location is the correct answer at the
end and 0 otherwise. To elaborate, if ST = (et , e1q , rq , e2q ) is the final state, then we receive a reward
of +1 if et = e2q else 0.=, i.e. R(ST ) = I{et = e2q }.
2.2
P OLICY N ETWORK
To solve the finite horizon deterministic partially observable Markov decision process described
above, we aim to design a randomized history-dependent policy π = (d1 , d2 , ..., dT−1 ), where dt :
3
Ht → P(ASt ) and history Ht = (Ht−1 , At−1 , Ot ) is just the sequence of observations and actions taken.
We restrict ourselves to the function class expressed by long short-term memory network (LSTM)
(Hochreiter & Schmidhuber, 1997) for learning the randomized history-dependent policy.
An agent based on LSTM encodes the history Ht as a continuous vector ht ∈ R2d . We also have
embedding matrix r ∈ R|R|×d and e ∈ R|E |×d for the binary relations and entities respectively. The
history embedding for Ht = (Ht−1 , At−1 , Ot ) is updated according to LSTM dynamics:
ht = LSTM (ht−1 , [at−1 ; ot ])
(1)
where at−1 ∈ Rd and ot ∈ Rd denote the vector representation for action/relation at time t − 1
and observation/entity at time t respectively and [; ] denote vector concatenation. To elucidate,
at−1 = rAt−1 , i.e. the embedding of the relation corresponding to label of the edge the agent chose at
time t − 1 and ot = eet if Ot = (et , e1q , rq ) i.e. the embedding of the entity corresponding to vertex
the agent is at time t.
Based on the history embedding ht , the policy network makes the decision to choose an action
from all available actions (ASt ) conditioned on the query relation. Recall that each possible action
represents an outgoing edge with information of the edge relation label l and destination vertex/entity
d. So embedding for each A ∈ ASt is [rl ; ed ], and stacking embeddings for all the outgoing edges
we obtain the matrix At . The network taking these as inputs is parameterized as a two-layer feedforward network with ReLU nonlinearity which takes in the current history representation ht and the
embedding for the query relation rq and outputs a probability distribution over the possible actions
from which a discrete action is sampled. In other words,
dt = softmax (At (W2 ReLU (W1 [ht ; ot ; rq ])))
At ∼ Categorical (dt )
Note that the nodes in G do not have a fixed ordering or number of edges coming out from them. The
size of matrix At is |ASt | × 2d, so the decision probabilities dt lies on simplex of size |ASt |. Also the
procedure above is invariant to order in which edges are presented as desired and falls in purview of
neural networks designed to be permutation invariant Zaheer et al. (2017). Finally, to summarise,
the parameters of the LSTM, the weights W1 , W2 , the corresponding biases (not shown above for
brevity), and the embedding matrices form the parameters θ of the policy network.
2.3
T RAINING
For the policy network (πθ ) described above, we want to find parameters θ that maximizes the
expected reward:
J(θ) = E(e1 ,r,e2 )∼D EA1 ,..,AT −1 ∼πθ [R(ST )|S1 = (e1 , e1 , r, e2 )]
where we assume there is a true underlying distribution (e1 , r, e2 ) ∼ D. To solve this optimization
problem, we employ REINFORCE (Williams, 1992) as follows:
• The first expectation is replaced with empirical average over the training dataset.
• For the second expectation, we approximate by running multiple rollouts for each training
example. The number of rollouts is fixed and for all our experiments we set this number to 20.
• For variance reduction, a common strategy is to use an additive control variate baseline
(Hammersley, 2013; Fishman, 2013; Evans & Swartz, 2000). We use a moving average of
the cumulative discounted reward as the baseline. We tune the weight of this moving average
as a hyperparameter. Note that in our experiments we found that learnt baseline performed
similarly, but we finally settled for cumulative discounted reward as the baseline owing to its
simplicity.
• To encourage the policy to sample more diverse paths rather than sticking with a few, we add
an entropy regularization term to our cost function after multiplying it by a constant (β). We
treat β as a hyperparameter to control the exploration exploitation trade-off.
Experimental Details We choose the relation and embedding dimension size as 200. The action
embedding is formed by concatenating the entity and relation embedding. We use a 3 layer LSTM
with dimension size of 400. The hidden layer size of MLP (weights W1 and W2 ) is set to 400. We
use Adam (Kingma & Ba, 2014) with the default parameters in REINFORCE for the update. The
best hyperparameter values can be found in appendix.
4
Dataset
COUNTRIES
UMLS
KINSHIP
WN 18 RR
NELL -995
FB 15 K -237
WikiMovies
#entities
272
135
104
40,945
75,492
14,505
43,230
#relations
2
49
26
15
200
237
9
#facts
1158
5,216
10686
86,835
154,213
272,115
196,453
#queries
24
661
1074
3134
3992
20,466
9952
Table 1: Statistics of various datasets used in experiments.
Task
S1
S2
S3
Metric
AUC-PR
ComplEx
99.37±0.4
87.95±2.8
48.44±6.3
Model
NTP
NTP-λ
90.83±15.4
100.0±0.0
87.4±11.7
93.04±0.4
56.68±17.6 77.26±17.0
MINERVA
100.0±0.0
91±0.01
93±0.01
Table 2: Performance on COUNTRIES dataset. MINERVA significantly outperforms baselines in the
challenging S3 task.
3
DATA
We test our model on the following query answering datasets. (a) COUNTRIES (Bouchard et al., 2015),
(b) Alyawarra kinship (KINSHIP), (c) Unified Medical Language Systems (UMLS) (Kok & Domingos,
2007) (d) WN 18 RR (Dettmers et al., 2017), (e) NELL -995, (f) FB15k-237 (g) WikiMovies (Miller
et al., 2016). We also test on a synthetic grid world dataset released by Yang et al. (2017) to test the
ability of the model to learn rules of long length.
The COUNTRIES dataset is carefully designed to explicitly test the logical rule learning and reasoning
capabilities of link prediction models. The dataset has 3 tasks (S1-3 in table 2) each requiring
reasoning steps of increasing length and difficulty (see Rocktäschel & Riedel (2017) for more details
about the tasks). We also test our model on existing large and challenging KG datasets ((d) - (f)).
WN 18 RR is created from the original WORDNET 18 dataset by removing test triples which can be
answered trivially, making the datasets more realistic and challenging. Additionally, we test our
model on a question answering dataset - WikiMovies (Miller et al., 2016) where the query is in
natural language but the answers can be found in an accompanying KB. Table 1 report the various
statistics of the datasets.
4
4.1
E XPERIMENTS
K NOWLEDGE G RAPH Q UERY A NSWERING
This section describes the experimental results on the various knowledge graph query answering
datasets. During inference, we do beam search with a beam width of 40 and rank entities by the
probability of the trajectory the model took to reach the entity.
COUNTRIES , KINSHIP, UMLS. We first test MINERVA on the
COUNTRIES dataset which is explicitly designed to test the ability
Model
NeuralLP
UMLS
KINSHIP
0.70
0.91
0.73
of models to learn logical rules. It contains countries, regions and
MINERVA
0.93
subregions as entities. The queries are of the form LocatedIn(c, Table 3: HITS@10 on UMLS and
?) and the answer is a region. For example, LocatedIn(Egypt, ?) KINSHIP
with the answer as Africa. Our experimental settings and scores
are directly comparable to NTP and ComplEx (Trouillon et al., 2016). NTP-λ is a NTP model trained
with an additional objective function of ComplEx. We also compare MINERVA against Neural LP
(Yang et al., 2017) on the UMLS and KINSHIP datasets.
The evaluation metric we report is HITS @k - which is the percentage of correct entities ranked in
top-k. For the COUNTRIES dataset, we report the area under the precision-recall curve for comparing
with the baselines. Following the design of the COUNTRIES dataset, for task S1 and S2, we set the
maximum path length T = 2 and for S3, we set T = 3.
5
DeepPath
0.960
0.711
0.742
0.957
0.738
0.795
0.890
0.790
0.750
MINERVA
0.970
0.825
0.851
0.985
0.846
0.793
0.895
0.946
0.824
Table 4: MAP scores for different query relations on
the NELL -995 dataset. Note that in this comparison,
MINERVA refers to only a single learnt model for all query
relations which is competitive with individual DeepPath
models trained separately for each query relation.
Accuracy
Task
athleteplaysinleague
worksfor
organizationhiredperson
athleteplayssport
teamplayssport
personborninlocation
athletehomestadium
organizationheadquarteredincity
athleteplaysforteam
1.0
0.9
0.8
0.7
0.6
0.5
0.4
MINERVA
Neural LP
2-4
4-6 6-8
Path length
8-10
Figure 2: Grid world experiment:
We significantly outperform NeuralLP for longer path lengths.
Table 2 shows that MINERVA outperforms all the baseline models except on the task S2 of COUNTRIES,
where the ensemble model NTP -λ outperforms it, albeit with a higher variance across runs. Our
gains are much more prominent in task S3, which is the hardest among all the tasks. We similarly
outperform NeuralLP on the UMLS and KINSHIP datasets.
N ELL-995 We also compare MINERVA to DeepPath. For a fair comparison, we only rank the answer
entities against the negative examples in the dataset used in their experiments2 and report the mean
average precision (MAP) scores for each query relation. DeepPath feeds the paths its agent gathers
as input features to the path ranking algorithm (PRA) (Lao et al., 2011), which trains a per-relation
classifier. But unlike them, we train one model which learns for all query relations. If our agent is not
able to reach the correct entity or one of the negative entities, the corresponding query gets a score of
negative infinity. As show in table 6, we outperform them or achieve comparable performance for all
the query relations For this experiment, we set the maximum length T = 3.
WN 18 RR Next we test MINERVA on another large KB
Model
HITS @1
HITS @3
HITS @10
dataset – WN 18 RR. On this dataset, we compare with
ConvE
0.306
0.360
0.411
three recently proposed latent factorization model – (a)
DistMult
0.389
0.439
0.491
ConvE (Dettmers et al., 2017), (b) DistMult (Yang et al., ComplEx
0.411
0.458
0.507
2015), (c) ComplEx (Trouillon et al., 2016). We report
MINERVA
0.413
0.456
0.513
HITS at various k and we compare favorably with the
Table 5: Performance on WN 18 RR
state-of-the-art results of ComplEx in all settings (table
5). For this experiment, we also set the maximum length T = 3.
FB15k-237 We test MINERVA on yet another popular KB dataset FB15K- Model
HITS @10
ConvE
0.458
237. The baselines are the same as before, however our implementation
DistMult
0.568
of DistMult gave a score of 56.8 HITS @10 which, to our knowledge, is
ComplEx
0.419
the highest score reported on this dataset3 .The performance of ConvE and
MINERVA
0.456
ComplEx are taken from Dettmers et al. (2017). Even though MINERVA
performs comparably to ConvE and ComplEx, the results are significantly Table 6: Performance
behind the performance of DistMult.
on FB 15 K -237
Upon delving more into the structure of knowledge graph derived from FB15K-237, we found few
interesting characteristics of the dataset. As a prelude, we would like to describe a long existing
concept in graph theory – clustering coefficient (Holland & Leinhardt, 1971; Watts & Strogatz,
1998). Clustering coefficient (τ) of a graph measures whether groups of nodes form ‘tightly knit’
communities - i.e. whether groups of nodes tend to cluster together. A high τ implies the presence
of higher number of densely connected groups of nodes. For instance, if we consider three nodes
A, B and C, a high τ means with high probability whenever three nodes are connected as A — B — C,
it implies nodes A — C are also connected forming a triangle. Intuitively, MINERVA can use such
closed shapes to learn paths such as (A — B — C) to predict the answer of the query, i.e. the third
node (C). The clustering coefficient also extends from triangles to cliques of arbitrary size (Watts &
2 We
are grateful to Xiong et al. (2017) for releasing the negative examples used in their experiments.
are aware of the high variance of DistMult scores reported on FB15k-237 by several papers, but to
ensure fairness we report the high scores our in-house implementation achieved.
3 We
6
6 × 10
1
4 × 10
3 × 10
1
2 × 10
1
NELL-995
FB15k-237
104
2 × 103
103
102
101
104
1
102
103
103
100
101
102
103
101
102
103
104
trie
s
Kin
sh
ip
101
un
Co
LL-
k-2
NE
15
FB
99
5
UM
LS
103
37
Clustering Coefficient
Kinship
3 × 103
100
Figure 3: Network avg. cluster coefficient of various
datasets
Figure 4: Count of number of unique path types of length 3 which
occur more than ‘x’ times in various datasets. In Kinship and NELL995, there are more than 103 path types which occur more than
103 times, however for FB15k-237, we see a sharp decrease as ‘x’
becomes higher.
Strogatz, 1998). Figure 3 plots τ for various datasets. We find that FB15k-237 has the least clustering
coefficient (0.19) among all datasets. This means that the dataset has sparse neighborhoods and hence
MINERVA finds it difficult to learn logical rules. We also check the frequency of occurrence of various
unique paths (types). We define a path type as the sequence of relations (ignoring the entities) in
a path. Intuitively, a predictive path which generalizes across queries will occur many number of
times in the graph. Figure 4 shows the plot. As we can see, the characteristics of FB15k-237 is quite
different from other datasets. Path types do not repeat that often, making it hard for MINERVA to
learn paths which generalizes. We also provide further analysis of the types of various query relation
in FB15k-237 in the appendix.
4.2
G RID W ORLD PATH F INDING
As we empirically find and also noted by previous work (Rocktäschel & Riedel, 2017; Das et al.,
2017; Yang et al., 2017), often the reasoning chains required to answer queries in KB is not too long
(restricted to 3 or 4 hops). To test if our model can learn long reasoning paths, we test our model on
a synthetic 16-by-16 grid world dataset created by Yang et al. (2017), where the task is to navigate
to a particular cell (answer entity) starting from a random cell (start entity) by following a set of
directions (query relation). The KB consists of atomic triples of the form ((2,1), North, (1,1)) – entity
(1,1) is north of entity (2,1). The queries consists of a sequence of directions (e.g. North, SouthWest,
East). The queries are classified into classes based on the path lengths. Figure 2 shows the accuracy
on varying path lengths. Compared to Neural LP, MINERVA is much more robust for queries which
require longer path lengths showing a very little degrade in performance for even the longest path
length in the dataset.
4.3
PARTIALLY S TRUCTURED Q UERIES
Queries in KB datasets are structured in the form of triples. Model
Accuracy
However, this is unsatisfactory since for most real applica- Memory Network
78.5
tions, the queries appear in natural language. As a first step
QA system
93.5
Key-Value Memory Network
93.9
in this direction, we extend MINERVA to take in “partially
Neural LP
94.6
structured” queries. We use the WikiMovies dataset (Miller
MINERVA
96.7
et al., 2016) which contains questions in natural language
albeit generated by templates created by human annotators.
Table 7: Performance on WikiMovies
An example question from the dataset is “Which is a film
written by Herb Freed?”. WikiMovies also has an accompanying KB which can be used to answer all
the questions.
We link the entity occurring in the question to the KB via simple string matching. To form the vector
representation of the query relation, we design a simple question encoder which computes the average
of the embeddings of the question words. The word embeddings are learned from scratch and we do
not use any pretrained embeddings. We compare our results with those reported in Yang et al. (2017)
(table 7). We got the best result using T = 1, suggesting that WikiMovies is not the best testbed for
multihop reasoning, but this experiment is a promising first step towards the realistic setup of having
textual queries and knowledge bases.
7
NBA
A
Pl thle
a
t
L ys e
p eag in
= 0 ue
.98
7
s
he
ac
Co eam 1
T 0.0
p=
LSU
NBA
p
LSU
A
Pl thle
Le ays te
= agu in
0. e
00
6
Coaches
Team
p = 0.898
Basketball
John Brady
WorksFor?
John Brady
AthletePlaysSport?
r
cto
dA 0
re -3
ar 1e
St p =
Pennies
From
Heaven
Louis
Armstrong
d
re
ar
St ctor 7
A 0.6
=
p
Pennies
From
Heaven
Dire
cted
By
p=0
.58
Normal
McLeod
Louis
Armstrong
Dire
cted
By
p=2
e-19
Normal
McLeod
Who directed the movie “Pennies from Heaven”?
Who starred in the movie “Pennies from Heaven”?
Figure 5: Based on the query relation our agent assigns different probabilities to different actions.
The dashed edges in the top row denote query relation. Examples in the bottom row are from the
WikiMovies dataset and hence the questions are partially structured.
5
A NALYSIS
Effectiveness of Remembering Path History. MINERVA encodes the history of decisions it has
taken in the past using LSTMs. To test the importance of remembering the sequence of decisions,
we did an ablation study in which the agent chose the next action based on only local information
i.e. current entity and query and did not have access to the history ht . For the KINSHIP dataset, we
observe a 27% points decrease in HITS@1 (0.184 v/s 0.46) and 13% decrease in HITS@10 (0.63
v/s 0.76). For grid-world, it is also not surprising that we see a big drop in performance. The final
accuracy is 0.23 for path lengths 2-4 and 0.04 for lengths 8-10. For NELL, the performance dropped
from 0.576 to 0.564 and for FB15k-237 the HITS@10 performance dropped from 0.456 to 0.408.
NO-OP and Inverse Relations. At each step, MINERVA can choose to take a NO - OP edge and
remain at the same node. This gives the agent the flexibility of taking paths of variable lengths.
Some questions are easier to answer than others and require lesser steps of reasoning and if the agent
reaches the answer early, it can choose to remain there. Example (i) in table 8 shows such an example.
Similarly inverse relation gives the agent the ability to recover from a potentially wrong decision it
has taken before. Example (ii) shows such an example, where the agent took a incorrect decision at
the first step but was able to revert the decision because of the presence of inverted edges.
Query based Decision Making. At each step before making a decision, our agent conditions on the
query relation. Figure 5 shows examples, where based on the query relation, the probabilities are
peaked on different actions. For example, when the query relation is WorksFor, MINERVA assigns
a much higher probability of taking the edge CoachesTeam than AthletePlaysInLeague. We also
see similar behavior on the WikiMovies dataset where the query consists of words instead of fixed
schema relation.
Inference Time. MINERVA is efficient at inference time since it has to essentially search for answer
entities in its local neighborhood, whereas previous methods rank all the entities in the dataset. For
instance, on the test dataset of WN 18 RR, the wall clock running time of MINERVA is 63 seconds
whereas that of a GPU implementation of DistMult is 211 seconds (with the maximum batch size).
6
R ELATED W ORK
Learning vector representations of entities and relations using tensor factorization (Nickel et al.,
2011; 2012; Bordes et al., 2013; Riedel et al., 2013; Nickel et al., 2014; Yang et al., 2015) or
neural methods (Socher et al., 2013; Toutanova et al., 2015; Verga et al., 2016) has been a popular
approach to reasoning with a knowledge base. However, these methods cannot capture more complex
reasoning patterns such as those found by following inference paths in KBs. Multi-hop link prediction
approaches (Lao et al., 2011; Neelakantan et al., 2015; Guu et al., 2015; Toutanova et al., 2016; Das
et al., 2017) address the problems above, but the reasoning paths that they operate on are gathered by
8
(i) Can learn general rules:
(S1) LocatedIn(X, Y) ← LocatedIn(X, Z) & LocatedIn(Z, Y)
(S2) LocatedIn(X, Y) ← NeighborOf(X, Z) & LocatedIn(Z, Y)
(S3) LocatedIn(X, Y) ← NeighborOf(X, Z) & NeighborOf(Z, W) & LocatedIn(W, Y)
WorksFor
(ii) Can learn shorter path: Richard F. Velky −−−−−→?
PersonLeadsOrg
NO-OP
NO-OP
Richard F. Velky −−−−−−−−−→ Schaghticokes −−−−→ Schaghticokes −−−−→ Schaghticokes
WorksFor
(iii) Can recover from mistakes: Donald Graham −−−−−→?
OrgTerminatedPerson
OrgTerminatedPerson−1
OrgHiredPerson
Donald Graham −−−−−−−−−−−−→ TNT Post −−−−−−−−−−−−−→ Donald Graham −−−−−−−−−→ Wash Post
Table 8: A few example of paths found by MINERVA on the COUNTRIES and NELL. MINERVA can
learn general rules as required by the COUNTRIES dataset (example (i)). It can learn shorter paths if
necessary (example (ii)) and has the ability to correct a previously taken decision (example (iii))
.
performing random walks independent of the type of query relation. Lao et al. (2011) further filters
paths from the set of sampled paths based on the restriction that the path must end at one of the target
entities in the training set and are within a maximum length. These constraints make them query
dependent but they are heuristic in nature. Our approach eliminates any necessity to pre-compute
paths and learns to efficiently search the graph conditioned on the input query relation.
Inductive Logic Programming (ILP) (Muggleton et al., 1992) aims to learn general purpose predicate
rules from examples and background knowledge. Early work in ILP such as FOIL (Quinlan, 1990),
PROGOL (Muggleton, 1995) are either rule-based or require negative examples which is often hard
to find in KBs (by design, KBs store true facts). Statistical relational learning methods (Getoor
& Taskar, 2007; Kok & Domingos, 2007; Schoenmackers et al., 2010) along with probabilistic
logic (Richardson & Domingos, 2006; Broecheler et al., 2010; Wang et al., 2013) combine machine
learning and logic but these approaches operate on symbols rather than vectors and hence do not
enjoy the generalization properties of embedding based approaches.
Neural Theorem Provers (NTP) (Rocktäschel & Riedel, 2017) and Neural LP (Yang et al., 2017)
are two recent methods in learning logical rules that can be trained end-to-end with gradient based
learning. NTPs are constructed by Prolog’s backward chaining inference method. It operates on
vectors rather than symbols, thereby providing a success score for each proof path. However, since a
score can be computed between any two vectors, the computation graph becomes quite large because
of such soft-matching during substitution step of backward chaining. For tractability, it resides to
heuristics such as only keeping the top-K scoring proof paths, but it loses any guarantee of computing
exact gradients. Also the efficacy of NTPs has yet to be shown on large KBs. Neural LP introduces
a differential rule learning system using operators defined in TensorLog (Cohen, 2016) and has a
LSTM based controller and a differentiable memory component (Graves et al., 2014; Sukhbaatar
et al., 2015) and the rule scores are calculated via attention. Even though, differentiable memory
allows the network to be trained end to end, it necessitates accessing the entire memory which can be
computationally expensive. RL approaches which can make hard selection of memory (Zaremba &
Sutskever, 2015) are computationally attractive. MINERVA uses a similar hard selection of relation
edges to walk on the graph. More importantly, MINERVA outperforms both these methods on their
respective benchmark datasets.
DeepPath (Xiong et al., 2017) uses RL based approaches to find paths in KBs. However, the state of
their MDP requires the target entity to be known in advance and hence their path finding strategy
is dependent on knowing the answer entity. MINERVA does not need any knowledge of the target
entity and instead learns to find the answer entity among all entities. DeepPath, additionally feeds its
gathered paths to Path Ranking Algorithm (Lao et al., 2011), whereas MINERVA is a complete system
9
trained to do query answering. DeepPath also uses fixed pretrained embeddings for its entity and
relations. Lastly, on comparing MINERVA with DeepPath in their experimental setting on the NELL
dataset, we match their performance or outperform them.
MINERVA is also similar to methods for learning to search for structured prediction (Collins & Roark,
2004; Daumé III & Marcu, 2005; Daumé III et al., 2009; Ross et al., 2011; Chang et al., 2015). These
methods are based on imitating a reference policy (oracle) which make near-optimal decision at
every step. In our problem setting, it is unclear what a good reference policy would be. For example,
a shortest path oracle between two entities would be bad, since the answer providing path should
depend on the query relation.
7
C ONCLUSION
We explored a new way of automated reasoning on large knowledge bases in which we use the
knowledge graphs representation of the knowledge base and train an agent to walk to the answer node
conditioned on the input query. We achieve state-of-the-art results on multiple benchmark knowledge
base completion tasks and we also show that our model is robust and can learn long chains-ofreasoning. Moreover it needs no pretraining or initial supervision. Future research directions include
applying more sophisticated RL techniques and working directly on textual queries and documents.
ACKNOWLEDGEMENTS
This work was supported in part by the Center for Data Science and the Center for Intelligent
Information Retrieval, in part by DARPA under agreement number FA8750-13-2-0020, in part by
Defense Advanced Research Agency (DARPA) contract number HR0011-15-2-0036, in part by the
National Science Foundation (NSF) grant numbers DMR-1534431 and IIS-1514053 and in part by
the Chan Zuckerberg Initiative under the project Scientific Knowledge Base Construction. The U.S.
Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon. Any opinions, findings and conclusions or recommendations
expressed in this material are those of the authors and do not necessarily reflect those of the sponsor.
10
R EFERENCES
Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. Freebase: A collaboratively created graph database for structuring human knowledge. In ICDM, 2008.
Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko.
Translating embeddings for modeling multi-relational data. In NIPS, 2013.
Guillaume Bouchard, Sameer Singh, and Theo Trouillon. On approximate reasoning capabilities of
low-rank vector spaces. AAAI Spring Symposium, 2015.
Matthias Broecheler, Lilyana Mihalkova, and Lise Getoor. Probabilistic similarity logic. In UAI,
2010.
Andrew Carlson, Justin Betteridge, Bryan Kisiel, Burr Settles, Estevam R. Hruschka, Jr., and Tom M.
Mitchell. Toward an Architecture for Never-ending Language Learning. In AAAI, 2010.
Kai-Wei Chang, Akshay Krishnamurthy, Alekh Agarwal, Hal Daume, and John Langford. Learning
to search better than your teacher. In ICML, 2015.
William Cohen. Tensorlog: A differentiable deductive database. arXiv:1605.06523, 2016.
Michael Collins and Brian Roark. Incremental parsing with the perceptron algorithm. In ACL, 2004.
Rajarshi Das, Arvind Neelakantan, David Belanger, and Andrew McCallum. Chains of reasoning
over entities, relations, and text using recurrent neural networks. In EACL, 2017.
Hal Daumé III and Daniel Marcu. Learning as search optimization: Approximate large margin
methods for structured prediction. In ICML, 2005.
Hal Daumé III, John Langford, and Daniel Marcu. Search-based structured prediction. Machine
learning, 2009.
Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. Convolutional 2d
knowledge graph embeddings. arXiv:1707.01476, 2017.
Michael Evans and Timothy Swartz. Approximating integrals via Monte Carlo and deterministic
methods. OUP Oxford, 2000.
George Fishman. Monte Carlo: concepts, algorithms, and applications. Springer Science & Business
Media, 2013.
Lise Getoor and Ben Taskar. Introduction to statistical relational learning. MIT press, 2007.
Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv:1410.5401, 2014.
Kelvin Guu, John Miller, and Percy Liang. Traversing knowledge graphs in vector space. In EMNLP,
2015.
John Hammersley. Monte carlo methods. Springer Science & Business Media, 2013.
Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 1997.
Paul W Holland and Samuel Leinhardt. Transitivity in structural models of small groups. Comparative
Group Studies, 1971.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980,
2014.
Stanley Kok and Pedro Domingos. Statistical predicate invention. In ICML, 2007.
Ni Lao, Tom Mitchell, and William Cohen. Random walk inference and learning in a large scale
knowledge base. In EMNLP, 2011.
John McCarthy. Programs with common sense. RLE and MIT Computation Center, 1960.
Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason
Weston. Key-value memory networks for directly reading documents. EMNLP, 2016.
11
Bonan Min, Ralph Grishman, Li Wan, Chang Wang, and David Gondek. Distant supervision for
relation extraction with an incomplete knowledge base. In HLT-NAACL, 2013.
Stephen Muggleton. Inverse entailment and progol. New generation computing, 1995.
Stephen Muggleton, Ramon Otero, and Alireza Tamaddoni-Nezhad. Inductive logic programming.
Springer, 1992.
Arvind Neelakantan, Benjamin Roth, and Andrew McCallum. Compositional vector space models
for knowledge base completion. In ACL, 2015.
Maximilian Nickel, Volker Tresp, and Hans-Peter Kriegel. A three-way model for collective learning
on multi-relational data. In ICML, 2011.
Maximilian Nickel, Volker Tresp, and Hans-Peter Kriegel. Factorizing yago: scalable machine
learning for linked data. In WWW, 2012.
Maximilian Nickel, Xueyan Jiang, and Volker Tresp. Reducing the rank in relational factorization
models by including observable patterns. In NIPS, 2014.
Nils J Nilsson. Logic and artificial intelligence. Artificial intelligence, 1991.
J Ross Quinlan. Learning logical definitions from relations. Machine learning, 1990.
Matthew Richardson and Pedro Domingos. Markov logic networks. Machine learning, 2006.
Sebastian Riedel, Limin Yao, Andrew McCallum, and Benjamin M. Marlin. Relation extraction with
matrix factorization and universal schemas. In NAACL, 2013.
Tim Rocktäschel and Sebastian Riedel. End-to-end differentiable proving. In NIPS, 2017.
Stéphane Ross, Geoffrey J Gordon, and Drew Bagnell. A reduction of imitation learning and
structured prediction to no-regret online learning. In AISTATS, 2011.
Stefan Schoenmackers, Oren Etzioni, Daniel Weld, and Jesse Davis. Learning first-order horn clauses
from web text. In EMNLP, 2010.
Yelong Shen, Po-Sen Huang, Jianfeng Gao, and Weizhu Chen. Reasonet: Learning to stop reading in
machine comprehension. In KDD, 2017.
Richard Socher, Danqi Chen, Christopher D Manning, and Andrew Ng. Reasoning with neural tensor
networks for knowledge base completion. In NIPS, 2013.
Fabian Suchanek, Gjergji Kasneci, and Gerhard Weikum. Yago: A core of semantic knowledge. In
WWW, 2007.
Sainbayar Sukhbaatar, Jason Weston, and Rob Fergus. End-to-end memory networks. In NIPS, 2015.
Kristina Toutanova, Danqi Chen, Patrick Pantel, Hoifung Poon, Pallavi Choudhury, and Michael
Gamon. Representing text for joint embedding of text and knowledge bases. In EMNLP, 2015.
Kristina Toutanova, Victoria Lin, Wen-tau Yih, Hoifung Poon, and Chris Quirk. Compositional
learning of embeddings for relation paths in knowledge base and text. In ACL, 2016.
Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. Complex
embeddings for simple link prediction. In ICML, 2016.
Patrick Verga, David Belanger, Emma Strubell, Benjamin Roth, and Andrew McCallum. Multilingual
relation extraction using compositional universal schema. In NAACL, 2016.
William Yang Wang, Kathryn Mazaitis, and William W Cohen. Programming with personalized
pagerank: a locally groundable first-order probabilistic logic. In CIKM, 2013.
Duncan J Watts and Steven H Strogatz. Collective dynamics of small-worldnetworks. nature, 1998.
Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement
learning. Machine learning, 1992.
12
Wenhan Xiong, Thien Hoang, and William Yang Wang. Deeppath: A reinforcement learning method
for knowledge graph reasoning. In EMNLP, 2017.
Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Embedding entities and
relations for learning and inference in knowledge bases. In ICLR, 2015.
Fan Yang, Zhilin Yang, and William W Cohen. Differentiable learning of logical rules for knowledge
base reasoning. In NIPS, 2017.
Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Ruslan Salakhutdinov, and
Alexander Smola. Deep sets. In NIPS, 2017.
Wojciech Zaremba and Ilya Sutskever.
arXiv:1505.00521, 2015.
Reinforcement learning neural turing machines.
13
(i) M to 1
team plays sport
Los Angeles Rams −−−−−−−−−→ American Football
country of origin
The Walking Dead −−−−−−−−−→ USA
(ii) 1 to M
job position in organization
CEO −−−−−−−−−−−−−−−→ Merck & Co.
cause of death
Traffic collision −−−−−−−−→ Albert Camus
instrument played by musician
Harmonica −−−−−−−−−−−−−−−−−→ Greg Graffin
Table 9: Few example facts belonging to m to 1, 1 to m relations in FB15k-237
.
Relation
/people/marriage union type/unions of this type./people/marriage/location of ceremony
/organization/role/leaders./organization/leadership/organization
/location/country/second level divisions
/user/ktrueman/default domain/international organization/member states
/base/marchmadness/ncaa basketball tournament/seeds./base/marchmadness/ncaa tournament seed/team
Table 10: Few example 1-to-M relations from FB15k-237 with high cardinality ratio of tail to head.
8
8.1
A PPENDIX
A NALYSIS OF QUERY RELATIONS OF FB15 K -237
We further did some query analysis on the FB15k-237 dataset. Following Bordes et al. (2013), we
categorized the query relations into (M)any to 1, 1 to M and 1 to 1 relations. An example of a M
to 1 relation would be ‘/people/profession’ (What is the profession of ‘X’?). An example of 1 to
M relation would be ‘/people/cause of death/people’. An example query of that relation would be
(Traffic collision, /people/cause of death/people, ?) ‘Who were killed in traffic collision accidents?’.
Another example would be /music/instrument/instrumentalists (Who plays the music instrument ‘X’?).
From a query answering point of view, the answers to this question is a list of entities. However,
during evaluation time, the model is evaluated based on whether it is able to predict the one target
entity which is in the query triple. Also since MINERVA outputs the end points of the paths as target
entities, it is sometimes possible that the particular target entity of the triple does not have a path
from the source entity (however there are paths to other ‘correct’ answer entities). Table 9 shows few
other examples of relations belonging to different classes.
Following Bordes et al. (2013), we classify a relation as 1-to-M if the ratio of cardinality of tail to head
entities is greater than 1.5 and as M-to-1 if it is lesser than 0.67. In the validation set of FB15k-237,
54% of the queries are 1-to-M, whereas only 26% are M-to-1. Contrasting it with NELL-995, 27% are
1-to-M and 36% are M-to-1 or UMLS or KINSHIP where only 18% and 32% of the relations are 1-to-M.
Table 10 shows relations from FB15k-237 dataset which have tail-to-head ratio. The average ratio
for 1- TO -M relations in FB15k-237 is 13.39 (substantially higher than 1.5). As explained before,
the current evaluation scheme is unfair when it comes to 1-to-M relations and the high percentage of
1-to-M relations in FB15k-237 also explains the sub optimal performance of MINERVA.
14
tail/head
129.75
65.15
49.18
36.5
33.6
Dataset
UMLS
KINSHIP
Countries S1
Countries S2
Countries S3
WN18RR
NELL-995
FB15K-237
WIKIMOVIES
β
0.05
0.05
0.01
0.02
0.01
0.05
0.05
0.02
0.15
λ
0
0.05
0.1
0.02
0.01
0.05
0.02
0.05
0
Path Length
2
2
2
2
3
3
3
3
1
Table 11: Best hyper parameters
8.2
H YPERPARAMETERS
In our experiments, we tune our model over two hyper parameters, viz., β which is the entropy
regularization constant and λ which is the moving average constant for the REINFORCE baseline.
The table 11 lists the best hyper parameters for all the datasets.
15
| 2cs.AI
|
How we can control the crack to propagate along the specified
path feasibly?
Zhenxing Cheng, Hu Wang*
State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body, Hunan
University, Changsha, 410082, P.R. Chin
Abstract A controllable crack propagation (CCP)
strategy is suggested. It is well known that crack always
leads the failure by crossing the critical domain in
engineering structure. Therefore, the CCP method is
proposed to control the crack to propagate along the
specified path, which is away from the critical domain.
To complete this strategy, two optimization methods are
engaged. Firstly, a back propagation neural network
(BPNN) assisted particle swarm optimization (PSO) is
suggested. In this method, to improve the efficiency of
CCP, the BPNN is used to build the metamodel instead
of the forward evaluation. Secondly, the popular PSO is
used. Considering the optimization iteration is a time
consuming process, an efficient reanalysis based
extended finite element methods (X-FEM) is used to
substitute the complete X-FEM solver to calculate the
crack propagation path. Moreover, an adaptive
subdomain partition strategy is suggested to improve
the fitting accuracy between real crack and specified
paths. Several typical numerical examples demonstrate
that both optimization methods can carry out the CCP.
The selection of them should be determined by the
tradeoff between efficiency and accuracy.
Keywords Crack propagation path, Reanalysis solver,
Back propagation neural network, Particle swarm
optimization, Extended finite element method
methods of simulating crack propagation. Such as finite
element method (FEM) (Bouchard et al., 2003; Branco
et al., 2015), extended finite element method (X-FEM)
(Belytschko et al., 2009; Zeng et al., 2016), edge-based
finite element method (ES-FEM) (G. R. Liu et al., 2011;
Nguyen-Xuan et al., 2013), meshless method (Gu et al.,
2011; Tanaka et al., 2015), and so on. The X-FEM might
be the most popular method for crack propagation
simulation due to its superiority of modeling both strong
and weak discontinuities. Belytschko and Black
proposed the initial idea of X-FEM at 1999 with
minimal re-mesh (Belytschko et al., 1999). Then, Moës
el al. (Belytschko et al., 2001) and Dolbow et al.
(Dolbow et al., 1999) adopted the Heaviside function to
enrichment function and 3D static crack was modeled
by Sukumar et al. (Sukumar et al., 2000). Sequentially,
the level set methods (LSMs) were applied to X-FEM
which could easily track both the crack position and tips
(Stolarska et al., 2001). Moreover, the X-FEM has much
more applications (Ahmed et al., 2012; Areias et al.,
2005; Belytschko et al., 2003; Chessa et al., 2002;
Huynh et al., 2009; J.-H. Song et al., 2006; Sukumar et
al., 2001; Zhuang et al., 2011; Zilian et al., 2008). More
details of the development of X-FEM can be found in
the literature (Abdelaziz et al., 2008; Belytschko et al.,
2009; Fries et al., 2010).
It is well known that the internal crack propagation
always leads the failure of engineering structure by
crossing the critical domain of the structure. Therefore,
if the crack doesn’t cross the critical domain, the failure
will not happen. Therefore, a controllable crack
propagation method is proposed to control the crack
propagation path and lead it propagate along the predefined path, so that the critical domain should not be
crossed by the crack and the failure will not happen. In
1 Introduction
Generally, the internal crack propagation is a critical
issue in the engineering practice due to its deep effect
on the quality and stability of engineering structures.
Therefore, predicting the path of crack propagation is
significant for guaranteeing the safety or reliability of
engineering structures. There are many numerical
1
this study, the particle swarm optimization (PSO)
method is used to obtain the suitable variables of design
and the artificial neural network is used to improve the
efficiency of PSO.
The PSO proposed by Kennedy and Eberhart is a
popular metaheuristic algorithm which inspired by the
social behavior of bird flocking (Kennedy et al., 1995).
Later Kennedy and Eberhart suggested a developed
version of PSO for discrete optimization (Kennedy et al.,
1997). Shi and Eberhart improved the PSO by inertia
weight (Shi et al., 1998). Recently, PSO has been
applied to many fields, such as structural optimization
(Vagelis et al., 2011), dynamic finite element model
updating (Shabbir et al., 2015), vehicle engineering
(Battaï
a et al., 2013), artificial neural network
(Chatterjee et al., 2016; W. Sun et al., 2016) and so on
(Amini et al., 2013; Amiri et al., 2012; Delice et al.,
2014). Much more studies on PSO can be found in the
literature (Eberhart et al., 2001; Ma et al., 2015; Poli et
al., 2007; Tyagi et al., 2011).
Considering the optimization iteration is a time
consuming process, an efficient reanalysis based XFEM is used to calculate the crack propagation path, in
which reanalysis methods are used to solve the
equilibrium equations efficiently. Reanalysis (Kirsch,
2002), as a fast computational method, was suggested to
predict the response of modified structures efficiently
instead of full analysis. In recent decades, reanalysis
methods have been well developed. Song et al. proposed
a direct reanalysis algorithm to update the triangular
factorization in sparse matrix solution (Q. Song et al.,
2014). Liu et al. applied the Cholesky factorization to
structural reanalysis (H. F. Liu et al., 2014). Huang and
Wang solved the large-scale problems with local
modification by the independent coefficient (IC)
method (Huang et al., 2013). Zuo et al. applied the
reanalysis method to the genetic algorithm (GA) (Zuo
et al., 2011). Sun et al. extended the reanalysis method
into a structural optimization process (R. Sun et al.,
2014). To improve the efficiency of reanalysis method,
He et al. developed a multiple-GPU based parallel IC
reanalysis method (He et al., 2015). Based on these
techniques, Wang et al. developed a CAD/CAE
integrated parallel reanalysis design system (H. Wang et
al., 2016).
In this study, a controllable crack propagation (CCP)
method is proposed to control the crack propagation
path and make the crack propagate along the specified
path, so that the critical domain of engineering structure
should not be destroyed by the crack. Moreover,
considering the optimization iteration is a time
consuming process, an efficient reanalysis based XFEM is used to calculate the crack propagation path, in
which the reanalysis solver is used to solve the
equilibrium equations efficiently. Then the BPNN
assisted PSO method should be used to obtain the
optimal design variables to make the real crack path
match with the specified path.
The rest of this paper is organized as follows. The
basic theory of X-FEM is briefly introduced in Section
2. The details of the CCP method can be found in
Section 3. Then, some numerical examples will be given
to investigate the performance of the CCP method in
Section 4. Finally, some conclusions are summarized in
Section 5.
2 Basics theories of XFEM
2.1 X-FEM approximation
In the X-FEM, the displacement approximation
consists of two parts: the standard finite element
approximation and partition of unity enrichment. Define
the displacement approximation of X-FEM as:
u (x)
h
N ( x )u ( x ) N
I
I
u s tan dard
I
J
( x )q J
(1)
J E
uenrich
where N I and uI denote the standard FEM shape
function and nodal degrees of freedom (DOF)
respectively. The （ x） means enrichment function
while the q J is the additional nodal degree of freedom.
2
where Kuu is the traditional finite element stiffness
matrix, Kua , K aa , K ab are components with Heaviside
enrichment and K ub , K ab , K bb are components with
crack tip enrichment.
2.2 Crack propagation model
Generally, the direction and magnitude of crack
propagation at each iteration are used to determine how
the crack will propagate. The direction of crack
propagation is found from the maximum
circumferential stress criterion and the crack will
Fig. 1 An arbitrary crack line in a structured mesh
Consider an arbitrary crack in a structured mesh as
shown in Fig. 1, and then Eq. (1) can be rewritten as
u (x)
h
N ( x )u H
I
I
I
I
propagate in the direction where σ θθ is maximum
( x ) N I ( x )a I
I S
,
4
(Erdogan et al., 1963). The angle of crack propagation
is defined as
(2)
I ,
( x ) N I ( x )b I
I T 1
2
KI
1 KI
,
(6)
θ 2arctan
sign K II
8
4 K II
K
II
where θ is defined in the crack tip coordinate system,
where is the solution domain, S is the domain
cut by crack, T is the domain which crack tip
located, H (x) is the shifted Heaviside enrichment
K I and K II are the mixed-mode stress intensity
and (x) is the shifted crack tip enrichment. The
factors. The details are given in the reference (Erdogan
et al., 1963).
There are two main patterns when modeling crack
growth. The first one assumes a constant increment of
crack growth at each cycle (Dolbow et al., 1999)while
the other option is to assume a constant number of
cycles and apply a fatigue crack growth law to predict
the crack growth increment for the fixed number of
cycles (Gravouil et al., 2002). In this study, a fixed
increment of crack growth a is considered.
details of H (x) and (x) are given as following:
1
H ( x)
1
Above crack
,
Below crack
(3)
{ (x)}4 1 r sin ,cos ,sin sin ,sin cos
2
2
2
2
.
(4)
The discrete X-FEM equations by substituting Eq.(2)
into the principle of virtual work can be written as
K
K
K
uu
T
ua
T
ub
K ua
K aa
T
K ab
K ub u
F
K a F ,
b F
K
3 Controllable crack propagation method
3.1 Framework of the CCP method
u
ab
a
bb
b
As mentioned above, the CCP method is proposed to
control the crack propagation path based on the X-FEM,
reanalysis, and BPNN assisted PSO methods. The
framework of CCP method is shown in Fig. 2.
(5)
3
Start
Generate
sample data
set (X, Y, R)
Initialize particles
with random position
and velocity vectors
Specified
crack path
Back
paopagation
neural
network
model
Evaluate the
Fitness
Function
Value
Filter out the invalid
data set
Pre-processing for XFEM
Evaluate the fitness of particles:
use BPNN model
Real
crack path
Find and update the local and
global best particle
Assemble the stiffness
matrix K
Find and update the velocity and
position of particles
Adaptive
subdomain
partition
strategy for
design space
Reanalysis Solver
No
Crack stop
extending?
Satisfy a stopping
criterion?
Yes
No
Yes
No
Reanalysis based X-FEM
Yes
BPNN assisted
PSO
Optimal solution
End
Fig. 2 The framework of the CCP method
It can be found that the CCP method mainly includes
two parts: reanalysis based X-FEM and BPNN assisted
PSO method. The first part, reanalysis based X-FEM is
used to calculate the real crack path and the reanalysis
solver is used to solve the equilibrium equations to
improve the efficiency of the X-FEM. The second part,
the BPNN assisted PSO is used to obtain the optimal
design variables to make the real crack path match with
the specified path. Moreover, the adaptive subdomain
partition strategy is used to improve the fitting accuracy
between real crack and specified paths. More details can
be found in Section 3.2, 3.3 and 3.4.
65
Ø13
17
F
67.5
120
F
13
Fig. 3 An edge crack in a plate
3.2 How to control the crack propagation path
Then if a hole was added in the plate as shown in the
left of Fig. 4, the crack propagation path will change as
the right of Fig. 4. Obviously, the crack propagation
path can be influenced by the size, position and number
of this hole. In order to clearly describe this property,
several different results are shown in Fig. 5. It can be
found that different arrangement of the holes obtained
different crack propagation paths and the paths can be
easily driven by arranging the holes.
Assume an edge crack in a plate as shown in the left
of Fig. 3, where the initial crack length is a0 10mm ,
the force F 2 10 4 N , and linear elastic material
behavior is assumed. The material is aluminum 7075T6 with E 7.17 104 MPa , 0.33 and a plane
strain state is considered. The increment of propagation
a 1mm . Calculate the crack propagation path by XFEM without reanalysis solver (full analysis), and the
crack path are shown in the right of Fig. 3.
4
numbers uniformly distributed in 0, i , xi , vi means
65
Ø13
the current position and velocity respectively, and
F
120
6.5
pi , p g means the previous best position and the global
best position.
Ø20
In this study, the BPNN assisted PSO is used to obtain
the optimal arrangement of holes and the flowchart is
shown in Fig. 2 where an adaptive subdomain partition
strategy is proposed to decide the number of holes
should be used. Moreover, the BPNN is used to
construct the model of fitness function value, so that the
fitness function value can be obtained efficiently. The
more details are shown in the following sections.
28.5
17
51
F
13
Fig. 4 An edge crack in a plate with a hole
3.3.1 Fitness function
The PSO is used to find the optimal design variables
(the size, position and number of holes) to make the real
crack path match with the specified path. Assume an
edge crack in a plate as shown in the left of Fig. 3, and
then a specified crack path is given as shown in Fig. 6.
Fig. 5 The crack propagation paths of different
arrangement of holes
Specified
crack path
3.3 The BPNN assisted PSO for crack path
x, y, r
2
2
Key point
2
Real crack path
Particle swarm optimization is a popular optimization
algorithm, and a general algorithm of PSO is described
as the following algorithm (Poli et al., 2007):
Optimized hole
Original hole
Algorithm: General algorithm for PSO
1: Initialize a population array of particles with
random position and velocity vectors;
2: Loop
3:
For each particle, evaluate the fitness by
fitness function;
4: Compare particle’s fitness evaluation with its
pbest . If current value is better than pbest , then update
the pbest and pi ;
5:
Update the velocity and position of the
particle according to the following equation:
Min:
x, y, r
x, y, r
1
1
1
Fig. 6 The description of the optimization
i
i
Obviously, the specified crack path is defined by a set
of key points and the ideal situation is to control the real
crack cross through these points. Although it is difficult
to be achieved, the optimal path might be found by
optimization method. Thus, this optimization problem
is formulated as
Minimize:
i
vi v U (0, 1 ) ( pi xi ) U (0, 2 ) ( p g xi ),
xi xi vi ;
6:
If a criterion is met (usually a sufficiently
good fitness or a maximum number of iteration), exit
loop;
7: End loop
In the algorithm, U (0, i ) means a vector of random
5
proposed to determine the number of holes should be
used. The main idea of the adaptive subdomain partition
strategy is to divide the design space into some
subdomains, and generate a hole in each subdomain.
Generally, the less holes used the more convenient for
processing, so the number of subdomains will be
increased gradually. In this study, the number of
subdomains will be started from one to two, three, four
and so on. The process of the adaptive subdomain
partition strategy is demonstrated in Fig. 7. First, a
threshold value needs to be defined, which is the
target value of fitness function c( x, y,r ) . Then the
n
c ( x, y , r )
ds(i)
i 1
n
,
(7)
Subject to:
KU F ,
ds (i ) f ( x, y, r ) ,
(8)
(9)
( x x1 ) 2 ( y y1 ) 2 r r1 ,
(10)
( x x2 )2 ( y y2 ) 2 r r2 ,
(11)
xmin r x xmax r ,
(12)
ymin r y ymax r ,
(13)
threshold value is used to guarantee the fitting
accuracy is available for engineering problems. When
the minimal value of objective function c( x, y, r ) , it
where c( x, y,r ) is the fitness function and the purpose
is to minimize the fitness function. Moreover, ds(i )
means the design space needs to be subdivided further,
otherwise, the optimal solution is obtained. The details
of this strategy are described as following:
means the minimum distance from the key point i to the
real crack path, and xmin , xmax , ymin , ymax are used to
The adaptive subdomain partition strategy
1: Initialize the number of subdomains in design
space n by n=1;
2: Loop
3: Calculate the fitness function c( x, y,r ) by
Eq.(7);
4:
Find the minimum value cmin of fitness
function;
5: If a criterion is met, exit loop. If not, n=n+1;
The criterion is
cmin min(c( x, y, r )) ;
6: Divide the design space into n subdomains;
7: End loop
define the design space. Moreover, the crack
propagation path should be calculated by the reanalysis
based X-FEM as shown in Fig. 2, so the optimization
must be subjected to the equilibrium equation and the
ds(i ) is determined by the size, position or number of
holes. It is obvious that the fitness function means the
average distance from the key points to the real crack
path, and it can be used to define the fitting accuracy
between real crack and specified paths.
3.3.2 Adaptive subdomain partition strategy for
design space
The adaptive subdomain partition strategy is
One subdomain
Two subdomains
Specified crack path
Three subdomains
Key point
Original hole
Four subdomains
Subdomain
Fig. 7 An illustration of the adaptive subdomain partition strategy
6
genetic algorithm (GA) (Irani et al., 2011), PSO (J.-R.
Zhang et al., 2007) and so on.
In this study, the BPNN is used to forecast the fitness
function value of PSO, so that the optimal solution can
be found more efficiently than the popular PSO.
Generally, a BPNN consists of an input layer, an output
layer and several hidden layers as shown in Fig. 8. A
systematic theory can be found in the literature (G.
Zhang et al., 1998; L. Zhang et al., 2002). Furthermore,
in order to forecast the fitness function value, a set of
sample data should be used to construct a BPNN, then
the BPNN need to be trained, and finally the trained
BPNN can be used to forecast the fitness function value,
the flowchart of BPNN assisted PSO is shown in Fig. 9.
3.3.3 Back propagation neural network
As mentioned above, the PSO needs to evaluate the
fitness of each particle in every generation. Usually, the
fitness could be evaluated by fitness function, such as
Eq.(7), but it is time consumed, because the fitness
needs to be evaluated by the X-FEM for each particle
and the X-FEM is an iterative process. Therefore, we
suggested a BPNN assisted PSO that the BPNN is used
to construct the model of fitness function due to its
flexible nonlinear modeling capability with strong
adaptability (Ticknor, 2013). The BPNN is one of the
popular artificial neural networks (L. Wang et al., 2006).
It’s a type of multi-layered feed-forward neural network
that minimizes an error backward while information is
transmitted forward (G. Zhang et al., 1998)and only one
single hidden BPNN layer is enough to approximate any
nonlinear function with arbitrary precision (Aslanargun
et al., 2007). Therefore, the BPNN has been widely used
in many fields and many intelligent evolution
algorithms have also been used to select the initial
connection weights and thresholds of BPNN, such as
Hidden Layer 1
Hidden Layer 2
Input
w
w
b
b
Output Layer
...
...
b
w
w
b
b
1
3
5
5
Fig. 8 Double hidden layers BPNN structure
Training
Constructing
Start
Load sample
data set
(X, Y, R)
Initialize particles
with random position
and velocity vectors
Construct a suitable
BPNN
Evaluate the fitness of particles:
use fitness function
Initialize the BPNN
Find and update the local and
global best particle
Train the BPNN
Find and update the velocity and
position of particles
Training
finished?
No
Satisfy a stopping
criterion?
Yes
Forecasting
Output
w
No
Yes
Test the BPNN
Optimal solution
Forecast data by BPNN
BPNN
PSO
Fig. 9 Flowchart of BPNN assisted PSO
7
End
1
3.4 The accuracy and efficiency of reanalysis based
X-FEM
Consider that the optimization need a large sample
points and the calculation of the crack propagation path
is an iterative process, the computational cost is
commonly expensive. Therefore, an efficient reanalysis
based X-FEM is used to calculate the crack propagation
path, in which a reanalysis solver is used to solve the
equilibrium equations efficiently. The reanalysis based
X-FEM is used to predict the response of the current
iteration by using the information of the first iteration.
It avoids the full analysis after the first iteration, and the
response of the subsequent iteration can be efficiently
obtained.
In this study, an exact reanalysis named decomposed
updating reanalysis (DUR) method is used to solve the
equilibrium equations, the briefly introduction of this
method is as following, more details can be found from
the reference (Cheng et al., 2017).
Assume that the equilibrium equation of the X-FEM
is
Ki Ui Fi ,
where U
i
If Δ j 0 , the j-th DOF is unbalanced, otherwise
the j-th DOF is balanced.
Therefore, the Eq.(17) can be rewritten as
i
K mm
i
K nm
i
K i 1K mn
B mm
,
Enn
i
K mn
y δn .
i 1
mm
be obtained by Eq.(23). Sequentially, the U
Consider that the increment of crack is very small in
each iteration, so only small part of δ should be
i
(24)
i
can be
obtained by Eq.(16).
In order to test the accuracy and efficiency of the
reanalysis based X-FEM, an edge crack in a plate with
a hole which mentioned in section 3.2 has been
calculated by the reanalysis based X-FEM. The
comparison between reanalysis, full analysis and
experimental results (Giner et al., 2009) as shown in Fig.
10 and Fig. 11. Moreover, the average errors and
computational cost are listed in Tab. 1. It can be found
that the reanalysis based X-FEM is accurate and it saves
(17)
into
two parts: unbalanced and balanced equations,
according to Eq.(18):
Δ sum Ki K 1 δ .
i
nm
Solve Eq.(24), y can be obtained, and then U can
(16)
then substitute Eq.(16) into Eq.(14), obtained
nonzero. Based on this property, divide the U
(23)
where y is a dimension-n vector. Then, substitute Eq.(23)
into Eq.(21) to find a unique solution of Eq.(17), obtain
i
nn
Ki U δ .
is
(22)
Define the general solution of Eq.(20) is
U By ,
K K K
,
1
where E nn is a rank-n unit matrix.
1
(21)
one of them. Calculate the fundamental solution system
of Eq. (20) by
Define that
i
Knm
Um Knni Un δn .
which has infinite solutions, and the solution U
(15)
δ F i K i U1 i 1
(20)
Obviously, Eq.(20) is a homogeneous equation set
is the displacement of the i-th iteration,
U U U i 1
i
i
Kmm
Um Kmn
Un 0
and
and the equilibrium equation of the first iteration can be
given as
i
(19)
where m is the number of balanced DOFs, and n is the
number of unbalanced DOFs.
Equation (19) can be rewritten as
(14)
K1 U1 F1 .
K mni U m 0
,
K nni U n δ n
(18)
8
Tab. 1 Performance comparison of edge crack in a plate with a hole by DUR and full analysis
Computational Time/s
Average Errors
DUR
Full analysis
Displacement
Von Mises Stress
2.047
26.781
4.8916e-13
1.7356e-12
about 13 times computational cost than the full analysis.
Therefore, the reanalysis based X-FEM is an accurate
method for crack quasi-static propagation problems
with high efficiency.
Young’s modulus E1 7.17 104 MPa , the Poisson’s
ratio 0.33 , and a plane strain state is considered.
Assume the increment of propagation a 1mm , then
the crack will propagate along a straight line like the
middle figure of Fig. 12 if there with no holes. However,
if we want to make the crack propagate along the
specified path as shown in the right figure of Fig. 12,
how could we realize it?
q
60
60
120
Fig. 10 The displacement comparison between
reanalysis, full analysis and experimental results
q
Fig. 12 An edge crack optimization in a plate
As mentioned above, some holes will be added to
control the crack path, and the CCP method should be
used to find the suitable number, position and size of
these holes. As shown in Fig. 9,we need obtain a sample
data set firstly. Then a BPNN should be constructed by
the training sample with 1000 groups of sample data.
After that, the BPNN will be trained and finally the
testing sample with 500 groups of sample data should
be used to test the performance of BPNN. Here a double
hidden layers BPNN is used and each hidden layer has
5 nodes. The regression of BPNN is shown in Fig. 13,
and the left is the regression of training sample while the
right is the regression of testing sample. It can be found
that the performance of BPNN is acceptable.
Fig. 11 The stress comparison between reanalysis
and full analysis
4 Numerical examples
In order to test feasibility of the CCP method, two
examples are tested by the proposed methods. These
two cases involve a simple case and an engineering case.
4.1 An edge crack optimization in a plate
As shown in the left figure of Fig. 12, an edge crack
plate is considered where the initial crack length
a0 10mm , and the uniformed load q 200 N / mm .
The parameters of material are given as following: the
9
Fig. 13 The regression of BPNN
After the BPNN has been trained, the optimal
solution should be obtained by the BPNN assisted PSO
and popular PSO with 40 particles. The optimization
procedure of each method is presented in Fig. 14. It can
be found that the BPNN assisted PSO reaches
convergence more quickly than the popular PSO while
the popular PSO can obtain a smaller fitness function
value. However, the BPNN assisted PSO can save much
more computational time than the popular PSO because
the fitness function value can be obtained by the BPNN
rather than X-FEM method. The optimal solutions of
them are listed in Tab. 2. Moreover, the results of crack
propagation path are shown in Fig. 15 and Fig. 16 where
the key points are used to define the specified crack path.
The results indicate that the crack did propagate along
the specified path.
Fig. 15 The displacement results of optimal solution
Fig. 16 Comparison of specified crack path and real
crack path
2.0
PSO
BPNN-PSO
4.2 The bottom plate of bearing pedestal
Fitness fuction
1.5
As mentioned before, the purpose of this study is to
control the crack keep away from the critical domain of
engineering structure to avoid failure. In this case, a
bottom plate of bearing pedestal is considered as shown
in Fig. 17 and all the parameters of material are the same
as section 4.1. Assume a crack is in the edge of bottom
plate as shown in Fig. 18, where the initial crack length
1.0
0.5
0.0
0
20
40
60
80
100
Generation
a0 5mm , and the uniformed load q 20 N / mm .
Fig. 14 Comparison between optimization procedures
of BPNN-PSO and popular PSO
Then the crack will propagate and cross the threaded
hole like the right figure of Fig. 18, and this will lead
the failure of bearing pedestal. Therefore, we need to
control the crack propagation path to make it not cross
the threaded holes just like Fig. 19, where the left figure
is path A while the right is path B.
10
Tab. 2 Optimal solutions of BPNN assisted PSO and popular PSO
Optimization
BPNN-PSO
PSO
Design variables (mm)
Fitness function
X
Y
R
26.5407
30.6038
26.8463
27.4947
12.0507
12.5502
Fig. 17 A bottom plate of bearing pedestal
0.4144
0.1084
Path A
q
Path B
Fig. 19 The specified crack paths of bottom plate
22
40
30
9
12
Ø5
q
Fig. 18 An edge crack in the bottom plate
Fig. 20 The regression of BPNN
For the path A, a BPNN has been constructed and
trained by the training samples with 1000 groups of data
and 500 groups of testing data have been tested the
performance of BPNN. A double hidden layers BPNN
is also used. The regression of BPNN is shown in Fig.
20. Then the optimal solution should be found by BPNN
assisted PSO and popular PSO with 40 particles, the
optimization procedure is shown in Fig. 21 and the
optimal solutions are listed in Tab. 3. Moreover, the
results of crack propagation path are shown in Fig. 22
and Fig. 23 where the key points are used to define the
specified crack path. The results indicate that the crack
did propagate along the specified path.
PSO
BPNN-PSO
1.2
Fitness fuction
1.0
0.8
0.6
0.4
0.2
0
20
40
60
80
100
Generation
Fig. 21 Comparison between optimization procedures
of BPNN-PSO and popular PSO of path A
11
Tab. 3 Optimal solutions of BPNN assisted PSO and popular PSO
Optimization
BPNN-PSO
PSO
Design variables (mm)
Fitness function
X
Y
R
7.1572
8.3984
3.1728
3.9024
2.7756
3.0252
0.2231
0.1453
28. The results indicate that the crack did propagate
along the specified path.
Fig. 22 Comparison of specified crack path and real
crack path
Fig. 24 The optimal result for only one hole
BPNN-PSO
PSO
Not satisfy constraints
Fig. 25 The regression of BPNN
Fig. 23 The displacement results of optimal solution
As for the path B, the same operations have been
carried out, and the optimal result by using only one
hole is shown in Fig. 24. It can be found that the solution
is not very nice by only one hole, so the adaptive
subdomain partition strategy divides the design space
into two sub-spaces, and for two holes, the optimal
solution is much better. For two holes optimization, the
regression of BPNN is shown in Fig. 25, where the size
of training sample is 1000 and the size of testing sample
is 500. It should be noted that the samples in center
domain is not satisfied with the constraints Eqs. (1013).Therefore, there are no training and testing samples.
The optimization procedure is shown in Fig. 26 and the
optimal solutions are listed in Tab. 4. Moreover, the
results of optimal solution are shown in Fig. 27 and Fig.
2.5
PSO
BPNN-PSO
Fitness fuction
2.0
1.5
1.0
0.5
0.0
0
20
40
60
80
100
Generation
Fig. 26 Comparison between optimization procedures
of BPNN-PSO and popular PSO of path B
12
Tab. 4 Optimal solutions of BPNN assisted PSO and popular PSO
Optimization
BPNN-PSO
PSO
Design variables (mm)
X1
Y1
R1
X2
Y2
R2
10.1674
10.3278
20.3498
20.4236
4.5053
4.5921
24.1869
24.7407
21.2344
21.2991
3.8063
3.6068
Fitness
function
0.66365
0.20189
in x64 Windows 7. Firstly, a comparison of
computational scale between the BPNN assisted PSO
and popular PSO is shown in Tab. 5, where the
computational data means the data set which needs to
be calculated during the optimization process and it can
be calculated by accumulating all particles in all
generations until the optimization converges. The
sample data is used to construct the BPNN model, so
there is no sample data for the popular PSO. For a
sample point, the computational time cost in solving the
equilibrium equations is listed in Tab. 6. It can be found
that the reanalysis solver DUR is more efficient than the
full analysis. Moreover, the modeling, optimization and
total time cost by the DUR and full analysis are listed in
Tab. 7 and Tab. 8 respectively, where the modeling time
is the cost of constructing BPNN model and the
optimization time is the cost of optimization process.
The comparison between the DUR and full analysis is
shown in Fig. 29.
From the above results, it can be found that the
accuracy of BPNN determine the performance of the
BPNN assisted PSO. However, in term of efficiency, it
prevails. Therefore, the tradeoff between efficiency and
accuracy is important for selection. Moreover, the
reanalysis solver DUR saves much computational cost
in solving the equilibrium equations. Furthermore, the
comparison of specified crack path and real crack path
indicates that the crack did propagate along the
specified path, so the CCP method is available to control
the crack propagation path.
Fig. 27 Comparison of specified crack path and real
crack path
Fig. 28 The displacement results of optimal solution
4.3 Analysis of computational cost
Two numerical examples have been tested in this
section by both the BPNN assisted PSO and popular
PSO methods. In order to compare the performance of
BPNN assisted and popular PSOs, the CPU running
time has been recorded and all simulations were
performed on an Intel(R) Core(TM) i7-5820K 3.30GHz
CPU with 32GB of memory within MATLAB R2016b
Tab. 5 Comparison of computational scale between the BPNN assisted PSO and popular PSO
Convergent generation
Case 1
Case 2-1
Case 2-2
Size of sample data
Size of computational data
BPNN-PSO
PSO
BPNN-PSO
PSO
BPNN-PSO
PSO
20
31
32
40
50
67
1000
1000
1000
----
800
1240
1280
1600
2000
2680
13
Tab. 6 Comparison of computational time cost between
DUR and full analysis for on data point
Case 2-2 (BPNN-PSO)
Computational time/s
Case 1
Case 2-1
Case 2-2
DUR
Full analysis
3.062
1.984
1.863
37.796
20.578
19.965
Full analysis
DUR
Case 2-2 (PSO)
Case 2-1(BPNN-PSO)
Case 2-1 (PSO)
Case 1 (BPNN-PSO)
Case 1 (PSO)
0
10000
20000
30000
40000
50000
60000
70000
Computational time (s)
Fig. 29 Comparison of computational time cost
between DUR and full analysis
Tab. 7 Computational time of the BPNN-PSO and PSO by reanalysis solver DUR
Modeling time/s
Case 1
Case 2-1
Case 2-2
Optimization time/s
Total time/s
BPNN-PSO
PSO
BPNN-PSO
PSO
BPNN-PSO
PSO
3062
1984
1863
----
76
121
129
4899
3968
4992
3138
2105
1992
4899
3968
4992
Tab. 8 Computational time of the BPNN-PSO and PSO by reanalysis solver full analysis
Modeling time/s
Case 1
Case 2-1
Case 2-2
Optimization time/s
Total time/s
BPNN-PSO
PSO
BPNN-PSO
PSO
BPNN-PSO
PSO
37796
20578
19965
----
81
118
132
60473
41156
53506
37877
20696
20097
60473
41156
53506
optimization methods are used to find the suitable
arrangement which including the number, position and
size of holes. Firstly, a BPNN assisted PSO is suggested.
In this method, the fitness function value of PSO can be
forecasted by the BPNN rather than X-FEM, so
efficiency of BPNN assisted PSO is much higher than
popular PSO. Moreover, considering the optimization
iteration is a time consuming process, an efficient
reanalysis based X-FEM is used to calculate the crack
propagation path, in which reanalysis methods are used
to solve the equilibrium equations efficiently.
Furthermore, in order to improve the fitting accuracy
between real crack path and specified crack path, the
adaptive subdomain partition strategy is suggested to
5 Conclusions
In this study, the controllable crack propagation (CCP)
method is proposed to control the crack propagation
path. Maybe the crack propagation is unavoidable, but
the crack propagation path can be specified by the CCP
method, so the critical domain of structure will not be
crossed. The main idea of CCP is to control the crack
propagation by arranging some holes in the design
domain, so the crack propagation path can be driven by
a suitable arrangement. The CCP method is a closed
loop optimization method which integrating the BPNN,
PSO, reanalysis based X-FEM, adaptive subdomain
partition strategy and other techniques. Two
14
decide the number of holes should be used. Numerical
examples indicate that the CCP method can control the
crack propagation along the specified path well, and the
fitting accuracy between real crack and specified paths
is available.
International Journal of Production Economics, 142(2),
259-277.
Belytschko, T., & Black, T. (1999). Elastic crack
growth in finite elements with minimal remeshing.
International Journal for Numerical Methods in
Engineering, 45(5), 601-620.
Belytschko, T., Chen, H., Xu, J., & Zi, G. (2003).
Dynamic crack propagation based on loss of
hyperbolicity and a new discontinuous enrichment.
International Journal for Numerical Methods in
Engineering, 58(12), 1873-1905.
Belytschko, T., Gracie, R., & Ventura, G. (2009). A
review of extended/generalized finite element methods
for material modeling. Modelling and Simulation in
Materials Science and Engineering, 17(4), 043001.
Belytschko, T., Moës, N., Usui, S., & Parimi, C.
(2001). Arbitrary discontinuities in finite elements.
International Journal for Numerical Methods in
Engineering, 50(4), 993-1013.
Bouchard, P. O., Bay, F., & Chastel, Y. (2003).
Numerical modelling of crack propagation: automatic
remeshing and comparison of different criteria.
Computer Methods in Applied Mechanics and
Engineering, 192(35-36), 3887-3908.
Branco, R., Antunes, F. V., & Costa, J. D. (2015). A
review on 3D-FE adaptive remeshing techniques for
crack growth modelling. Engineering Fracture
Mechanics, 141, 170-195.
Chatterjee, S., Sarkar, S., Hore, S., Dey, N., Ashour,
A. S., & Balas, V. E. (2016). Particle swarm
optimization trained neural network for structural
failure prediction of multistoried RC buildings. Neural
Computing and Applications, 28(8), 2005-2016.
Cheng, Z., & Wang, H. (2017). A fast and exact
computational method for crack propagation based on
extended finite element method. arXiv preprint
arXiv:1708.01610.
Chessa, J., Smolinski, P., & Belytschko, T. (2002).
The extended finite element method (XFEM) for
solidification problems. International Journal for
Numerical Methods in Engineering, 53(8), 1959-1977.
Delice, Y., Kızılkaya Aydoğan, E., Özcan, U., & İlkay,
M. S. (2014). A modified particle swarm optimization
algorithm to mixed-model two-sided assembly line
Acknowledgements
This work has been supported by Project of the
National
Key
R&D
Program
of
China
2017YFB0203701 and the National Natural Science
Foundation of China under the Grant Numbers
11572120.
References
Abdelaziz, Y., & Hamouine, A. (2008). A survey of
the extended finite element. Computers & Structures,
86(11-12), 1141-1151.
Ahmed, A., van der Meer, F. P., & Sluys, L. J. (2012).
A geometrically nonlinear discontinuous solid-like shell
element (DSLS) for thin shell structures. Computer
Methods in Applied Mechanics and Engineering, 201–
204, 191-207.
Amini, F., Hazaveh, N. K., & Rad, A. A. (2013).
Wavelet PSO‐Based LQR Algorithm for Optimal
Structural Control Using Active Tuned Mass Dampers.
Computer-Aided Civil and Infrastructure Engineering,
28(7), 542-557.
Amiri, G. G., Rad, A. A., Aghajari, S., & Hazaveh, N.
K. (2012). Generation of Near-Field Artificial Ground
Motions Compatible with Median-Predicted Spectra
Using PSO-Based Neural Network and Wavelet
Analysis. Computer-Aided Civil and Infrastructure
Engineering, 27(9), 711–730.
Areias, P., & Belytschko, T. (2005). Non‐linear
analysis of shells with arbitrary evolving cracks using
XFEM. International Journal for Numerical Methods in
Engineering, 62(3), 384-415.
Aslanargun, A., Mammadov, M., Yazici, B., &
Yolacan, S. (2007). Comparison of ARIMA, neural
networks and hybrid models in time series: tourist
arrival forecasting. Journal of Statistical Computation
and Simulation, 77(1), 29-53.
Battaï
a, O., & Dolgui, A. (2013). A taxonomy of line
balancing problems and their solutionapproaches.
15
balancing. Journal of Intelligent Manufacturing, 28(1),
23-36.
Dolbow, J., & Belytschko, T. (1999). A finite element
method for crack growth without remeshing.
International Journal for Numerical Methods in
Engineering, 46(1), 131-150.
Eberhart, & Shi, Y. (2001). Particle swarm
optimization: developments, applications and resources.
Paper presented at the Evolutionary Computation, 2001.
Proceedings of the 2001 Congress on.
Erdogan, F., & Sih, G. (1963). On the crack extension
in plates under plane loading and transverse shear.
Journal of basic engineering, 85(4), 519-527.
Fries, T. P., & Belytschko, T. (2010). The
extended/generalized finite element method: an
overview of the method and its applications.
International Journal for Numerical Methods in
Engineering, 84(3), 253-304.
Giner, E., Sukumar, N., Tarancón, J. E., & Fuenmayor,
F. J. (2009). An Abaqus implementation of the extended
finite element method. Engineering Fracture
Mechanics, 76(3), 347-368.
Gravouil, A., Moës, N., & Belytschko, T. (2002).
Non‐planar 3D crack growth by the extended finite
element and level sets—Part II: Level set update.
International Journal for Numerical Methods in
Engineering, 53(11), 2569-2586.
Gu, Y. T., Wang, W., Zhang, L. C., & Feng, X. Q.
(2011). An enriched radial point interpolation method
(e-RPIM) for analysis of crack tip fields. Engineering
Fracture Mechanics, 78(1), 175-190.
He, G., Wang, H., Li, E., Huang, G., & Li, G. (2015).
A multiple-GPU based parallel independent coefficient
reanalysis method and applications for vehicle design.
Advances in Engineering Software, 85, 108-124.
Huang, G., Wang, H., & Li, G. (2013). A reanalysis
method for local modification and the application in
large-scale problems. Structural and Multidisciplinary
Optimization, 49(6), 915-930.
Huynh, D., & Belytschko, T. (2009). The extended
finite element method for fracture in composite
materials. International Journal for Numerical Methods
in Engineering, 77(2), 214-239.
Irani, R., & Nasimi, R. (2011). Evolving neural
network using real coded genetic algorithm for
permeability estimation of the reservoir. Expert Systems
with Applications, 38(8), 9862-9866.
Kennedy, J., & Eberhart, R. (1995). Particle swarm
optimization. Paper presented at the IEEE International
Conference on Neural Networks, 1995. Proceedings.
Kennedy, J., & Eberhart, R. C. (1997). A discrete
binary version of the particle swarm algorithm. Paper
presented at the IEEE International Conference on
Systems, Man, and Cybernetics, 1997. Computational
Cybernetics and Simulation.
Kirsch, U. (2002). Design-Oriented Analysis of
Structures. Dordrecht: Kluwer Academic Publishers.
Liu, G. R., Nourbakhshnia, N., & Zhang, Y. W.
(2011). A novel singular ES-FEM method for
simulating singular stress fields near the crack tips for
linear fracture problems. Engineering Fracture
Mechanics, 78(6), 863-876.
Liu, H. F., Wu, B. S., & Li, Z. G. (2014). Method of
Updating the Cholesky Factorization for Structural
Reanalysis with Added Degrees of Freedom. Journal of
Engineering Mechanics, 140(2), 384-392.
Ma, W., Wang, M., & Zhu, X. (2015). Hybrid particle
swarm optimization and differential evolution
algorithm for bi-level programming problem and its
application to pricing and lot-sizing decisions. Journal
of Intelligent Manufacturing, 26(3), 471-483.
Nguyen-Xuan, H., Liu, G. R., Bordas, S., Natarajan,
S., & Rabczuk, T. (2013). An adaptive singular ES-FEM
for mechanics problems with singular field of arbitrary
order. Computer Methods in Applied Mechanics and
Engineering, 253, 252-273.
Poli, R., Kennedy, J., & Blackwell, T. (2007). Particle
swarm optimization. Swarm Intelligence, 1(1), 33-57.
Shabbir, F., & Omenzetter, P. (2015). Particle Swarm
Optimization with Sequential Niche Technique for
Dynamic Finite Element Model Updating. ComputerAided Civil and Infrastructure Engineering, 30(5), 359375.
Shi, Y., & Eberhart, R. (1998). A modified particle
swarm optimizer: Springer Berlin Heidelberg.
Song, J.-H., Areias, P. M. A., & Belytschko, T. (2006).
A method for dynamic crack and shear band
propagation with phantom nodes. International Journal
16
for Numerical Methods in Engineering, 67(6), 868-893.
Song, Q., Chen, P., & Sun, S. (2014). An exact
reanalysis algorithm for local non-topological high-rank
structural modifications in finite element analysis.
Computers & Structures, 143, 60-72.
Stolarska, M., Chopp, D., Moës, N., & Belytschko, T.
(2001). Modelling crack growth by level sets in the
extended finite element method. International Journal
for Numerical Methods in Engineering, 51(8), 943-960.
Sukumar, N., Chopp, D. L., Moës, N., & Belytschko,
T. (2001). Modeling holes and inclusions by level sets
in the extended finite-element method. Computer
Methods in Applied Mechanics and Engineering,
190(46), 6183-6200.
Sukumar, N., Moës, N., Moran, B., & Belytschko, T.
(2000). Extended finite element method for threedimensional crack modelling. International Journal for
Numerical Methods in Engineering, 48(11), 1549-1570.
Sun, R., Liu, D., Xu, T., Zhang, H., & Zuo, W. (2014).
New Adaptive Technique of Kirsch Method for
Structural Reanalysis. AIAA Journal, 52(3), 486-495.
Sun, W., & Xu, Y. (2016). Using a back propagation
neural network based on improved particle swarm
optimization to study the influential factors of carbon
dioxide emissions in Hebei Province, China. Journal of
Cleaner Production, 112, 1282-1291.
Tanaka, S., Suzuki, H., Sadamoto, S., Imachi, M., &
Bui, T. Q. (2015). Analysis of cracked shear deformable
plates by an effective meshfree plate formulation.
Engineering Fracture Mechanics, 144, 142-157.
Ticknor, J. L. (2013). A Bayesian regularized
artificial neural network for stock market forecasting.
Expert Systems with Applications, 40(14), 5501-5506.
Tyagi, S. K., Yang, K., Tyagi, A., & Dwivedi, S. N.
(2011). Development of a fuzzy goal programming
model for optimization of lead time and cost in an
overlapped product development project using a
Gaussian Adaptive Particle Swarm Optimization-based
approach. Engineering Applications of Artificial
Intelligence, 24(5), 866-879.
Vagelis, P., & Manolis, P. (2011). A Hybrid Particle
Swarm - Gradient Algorithm for Global Structural
Optimization. Computer-Aided Civil and Infrastructure
Engineering, 26(1), 48-68.
Wang, H., Zeng, Y., Li, E., Huang, G., Gao, G., & Li,
G. (2016). “Seen Is Solution” a CAD/CAE integrated
parallel reanalysis design system. Computer Methods in
Applied Mechanics and Engineering, 299, 187-214.
Wang, L., Zeng, Y., Zhang, J., Huang, W., & Bao, Y.
(2006). The criticality of spare parts evaluating model
using artificial neural network approach. Computational
Science–ICCS 2006, 728-735.
Zeng, Q., Liu, Z., Xu, D., Wang, H., & Zhuang, Z.
(2016). Modeling arbitrary crack propagation in
coupled shell/solid structures with X-FEM.
International Journal for Numerical Methods in
Engineering, 106(12), 1018-1040.
Zhang, G., Patuwo, B. E., & Hu, M. Y. (1998).
Forecasting with artificial neural networks:: The state of
the art. International journal of forecasting, 14(1), 3562.
Zhang, J.-R., Zhang, J., Lok, T.-M., & Lyu, M. R.
(2007). A hybrid particle swarm optimization–backpropagation algorithm for feedforward neural network
training. Applied Mathematics and Computation,
185(2), 1026-1037.
Zhang, L., & Subbarayan, G. (2002). An evaluation
of back-propagation neural networks for the optimal
design of structural systems: Part II. Numerical
evaluation. Computer Methods in Applied Mechanics
and Engineering, 191(25), 2887-2904.
Zhuang, Z., & Cheng, B. B. (2011). Equilibrium state
of mode-I sub-interfacial crack growth in bi-materials.
International Journal of Fracture, 170(1), 27-36.
Zilian, A., & Legay, A. (2008). The enriched space–
time finite element method (EST) for simultaneous
solution of fluid–structure interaction. International
Journal for Numerical Methods in Biomedical
Engineering, 75(3).
Zuo, W., Xu, T., Zhang, H., & Xu, T. (2011). Fast
structural optimization with frequency constraints by
genetic algorithm using adaptive eigenvalue reanalysis
methods. Structural and Multidisciplinary Optimization,
43(6), 799-810.
17
| 9cs.NE
|
arXiv:1504.06185v3 [math.ST] 7 Nov 2016
On Locally Dyadic Stationary Processes
Theodoros Moysiadis
Institute of Applied Biosciences, Centre for Research and Technology Hellas
and
Konstantinos Fokianos
Department of Mathematics & Statistics, University of Cyprus
Submitted: April 2015
First Revision: April 2016
Second Revision: September 2016
Abstract
We introduce the concept of local dyadic stationarity, to account for non-stationary time series, within the framework of Walsh-Fourier analysis. We define and study the time varying
dyadic ARMA models (tvDARMA). It is proven that the general tvDARMA process can be
approximated locally by either a tvDMA and a tvDAR process.
Keywords: dyadic stationarity, local stationarity, spectral density, stationarity, Walsh functions,
Walsh-Fourier analysis.
1
1 Introduction
The concept of stationarity is crucial in the statistical theory of time series analysis, especially
for the development of asymptotic theory. However, the assumption of stationarity is often not
realistic in applications. For example, a time series can display significant changes through time
and therefore stationarity is a questionable assumption. One of the most important consequences,
is that attempts to develop asymptotic results are, generally speaking, groundless, since future
information of the process does not necessarily contain any information regarding the present of
the process. In addition, there is no natural generalization of stationarity to non-stationarity, since
non-stationary processes might exhibit trend or/and periodicity and other types of non-standard
behaviour.
Priestley (1965) considered non-stationary processes whose characteristics are changing slowly
over time and developed the theory of evolutionary spectra (see Priestley (1981, 1988)). However,
such an approach makes it difficult to obtain asymptotic results, which are needed for developing estimation theory. In order to apply standard asymptotic theory for non-stationary processes,
Dahlhaus, in a series of contributions, introduced an appropriate theoretical framework, based
on the concept of local stationarity (see, for example, Dahlhaus (1996b, 1997, 2000)). The definition of local stationarity is based on the existence of a time varying spectral representation
(Dahlhaus (1996b)). Dahlhaus (2012) gives an excellent and detailed overview of the theory of
locally stationary processes. A comparison between the methodology developed by Dahlhaus
and Priestley is discussed in Dahlhaus (1996a). Some other works related to locally stationary
time series include the works by Granger and Hatanaka (1964), Tjøstheim (1976), Martin (1981),
Melard and Schutter (1989), Neumann and Von Sachs (1997), Nason et al. (2000), Ombao et al.
(2002), Sakiyama and Taniguchi (2004) and Davis et al. (2006), among others.
The main goal of this contribution is to utilize the idea of local stationarity, in the sense of the
above mentioned papers, for studying the spectral behaviour of time series, based on the system of
2
Walsh functions. These functions led to the development of Walsh-Fourier (square wave) analysis,
just like the sinusoidal functions led to Fourier (trigonometric) analysis. The motivation behind
Walsh-Fourier analysis was the need to approximate stationary time series, which display square
waveforms with abrupt switches (e.g. in communications and engineering), see Stankovic̀ et al.
(2005) for instance. We introduce the concept of local stationarity but based on the orthogonal
system of Walsh functions to account for such phenomena that exhibit, in addition, non-stationary
behaviour. We study important general classes of time series, similar in concept with the time varying ARMA (tvARMA) process- see Dahlhaus (2012). We anticipate that our theory and methods
will be applicable to non-stationary data observed in diverse applications like pattern recognition
for binary images, linear system theory and other (see Stankovic̀ et al. (2005), for more).
The Walsh functions were introduced by Walsh (1923). They take only two values, +1 and −1,
and have similar properties with the trigonometric functions (although they are not periodic). The
introduction of Walsh functions has been followed by a series of papers, related to their mathematical properties and generalizations (Fine (1949)), which provided the theoretical framework for
various applications, see e.g. Beauchamp (1984), Stoffer (1991) and Abbasi et al. (2012), among
others. Stoffer (1991) gives an excellent account of the history of Walsh functions and a comparison between Walsh and Fourier analysis.
The statistical analysis of stationary time series via Walsh functions has been based on real
and dyadic time. The dyadic time is based on the concept of dyadic addition (see Subsection
2.1). For time points m, n, the real time sum m + n is now replaced by the dyadic sum m ⊕ n.
Morettin (1981) reviewed work on Walsh spectral analysis in both time scales. Walsh-Fourier
analysis of real time stationary processes has been studied by Kohn (1980a,b), Morettin (1983)
and Stoffer (1985, 1987, 1990), among others. The dyadic time stationarity is defined in respect to
the real time stationarity as in Subsection 2.2 (see also Nagai (1977)). Further references related to
the Walsh-Fourier analysis of dyadic stationary processes are Morettin (1974, 1978, 1981), Nagai
3
(1980), Nagai and Taniguchi (1987) and Taniguchi et al. (1989). In particular, Morettin (1974,
1978) studied the finite Walsh transform, considered the Walsh periodogram as an estimator of the
Walsh spectrum and studied its theoretical properties. Nagai (1977) proved that a dyadic stationary
process has always unique spectral representation in terms of the system of Walsh functions and
studied the dyadic linear process (see also Morettin (1974)). Nagai (1980) also studied dyadic
autoregressive and moving average processes and their relation.
In this article, we introduce the concept of local dyadic stationarity and discuss the advantages
and the perspectives of such consideration in the framework of Walsh functions. In Section 2 of
this article, we recall some definitions and review some fundamental results for dyadic stationary processes. In Section 3, we introduce the concept of local dyadic stationarity and study the
time varying dyadic moving average process. In Section 4, we define the general class of time
varying dyadic autoregressive moving average processes and show that they exhibit locally dyadic
stationarity. The article concludes with several remarks concerning further research in this topic.
2 Preliminaries
2.1 Dyadic addition
We recall the definition of dyadic addition and of a dyadic process following Kohn (1980a). Consider m and n to be non-negative integers that have the following dyadic expansions
m=
f
X
k=0
mk 2k , n =
f
X
nk 2k ,
where mk , nk ∈ {0, 1}.
k=0
Then, the dyadic sum m ⊕ n is defined as
m⊕n=
f
X
|mk − nk |2k .
k=0
4
Consider now x and y to be real numbers that belong to the interval I ≡ [0, 1). We write
x=
∞
X
xk 2−k , y =
∞
X
yk 2−k ,
where xk , yk ∈ {0, 1}.
k=1
k=1
In general, each of the above representations is not unique. We follow the convention that if,
e.g. x, can be written both through a finite or an infinite order representation, then choose the
representation where xk = 0, ∀ k > k0 . With this convention, the dyadic sum x ⊕ y is defined as
x⊕y =
∞
X
|xk − yk |2−k .
k=1
Recall that the k-th Rademacher function is φk (x) = (−1)xk+1 , ∀x ∈ I, ∀ k ≥ 0. Then the
system of Walsh functions, W (n, x), n = 0, 1, 2, . . . , x ∈ I, is defined as follows. If n = 0, set
W (0, x) = 1, ∀x ∈ I. For n > 0, let n = 2n1 + 2n2 + . . . + 2nν , where 0 ≤ n1 < n2 . . . < nν .
Then
W (n, x) =
1,
n = 0,
∀x ∈ I.
φ (x)φ (x) · · · φ (x), n > 0,
n1
n2
nν
We mention briefly some characteristic properties of the Walsh functions.
(i) The system of Walsh functions is orthonormal in I, that is
Z 1
1 for
W (n, x)W (m, x)dx =
0 for
0
n = m,
n 6= m,
and constitutes a complete set. If f (x), x ∈ I is a square integrable function, then it can be
expanded in a Walsh-Fourier series, i.e.
f (x) =
∞
X
cn W (n, x),
n=0
with cn =
R1
0
f (x)W (n, x)dx.
(ii) ∀n, m ∈ N, x, y ∈ I, W (n, x)W (m, x) = W (n⊕m, x) and W (n, x)W (n, y) = W (n, x⊕y).
5
The above properties of Walsh functions motivate the study of stationary time series in terms
of this basis. It is well known that a second order stationary process {Xt , t ∈ N} is represented as
Xt =
Z
π
eiλt dZ(λ),
−π
with {Z(λ), λ ∈ (−∞, ∞)} an orthogonal-increment process such that
E dZ(λ)dZ(µ) = η(λ − µ)dF (λ)dµ
where η(·) is the Dirac function periodically extended to R with period 2π and F (·) is the spectral
distribution function (see Brillinger (1974), for instance).
The Walsh functions can be used instead to represent Xt under the concept of dyadic stationarity. There are differences though between real and dyadic stationary processes; see Morettin
(1981) for further discussion. The concept of dyadic stationarity is explained briefly next.
2.2 Dyadic stationarity
We call a stochastic process {Xt , t ∈ N} dyadic stationary if it has constant mean, finite second
moment and its covariance function
R(n, m) = cov(Xn , Xm ) = E (Xn − E[Xn ])(Xm − E[Xm ]) ,
n, m ∈ N,
is invariant under dyadic addition, i.e. it depends only on n ⊕ m. Hence, we write for notational
convenience R(τ ) = R(n, n ⊕ τ ). In the following assume that E[Xt ] = 0, E[Xt2 ] = 1, ∀t ∈ N.
We recall some important results about dyadic stationary processes.
A dyadic stationary process Xt has a dyadic spectral representation given by (Morettin (1974,
p. 193))
Xt =
Z
1
W (t, x)dZX (x),
0
6
t ∈ N,
where {ZX (x), x ∈ I} is a real random process with orthogonal increments, such that
E[dZX (x1 )dZX (x2 )] = η(x1 ⊕ x2 )dGX (x1 )dx2 ,
where η(·) is the Dirac function periodically extended to R with period 1. The function GX (·),
defined on I, is a unique distribution function, which is called the dyadic spectral distribution of
the process Xt . In addition
R(τ ) =
Z
1
W (τ, x)dGX (x).
0
If GX (·) is absolutely continuous, then dGX (x) = gX (x)dx, where gX (x) is called the dyadic
spectral density of Xt .
Example 1. A simple example of a dyadic stationary process is a sequence {εt , t ∈ N} of independent random variables with E(εt ) = 0 and E(ε2t ) = σ 2 , ∀ t ∈ N. It is straightforward to show
that its covariance function is
σ 2 , if τ = 0,
E (εn εn⊕τ ) = R(τ ) =
0, if τ 6= 0.
Since the sequence εt is dyadic stationary, it has a dyadic spectral representation of the form
Z 1
εt =
W (t, x)dZε (x), t ∈ N,
0
with
E[(dZε (x))2 ] = dGε (x) = σ 2 dx,
x ∈ I.
This example illustrates the analogy of dyadic and real time stationary processes. Indeed, it is well
known that a white noise real time process possesses a flat spectrum; the same is true under dyadic
stationarity.
A stochastic process {Xt , t ∈ N} is a linear dyadic process if it can be represented as (Morettin
(1974))
Xt =
∞
X
k=0
7
ak εt⊕k ,
(1)
where εt is the sequence of i.i.d. variables, as in Example 1, and {ak , k ∈ N} are real numbers,
P
2
which satisfy ∞
k=0 ak < ∞. This definition is similar in spirit to the definition of the general
linear process Priestley (1981, p.415).
It can be shown that a linear dyadic process of the form (1) is dyadic stationary, because
R(τ ) =
Z
1
W (τ, x) σ
0
∞
X
ak W (k, x)
k=0
!2
dx.
(2)
In addition, it has an absolutely continuous dyadic spectral distribution function and its dyadic
spectral density function has the form
g(x) = σ 2
∞
X
ak W (k, x)
k=0
where A(x) =
P∞
k=0 ak W (k, x).
!2
= σ 2 A2 (x),
In this case, note that G(x) = σ 2
Rx
0
(3)
A2 (y)dy, for x ∈ I. Again,
we note the analogy between real time and dyadic stationarity; the above formula is identical to the
formula obtained in the real time linear process model (Priestley (1981)). Furthermore, if aq 6= 0
and ak = 0, ∀ k > q in (1), then Xt is said to be a dyadic moving average process of order q,
abbreviated as DMA(q). In general, the process Xt defined by (1) is called a DMA(∞) process.
3 Local Dyadic Stationarity
Consider now (3), and suppose, for example, that the function A(·) depends upon time, i.e. it has
the form At,T (·), where T denotes the sample size. Xt is now reexpressed as a triangular array
Xt,T . We rescale At,T (·) from the axis of the first T non negative integers (t = 1, 2, . . . , T ) to the
unit interval I. The reason for this rescaling will be clear later on. The rescaled form of At,T (·) is
denoted by A (t/T, ·). We give a general definition regarding local dyadic stationarity for a process
Xt,T , in the spirit of Dahlhaus (e.g Dahlhaus (1996b, 1997)).
8
Definition 1. A sequence of stochastic processes {Xt,T , t = 1, 2, . . . , T } is called locally dyadic
stationary with transfer function At,T (·) and trend function µ(·), where At,T (·) and µ(·) are deterministic functions, if there exists a representation
Z 1
t
Xt,T = µ
+
W (t, x)At,T (x)dU(x),
T
0
(4)
where the following hold:
(i) U(x) is a real-valued stochastic process on I and
cum{dU(x1 ), . . . , dU(xk )} = η(x1 ⊕ x2 ⊕ . . . ⊕ xk )gk (x1 , . . . , xk−1 )dx1 , . . . , dxk ,
where cum{. . .} denotes the k-th order cumulant, g1 ≡ 0, g2 (x1 ) ≡ 1, |gk (x1 , . . . , xk−1 )| are
bounded for all k and η(·) denotes the Dirac delta function periodically extended to R with period
1.
(ii) There exists a constant K and a function A : [0, 1] × R → R such that
K
t
, x ≤ , ∀ T.
sup At,T (x) − A
T
T
t,x
(5)
The functions A(u, x) and µ(u) are assumed to be continuous with respect to u = t/T .
The above definition is analogous to the definition given by Dahlhaus (1996b). The first condition states that the U(x) has moments of order k; the functions gk (·) are the (k−1)-th polyspectrum
of U(x) following Brillinger (1965). The second assumption requires that the transfer function
At,T (·) is approximated locally by a function A(t/T, ·), which is the transfer function of a dyadic
stationary process. Note that the continuity of A(u, x) and µ(u) in u is required for the process
Xt,T to exhibit locally dyadic stationary behaviour. Furthermore, without loss of generality, we
assume that g2 (x) ≡ 1 because the transfer function can be always rescaled such that the process
{U(x), x ∈ I} is white noise. Indeed, the boundedness assumption of g2 (.) implies again (5) for
the rescaled transfer function.
9
Example 2. Suppose Yt is a dyadic stationary process with dyadic spectral representation
Z 1
Yt =
W (t, x)A(x)dZ(x),
0
where E|Z k (x)| < ∞, ∀ k > 0 Define Xt,T by
t
t
+σ
Yt ,
Xt,T = µ
T
T
where µ(·), σ(·) are continuous functions defined on I → R. Then
Z 1
t
Xt,T = µ
+
W (t, x)At,T (x)dZ(x),
T
0
where At,T (x) = A(t/T, x) = σ(t/T )A(x) and the assumptions (i) and (ii) are satisfied. Hence
Xt,T is locally a dyadic stationary process.
Consider now the process {Xt,T , t = 1, 2, . . . , T }
Xt,T =
∞
X
ak,t,T εt⊕k .
(6)
k=0
where εt is an i.i.d. sequence and {ak,t,T , k ∈ N} is a time-dependent process of real numbers
P∞ 2
such that ∀t,
k=0 ak,t,T < ∞. We call this process a time varying dyadic moving average
process of infinite order (tvDMA(∞)). If we set in (6) aq,t,T 6= 0 and ak,t,T = 0, ∀ k > q, ∀t,
then we call Xt,T a time varying dyadic moving average process of order q (tvDMA(q)). We
rescale now the parameter curves ak,t,T to the unit interval I, assuming that there exist functions
ak (t/T ) : I → R that satisfy ak,t,T ≈ ak (t/T ). We further assume that ak (·) satisfy some regularity
conditions (see Remark 2). The reasons for the rescaling are described in detail, e.g. in Dahlhaus
(2012, Sec.2). Briefly, suppose that we choose ak,t,T to be polynomials of t. Then, as t → ∞,
P
2
ak,t,T → ∞ as well, which violates the condition ∞
k=0 ak,t,T < ∞. In addition, rescaling enables
us to impose smoothing conditions through the continuity of the functions ak (·), ensuring that the
process exhibits locally dyadic stationary behaviour. Indeed, the number of observations within
10
the neighbourhood of a fixed point u0 ∈ I increases as T → ∞ enabling to develop and apply
locally for Xt,T asymptotic results for dyadic stationary processes. Suppose that the process Xt,T
defined by (6) is written as
Xt,T
∞
X
t
εt⊕k .
=
ak
T
k=0
We assume that ak (u) = ak (0) for u < 0 and ak (u) = ak (1) for u > 1 and that the functions ak (·)
satisfy some standard smoothness conditions; see Dahlhaus (1997). Consider now a fixed point
u0 = t0 /T and its neighborhood [u0 ± ǫ]. If the length of this segment is sufficiently small, the
process Xt,T can be approximated by the process X̃t (u0 ), which is defined as
X̃t (u0 ) =
∞
X
ak (u0 )εt⊕k ,
k=0
where ak (u0 ) are constants, with u0 indicating their dependence from the fixed point u0 (see also
Dahlhaus (2012)). X̃t (u0 ) is dyadic stationary. Indeed, we can write
!
Z 1
Z 1
∞
X
X̃t (u0) =
W (t, x)
ak (u0 )W (k, x) dZε (x) =
W (t, x)A(u0 , x)dZε(x), (7)
0
where A(u0 , x) =
0
k=0
P∞
k=0
ak (u0 )W (k, x).
From equations (2) and (3), X̃t (u0 ) has covariance function
R(u0, τ ) =
Z
1
W (τ, x) σ
0
∞
X
ak (u0 )W (k, x)
k=0
!2
dx,
and a unique dyadic spectral density function given by
g (u0 , x) =
σ
∞
X
ak (u0 ) W (k, x)
k=0
!2
= σ 2 A2 (u0 , x) ,
where u0 indicates the dependence from a fixed point.
We can show that for {u = t/T : |t/T − u0 | ≤ ǫ}, it holds |Xt,T − X̃t (u0 )| = OP (1/T ), see
Corollary 1. Therefore, we can say that Xt,T has locally the same covariance and dyadic spectral
11
density function as X̃t (u0 ) and therefore exhibits locally dyadic stationary behaviour. Note that
the tvDMA(∞) process Xt,T in (6) is locally dyadic stationary due to Definition 1, since it has a
time varying spectral representation as in (4). Indeed, we have
Xt,T
Z
∞
X
ak,t,T
=
k=0
where At,T (x) =
P∞
k=0
1
W (t ⊕ k, x)dZε (x)
0
=
Z
1
W (t, x)At,T (x)dZε (x),
0
ak,t,T W (k, x) is the time varying transfer function. We show in Theorem 1
that, in general, a locally dyadic stationary process is approximated by a dyadic stationary process
within a given interval.
Theorem 1. Suppose that {Xt,T , t = 1, 2, . . . , T } is a sequence of stochastic processes which
satisfy a representation of the form (4) where At,T (x) is the time varying transfer function (set
µ (t/T ) = 0). Suppose that {X̃t (u0 ), t = 1, 2, . . . , T } is a dyadic stationary process with
X̃t (u0) =
Z
1
W (t, x)A(u0 , x)dU(x),
(8)
0
where A(u0 , x) depends on the fixed point u0 ∈ I. Then within an interval (u0 ± ǫ) and under the
assumptions of Definition 1 it holds that
|Xt,T − X̃t (u0 )| = OP (1/T ).
Proof. From equations (4) and (8) we have that
|Xt,T − X̃t (u0)| =
≤
Z
Z
1
W (t, x)At,T (x)dU(x) −
0
1
Z
1
W (t, x)A(u0 , x)dU(x)
0
|W (t, x)| · |At,T (x) − A(u0 , x)| dU(x)
0
=
Z
1
|At,T (x) − A(u0 , x)| dU(x),
0
12
(9)
since W (t, x) ∈ {−1, 1}. In addition
|At,T (x) − A(u0 , x)| ≤
≤
At,T (x) − A
t
t
,x + A
, x − A(u0 , x)
T
T
K
+ |A (u, x) − A(u0 , x)| ,
T
(10)
from (5) in assumption (ii) of Definition 1. However, the same assumption states that A(u, x) is
continuous. Therefore, since {u = t/T : |t/T − u0 | ≤ ǫ} and for any ǫ′ > 0 we can choose ǫ > 0
to be such that |A (u, x) − A(u0 , x)| < ǫ′ , (10) becomes
|At,T (x) − A(u0 , x)| ≤
K∗
,
T
(11)
for some positive constant K ∗ . Finally, from (9) and (11), we obtain that
E|Xt,T − X̃t (u0 )| = O(1/T ),
and hence we have the desired result.
Corollary 1. Theorem 1 holds for the tvDMA(∞) process Xt,T , defined by (6), since this process
satisfies the assumptions of Theorem 1, for X̃t (u0 ) given by (7).
Remark 1. Theorem 1 implies that a locally dyadic stationary process could be approximated by
dyadic stationary processes within different intervals in I (that may overlap). Thus its behaviour
could be described via the behaviour of those dyadic stationary processes.
Remark 2. Equation (5) implies a similar assumption for the sup |ak,t,T − ak (t/T )| and the above
t,x
discussion still holds.
Example 3. Consider, for example, the infinite time varying MA (tvMA(∞)) representation Xt,T =
P∞
k=0 ak,t,T εt−k , in the real time. Then its time varying spectral density function is given by
!2
∞
X
2
f (u, λ) = (σ /2π)
ak (u) exp(−iλk) .
k=0
13
0.8
0.6
0.4
0.2
5
4
3
2
1
1.0
1.0
0.8
0.8
le
ca
res
le
ca
res
0.6
0.6
3.0
dt
0.4
1.0
0.6
cy
uen
freq
0.2
0.5
0.0
0.0
0.0
eq
h fr
0.4
0.2
cy
uen
e
e
1.5
0.2
1.0
0.8
0.4
im
2.0
im
dt
2.5
ls
Wa
0.0
Figure 1: The time varying spectral density function for the tvMA(1) process (left) and the time
varying dyadic spectral density function for the tvDMA(1) process (right).
Respectively, the time varying dyadic spectral density function of the tvDMA(∞) is given by
g (u, x) =
σ
∞
X
ak (u)W (k, x)
k=0
!2
.
We compare the behaviour of functions g(u, x) and f (u, λ) for the same order of the respective
processes and for the same representation of the time varying coefficients ak (u) (set σ 2 = 1).
Figure 1 shows the spectral density function of a tvMA(1) and tvDMA(1) processes. We set
a0 (u) = −1.8 cos(1.5 − cos(4πu)) and a1 (u) = 0.81. Figure 2 shows the spectral density function
of a tvMA(2) and tvDMA(2) processes. In this case we set a0 (u) = 1.2 cos(2πu), a1 (u) =
2 cos(1.5 − cos(8πu)) and a2 (u) = u. Both figures reveal the differences between real and dyadic
stationarity. The square waveform of Walsh functions allows a more oscillatory behaviour of the
dyadic spectral density function.
14
15
2.0
1.5
1.0
0.5
10
5
1.0
1.0
0.8
0.8
le
ca
res
le
ca
res
0.6
0.6
3.0
dt
0.4
cy
uen
0.2
0.5
0.0
0.0
0.0
cy
uen
eq
h fr
0.4
0.2
freq
1.0
0.6
e
e
1.5
0.2
1.0
0.8
0.4
im
2.0
im
dt
2.5
ls
Wa
0.0
Figure 2: The time varying spectral density function for the tvMA(2) process (left) and the time
varying dyadic spectral density function for the tvDMA(2) process (right).
4 tvDARMA processes
It is well known that autoregressive, moving average, and ARMA models can be regarded as
special cases of the general linear process. Nagai (1980) shows that a dyadic autoregressive process
of finite order is always inverted into a dyadic moving average process of finite order, and vice
versa. We obtain similar results, but within a time varying framework. We define the time varying,
dyadic, autoregressive, moving average (tvDARMA) process as follows.
Definition 2. A stochastic process {Xt , t = 1, 2, . . . , T } is called tvDARMA(p, r) if it is locally
dyadic stationary and can be represented by
p
X
bk,t,T Xt⊕k,T =
m
an,t,T εt⊕n ,
(12)
n=0
k=0
+
r
X
f
where p, r ∈ Z with p = 2 − 1, r = 2 − 1, the sequences of parameters {bk,t,T }k=0,1,...,p ,
15
{an,t,T }n=0,1,...,p are real numbers with at least two non-zero parameters bk0 ,t,T , an0 ,t,T for 2m−1 ≤
k0 ≤ 2m − 1 and 2f −1 ≤ n0 ≤ 2f − 1. In addition, {εt , t = 1, 2, . . . , T } is an i.i.d. sequence with
E(εt ) = 0 and E(ε2t ) = σ 2 .
Assume that b0,t,T = a0,t,T = 1. If we set in (12), p = 0, then the tvDMA process arises as in
(6), but for a finite order r. In case we set in (12) r = 0, then (12) becomes
p
X
bk,t,T Xt⊕k,T = εt .
(13)
k=0
We call Xt,T in (13) a time varying, dyadic, autoregressive process of order p (tvDAR(p)). We
show that a tvDAR process, and even more generally, a tvDARMA process, can be approximated
by a tvDMA process. Following Nagai and Taniguchi (1987), who study multivariate dyadic stationary processes, set
φt,T (x) =
p
X
bj,t,T W (j, x), x ∈ I,
(14)
j=0
where p = 2m −1, m ∈ N and {bj,t,T }j=0,1,...,p are real numbers. Denote by Σt,T the (p+1)×(p+1)
matrix, which is given by
Σt,T
b
b
· · · b0⊕p,t,T
0⊕0,t,T 0⊕1,t,T
b1⊕0,t,T b1⊕1,t,T · · · b1⊕p,t,T
=
..
..
..
..
.
.
.
.
bp⊕0,t,T bp⊕1,t,T · · · bp⊕p,t,T
Lemma 1. The following equation holds
det[Σt,T ] =
p
Y
.
φt,T (xj ),
j=0
where xj = j/(p+1), j = 0, 1, . . . , p. Therefore the function φt,T (x) 6= 0 if and only if det[Σt,T ] 6=
0.
16
Lemma 2. Assume that φt,T (x) 6= 0 in (14). Then there exists a function ηt,T (x), which is defined
by
ηt,T (x) =
p
X
dm,t,T W (m, x),
{dm,t,T }m=0,1,...,p ∈ R, x ∈ I,
m=0
and satisfies φt,T (x)ηt,T (x) = 1. The coefficients dm,t,T are uniquely determined by
Σt,T
d0,t,T d1,t,T · · · dp,t,T
′
=
1 0 ··· 0
′
.
Define St,T to be the (r + 1) × (r + 1) matrix
St,T
a
a
· · · a0⊕r,t,T
0⊕0,t,T 0⊕1,t,T
a1⊕0,t,T a1⊕1,t,T · · · a1⊕r,t,T
=
..
..
..
..
.
.
.
.
ar⊕0,t,T ar⊕1,t,T · · · ar⊕r,t,T
.
The following theorem states that a tvDARMA process can be approximated locally by a
tvDMA and a tvDAR process.
Theorem 2. Suppose that {Xt,T , t = 1, 2, . . . , T } is a tvDARMA(p, r) as in (12). Set µ =
max(p, r). Then the following hold
(i) If det[Σt,T ] 6= 0, then Xt,T can be approximated locally by a tvDMA(µ) process.
(ii) If det[St,T ] 6= 0, then Xt,T can be approximated locally by a tvDAR(µ) process.
Corollary 2. Suppose that {Xt,T , t = 1, 2, . . . , T } is a tvDAR(p) as in (13). If det[Σt,T ] 6= 0, Xt,T
can be approximated locally by a tvDMA(p) process.
Proofs of these results are given in the appendix.
17
5 Conclusions
We anticipate that the above results will be useful for future work in the field of applications. In
this direction, the concepts of Walsh spectrum and Walsh transform will be studied. The Walsh
spectrum for a real-valued dyadic stationary process Xt is defined by
f (x) =
∞
X
R(τ )W (τ, x),
0 ≤ x < ∞,
(15)
τ =0
where the covariance function R(·) satisfies
P∞
τ =0
|R(τ )| < ∞ (see e.g. Morettin (1974, 1978,
1981)). Inverting (15), the covariance is given by
Z 1
R(τ ) =
W (τ, x)f (x)dx.
0
The finite Walsh transform is given by
d (N ) (x) =
N
−1
X
Xn W (n, x),
x ∈ I.
n=0
To estimate the Walsh spectrum, Morettin (1981) defined the Walsh periodogram, by
I (N ) (x) = N −1 [d (N ) (x)]2 ,
and showed that I (N ) (x) is asymptotically an unbiased, but inconsistent, estimator of f (x). He
also considered the smooth Walsh periodogram and other classes of estimates. Dyadic stationarity is necessary to estimate the time-varying Walsh spectrum. Therefore, in the case of local
dyadic stationarity, it would be reasonable to divide the rescaled interval I into subintervals and
estimate the Walsh spectrum within each subinterval, where local dyadic stationarity is satisfied.
The number of the observations within each subinterval {u = t/T : |t/T − u0 | ≤ ǫ} increases as
T tends to infinity and the above asymptotical results still hold. A similar method is applied by
Dahlhaus and Giraitis (1998) for real-time stationary processes.
18
Kohn (1980a,b) studied the system of Walsh functions for real time stationary processes. He
defined the j-th logical autocovariance, τ (j), and the corresponding Walsh-Fourier spectral density
function F (x). This notation replaces the previous notation used for R(·) and f (·) above. He
considered the finite Walsh-Fourier transform and studied its asymptotic properties. A class of
estimators for F (x) was obtained, the average Walsh periodogram being a member of this class.
The concept of local stationarity could also be applied in the real time setting and we conjecture
that similar results could be obtained also in this case.
Acknowledgements
We cordially thank Prof. I. Nikiforov and four anonymous reviewers for several constructive comments that improved considerably an earlier version of the manuscript.
19
Appendix
Proof of Lemma 1 Recall that p = 2m − 1. The Walsh-ordered Hadamard matrix HW (m) is a
(2m × 2m ) matrix with elements of the form W (n, xj ), xj = j/(p + 1), j, n = 0, 1, . . . , p, see
also Stoffer (1991). Then the following relations hold:
b0⊕0,t,T b0⊕1,t,T · · · b0⊕p,t,T
W (0, x0 ) W (0, x1 ) · · · W (0, xp )
b1⊕0,t,T b1⊕1,t,T · · · b1⊕p,t,T W (1, x0 ) W (1, x1 ) · · · W (1, xp )
·
Σt,T HW (m) =
..
..
..
..
..
..
..
..
.
.
.
.
.
.
.
.
W (p, x0 ) W (p, x1 ) · · · W (p, xp )
bp⊕0,t,T bp⊕1,t,T · · · bp⊕p,t,T
( p
)
X
=
b(i−1)⊕l,t,T W (l, xj−1 )
= φt,T (xj−1 )W (i − 1, xj−1)
l=0
(i,j)
(i,j)
= HW (m) · diag[φt,T (x0 ), φt,T (x1 ), . . . , φt,T (xp )].
(A-1)
But, since det[HW (m)] = (p + 1)(p+1)/2 6= 0, we get from (A-1) that
p
Y
φt,T (xj ).
det[Σt,T ] = det diag(φt,T (x0 ), φt,T (x1 ), . . . , φt,T (xp )) =
j=0
For the second argument of Lemma 1, note that φt,T (x) =
Pp
j=0 bj,t,T W (j, x),
x ∈ I and
every Walsh function W (n, x), n = 0, 1, . . . , p remains invariant for x ∈ [xj , xj+1 ), xj = j/(p +
1), j, n = 0, 1, . . . , p and equal to W (n, xj ). Therefore φt,T (x) = φt,T (xj ), ∀x ∈ [xj , xj+1 ).
Proof of Lemma 2
φt,T (x)ηt,T (x) =
=
p
X
bl,t,T W (l, x)
l=0
" p
p
X
X
h=0
j=0
!
p
X
dm,t,T W (m, x)
m=0
#
bj⊕h,t,T dj,t,T W (h, x).
20
!
=
p
p
X
X
bl,t,T dm,t,T W (l ⊕ m, x)
l=0 m=0
(A-2)
In order for φt,T (x)ηt,T (x) = 1 to hold in I, we have from equation (A-2) that
" p
#
p
X
X
bj⊕h,t,T dj,t,T W (h, xl ) = 1, l = 0, 1, . . . , p,
h=0
j=0
which is equivalently written in matrix notation as
P
p
j=0 bj,t,T dj,t,T
Pp
j=0 bj⊕1,t,T dj,t,T
′
HW
(m)
..
.
P
p
j=0 bj⊕p,t,T dj,t,T
=
1
1
..
.
1
.
(A-3)
′
But HW (m)HW
(m) = 2m I2m . Hence, since from assumption we have that det [Σt,T ] 6= 0, equa-
tion (A-3) gives that
P
p
j=0 bj,t,T dj,t,T
Pp
j=0 bj⊕1,t,T dj,t,T
2m I2m
..
.
P
p
j=0 bj⊕p,t,T dj,t,T
= HW (m)
1
1
..
.
1
m
=2
1
0
..
.
0
d
0,t,T
d
=⇒ 1,t,T
..
.
dp,t,T
−1
=Σ
t,T
1
0
..
.
0
Proof of Theorem 2 Since Xt,T is locally dyadic stationary it has a Walsh spectral representation
Z 1
Xt,T =
W (t, x)At,T (x)dU(x),
0
while εt is dyadic stationary and represented by
Z 1
εt =
W (t, x)dZε (x).
0
Then the LHS and RHS of equation (12) can be written as
LHS =
p
X
k=0
bk,t,T Xt⊕k,T =
Z
1
W (t, x)
0
p
X
k=0
21
!
bk,t,T At⊕k,T (x)W (k, x) dU(x),
(A-4)
.
and
RHS =
r
X
an,t,T εt⊕n =
Set
Z
1
W (t, x)
0
|LHS − LHS′ | =
1
Z
1
|W (t, x)|
0
=
W (t, x)
r
X
!
an,t,T W (n, x) dZε (x).
n=0
p
X
!
bk,t,T At,T (x)W (k, x) dU(x).
k=0
Then
Z
1
0
n=0
LHS′ =
Z
p
X
!
|bk,t,T | · |At⊕k,T (x) − At,T (x)| · |W (k, x)| dU(x)
k=0
p
X
0
(A-5)
!
|bk,t,T | · |At⊕k,T (x) − At,T (x)| dU(x).
k=0
(A-6)
Consider the interval A = {|(t ⊕ k/T ) − t/T | ≤ ε}. Then, ∀ε, we can assume that T is large
enough, such that t ⊕ k/T ∈ A, ∀ k = 0, 1, . . . , p. From the continuity of the function A(t/T, ·)
and assumption (5) we have that
t⊕k
t⊕k
At⊕k,T (x) − A
,x + A
, x − At,T (x)
T
T
t⊕k
t⊕k
,x + A
, x − At,T (x)
At⊕k,T (x) − A
T
T
t
t⊕k
K
t
+ A
,x −A
,x + A
, x − At,T (x)
T
T
T
T
2K
+ ε′ = ε′′ .
(A-7)
T
|At⊕k,T (x) − At,T (x)| =
≤
≤
≤
From (A-6) and (A-7), we have that
′
|LHS − LHS | ≤ ε
′′
p
X
k=0
since
Pp
k=0
|bk,t,T | < M ′ and
R1
0
|bk,t,T |
Z
1
dU(x) ≤ ε′′′ ,
0
dU(x) < M ′′ , with M ′ and M ′′ real constants.
22
Set φ1,t,T (x) =
Pp
k=0 bk,t,T W (k, x) and
φ2,t,T (x) =
Pr
n=0
an,t,T W (n, x). Assume that LHS′ =
RHS. Then, since the system of Walsh functions is complete and equation (12) holds for t =
1, 2, . . . , T , we have that φ1,t,T (x)At,T (x)dU(x) = φ2,t,T (x)dZε (x).
(i) Since det[Σt,T ] 6= 0, from Lemma 1 we have that φ1,t,T (x) 6= 0 and from Lemma 2 there exists
P
a function η1,t,T (x) = pk=0 gk,t,T W (k, x) such that
Hence,
At,T (x)dU(x) = η1,t,T (x)φ2,t,T (x)dZε (x)
( Pµ Pp
g
a
j=0
l=0 l,t,T l⊕j,t,T W (j, x)dZε (x), p ≤ r,
=
Pr
Pµ
g
a
l=0 l⊕j,t,T l,t,T W (j, x)dZε (x), p > r.
j=0
At,T (x)dU(x) =
µ
X
Kj,t,T W (j, x)dZε (x),
j=0
where
Kj,t,T =
Therefore,
Xt,T =
( Pp
s=0
gs,t,T as⊕j,t,T , p ≤ r,
Pr
j = 0, 1, . . . , p.
s=0 gs⊕j,t,T as,t,T , p > r,
µ
X
j=0
Kj,t,T
Z
1
W (t ⊕ j, x)dZε (x) =
0
µ
X
Kj,t,T εt⊕j .
j=0
(ii) Similarly with (i).
Proof of Corollary 2
Suppose that Xt,T =
R1
0
At,T (x)W (t, x)dU(x). The LHS of (13) is given
by (A-4) and the LHS′ by (A-5). From Lemma 2, since by assumption det[Σt,T ] 6= 0, ∃ ηt,T (x) =
Pp
m=0 dm,t,T W (m, x), dm,t,T ∈ R, such that φ1,t,T (x)ηt,T (x) = 1. Therefore, and since the system
of Walsh functions is complete, we have that
At,T (x)dU(x) = ηt,T (x)dZε (x) =
p
X
dm,t,T W (m, x)dZε(x).
m=0
Hence, Xt,T =
Pp
m=0
dm,t,T
R1
0
W (t ⊕ m, x)dZε (x) =
23
Pp
m=0
dm,t,T εt⊕m .
References
Abbasi, S. A., A. Alamoud, et al. (2012). FPGA based Walsh and inverse Walsh transforms for
signal processing. Elektronika ir Elektrotechnika 18, 3–8.
Beauchamp, K. G. (1984). Walsh functions and their applications. Academic Press.
Brillinger, D. R. (1965). An introduction to polyspectra. The Annals of mathematical statistics,
1351–1374.
Brillinger, D. R. (1974). Fourier analysis of stationary processes. Proc. IEEE 62, 1628–1643.
Dahlhaus, R. (1996a). Asymptotic statistical inference for nonstationary processes with evolutionary spectra. In Athens conference on applied probability and time series analysis, pp. 145–159.
Springer.
Dahlhaus, R. (1996b). On the Kullback-Leibler information divergence of locally stationary processes. Stochastic Processes and their Applications 62, 139–168.
Dahlhaus, R. (1997). Fitting time series models to nonstationary processes. The Annals of Statistics 25, 1–37.
Dahlhaus, R. (2000). A likelihood approximation for locally stationary processes. Annals of
Statistics 28, 1762–1794.
Dahlhaus, R. (2012). Locally stationary processes. In T. S. Rao, S. S. Rao, and C. R. Rao (Eds.),
Handbook of Statistics: Time Series Analysis–Methods and Applications, Volume 30, pp. 351–
408. Amsterdam: Elsevier B. V.
Dahlhaus, R. and L. Giraitis (1998). On the optimal segment length for parameter estimates for
locally stationary time series. Journal of Time Series Analysis 19, 629–655.
24
Davis, R. A., T. C. M. Lee, and G. A. Rodriguez-Yam (2006). Structural break estimation for
nonstationary time series models. Journal of the American Statistical Association 101, 223–
239.
Fine, N. J. (1949). On the Walsh functions. Transactions of the American Mathematical Society 65,
372–414.
Granger, C. W. J. and M. Hatanaka (1964). Spectral analysis of economic time series. Princeton:
Univ. Press.
Kohn, R. (1980a). On the spectral decomposition of stationary time series using Walsh functions.
I. Advances in Applied Probability, 183–199.
Kohn, R. (1980b). On the spectral decomposition of stationary time series using Walsh functions.
II. Advances in Applied Probability, 462–474.
Martin, W. (1981). Line tracking in nonstationary processes. Signal Processing 3, 147–155.
Melard, G. and A. H.-d. Schutter (1989). Contributions to evolutionary spectral theory. Journal of
Time Series Analysis 10, 41–63.
Morettin, P. A. (1974). Walsh-function analysis of a certain class of time series. Stochastic processes and their applications 2, 183–193.
Morettin, P. A. (1978). Estimation of the spectrum and of the covariance function of a dyadicstationary series. Bulletin of the Brazilian Mathematical Society 9, 83–88.
Morettin, P. A. (1981). Walsh spectral analysis. SIAM Review 23, 279–291.
Morettin, P. A. (1983). A note on a central limit theorem for stationary processes. Journal of Time
Series Analysis 4, 49–52.
25
Nagai, T. (1977). Dyadic stationary processes and their spectral representations. Bulletin of Mathematical Statistics 17, 65–73.
Nagai, T. (1980). On finite parametric linear models of dyadic stationary processes. Bulletin of
Mathematical Statistics 19, 45–53.
Nagai, T. and M. Taniguchi (1987). Walsh spectral analysis of multiple dyadic stationary processes
and its applications. Stochastic Processes and their Applications 24, 19–30.
Nason, G. P., R. Von Sachs, and G. Kroisandt (2000). Wavelet processes and adaptive estimation of
the evolutionary wavelet spectrum. Journal of the Royal Statistical Society: Series B (Statistical
Methodology) 62, 271–292.
Neumann, M. H. and R. Von Sachs (1997). Wavelet thresholding in anisotropic function classes
and application to adaptive estimation of evolutionary spectra. The Annals of Statistics 25, 38–
76.
Ombao, H., J. Raz, R. von Sachs, and W. Guo (2002). The slex model of a non-stationary random
process. Annals of the Institute of Statistical Mathematics 54(1), 171–200.
Priestley, M. B. (1965). Evolutionary spectra and non-stationary processes. Journal of the Royal
Statistical Society. Series B (Methodological) 27, 204–237.
Priestley, M. B. (1981). Spectral analysis and time series. Academic Press, London.
Priestley, M. B. (1988). Non-linear and non-stationary time series analysis. Academic Press,
London.
Sakiyama, K. and M. Taniguchi (2004). Discriminant analysis for locally stationary processes.
Journal of Multivariate Analysis 90, 282–300.
26
Stankovic̀, R. S., C. Moraga, and J. Astola (2005). Fourier Analysis on Finite Groups with Applications in Signal Processing and System Design. Hoboken, NJ: Wiley-IEEE.
Stoffer, D. S. (1985). Central limit theorems for finiteWalsh-Fourier transforms of weakly stationary time series. Journal of Time Series Analysis 6, 261–267.
Stoffer, D. S. (1987). Walsh-Fourier analysis of discrete-valued time series. Journal of Time Series
Analysis 8, 449–467.
Stoffer, D. S. (1990). Multivariate Walsh-Fourier analysis. Journal of Time Series Analysis 11,
57–73.
Stoffer, D. S. (1991). Walsh-Fourier analysis and its statistical applications. Journal of the American Statistical Association 86, 461–479.
Taniguchi, M., L. Zhao, P. Krishnaiah, and Z. Bai (1989). Statistical analysis of dyadic stationary
processes. Annals of the Institute of Statistical Mathematics 41, 205–225.
Tjøstheim, D. (1976). Spectral generating operators for non-stationary processes. Advances in
Applied Probability 8, 831–846.
Walsh, J. (1923). A closed set of normal orthogonal functions. American Journal of Mathematics 45, 5–24.
27
| 10math.ST
|
arXiv:1105.3869v2 [math.CT] 15 Aug 2013
ON THE IMAGE OF THE TOTALING FUNCTOR
KRISTEN A. BECK
Abstract. Let A be a DG algebra with a trivial differential over a commutative unital ring. This paper investigates the image of the totaling functor,
defined from the category of complexes of graded A-modules to the category
of DG A-modules. Specifically, we exhibit a special class of semifree DG Amodules which can always be expressed as the totaling of some complex of
graded free A-modules. As a corollary, we also provide results concerning the
image of the totaling functor when A is a polynomial ring over a field.
Introduction
Let A be a DG algebra over a commutative unital ring. Furthermore, let DG(A)
denote the category of DG A-modules and their degree zero chain maps, and let
ChGr(A♮ ) denote the category of (co)chain complexes of graded A♮ -modules and
their degree zero chain maps. The motivation for the work in this paper is to better
understand the difference between these categories. In the most general case, one
can easily see that DG(A) is a much ‘richer’ category than ChGr(A♮ ) — indeed,
its objects take into account two differentials rather than just one. However, when
A has a trivial differential (so that A = A♮ ), the difference between DG(A) and
ChGr(A♮ ) is not so striking.
The goal of this paper is to address the latter scenario by studying the image of
the so-called totaling functor Tot : ChGr(A♮ ) → DG(A) in the case that A has a
trivial differential. Given a cochain complex X of graded modules over such a DG
algebra A, one defines Tot X to be the complex whose underlying graded A♮ -module
structure is given by
M
(Tot X)♮ :=
Σ−i X i
i∈Z
and whose differential follows in a natural way from that of X. Upon defining an
A-action on this complex, Tot X admits the structure of a DG A-module. The
primary question we consider is whether the totaling functor is surjective.
In Theorem 2.4, we provide a necessary and sufficient condition for a DG module
over a DG algebra A with a trivial differential to be equal to the totaling of some
complex of graded free A♮ -modules. The constructive nature of the proof of this
result furthermore allows us to express a ‘totaling pre-image’ for DG modules which
lie in the image of the totaling functor. To answer the question of whether the
totaling functor is surjective, we restrict to the case that A is a polynomial ring
over a field. In Example 2.5, we exhibit a DG module over a polynomial ring in
Date. November 9, 2017.
2010 Mathematics Subject Classification. 13D09, 18E30.
Key words and phrases. Totaling, differential graded algebra, derived category.
This research was partially supported by NSA Grant H98230-07-1-0197 and NSF GK-12 Program Grant 0841400.
1
2
K. A. BECK
two (or more) variables which does not satisfy the condition specified by Theorem
2.4, and therefore does not lie in the image of the functor. Moreover, in Theorem
2.11, we illustrate that every DG module over a polynomial ring in one variable is
quasiisomorphic to the totaling of some complex of graded A♮ -modules.
1. Background
The results in this paper will assume a working understanding of differential
graded (DG) algebras and their modules. For a thorough treatment of this subject,
the reader is referred to [1], [2], or [4]. In what follows, A is assumed to be a DG
algebra over a commutative unital ring.
1.1. Semifree DG modules. We begin by introducing a class of DG modules
which generalize free modules over a ring. They will form the basic structures
necessary for the construction of (counter-)examples in the sequel, and will also be
essential to the statement of our main result.
Definition 1.1.1. Let M be a DG A-module. A subset E ⊆ M ♮ is called a
semibasis for M if
(1) E is F
a basis for M ♮ over A♮ , and
(2) E = d∈N Ed (a disjoint union) such that
!
G
∂(Ed ) ⊆ A
Ei
i<d
for all d ∈ N.
A DG module that possesses a semibasis is said to be semifree.
Proposition 1.1.2. [2, 8.2.3] Let M be a DG A-module. The following are equivalent.
(1) M is semifree.
(2) M ♮ has a well-ordered basis E over A♮ such that for each e ∈ E
∂(e) ∈ A ({e′ ∈ E | e′ < e}) .
F
Proof. To show (1) ⇒ (2), let E = d∈N Ed be a semibasis for M over A. For each
d ∈ N, impose an ordering on Ed , and further suppose that whenever d′ < d, e′ < e
for every e′ ∈ Ed′ and e ∈ Ed . This implies that E has an ordering with the desired
property.
On the other hand, in order to show that (2) ⇒ (1), suppose that E is a wellordered basis for M ♮ over A♮ such that ∂(e) ∈ A ({e′ ∈ E | e′ < e}) for every e ∈ E.
Set E−1 = ∅ and M−1 = {0}, and for each d ∈ N, recursively define Ed and Md in
the following manner.
S
Ed := e ∈ E \ i<d Ei | ∂(e) ∈ Md−1
[
Md := A Ei
i≤d
By this construction, the
F Ed are mutually disjoint. Furthermore, the well-ordering
of E implies that E = d∈N Ed , and the result follows.
ON THE IMAGE OF THE TOTALING FUNCTOR
3
Remark 1.1.3. A semifree resolution of a DG A-module M is a quasiisomorphism
π : F → M of DG A-modules where F is semifree. In [2], Avramov, Foxby, and
Halperin show that every DG module possesses a semifree resolution. This fact
makes possible the study of differential graded homological algebra.
1.2. The totaling functor. Throughout this section, suppose that A has a trivial
differential.1 Let X be a cochain complex in ChGr(A♮ ) = L
ChGr(A), and denote
the internal grading of X i with a subscript; that is, X i = j∈Z Xji . Now define
the totaling of X to be the complex Tot X = (Tot X)♮ , ∂ Tot X , whose underlying
graded structure is given by
M
♮
(Tot X) :=
Σ−i X i .
i∈Z
Therefore, the dth homological component of Tot X is given by
M
i
(Tot X)d =
Xd+i
i∈Z
and one defines the A-module structure on Tot X according to
a(σ −i xi )i∈Z := σ −i ((−1)|a|i axi )
i∈Z
−i i
for each a ∈ A and (σ x )i∈Z ∈ Tot X. Moreover, the differential on Tot X is
defined by
i
∂ Tot X (σ −i xi )i∈Z := σ −i−1 ∂X
(xi ) i∈Z
for each (σ −i xi )i∈Z ∈ Tot X.
It is now straightforward to check that Tot X is in fact a DG module over A.
Indeed, for any a ∈ A and (σ −i xi )i∈Z ∈ Tot X, one has
∂ Tot X a(σ −i xi )i∈Z = ∂ Tot X σ −i ((−1)|a|i axi )
i∈Z
−i−1 i
|a|i
i
= σ
∂X ((−1) ax )
i∈Z
|a|i −i−1
i
i
= (−1) σ
(a ∂X (x ))
i∈Z
|a|i
|a|(i+1)
−i−1 i
= (−1) (−1)
aσ
∂X (xi )
i∈Z
i
= (−1)|a| a σ −i−1 ∂X
(xi ) i∈Z
= (−1)|a| a ∂ Tot X (σ −i xi )i∈Z
so that the Leibniz rule is satisfied.
For any morphism µ : X → Y of complexes of graded A-modules, define the
following map.
Tot µ : Tot X → Tot Y
(σ −i xi )i∈Z 7→ (σ −i µi (xi ))i∈Z
1Indeed, Tot can be defined on the category ChDG(A) over an arbitrary DG algebra A (see
[2, Section 7.2] for details). However, for the purposes of our work, it suffices to define the functor
in the present setting.
4
K. A. BECK
To see that Tot µ is a morphism of DG A-modules, note that for each xi ∈ Xji ,
µi (xi ) ∈ Yji . It follows that totaling is functorial. Thus, Tot : ChGr(A) → DG(A)
is called the totaling functor.
Remark 1.2.1.
L totaling of a chain complex in ChGr(A), one should
L To define the
note that i∈Z Σ−i X i ∼
= i∈Z Σi Xi .
2. Results
In this section, we turn our attention to the image of the totaling functor. Unless
otherwise stated, A shall denote a DG algebra with a trivial differential and k shall
denote a field. Furthermore, one may assume that all polynomial rings have the
standard grading.
Definition 2.1. Let A be an arbitrary DG algebra, and M a semifree DG module
with semibasis E over A. Define a family of disjoint sets recursively by
E0 := {e ∈ E | ∂(e) = 0}
(2.1.1)
Eℓ := {e ∈ E | 0 6= ∂(e) ∈ AEℓ−1 }
for all ℓ ∈ Z+ . Notice that since M is semifree, E0 is nonempty. Now consider the
following containment of sets.
G
(2.1.2)
Eℓ ⊆ E
ℓ∈N
If the containment in (2.1.2) is strict, we say that E has crossing.
Remark 2.2. One can determine whether a particular semibasis for a finitelygenerated DG module has crossing by simply examining the matrix representing ∂.
Indeed, if D is the matrix representing ∂ with respect to the semibasis E, then D
must a priori be strictly upper-triangular. ThisFfollows directly from the definition
m
of semibasis. If one further supposes that E = ℓ=0 Eℓ has no crossing, then D will
take a block super-diagonal form. Specifically,
Z0 D0
Z1 D1
Z2
D=
..
. D
m−1
Zm
where Zi is a |Ei | × |Ei | zero matrix, Di is a |Ei | × |Ei+1 | matrix with entries in A♮ ,
and the non-specified entries are zeros.
To illustrate the concept of crossing, we provide an example which exhibits
semibases with and without crossing for the same semifree DG module.
Example 2.3. Let A = k[x, y, z] and consider the rank four semifree DG A-module
M with semibasis E = {e1 , e2 , e3 , e4 } such that |e1 | = 0, |e2 | = 2, |e3 | = 3, |e4 | = 5,
and
∂(e1 ) = 0
∂(e2 ) = xe1
∂(e3 ) = yze1
∂(e4 ) = xz 3 e1 + yze2 − xe3 .
ON THE IMAGE OF THE TOTALING FUNCTOR
5
Then E0 = {e1 }, E1 = {e2 , e3 }, and Eℓ = ∅ for ℓ ≥ 2. Thus, the strict containment
G
Eℓ ( E
ℓ∈N
implies that E has crossing.
On the other hand, define a semibasis E ′ for M by e′i = ei for 1 ≤ i ≤ 3 and
′
e4 = e4 − z 3 e2 . The action of ∂ M on E ′ is given by
∂(e′1 ) = 0
∂(e′2 ) = xe′1
∂(e′3 ) = yze′1
∂(e′4 ) = yze′2 − xe′3 .
In this case, we obtain E0 = {e′1 }, E1 = {e′2 , e′3 }, E2 = {e′4 }, and Eℓ = ∅ for ℓ ≥ 3.
Therefore, we have an equality of sets
G
Eℓ = E ′
ℓ∈N
which implies that E has no crossing.
Next we state our main result, which gives a necessary and sufficient condition
for a DG A-module to be equal to the totaling of some complex of graded free
A-modules.
Theorem 2.4. Let M be a semifree DG module over a DG algebra A with trivial
differential. Then M has a semibasis without crossing if and only if there exists a
bounded-below complex X of graded free A-modules such that Tot X = M .
F
Proof. Let E = d∈N Ed be aFsemibasis of M over A. First suppose that E has no
crossing, implying that E = ℓ∈N Eℓ , where
E0 = {e ∈ E |∂ M (e) = 0}
Eℓ = e ∈ E | 0 6= ∂ M (e) ∈ AEℓ−1
for all ℓ ∈ Z+ . Now define a (possibly infinite) sequence X of homomorphisms of
graded A-modules by
X:
∂X
∂X
· · · → Σ−2 AE2 −−2→ Σ−1 AE1 −−1→ AE0 → 0
where, for each ℓ ∈ N and e ∈ Eℓ , one has that
∂ℓX (0, . . . , 0, σ −ℓ e, 0, . . . , 0) = σ −ℓ+1 ∂ M (e) ⊆ Σ−ℓ+1 AEℓ−1 .
X
By construction, ∂ℓ+1
◦ ∂ℓX = 0 for all ℓ ∈ N, implying that X ∈ ChGr(A). To see
that Tot X = M , notice that
M
(Tot X)♮ =
Σℓ X ℓ
ℓ∈N
M
=
Σℓ Σ−ℓ AEℓ
ℓ∈N
∼
=
M
ℓ∈N
♮
=M
AEℓ
6
K. A. BECK
and that
∂ Tot X (σ ℓ σ −ℓ e) = σ ℓ−1 ∂ℓX (σ −ℓ e)
= σ ℓ−1 σ −ℓ+1 ∂ M (e)
= ∂ M (e)
for all ℓ ∈ N and e ∈ Eℓ . The result follows.
On the other hand, suppose that X is a bounded-below complex of graded free
ei be a basis for Xi over A. That is,
A-modules such that Tot X = M , and let E
where
♮
M ♮ = (Tot X) =
∂X
∂X
e0 → 0
e1 −−1→ AE
e2 −−2→ AE
· · · → AE
X:
M
i∈N
ei
Σi AE
and
∂ M (σ i e)i∈N = σ i−1 ∂iX (e) i∈N
ei . Now define
for each e ∈ E
o
Gn
ei | ∂iX (e) = 0
Ee0 :=
e∈E
i∈Z
Eeℓ :=
Gn
i∈Z
ei | 0 6= ∂ X (e) ∈ AEeℓ−1
e∈E
i
o
F
ei for each
for each ℓ ∈ Z+ . Note that one can write Eeℓ = i∈N Eei,ℓ where Eei,ℓ ⊆ E
F
L
ie
ℓ ∈ N. Let Eℓ := i∈N Σ Ei,ℓ . We will show that E := ℓ∈N Eℓ is a semibasis for
M which does not have crossing.
To this end, we first note that E is a basis for M ♮ over A by construction.
Furthermore, if (σ i e)i∈N ∈ Eℓ , then
M
Σi−1 AEei−1,ℓ−1 = AEℓ−1
∂ M (σ i e)i∈N = σ i−1 ∂iX (e) i∈N ∈
i∈N
since e ∈ Eei,ℓ . This implies that E is a semibasis for M without crossing.
To see that the statement of Theorem 2.4 is not trivial, consider the following
example.
Example 2.5. Let A = k[x1 , . . . , xd ] where d ≥ 2, and consider the rank four
semifree DG A-module M given by M ♮ = Ae1 ⊕ Ae2 ⊕ Ae3 ⊕ Ae4 such that |e1 | = 0,
|e2 | = 3, |e3 | = 4, |e4 | = 8, and
∂(e1 ) = 0
∂(e2 ) = x1 x2 e1
∂(e3 ) = x32 e1
∂(e4 ) = x71 e1 − x42 e2 + x1 x22 e3 .
Note that E = {e1 , e2 , e3 , e4 } has crossing.
matrix representation
0 x1 x2
0
0
D=
0
0
0
0
This can be verified by examining the
x32
0
0
0
x71
−x42
x1 x22
0
ON THE IMAGE OF THE TOTALING FUNCTOR
7
of ∂ with respect to E. We will show that there does not exist a semibasis for M
without crossing.
To this end, suppose that E ′ is another semibasis for M , and denote by P
the change-of-basis matrix; that is, P : E-coordinates → E ′ -coordinates. If D′ is
′
′
the matrix representing ∂ with respect to
E , then D = P D ′must take the form
prescribed by Remark 2.2. Letting P = pij , one can write D as follows.
0 p11 x1 x2 p11 x32 p11 x71 − p12 x42 + p13 x1 x22
0 p21 x1 x2 p21 x32 p21 x71 − p22 x42 + p23 x1 x22
D′ = P D =
0 p31 x1 x2 p31 x32 p31 x71 − p32 x42 + p33 x1 x22
0 p41 x1 x2 p41 x32 p41 x71 − p42 x42 + p43 x1 x22
Now, since E ′ is a priori a semibasis for M , it should be true that D′ is strictly
upper-triangular. This implies that p21 = p31 = p41 = 0 and p41 x71 − p42 x42 +
p43 x1 x22 = −p42 x42 + p43 x1 x22 = 0. Furthermore, since P needs to be invertible, p11
must be a unit. With these relations given, D′ now takes the form
0 αx1 x2 αx32 αx71 − p12 x42 + p13 x1 x22
0
0
0
p21 x71 − p22 x42 + p23 x1 x22
D′ =
0
0
0
p31 x71 − p32 x42 + p33 x1 x22
0
0
0
0
for some α ∈ k. Supposing that E ′ has no crossing, it follows by Remark 2.2 that
αx71 − p12 x42 + p13 x1 x22 = 0. But this implies that px42 − qx1 x22 = x71 for some
p, q ∈ k[x1 , . . . , xd ], which cannot be true. Therefore, E ′ must have crossing.
Remark 2.6. A result of Avramov and Jorgensen [3] can furthermore be used to
show that the semifree DG module exhibited in Example 2.5 is, more generally, not
in the image of the totaling functor Tot : DGr(A) → DDG(A) defined on the respective derived categories. The reader is referred to [5, Chapter 10] for a constructive
definition of derived categories.
Corollary 2.7. et A be a polynomial ring in more than one variable over a field.
Then the functor Tot : ChGr(A) → DG(A) is not surjective.
The constructive nature of the proof of Theorem 2.4 makes it possible for one to
cook up a ‘totaling pre-image’ for any semifree DG module which admits a basis
without crossing. The following example demonstrates this fact.
Example 2.8. Let M be the rank four semifree DG module with semibasis E over
A = k[x, y, z] defined in Example 2.3. As E has no crossing, the construction in
the proof to Theorem 2.4 yields a complex
X:
"
yz
−x
#
[ x yz ]
0 → Σ−2 Ae4 −−−−→ Σ−1 (Ae2 ⊕ Ae3 ) −−−−−−→ Ae1 → 0
of graded A-modules such that Tot X = M .
As one might have guessed, it is not always possible to cook up a semifree DG
module whose semibases are guaranteed to have crossing. For example, if a DG Amodule has rank no more than three over A, it is impossible to define a differential
which gives the the semibasis crossing. The following corollary makes this precise.
Corollary 2.9. Let A be a graded domain, and suppose that M is a semifree DG
A-module. If rankA M ≤ 3 then there exists a complex X of graded A-modules such
that Tot X = M .
8
K. A. BECK
Proof. Let E = {e1 , . . . , ed } be a well-ordered basis for M ♮ over A. Since M is
semifree, we obtain the following possible (non-trivial) forms for its differential,
where aij ∈ A for each i, j.
d = 1 ⇒ ∂(e1 ) = 0
d = 2 ⇒ ∂(e1 ) = 0
∂(e2 ) = a12 e1
d = 3 ⇒ ∂(e1 ) = 0
∂(e2 ) = a12 e1
∂(e3 ) = a13 e1
or
∂(e1 ) = 0
∂(e2 ) = 0
∂(e3 ) = a13 e1 + a23 e2
Note that in each case, E has no crossing. The result follows by Theorem 2.4.
We now turn our attention to the image of the totaling functor in the case that A
is a polynomial ring in one variable. In this setting, it turns out that using Theorem
2.4 to find semifree DG modules which are not in the image of the totaling functor
is not so straightforward. Accordingly, we take a slightly different approach here.
The next lemma, which will follows directly from the structure theorem for
finitely generated modules over a principal ideal domain.
Lemma 2.10. Let M be a DG module which is n-generated over A = k[x]. Then
there exist integers 0 ≤ s ≤ d and 1 ≤ t ≤ d such that H(M ) has a graded minimal
free resolution over A given by
0→
s
M
0
h1
..
.
0
0
hs
Σcj A −−−−−−−−−−−−−→
t
M
Σri A → H(M ) → 0
i=1
j=1
for some integers ri , cj for 1 ≤ i ≤ t and 1 ≤ j ≤ s, and where each hi = xci −ri ∈
Ax for each 1 ≤ i ≤ s.
Theorem 2.11. Let A be a polynomial ring in one variable over a field. Then
every semifree DG A-module is quasiisomorphic to the totaling of some complex
of graded free A-modules. Specifically, the functor Tot : DGr(A) → DDG(A) is
surjective.
Proof. Let M be a rank d semifree DG module over A = k[x]. By Lemma 2.10,
one can assume the homology of M to have the form
(2.11.1)
H(M ) ∼
=
s
t
M
M
Σr i A
⊕
Σr i A
ci A
h
Σ
i
i=1
i=s+1
for some 1 ≤ t ≤ d and 0 ≤ s ≤ t, and positive integers ri , cj . (Of course, s = 0
corresponds to the case that H(M ) is a free A-module.) Consider the deleted
ON THE IMAGE OF THE TOTALING FUNCTOR
9
minimal graded free resolution F of H(M ) given as follows.
0
0
F:
// Σc1 A
⊕
..
.
⊕
// Σcs A
h1
// Σr1 A
⊕
..
.
⊕
0
⊕
..
.
⊕
// Σrs A
⊕
// Σrs+1 A
⊕
..
.
⊕
0
⊕
// Σrt A
hs
// 0
// 0
// 0
// 0
To complete the proof, we will show that Tot F ≃ M . While it is clear that
H(Tot F ) ∼
= H(M ), it remains to exhibit a chain map (or sequence thereof) which
induces this isomorphism. To this end, for each 1 ≤ i ≤ t, let Gi be the subcomplex
which is given by the ith summand of F .
Gi :
(
h
i
Σr i A → 0
0 → Σci A −→
0 → Σr i A → 0
if 1 ≤ i ≤ s
if s < i ≤ t
Our goal is to define a family of chain maps µi : Tot Gi → M such that the chain
map µ : Tot F → M given by
(2.11.2)
µ = (µi ) :
t
M
Tot Gi → M
i=1
induces an isomorphism in homology. For each 1 ≤ i ≤ t, define µi by a family
µij : (Tot Gi )j → Mj of homomorphisms of vector spaces over k in such a way that
each µi is A-linear and furthermore commutes with the differentials of Tot Gi and
M.
For the sake of clarity, we include the following diagram, which illustrates the
action of µi on Tot Gi . Note that the complexes are expressed vertically and that
the diagonal maps represent the nontrivial component of ∂ Tot Gi .
10
K. A. BECK
..
.
..
.
⊕
⊕
(Σci +1 A)ci +2 ⊕ (Σri A)ci +2
▼▼▼
▼▼▼hi
▼▼▼
⊕
⊕
▼▼&&
(Σci +1 A)ci +1 ⊕ (Σri A)ci +1
▼▼▼
▼▼▼hi
▼▼▼
⊕
⊕
▼▼&&
0
⊕ (Σri A)ci
(2.11.3)
⊕
..
.
..
.
µic
i +2
// Mci +2
∂cM+2
i
µic +1
i
// Mci +1
∂cM
i +1
µic
i
⊕
..
.
// Mci
M
∂ci
..
.
∂rM
i −1
⊕
(Σri A)ri
µir
i
// Mri
∂rM
i
⊕
0
⊕
..
.
µir −1
i
// Mri −1
..
.
Although this diagram is restricted to indices given by 1 ≤ i ≤ s, the situation
is straightforward for i > s; indeed, these cases represent the torsion-free part of
H(M ). Thus, for s < i ≤ t, one has the following.
(
0 → Mj
if j < ri
i
µj :
ri
if j ≥ ri
(Σ A)j → Mj
Now consider the following isomorphism of graded A-modules.
ϕ:
t
s
M
M
Σr i A
Σri A → H(M )
⊕
ci A
h
Σ
i
i=s+1
i=1
in such a way that ϕ (σ ri 1) =
Fixing 1 ≤ i ≤ t, let 0 6= zi ∈ Mri be a cycle defined
ℓ ri
cls(zi ) ∈ Hri (M ). Since ϕ is A-linear, ϕ x σ 1 = cls(xℓ zi ) for ℓ ≥ 0. Notice that,
for small enough values of ℓ, these classes are nonzero. To be precise, cls(xℓ zi ) = 0
if and only if 1 ≤ i ≤ s and ℓ ≥ ci − ri . Therefore, for each 1 ≤ i ≤ s, there exists
mi ∈ Mci +1 such that ∂cMi +1 (mi ) = xci −ri zi .
We now proceed to define, for each i and j, a basis Uji for (Tot Gi )j over k.
{xj−ri σ ri 1, (−1)j−ci −1 xj−ci −1 σ ci +1 1} if 1 ≤ i ≤ s and j > ci
Uji ⊇
otherwise
{xj−ri σ ri 1}
ON THE IMAGE OF THE TOTALING FUNCTOR
11
Finally, we define µij : (Tot Gi )j → Mj in the following way.
j−r
i
zi
if u = xj−ri σ ri 1
x
i
j−ci −1
x
mi
if u = (−1)j−ci −1 xj−ci −1 σ ci +1 1
µj (u) =
0
otherwise
By construction, µi = (µij ) is an A-linear degree zero chain map between Tot Gi
and M . Further, one can easily check that the Leibniz rule is satisfied, so that
the map µ : Tot F → M given in (2.11.2) is a morphism of DG modules. The
above construction also guarantees that µ establishes
L a one-to-one correspondence
between the generators of homology of Tot F = m
i=1 Tot Gi and that of M . The
result follows.
The following example illustrates the practical use of the construction used in
the proof of Theorem 2.11.
Example 2.12. Let M be the rank five semifree DG module over A = k[x] with
semibasis given by {e1 , e2 , e3 , e4 , e5 } such that |e1 | = 0, |e2 | = 2, |e3 | = 4, |e4 | = 8,
|e5 | = 9, and where the differential of M is defined by the following.
∂(e1 ) = 0
∂(e2 ) = 0
∂(e3 ) = 0
∂(e4 ) = x7 e1 + x5 e2
∂(e5 ) = x4 e3
The homology of M can be decomposed
Ae1 ⊕ Ae2 ⊕ Ae3
H(M ) =
A(x7 e1 + x5 e2 ) ⊕ Ax4 e3
A(x2 e1 + e2 )
Ae3
∼
⊕ Ae1
⊕
=
A(x7 e1 + x5 e2 ) Ax4 e3
(2.12.1)
whence one obtains the following deleted minimal free resolution F of H(M ).
0
F:
0
x5
// Σ7 A
⊕
// Σ8 A
⊕
0
x4
// Σ2 A
⊕
// Σ4 A
⊕
// A
// 0
// 0
// 0
We shall now utilize the proof of Theorem 2.11 to show that Tot F ≃ M . From
F we obtain the following subcomplexes.
x5
G1 :
0 → Σ7 A −→ Σ2 A → 0
G2 :
0 → Σ8 A −→ Σ4 A → 0
G3 :
0→A→0
x4
Referring to the decomposition of homology in (2.12.1), one has that the cycles
generating H(M ) over A are z1 = x2 e1 + e2 , z2 = e3 , and z3 = e1 . Then, for
12
K. A. BECK
each j ∈ N, a basis Uj1 of (Tot G3 )j over k must be chosen to contain {xj σ 0 1}.
Furthermore, the respective bases of (Tot G1 )j and (Tot G2 )j over k are given as
follows.
{xj−2 σ 2 1}
if j < 8
Uj1 ⊇
j−2 2
j−8 j−8 8
{x σ 1, (−1) x σ 1} if j ≥ 8
{xj−4 σ 4 1}
if j < 9
Uj2 ⊇
{xj−4 σ 4 1, (−1)j−9 xj−9 σ 9 1} if j ≥ 9
Using these bases, one can now define chain maps µi = (µij ) : (Tot Gi )j → Mj .
j
j−2
if u = xj−2 σ 2 1
x e1 + x e2
xj−8 e4
if u = (−1)j−8 xj−8 σ 8 1
µ1j (u) =
0
otherwise
j−4
if u = xj−4 σ 4 1
x e3
xj−9 e5
if u = (−1)j−9 xj−9 σ 9 1
µ2j (u) =
0
otherwise
(
xj e1
if u = xj σ 0 1
µ3j (u) =
0
otherwise
Finally, µ : Tot F → M is given by µ(x) = µ1 (x), µ2 (x), µ3 (x) for each x ∈ Tot F .
Acknowledgments
The author would like to thank Dave Jorgensen and Sean Sather-Wagstaff for
both piquing and cultivating her interest in DG algebra. Thanks also to the referee
for many helpful suggestions which greatly improved the overall quality of this
manuscript.
References
[1] L. L. Avramov, Infinite free resolutions, Six lectures on commutative algebra, 2010, pp. 1–118.
MR2641236
[2] L. L. Avramov, H.-B. Foxby, and S. Halperin, Differential graded homological algebra, (in
preparation).
[3] L. L. Avramov and D. A. Jorgensen, Realization of cohomology over a complete intersection,
(in preparation).
[4] B. Keller, On differential graded categories, International Congress of Mathematicians. Vol.
II, 2006, pp. 151–190. MR2275593 (2008g:18015)
[5] C. A. Weibel, An introduction to homological algebra, Cambridge Studies in Advanced Mathematics, vol. 38, Cambridge University Press, Cambridge, 1994. MR1269324
Kristen A. Beck, Department of Mathematics, University of Arizona, 617 N. Santa
Rita Ave., Tucson, AZ, 85719, U.S.A.
E-mail address: [email protected]
| 0math.AC
|
Consensus analysis of systems with time-varying
interactions : An event-triggered approach
S. Arun Kumar ∗ N. R. Chowdhury ∗ S. Srikant ∗ J. Raisch ∗∗,∗∗∗
∗
arXiv:1705.00537v2 [cs.SY] 2 May 2017
Department of Systems and Control Engineering, Indian Institute of
Technology Bombay, Mumbai -400076, India.
(e-mail:{arunkumar92, nilanjan, srikant}@sc.iitb.ac.in)
∗∗
Fachgebiet Regelungssysteme, Technische Universität Berlin, Berlin,
10587 Germany (e-mail: [email protected]).
∗∗∗
Systems and Control Theory Group, Max Planck Institute for Dynamics of
Complex Technical Systems.
Abstract: We present consensus analysis of systems with single integrator dynamics interacting via
time-varying graphs under the event-triggered control paradigm. Event-triggered control sparsifies the
control applied, thus reducing the control effort expended. Initially, we consider a multi-agent system
with persistently exciting interactions and study the behaviour under the application of event-triggered
control with two types of trigger functions- static and dynamic trigger. We show that while in the case of
static trigger, the edge-states converge to a ball around the origin, the dynamic trigger function forces the
states to reach consensus exponentially. Finally, we extend these results to a more general setting where
we consider switching topologies. We show that similar results can be obtained for agents interacting
via switching topologies and validate our results by means of simulations.
1
Introduction
A flock of birds, a swarm of bees, a school of fishes, a colony of
ants -all display a wonderful coordination and complex patterns
which have caught our attention since time immemorial. On
closer observation we realize that these seemingly complex
patterns emerge out of simple yet powerful, local rules on
each agent of the group based on interactions with only the
neighbours. When we attempt to algorithmize such a ’decentralized’ behaviour, two types of questions can be posed- what
would be the result of a particular decentralized control law or
what decentralized control law would result in a desired global
pattern or formation. Consensus is one of the most common and
powerful ’global’ behaviours usually studied, primarily because
it can be easily extended to other problems like formation, rendezvous, flocking etc. Vicsek et al. (1995) proposed an averagebased decentralized control law for a multi-agent system based
on only local information and observed that the agents attain
consensus while Jadbabaie et al. (2003) gave a proof of the
same assuming the agents interact via connected graphs that
can switch at different time instants. Various results have been
proposed by Ren and Beard (2005), Ren and Atkins (2005),
Olfati-Saber and Shamma (2005), and others on consensus of
a multi-agent system under directed or undirected graphs, with
switching topologies and also with time delays.
Studying consensus behaviour for systems with switching or
time-varying graphs is naturally of interest as in real-life scenarios it is not possible to assume that each agent has the same
set of neighbours at all times. Time-varying graphs are useful
when the information from each neighbour is assigned a weight
proportional to the reliability of the information or the distance
between them. Martin and Girard (2013) proved consensus under the assumptions of persistent-connectivity and cut-balance
interactions. Chowdhury et al. (2016) obtained bounds on the
rate of convergence for single-integrator and double integrator
dynamics under persistent interactions.
With control laws being implemented on digital computers,
developing discrete-time counterparts to continuous-time control laws is an eventuality. Among discrete control laws eventtriggered control is preferred over time-triggered control as it
comes into play only when an ’event’ is triggered, thus sparsifying control. Event-triggered control of multi-agent systems
has been studied by many. Dimarogonas et al. (2012), Seyboth
et al. (2011) prove consensus under a connected time-invariant
graph for single integrator dynamics, while Yu et al. (2015),
Zhu et al. (2014) have extended the results for agents with
general linear systems dynamics. Chen and Dai (2016) has
studied the consensus of time varying systems with non-linear
dynamics under event-triggered control but under a constant
interaction topology. Our work considers the broader case of
time-varying graphs that can have different spanning-trees and
we prove consensus of single-integrator systems under eventtriggered control structure.
The paper is organized as follows. We brush up on graph theory
and persistent excitation in section 2. The system dynamics for
single integrator systems are introduced in section 3. In section
4, we introduce the notion of event-triggered control, define
the trigger conditions and evaluate convergence under eventtriggered control in section 5. We extend the results to switching
graphs in section 5.1 and show simulation results in section 6.
The results that we obtained are summarized in section 7.
2
2.1
Preliminaries
Notions of graph theory
In this work we consider agent interactions represented by undirected graphs G = (V, E), where V = {v1 , v2 , · · · vn } denotes
2
a non-empty set of nodes and [V ] ⊇ E = {e1 , e2 , · · · em } is
2
the edge set, where [V ] is the set of all subsets of V containing
two elements. Each node represents an agent of the system and
each edge (vi , vj ) signifies that the agents occupying nodes vi
and vj can exchange information with each other. We define
D (G) (= [dij ]) ∈ Rn×m to be the incidence matrix associated
with the graph G by arbitrarily assigning orientation to each
edge ej ∈ E. Then [dij ] = −1 if vi is the tail of ej ,[dij ] = 1 if
vi is the head of ej , [dij ] = 0 otherwise. The Laplacian of G,
>
L (G) = D (G) D (G)
(1)
is a symmetric square matrix that captures the inter-connections
between each pairs of nodes. We model the time-varying interactions between the agents by assuming a constant, underlying
graph G with edge weights that can be time-varying. The Laplacian can be tweaked to reflect the resultant time-varying graph
G̃ (t) as,
>
L G̃ (t) = D (G) W (t) D (G) .
(2)
where W (t) ∈ Rm×m is a diagonal matrix which captures the
time-varying nature of each interaction and wii ≥ 0 ∀t. We
assume that the underlying graph G is connected and therefore
contains a spanning-tree, i.e. the edge set E of G can be partitioned into two subsets E = Eτ ∪ Ec , where Eτ consists of
the spanning-tree edges and Ec contains the cycle edges. We
also assume that the time-varying edge weight matrix W (t)
is piece-wise continuous and satisfies the persistent excitation
condition with constants (µ1 , µ2 , T ).
2.2
Persistence of excitation
Definition 1. (Sastry and Bodson, 2011, p. 72) The signal
g(·) : R≥0 → Rn×m is Persistently Exciting (PE) if there exist
finite positive constants µ1 , µ2 , T such that,
Z t+T
µ2 In ≥
g(τ )g(τ )T dτ ≥ µ1 In
∀t ≥ t0 (3)
t
A function g(·) that satisfies the condition (3) is said to be
persistently exciting with constants (µ1 , µ2 , T ).
We say that a graph G̃ is persistently exciting if its associated
edge-weight matrix W (t) is persistently exciting.
2.3
Other Conventions
k·k denotes the frobenius 2-norm on vectors and the induced
2-norm on matrices. λmin (·) and λmin (·) operate on square
matrices and return the minimum and maximum eigenvalues
respectively of the said matrix. Boldfaced 1 and 0 (1, 0) represent vectors with all ones and all zeroes respectively and I
is used to denote the identity matrix. Their dimensions can be
inferred contextually if not mentioned explicitly.
3
Network Models
In this section we state the single integrator dynamical equations and perform a series of linear transformations, to bring
them to a form that we could work with later on. This section
has been taken from Chowdhury et al. (2016) and presented
here for reference of the readers.
3.1
Single Integrator
Consider a multi-agent system with states xi ∈ R for i =
1, 2, 3 · · · n with a connected underlying
graph
G and the Laplacian of the time-varying graphs L G̃ (t) = [lij (t)]. The
dynamics of each state with control ui ∈ R can be written as,
ẋi =ui
n
X
ui = − k
lij (t)xj
(4)
j=1
for t > t0 with initial condition xi (t0 ) ∈ R for i = 1, 2, 3.....n
and positive control gain k ∈ R. The augmented dynamics
of the system can be written in terms of the state vectors
>
x = [x1 , x2 , · · · xn ] ∈ Rn as,
ẋ(t) = − kL G̃(t) x
>
= − kD (G) W (t)D (G) x
(5)
Taking cue from Zelazo and Mesbahi (2011), Mesbahi and
Egerstedt, 2010, p. 77-81 we transform the consensus problem
of equation (5) into a stabilization problem by considering the
edge states instead of node states. We effect the conversion
through the following transformation,
>
xe =D (G) x.
(6)
The dynamics of the edge states, on differentiation of (6) yields,
ẋe = − kLe (G) W (t) xe
(7)
m×m
where Le (G) ∈ R
represents the edge-Laplacian of the
>
graph G and can be expressed as, Le (G) = D (G) D (G) as
shown in Zelazo and Burger (2014). Also, using the fact that
the underlying graph contains a spanning-tree Gτ we partition
the edge states after a suitable permutation as follows,
x
xe = τ
(8)
xc
for xτ ∈ Rp and xc ∈ Rm−p where p is the number
of edges in Gτ . Similarly we can partition
D (G), W(t) as
Wτ (t) 0
D (G) = [D (Gτ ) D (Gc )], W (t) =
. The
0
Wc (t)
edge-Laplacian, in terms of the partitions of D (G) is then,
>
Le (G) = [D (Gτ ) D (Gc )] [D (Gτ ) D (Gc )]
>
Le (Gτ )
D (Gτ ) D (Gc )
=
.
(9)
>
D (Gc ) D (Gτ )
Le (Gc )
We know that for connected graphs, xc can always be written
as xc = Z > xτ as shown in Sandhu et al. (2005) where Z =
−1
>
(Le (Gτ )) D (Gτ ) D (Gc ). This is because the spanning-tree
edges essentially capture the behaviour of all the edges. So
we can focus our attention only on the spanning-tree edges.
Using (7), (8) and (9), we get ẋτ = −kLe (Gτ ) RW (t) R> xτ
where R = [Ip Z] ∈ Rp×m . The preceding transformations
that helped us re-write the system dynamics in terms of xτ
are along the lines of Zelazo and Mesbahi (2011) and Mesbahi
and Egerstedt, 2010, p. 77-81. Since Le (Gτ ) is symmetric and
positive definite, it can be diagonalized as Le (Gτ ) = ΓΛΓ>
for some orthogonal matrix Γ ∈ Rp×p and diagonal matrix
Λ ∈ Rp×p . Consider a change of variable by the transformation
Υ = Γ> xτ . The above equation becomes,
Υ̇ = − kΛM (t) Υ
(10)
where M (t) = Γ> RW (t) R> Γ ∈ Rp×p . As stated earlier, the
consensus problem of equation (5) is equivalent to stabilization
(1) Static Trigger Function
fi (ei (t)) = kei (t)k − c
(2) Dynamic Trigger Function
problem of equation (10). For the sake of further reference,
we denote the preceding set of transformations from x to Υ
>
by ψ := Γ> [Ip 0m−p ] D (G) . So we have, Υ = ψx. We
state (Chowdhury et al., 2016, Theorem 5) and use the result
to obtain the rate of convergence to consensus of a single
integrator system defined by equations (5).
Theorem 1. ((Chowdhury et al., 2016, Theorem 5)) Consider
the closed-loop consensus dynamics (5). Assume that, the underlying graph G is connected. The states of the closed-loop dynamics x(t) with time-varying communication topology characterized by W (t), achieve consensus exponentially, if there
exists a spanning tree with corresponding edge-weight matrix
Wτ (t) that is persistently exciting. Further the convergence rate
αv to consensus is bounded below by,
1
1
αv ≥
ln
2T
1 − 2kλ√min (Λ)µ1 2
(1+k pkΛkµ2 )
where, T, µ1 and µ2 are the constants appearing in Definition 1
and Λ is a diagonal matrix containing the eigenvalues of the
spanning tree edge Laplacian matrix.
The static trigger function can be seen as a special case of the
dynamic trigger function with β = 0. So we will perform our
analysis with (15) as the trigger function, and substitute β = 0
when we want to evaluate the static-trigger case. The dynamics
of single integrator systems under event-triggered control can
be written as,
ẋ(t) = − kL G̃ (t) (x + e)
(16)
Further,
4.1
kΥ(t)k ≤mv e−αv (t−t0 ) kΥ(t0 )k
(11)
where αv and mv can be calculated from the underlying graph
and control gains. Note: The relationship between kxe k and
kΥk can be established in the following way.
q
2
2
kxe k = kxτ k + kxc k
q
2
≤ kxτ k 1 + kZ > k
Using the fact that Γ is orthogonal, xτ = ΓΥ,
kxe k ≤ρ kΥk
(12)
q
2
where ρ = kΓk 1 + kZ > k . The induced 2-norm of Z can
1
be calculated as λmax Z > Z 2 .
4
Event-Triggered Control
where c > 0
j=1
To define an event, we introduce error variables ei (t) =
x̂i (t) − xi (t) which denote the difference between the broadcasted value and the current value of the state for each agent.
The trigger condition that updates the broadcasted states x̂ effectively shapes the behaviour of the system. We define the trigger condition using a trigger function for each state fi (t, ei ) :
R → R . An event is said to be ’triggered’ when fi > 0. Once
an event is ’triggered’ say at time t∗ , the broadcast value is
0
updated i.e. x̂i (t) = xi (t∗ ) =⇒ ei (t∗ ) = 0 for t∗ ≤ t < t
0
where t is the time instant when the next subsequent event is
triggered. We define two trigger functions as shown in Seyboth
et al. (2011)- the static trigger and the dynamic trigger.
(15)
where e = [e1 , e2 · · · en ]> ∈ Rn . The equivalent stabilization
problem under event triggered control for single integrator will
then be,
Υ̇ = − kΛM (t) (Υ + ẽ)
(17)
where ẽ = ψe.
Bounds on the error variables
In this section, we obtain bounds on ẽ in terms of the bounds
on e. The trigger function is so designed that each ei is always
upper-bounded. We can see that,
kei k ≤ce−βt
(18)
We can relate the bounds on e to bounds on ẽ in the following
>
way. We know that ẽ = ψ e. Let us define ē := D (G) e. From
the structure of D (G), we get
ēi = ej − ek
(19)
for i = 1, 2, · · · m and j, k chosen on the basis of D (G). Using
(18), (19) can be rewritten as,
kēi k = kej − ek k
≤ kej k + kek k
Also,
In this section we outline the event-triggered control strategies and show how it modifies the single-integrator dynamics defined by equation (5). Under the event-triggered control
paradigm each agent broadcasts a (piecewise)constant value, x̂i
which is updated to the current value of the state whenever an
’event is triggered’. The control applied by each agent would
then be,
n
X
ui = − k
lij (t)x̂j
(13)
fi (t, ei (t)) = kei (t)k − ce−β(t)
(14)
≤ 2ce−βt .
(20)
ẽ = ψe
= Γ> [Ip 0m−p ] ē
>
= Γ> [ē1 ē2 · · · ēp ] .
(21)
Using the bound on each ēi from (20),
kẽk ≤ kΓk k[ē1 ē2 · · · ēp ]k
√
≤ pkΓkkēi k
√
≤ 2c pkΓke−βt .
(22)
√
Defining C := 2c p kΓk, we can write the above inequality to
be,
kẽ(t)k ≤ Ce−βt .
(23)
5
Consensus Analysis
Theorem 2. Consider a multi-agent system with single integrator dynamics as defined by equation (5). Assume that the
underlying graph (G), representing the interaction between the
agents be connected, with p edges in the spanning-tree. If the
spanning tree edge weight matrix Wτ (t) is persistently exciting
with constants (µ1 ,µ2 ,T ), then
(1) on application of event-triggered control with a static
trigger function defined by equation (14), the edge states
xe of the system exponentially converge to a ball around
the origin defined by kxe k ≤ ρκm
2 .
(2) on application of event-triggered control with a dynamic
trigger function defined by equation (15), the edge states
xe of the system exponentially converge to origin. Also
the rate of convergence is lower bounded by β as defined
in (15).
αv t0 +2αv T
µ2
where, κ2 = kkΛkmv eCe
and ρ is defined as in (12).
αv T −1
Also the closed loop systems in the cases of static and dynamic
triggers does not exhibit zeno behaviour when β is chosen to be
greater than αv .
Note: When a connected graph has multiple spanning-trees, κM
2
and κm
2 are the largest and smallest values of κ2 that can be
calculated considering each different spanning-tree.
Proof. Let φ (t, t0 ) be the state transition matrix corresponding
to system defined by equation (10) . The solution of the system
can be written using φ (t, t0 ) as,
Υ(t) =φ(t, t0 )Υ(t0 )
We obtain a bound on kφ(t, t0 )k by the following steps.
kΥ(t)k =kφ(t, t0 )Υ(t0 )k
kφ(t, t0 )Υ(t0 )k
=
kΥ(t0 )k
(24)
kΥ(t0 )k
Comparing (24) and (11) we can conclude that,
kφ(t, t0 )Υ(t0 )k
≤ mv e−αv (t−t0 )
kΥ(t0 )k
kφ(t, t0 )Υ(t0 )k
sup
= kφ(t, t0 )k ≤ mv e−αv (t−t0 ) . (25)
kΥ(t0 )k
Υ(t0 )
The last statement holds true because Υ (t0 ) can be arbitrary.
Consider the single integrator multi-agent system with eventtriggered control defined by the equation (17). The solution of
this system can be expressed as ,
Z t
Υ(t) =φ(t, t0 )Υ(t0 ) − k
φ(t, τ )ΛM (τ )ẽ(τ )dτ
(26)
the fact that M (τ ) is persistently exciting and β < αv we can
obtain the following bound on each interval,
Z t0 +T
e−(β−αv )τ kM (τ )k dτ ≤µ2 supτ ∈[t0 ,t0 +T ] e−(β−αv )τ
t0
0
≤µ2 e−(β−αv )(t +T )
Using this upper bound on each integral we get
Z t
e−(β−αv )τ kM (τ )k dτ ≤ µ2 e−(β−αv )T e−(β−αv )t0 +
t0
e−(β−αv )(t0 +T ) + · · · e−(β−αv )(t0 +θT ) )
Using the property of the sum of terms in a geometric progression, we get
Z t
e−(β−αv )τ kM (τ )kdτ ≤
t0
1 − e−(β−αv )θT
µ2 e−(β−αv )(t0 +T )
1 − e−(β−αv )T
(28)
Using inequality (28) in inequality (27),
kΥ(t)k ≤ mv e−αv (t−t0 ) kΥ(t0 )k +
−(β−αv )(t0 +T )
k kΛk mv Ce
µ2 e
−αv t
1 − e−(β−αv )θT
1 − e−(β−αv )T
(29)
We define κ1 = eαv t0 mv kΥ(t0 )k − κ3 , κ2 = e(αv +β)T κ3 ,
k kΛk mv Ce−(β−αv )(t0 +T ) µ2
e−(β−αv )T − 1
to be able to express (29) as,
κ3 =
kΥ(t)k ≤ κ1 e−αv t + κ3 e−βθT −αv (θT −t)
(30)
Using the properties of floor function, we can simplify the
above inequality further,
kΥ(t)k ≤ κ1 e−αv t + κ3 e−β(θT −t)−αv (θT −t)−βt
≤ κ1 e−αv t + κ2 e−βt
−αv t
−βt
(31)
=⇒ kxe (t)k ≤ ρ κ1 e
+ κ2 e
.
(32)
From the above expression it can clearly be seen that for the
We get the following inequality from the above equation,
dynamic trigger case, limt→∞ kxe (t)k = 0. This guarantees
consensus of the states x. Also as β < αv by choice, the rate of
kΥ(t)k ≤ kφ(t, t0 )k kΥ(t0 )k +
Z t
decay of the RHS of (32) will be dominated by β. For the static
trigger case, β = 0
k kΛk
kφ(t, τ )k kM (τ )k kẽ(τ )k dτ
t0
kxe (t)k ≤ ρ κ1 e−αv t + κ2
Using the bounds on φ from equation (25) and ẽ (τ ) from
limt→∞ kxe (t)k ≤ limt→∞ ρκ1 e−αv t + limt→∞ ρκ2
equation (23) we get,
≤ ρκ2
kΥ(t)k ≤mv e−αv (t−t0 ) kΥ(t0 )k +
It can be seen that the edge-states converge to a ball around the
Z t
k kΛk mv Ce−αv t
e−(β−αv )τ kM (τ )k dτ (27) origin kxe k ≤ ρκ2 exponentially with a rate of convergence
bounded below by αv .
t0
To obtain a bound on the integral in the above inequal- Ruling out zeno behaviour in closed loop systems
The system states xi and error states ei together form a hybrid
ity, we divide the interval [t0 , t] into partitions of size T
i.e. [t0 , t0 + T ], [t0 + T, t0 + 2T ] and so on . The number system, which makes it necessary for us to ensure that zeno
th
of such partitions
of length T possible will be given by behaviour does not occur. Say that for the i agent, an event
t−t0
is
triggered
at
a
time
instant
t
and
another
consecutive event
1
, where b·c is the floor function . The last
θ =
T
is
triggered
at
time
t
for
some
t
≤
t
<
t
2
0
1
2 . To rule out the
partition can then be written as [t0 + θT, t]. The integral in
R t0 +T −(β−α )τ
occurrence
of
zeno
behaviour,
it
is
sufficient
to
show that there
v
(27) can then be written as, t0
e
kM (τ )k dτ +
exists a positive, non-zero lower bound on γ := t2 − t1 . We
R t0 +2T −(β−α )τ
Rt
v
e
kM (τ )k dτ · · ·+ t0 +θT e−(β−αv )τ kM (τ )k dτ .have ėi = −ẋi . Consider the following set of inequalities for
t0 +T
Applying Hölder inequality with p = ∞ and q = 1 and using time t1 ≤ t < t2 .
t0
>
kėi k = kẋi k ≤ kẋk = kkD (G) W (t)D (G) x (t1 )k
≤ kkD (G)k kW (t)k kxe (t1 )k
Using (32) we can rewrite the above inequality as,
kėi k ≤ kρkD (G)k kW (t)k κ1 e−αv t1 + κ2 e−βt1 . (33)
Also,
Z t2
Z t2
(34)
ėi dt = kei (t2 )k.
kėi kdt ≥
t1
t1
We substitute kei (t2 )k = ce−βt2 as an event is triggered at
t2 . Integrating the inequality (33) with limits t1 and t2 and
using the fact that the edge weights are bounded such that
kW (t)k ≤ ω and (34) we get,
Z t2
kρωkD (G)k κ1 e−αv t1 + κ2 e−βt1 dt
kei (t2 )k ≤
t1
−βt2
≤ kρωkD (G)k κ1 e−αv t1 + κ2 e−βt1 γ
ce
ce−βγ ≤ kρωkD (G)k κ1 e−(αv −β)t1 + κ2 γ
Rearranging the above inequality we get,
c
γ eβγ ≥
(35)
kρωkD (G)k κ1 e−(αv −β)t1 + κ2
It is evident that γ > 0 as the RHS of (35) is strictly positive.
The minimum value of γ is a solution of the following equation,
c
(36)
γ eβγ =
kρωkD (G)k (κ1 + κ2 )
This proves that zeno behaviour does not occur in the dynanic
trigger case. When β = 0, the lower bound on gamma is the
RHS of equation (36), which rules out zeno behaviour in static
trigger case as well. The preceding arguments on ruling out
zeno behaviour are similar to those provided in Seyboth et al.
(2011) for time-invariant graphs.
5.1
(Graham et al., 1980, p. 29) it is possible to find N such that
in the interval (ti1 , tiN ), the same spanning-tree occurs in each
τ (1 + (j − 1)d, 1) for j = 1, 2, · · · k and some d > 0. This
allows us to select a persistence window T > (d + 1)tmax
where tmax = maxj τ (j, 1) for j = 1, 2, · · · N . With the
aforementioned selection of T , Theorem 2 can be invoked to
prove the convergence of the edge states xe to a ball around
origin in the case of static trigger and to origin in the case
of dynamic trigger. As we cannot predict which particular
spanning-tree repeats in each interval τ (1 + (i − 1)d, 1), the
least conservative bounds are chosen.
6
Simulations
A multi-agent system with four agents under switching graphs
was simulated using Matlab c for the static and dynamic trigger
cases. The underlying communication topology considered is
shown in figure 1 and the different spanning-trees that were
switched between are shown in figure 2.
1
g1
2
g4
g6
g2
g5
4
g3
3
Fig. 1. Underlying graph of arbitrary orientation
Consensus in Switching Topologies
Theorem 2 is not valid for switching topologies because our
spanning tree and thereby the edge states xe itself can be
changing. The following corollary extends the aforementioned
theorem to agents interacting via switching topologies.
Corollary 1. Let t1 , t2 , · · · be the infinite time sequence of
graph switching instants with, ti+1 − ti ≥ tL for some
positive tL , and i = 0, 1, · · · . Consider the agent dynamics
(16) corresponding to a single integrator system with eventtriggered control. If there exists an infinite sequence of contiguous, non-emptyand uniformly
bounded time intervals τ (j, 1)
Let τ (j, l) = tij , tij+l ; j = 1, 2, · · · starting at ti1 = t0 ,
with the property that the union of the undirected graphs across
each such interval has a spanning tree, then
(1) with static-trigger, the edge-state xe converges to a ball of
radius kxe k ≤ ρκM
2 .
(2) with dynamic-trigger, the edge-state xe converges exponentially to origin at rate greater than β.
Proof. The case of switching topologies differs from the setting of theorem 2 by the fact that the union of graphs in each
time interval τ (j, 1) contains a different spanning-tree compared to the occurrence of the same spanning tree in theorem
2. We show that Corollary 1 can be treated as an extension
of Theorem 2 using the results of Van der Waerden’s theorem ((Graham et al., 1980, p. 29)). The collection of possible spanning-tress forms a finite, non-empty set. Taking each
possible spanning-tree to be a different colour, by Theorem
Fig. 2. Spanning Trees Considered
The edge-weights were chosen as gi = square(4 ∗ t, 20 − (i −
1)0.1π) + 1). ∗ sin(5 ∗ t) for i = 1, 2 · · · 4 and g6 = 0, where
the function square(at, b) for a, b ∈ R generates
a square
wave
Ton
2π
of unit amplitude, period a and duty-cycle Toff +Ton b. The
aforementioned gi are defined to emulate real scenarios where
there might be instances when no edges are active. The initial
>
value of the states was taken to be x0 = [1 2 0.3 0.4] Plots 3
and 4 show the evolution of the system states x and the norm of
the edge states kxe k under the static trigger with c = 0.5 (refer
equation (14)). The evolution of the states with dynamic trigger
function with β = 0.06 and c = 0.5 (refer equation (15)) is
plotted in figures (5) and (6). The bounds were calculated using
inequalities presented in Corollary 1 and plotted along with the
norm of the edge states. These plots show the convergence of
the edge states xe to a ball around the origin in case of static
trigger and consensus of states x in the case of dynamic trigger.
3
12
x1
x2
x3
x4
kxe k
Envelope
Norm of edge states
10
States
2
1
8
6
4
2
0
0
10
20
30
0
40
Time[s]
0
10
20
30
40
50
Time[s]
Fig. 3. Evolution of states under static trigger with the graph switching
between spanning trees G1 and G2 in contiguous intervals
Fig. 6. Evolution of norm of edge-states under dynamic trigger with the graph
switching between spanning trees G1 and G2 in contiguous intervals
References
4
Norm of edge states
kxe k
Envelope
3
2
1
0
0
10
20
30
40
Time[s]
Fig. 4. Evolution of norm of edge-states under static trigger with the graph
switching between spanning trees G1 and G2 in contiguous intervals
3
x1
x2
x3
x4
States
2
1
0
0
10
20
30
40
50
Time[s]
Fig. 5. Evolution of states under dynamic trigger with the graph switching
between spanning trees G1 and G2 in contiguous intervals
7
Conclusions
The application of event-triggered control to classical consensus algorithms with time-varying, persistently exciting topologies guarantees consensus with dynamic trigger function. Under the more practically implementable static trigger function,
the edge-states converge to a ball around the origin. For switching topologies, we utilize the work of Chowdhury et al. (2016)
to show that we can extend the results of the persistent, continuously varying graphs to the case of switching topologies.
The convergence bounds thus obtained depend on the ’slowest’
spanning tree.
Chen, W. and Dai, H. (2016). Event-triggered consensus of time-varying
complex system. In 2016 Chinese Control and Decision Conference
(CCDC), 231–234. IEEE.
Chowdhury, N.R., Sukumar, S., and Balachandran, N. (2016). Persistence
based convergence rate analysis of consensus protocols for dynamic graph
networks. European Journal of Control, 29, 33–43.
Dimarogonas, D.V., Frazzoli, E., and Johansson, K.H. (2012). Distributed
event-triggered control for multi-agent systems. IEEE Transactions on
Automatic Control, 57(5), 1291–1297.
Graham, R.L., Rothschild, B.L., and Spencer, J.H. (1980). Ramsey theory,
volume 2. Wiley New York.
Jadbabaie, A., Lin, J., and Morse, A.S. (2003). Coordination of groups of
mobile autonomous agents using nearest neighbor rules. Automatic Control,
IEEE Transactions on, 48(6), 988–1001.
Martin, S. and Girard, A. (2013). Continuous-time consensus under persistent
connectivity and slow divergence of reciprocal interaction weights. SIAM
Journal on Control and Optimization, 51(3), 2568–2584.
Mesbahi, M. and Egerstedt, M. (2010). Graph theoretic methods in multiagent
networks. Princeton University Press.
Olfati-Saber, R. and Shamma, J.S. (2005). Consensus filters for sensor networks and distributed sensor fusion. In 44th IEEE Conference on, Decision
and Control, ECC. CDC-ECC’05. 2005, 6698–6703. IEEE.
Ren, W. and Atkins, E. (2005). Second-order consensus protocols in multiple
vehicle systems with local interactions. In AIAA Guidance, Navigation, and
Control Conference and Exhibit, 15–18.
Ren, W. and Beard, R.W. (2005). Consensus seeking in multiagent systems
under dynamically changing interaction topologies. IEEE Transactions on,
Automatic Control, 50(5), 655–661.
Sandhu, J., Mesbahi, M., and Tsukamaki, T. (2005). Cuts and cycles in relative
sensing and control of spatially distributed systems. In Proceedings of the
American Control Conference, volume 1, 73.
Sastry, S. and Bodson, M. (2011). Adaptive control: stability, convergence and
robustness. Courier Dover Publications.
Seyboth, G.S., Dimarogonas, D.V., and Johansson, K.H. (2011). Control of
multi-agent systems via event-based communication. IFAC Proceedings
Volumes, 44(1), 10086–10091.
Vicsek, T., Czirók, A., Ben-Jacob, E., Cohen, I., and Shochet, O. (1995). Novel
type of phase transition in a system of self-driven particles. Physical Review
Letters, 75(6), 1226–1229.
Yu, P., Ding, L., Liu, Z.W., and Guan, Z.H. (2015). A distributed eventtriggered transmission strategy for exponential consensus of general linear
multi-agent systems with directed topology. Journal of the Franklin Institute, 352(12), 5866–5881.
Zelazo, D. and Burger, M. (2014). On the definiteness of the weighted laplacian
and its connection to effective resistance. In Decision and Control (CDC),
2014 IEEE 53rd Annual Conference on, 2895–2900. IEEE.
Zelazo, D. and Mesbahi, M. (2011). Edge agreement: Graph-theoretic performance bounds and passivity analysis. IEEE TAC, 56(3), 544–555.
Zhu, W., Jiang, Z.P., and Feng, G. (2014). Event-based consensus of multiagent systems with general linear models. Automatica, 50(2), 552–558.
| 3cs.SY
|
On the Formal Semantics of the Cognitive
Middleware AWDRAT∗
arXiv:1412.3588v2 [cs.PL] 14 Dec 2014
Muhammad Taimoor Khan† , Dimitrios Serpanos† and Howard
Shrobe*
†
{mtkhan, dserpanos}@qf.org.qa
*
[email protected]
†
QCRI, Qatar
*
CSAIL, MIT, USA
March 23, 2018
Abstract
The purpose of this work is two fold: on one hand we want to formalize the behavior of critical components of the self generating and
adapting cognitive middleware AWDRAT such that the formalism not
only helps to understand the semantics and technical details of the
middleware but also opens an opportunity to extend the middleware
to support other complex application domains of cybersecurity; on the
other hand, the formalism serves as a pre-requisite for our proof of the
behavioral correctness of the critical components to ensure the safety
of the middleware itself. However, here we focus only on the core and
critical component of the middleware, i.e. Execution Monitor which
is a part of the module “Architectural Differencer” of AWDRAT. The
role of the execution monitor is to identify inconsistencies between runtime observations of the target system and predictions of the System
Architectural Model. Therefore, to achieve this goal, we first define the
formal (denotational) semantics of the observations (runtime events)
and predictions (executable specifications as of System Architectural
Model); then based on the aforementioned formal semantices, we formalize the behavior of the “Execution Monitor” of the middleware.
∗
This is the draft version of the subject technical report.
1
Contents
1 Calculus of AWDRAT
4
2 Syntax of System Architectural Model
5
3 Semantics of System Architectural Model
3.1 Semantic Algebras . . . . . . . . . . . . .
3.1.1 Truth Values . . . . . . . . . . . .
3.1.2 Numeral Values . . . . . . . . . . .
3.1.3 Environment Values . . . . . . . .
3.1.4 State Values . . . . . . . . . . . .
3.1.5 Semantic Values . . . . . . . . . .
3.1.6 Character String Values . . . . . .
3.1.7 Lifted Values . . . . . . . . . . . .
3.1.8 (Registered) Event Values . . . . .
3.1.9 (Observed) Event Values . . . . .
3.1.10 (Runtime) Event Values . . . . . .
3.1.11 Resource Values . . . . . . . . . .
3.1.12 Function Values . . . . . . . . . .
3.1.13 Component Values . . . . . . . . .
3.1.14 Split Values . . . . . . . . . . . . .
3.1.15 Attack Values . . . . . . . . . . . .
3.2 Signatures of Valuation Functions . . . . .
3.2.1 System Architectural Model . . . .
3.2.2 Behavioral Models . . . . . . . . .
3.3 Auxiliary Predicates and Functions . . . .
3.4 Definition of Valuation Functions . . . . .
3.4.1 System Architectural Model . . . .
3.4.2 Register Model . . . . . . . . . . .
3.4.3 Register Model Sequence . . . . .
3.4.4 Structural Model . . . . . . . . . .
3.4.5 Structural Model Sequence . . . .
3.4.6 Behavioral Model . . . . . . . . . .
3.4.7 Behavioral Model Sequence . . . .
3.4.8 Split Model . . . . . . . . . . . . .
3.4.9 Split Model Sequence . . . . . . .
3.4.10 Attack Model . . . . . . . . . . . .
3.4.11 Attack Model Sequence . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
12
12
12
12
12
14
15
15
15
15
16
16
16
17
17
18
18
18
18
18
19
22
22
23
24
24
29
30
31
32
33
33
35
4 Execution Monitor
35
4.1 Observation Model . . . . . . . . . . . . . . . . . . . . . . . . 36
4.1.1 Observations . . . . . . . . . . . . . . . . . . . . . . . 37
2
5 Semantics of the Execution Monitor
37
6 Conclusions and Future Work
39
Appendices
42
A Formal Syntax of System Architectural Model
42
A.1 Declaration of Syntactic Domains . . . . . . . . . . . . . . . . 42
A.2 Grammar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
B An Example of a System Architectural Model
49
B.1 MAF Editor Model . . . . . . . . . . . . . . . . . . . . . . . . 50
3
1
Calculus of AWDRAT
Defending systems against cyber attack requires us to be able to rapidly and
accurately detect that an attack has occurred. Today’s detection systems
are woefully inadequate suffering from both high false positive and false
negative rates. There are two key reasons for this: First, the systems do not
understand the complete behavior of the system they are protecting. The
second is that they do not understand what an attacker is trying to achieve.
Most such systems, in fact, are retrospective, that is they understand some
surface signatures of previous attacks and attempt to recognize the same
signature in current traffic. Furthermore, they are passive in character,
they sit back and wait for something similar to what has already happened
to recur. Attackers, of course, respond by varying their attacks so as to
avoid detection.
AWDRAT [5] is a representative of a new class of protection systems
that employ a different, active form of perception, one that is informed
both by knowledge of what the protected application is trying to do and
by knowledge of how attackers think. It employs both bottom-up reasoning
(going from sensors data to conclusions about what attacks might be in
progress) as well as top-down reasoning (given a set of hypotheses about
what attacks might be in progress, in focuses its attention to those events
most likely to significantly help in discerning the ground truth).
There are two dimensions along which detection systems can be characterized. The first is the distinction between profile and model based
approaches. The other dimension is the distinction between looking for
matches to bad behavior or deviations from good. This gives four quadrants, each with unique strengths and weaknesses. For example, the bulk of
our sensors are model-based and look for matches to bad behavior; signature
based systems are in this category. The advantage is that when a match occurs, you know what has happened; i.e. these systems have high diagnostic
resolution. But they also lack robustness; if they don’t have a model of an
attack, and there are always novel attacks, then they will fail to detect it.
On the other hand, there are a class of detectors that use employ machine
learning techniques on labeled training data to build statistical profiles of
attacks. These systems tend to be a bit more robust than model based
systems, since the machine learning techniques tend to generalize from the
data presented. However, they make up for this by a loss of diagnostic resolution. The third quadrant involves building a statistical profile of normal
behavior, detecting deviations from the profile. Such anomaly detectors are
yet more robust, since they don’t depend on prior knowledge of the form of
the attack, but they afford even less diagnostic resolution. When things go
wrong, all you know is that something out of the ordinary has happened;
whether that something is malicious or not isn’t known.
AWDRAT sits in the fourth quadrant: It has a model of normal be4
havior; when the application deviates from the behavior prescribed by that
model, it employs diagnostic reasoning techniques [6] to further isolate and
characterize the failure. It has both greater robustness and higher diagnostic
resolution. But it achieves this only through the construction of a far more
complex model.
AWDRAT has an active model of normal behavior, namely an executable
specification (aka System Architectural Model) of the application [5]. This
executable specification consists of a decomposition into sub-modules and
pre- and post-conditions for each sub-module. In addition, data-flow and
control-flow links connect the sub-modules, specifying the expected flow of
values and of control. The pre- and post-conditions are arbitrary first-order
statements about the set of data values that flow into and out of the submodules.
AWDRAT runs this executable specification in parallel with the actual
application code, comparing their results at the granularity and abstraction
level of the executable specification. (This is therefore a special case of
the standard fault tolerance technique of running multiple versions of the
same code and comparing their results.) The executable specification is
hierarchical, allowing flexibility in the granularity of the monitoring. When
threats are not expected, the executable specification is run at a high level of
abstraction, incurring less overhead, but requiring more diagnostic reasoning
should the program diverge from the prescribed behavior of the executable
specification. In times of heightened threat, the executable specification can
be elaborated to a greater degree, incurring more overhead, but providing
more containment.
Optionally, the model can also include models for suspected incorrect
behaviors of a component, allowing the diagnostic reasoning to characterize
the way in which a component might have misbehaved. A diagnosis is then
a selection of behavioral modes for each component of the specification such
that the specification predicts the observed misbehavior of the system.
The rest of the paper is organized as follows: in Section 2 we discuss
syntax of the System Architectural Model followed by the formalization of
semantics of the critical syntactic domains model in Section 3. Section 5
formalizes the semantics of the execution monitor. Finally, we conclude in
Section 6. Appendices A and B give the formal syntactic grammar and an
example System Architectural Model respectively.
2
Syntax of System Architectural Model
An AWDRAT model is built from several related following forms which
represent corresponding high-level syntactical domains of the model. Note,
we only discuss selected domains here, for complete syntactic domains and
their elements, please see Appendix A.
5
1. A description of a component type consists of
(a) its interface
• a list of inputs
• a list of its outputs
• a list of the resources it uses (e.g. files it reads, the code in
memory that represents this component, etc)
• list of subcomponents required for the execution of the subject component
• a list of events that represent entry into the component (usually just one)
• a list of events that represent exit from the component (usually just one)
• a list of events that are allowed to occur during any execution
of this component
• a set of conditional probabilities between the possible modes
of the resources and the possible modes of the whole component
• a list of known vulnerabilities occurred to the component
(b) and a structural model which is a list of sub-components some of
which might be splits or joins of
• data-flows between linking ports of the sub-components (outputs of one to inputs of another)
• control-flow links between cases of a branch and a component
that will be enabled if that branch is taken
The description of the component type is represented by syntactical
domain “StrMod” which is defined as follows:
StrMod ::= define-ensemble CompName
:entry-events :auto | (EvntSeq)
:exit-events (EvntSeq)
:allowable-events (EvntSeq)
:inputs (ObjNameSeq)
:outputs (ObjNameSeq)
:components (CompSeq)
:controlflows (CtrlFlowSeq)
:splits (SpltCFSeq)
:joins (JoinCFSeq)
:dataflows (DataFlowSeq)
:resources (ResSeq)
:resource-mapping (ResMapSeq)
6
:model-mappings (ModMapSeq)
:vulnerabilities (VulnrabltySeq)
Example 1: The specification of the component maf-editor is given
below. In detail, the specification says that the component
• is top level component and hence starts automatically and
thus requires no entry-event,
• requires no inputs
• results in the-model as an output
• has four subcomponents, i.e. startup, create-model, create-events
and save which have corresponding types and also have both
normal and compromised behaviors
• has control and data flows as described
• has an access to two resources, i.e. imagery and code-files
which have corresponding probabilities of being in a normal
and hacked mode
• has model mappings of the above resources to the subcomponents as described in model-mappings and
• has two vulnerabilities, i.e. reads-complex-imagery and
loads-code for the resources imagery and code-files respectively.
(define-ensemble maf-editor
:entry-events :auto
:inputs ()
:outputs (the-model)
:components
((startup :type maf-startup :models (normal compromised))
(create-model :type maf-create-model :models (normal compromised))
(create-events :type maf-create-events :models (normal compromised))
(save :type maf-save :models (normal compromised)))
:controlflows ((before maf-editor before startup)
(after startup before create-model))
:dataflows ((the-model create-model the-model create-events)
(the-model create-events the-model save)
(the-model save the-model maf-save-model))
:resources ((imagery image-file (normal .7) (hacked .3))
(code-files loadable-files (normal .8) (hacked .2)))
:resource-mappings ((startup imagery)
7
(create-model code-files)
(create-events code-files)
(save-model code-files))
:model-mappings ((startup normal ((imagery normal)) .99)
(startup compromised ((imagery normal)) .01)
(startup normal ((imagery hacked)) .9)
(startup compromised ((imagery hacked)) .1)
(create-model
(create-model
(create-model
(create-model
normal ((code-files normal)) .99)
compromised ((code-files normal)) .01)
normal ((code-files hacked)) .9)
compromised ((code-files hacked)) .1)
(create-events normal ((code-files normal)) .99)
(create-events compromised ((code-files normal)) .01)
(create-events normal ((code-files hacked)) .9)
(create-events compromised ((code-files hacked)) .1)
(save
(save
(save
(save
normal ((code-files normal)) .99)
compromised ((code-files normal)) .001)
normal ((code-files hacked)) .01)
compromised ((code-files hacked)) .999))
:vulnerabilities ((imagery reads-complex-imagery)
(code-files loads-code)
))
2. Behavioral specification of a component (a component type may have
one normal behavioral specification and many abnormal behavioral
specifications, each one representing some failure mode) which has
• inputs and outputs
• preconditions on the inputs (logical expressions involving one or
more of the inputs)
• postconditions (logical expressions involving one or more of the
outputs and the inputs)
• allowable events during the execution in this mode
The behavioral specification of a component is represented by a corresponding syntactical domain “BehMod” as follows:
BehMod ::= defbehavior-model (CompName normal | compromised)
8
:inputs (ObjNameSeq)
:outputs (ObjNameSeq)
:allowable-events (EvntSeq)
:prerequisites (BehCondSeq)
:post-conditions (BehCondSeq)
Example 2: In the following first we give the structure of a component maf-create-model (which is one of the submodule as
stated in the previous specification example) and then give the
behavioral specification of the component. The structure of the
component is defined as follows:
(define-ensemble maf-create-model
:entry-events (create-mission-action-action-performed)
:exit-events (mission-builder-submit)
:allowable-events (create-mission-builder-with-client-panel
create-mission-builder
create-mission-builder-with-hash-table
mission-builder-submit
(set-initial-info exit (the-model nil))
create-mission-action-action-performed
retrieve-info
create-mission-action-action-performed
(set-initial-info entry)
)
:inputs ()
:outputs (the-model))
In the following we define the legal and illegal (compromised)
behaviors of the component. For example, the specification of a
legal (normal) behavior of the component says that as a normal
behavior the component
•
•
•
•
requires no input as specified by the clause inputs
has the-model output and also
no prerequisite of the component but
guarantees that the object mission-builder of the-model
are consistent.
The corresponding normal behavior is defined as:
(defbehavior-model (maf-create-model normal)
:inputs ()
:outputs (the-model)
:prerequisites ()
9
:post-conditions ([dscs ?the-model mission-builder good])
)
(defbehavior-model (maf-create-model compromised)
:inputs ()
:outputs (the-model)
:prerequisites ()
:post-conditions ([not [dscs ?the-model mission-builder good]])
)
Similarly, the compromised behavior of the component is also described above. For further details on the behavioral specification
of the other components, please see Appendix A.
3. Model of a resource type contains
• possible modes
• prior probabilities of being in each mode
• attack types to which it is vulnerable
The syntactical domain “ResModMap” represents the model of a resource type
ResModMap ::= ResName normal | hacked FVal
| ((ResName normal | hacked)) FVal
where “FVal” represents the float values for probabilities.
The trust model of the resources is specified in example 1 above by
the clauses :resources and :resource-mappings.
4. Attack Model
• a list of types of attacks that are being anticipated and the prior
probability of each
• a list describing how each attack type can effect that mode of a
resource
• a set of logical rules expressing the conditional probabilities between attack types and resource modes
The attack models are presented by the syntactic domain “AtkMod”
while the corresponding attack rules are specified by the syntactic
domain “AtkRule” as given below respectively:
10
AtkMod ::= define-attack-model AtkModName
:attack-types (AtkTypeSeq)
:vulnerability-mapping (AtkVulnrabltyMapSeq)
AtkRule ::= defrule AtkRulName (:forward)
if AtkCondSeq
then AtkConsSeq
Example 3: The example attack model maf-attacks specifies the
two attacks hacked-image-file-attack and hacked-code-file-attack
with some probabilities as specified in the following.
(define-attack-model maf-attacks
:attack-types ( (hacked-image-file-attack .3)
(hacked-code-file-attack .5))
:vulnerability-mapping
((reads-complex-imagery hacked-image-file-attack)
(loads-code hacked-code-file-attack)))
Furthermore, the two attacks are mapped to the corresponding
vulnerabilities reads-complex-imagery and loads-code respectively.
Additionally, the corresponding one attack rule bad-image-file-takeover
says that if we have the contextual resource ?ensemble and
type-of-resource is image-file and the resource-might-have-been-attacked
with hacked-image-file-attack then it is highly probable (.9)
that the resource has been hacked by hacked-image-file-attack
as given below:
(defrule bad-image-file-takeover (:forward)
if [and [resource ?ensemble ?resource-name ?resource]
[resource-type-of ?resource image-file]
[resource-might-have-been-attacked ?resource
hacked-image-file-attack]]
then [and [attack-implies-compromised-mode
hacked-image-file-attack ?resource hacked .9 ]
[attack-implies-compromised-mode
hacked-image-file-attack ?resource normal .1 ]])
Further details on the example of the model, please see Appendix B.
However, the corresponding syntactic details of the elements of the above
syntactic domains are explained in the corresponding subsections of the
11
next section. However, for the general syntax of the domain, please see
Appendix A.
3
Semantics of System Architectural Model
In this section, we first give the definition of semantic algebras and then
discuss informal description and the formal semantics of the core constructs
of the System Architectural Model.
3.1
Semantic Algebras
The definition of a formal denotational semantics is based on a collection
of data structures. Semantic domains represent set of elements that share
some common properties. A semantic domain is accompanied by a set of operations as functions over the domain. A domain and its operations together
form a semantic algebra [4]. In the following we enlist the semantic domains
and their corresponding operations. Some operations are defined and some
are just declared for the purpose of completeness of this document.
3.1.1
Truth Values
The truth values are represented by the semantic domain “Bool” which is
defined as follows:
Domain: Bool
Operations:
• true: Bool
• false: Bool
• and: Bool × Bool → Bool
• or: Bool × Bool → Bool
• not: Bool × Bool → Bool
3.1.2
Numeral Values
Here we consider typical domains to represent integer and float values (e.g.
Q, N).
3.1.3
Environment Values
The domain Environment holds the environment values of the System Architectural Model. Environment is formalized as a tuple of domains Context
12
and Space. The domain Context is a mapping of identifiers to the environment values (Variable, Component, Resource, RTEvent and Function), while
the domain Space models the memory space.
Domain: Environment
Environment := Context × Space
Context := Identifier → EnvValue
EnvValue := Variable + Component + Resource + RTEvent
+ Function + AtkModel
Space := P(Variable)
Variable := n, where n ∈ N represents locations
Operations:
• space: Environment → Space
space(<c,s>) = s
• context: Environment → Context
context(<c,s>) = c
• environment: Context × Space → Environment
environment(c,s) = <c, s>
• take: Space → Identifier × Space
take(s) = LET x = SUCH x: x ∈ s IN <x, s\{x}>
• push: Environment × Identifier → Environment
push(e, I) = LET <x, s’>= take(space(e)) IN
environment(context(e)[I 7→ inVariable(x)], s’)
• push: Environment × Identifier × Component → Environment
push(e, I, c) = LET <x, s’>= take(space(e)) IN
environment(context(e)[I 7→ inComponent(c)], s’)
• push: Environment × Identifier × AtkModel → Environment
push(e, I, m) = LET <x, s’>= take(space(e)) IN
environment(context(e)[I 7→ inAtkModel(m)], s’)
13
3.1.4
State Values
This section defines the domain for the State of the execution of program.
A Store is the most important part of the state and holds for every Variable
a Value. The value can be read and modified. The Data of the state is a
tuple of a Flag that represents the current status of the state and a Mode to
represent the current mode of execution of the state respectively component.
Domain: State
State := Store × Data
Store := Variable → Value
Data := Flag × Mode
Flag := {running, ready, completed}
Mode := {normal, compromised}
Operations:
• state: Store × Flag → State
state(s,f) = <s,f>
• store: State → Store
store(<s,f>) = s
• data: State → Data
data(<s,d>) = d
• flag: Data → Flag
flag(<f,m>) = f
• mode: Data → Mode
mode(<f,m>) = m
• setFlag: State × Flag → State
setFlag(s, f) = LET d = <f, mode(data(s))>IN <s, d>
• setMode: State × Mode → State
setMode(s, m) = LET d = <flag(data(s)), m>IN <s, d>
• eqFlag: State × Flag → Bool
eqFlag(s, f) = IF equals(flag(data(s)), f) THEN true ELSE false
END
• eqMode: State × Mode → Bool
eqMode(s, m) = IF equals(mode(data(s)), m) THEN true ELSE false
END
• update: State × Variable × Value → State
update(s, var, val) = state(store(s)[var 7→ val], flag(s))
14
3.1.5
Semantic Values
Value is a disjunctive union domain and note that the domain Value is a
recursive domain, e.g. List is defined by Value* as discussed in the next
section.
Domain: Value
Value := Event + ObsEvent + RTEvent + Function + Component + Split +
Resource + AtkModel + String + List + ... + Value∗
Operations:
• equals: Value × Value → Bool
3.1.6
Character String Values
Character strings are defined as a semantic domain String.
3.1.7
Lifted Values
The evaluation of some semantic domains might result as unsafe. To address
these unsafe evaluations we lifted the domains of State and Value to domains
State⊥ and Value⊥ , which are disjoint sums of the basic domains and the
domain ⊥.
In order to capture different kind of events we need different semantic
domain to model each of them. The three kind of events are:
1. Registered Events are the events of interest for monitoring. These
events are defined by the user at the top of the System Architectural Model by the syntactic domain “RegModSeq” as discussed in
Appendix A. AWDRAT register these events for the monitoring purposes.
2. Observed Events are the entry, exit and allowable events as defined by
the syntactic domain “Event” of the System Architectural Model.
3. Run Time Events are the runtime events that are generated by the
monitor from the target system. These events are also called observations.
In the following, we give definitions of the corresponding semantic domains respectively.
3.1.8
(Registered) Event Values
The semantics domain Event defines the registered events as a predicate
over a sequence of input values, sequence of output values, a pre-state and
a corresponding post-state as follows:
Event := P(Value∗ × Value∗ × State × State⊥ )
15
3.1.9
(Observed) Event Values
The semantics domain ObsEvent formalizes the observed events of System
Architectural Model. An ObsEvent is defined as a predicate over a sequence
of input values, a pre-state and a post-state as follows:
ObsEvent := P(Value∗ × State × State⊥ )
Note that the observed events do not capture output values because they
just work as placeholders for runtime and registered events.
3.1.10
(Runtime) Event Values
The runtime events of System Architectural Model are formalized with the
help of a semantic domain RTEvent. The semantic domain RTEvent is
defined as a predicate over a sequence of input values, sequence of output
values, a pre-state, post-state and event data as follows::
RTEvent := P(Value∗ ⊥ × Value∗ × State × State⊥ × EventData)
where
EventData := Tag × TimeStamp × ProcessID
Tag := {entry, exit}
TimeStamp := date and time of the event execution
ProcessID := operating system process id for the event
An EventData captures the type of an event, time of event generation
and an operating system level process id for this event. Note, process identification provides more low-level information about the event which is helpful
to detect any misbehavior of the event correspondingly component.
3.1.11
Resource Values
The semantic domain Resource is one of the complex domains because semantically this domain depends on the runtime behavior of an associated
components as well. The semantics domain Resource formalizes different
kind of resources used by computational modules of System Architectural
Model and is also defined as a predicate over a
• map which is further a predicate over
– a mode,
– its likelihood value being in normal mode,
– corresponding likelihood being in hacked mode and
– an associated vulnerability,
• current mode of the resource,
16
• probability of the resource being in the current mode,
• name of the running component associated with the resource,
• mode of the running associated component,
• a pre-state and a post-state of the program.
The predicate Resource is mathematically defined as:
Resource := P(ModeMap × Mode × FVal × I × Mode × State × State⊥ )
where
Mode := {normal, compromised}
ModeMap := P(Mode × Fval × FVal × Vulnerability)
3.1.12
Function Values
The semantics domain Function defines and formalizes a specification function of System Architectural Model and can be defined mathematically as:
S
F unction = n∈N F unctionn
where
F unctionn = V aluen → V alue
3.1.13
Component Values
The semantics domain Component formalizes the model of the components
of the target system which are specified by the corresponding behaviors in
the System Architectural Model. A Component is defined as a predicate over
a structural behavior of the component, a normal behavior of the component,
its corresponding compromised behavior, a pre-state and a post-state of the
program as follows:
Component = P(SBehavior × NBehavior × CBehavior × State × State⊥ )
where
SBehavior := P(Value∗ × Value∗ × Value∗ × State × State⊥ )
NBehavior = CBehavior := P(Value∗ × Value∗ × State × State⊥ )
Furthermore, a structural behavior is defined as a predicate over a sequence
of input values, sequence of output value, sequence of allowable values (as
a consequence of allowable events), a pre-state and a post-state of the behavior. Also, a normal (functional) behavior and corresponding compromised behavior are defined as a predicates NBehavior and CBehavior over
a sequence of input values, sequence of output values, a pre-state and a
corresponding post-state respectively. Note that the two predicates are the
valuation functions of corresponding syntactic domains.
17
3.1.14
Split Values
The semantics domain Split formalizes the control flow behavior of a certain
unit of a computational module of System Architectural Model and is defined
as a predicate over a sequence of parameter values of the split as follows:
Split := P(Value∗ )
3.1.15
Attack Values
The semantics domain AtkModel formalizes the attack model and is defined
as a predicate over an attack name, probability of the attack and the corresponding vulnerability causing the attack; the attack model is formulated
as follows:
AtkModel := P(Identifier × FVal × Vulnerability)
These values are the result of the valuation function for the corresponding
syntactic domain.
3.2
Signatures of Valuation Functions
A valuation function defines a mapping of a language’s abstract syntax
structures to its corresponding meanings (semantic algebras) [4]. A valuation function VF for a syntax domain VF is usually formalized by a set of
equations, one per alternative in the corresponding BNF Register for each
syntactic domain of specification expression.
We define the result of valuation function as a predicate. In this section
we first give the definitions of various relations and functions that are used
in the definition of valuation functions. For example the behavioral relation
(BehRelation) is defined as a predicate over an environment, a pre-state and
a post-state. The corresponding relation is defined as follows:
BehRelation := P(Environment × State × State⊥ )
3.2.1
System Architectural Model
The valuation function for the abstract syntax domain system architectural
model values of SAM is defined as follows:
〚SAM〛: Environment → BehRelation
3.2.2
Behavioral Models
The valuation functions for abstract syntax domains of Register, structural,
behavioral and split model values (RuleMod, StrMod, BehMod and SpltMod
respectively) are the same and can be defined similarly; however in the
following we give only the signature of valuation function for the behavioral
model:
18
〚BehMod〛: Environment → BehRelation
In the following section we define the auxiliary functions and predicates
used in the formal semantics of the specification language (and associated
domains).
3.3
Auxiliary Predicates and Functions
In the following subsections auxiliary functions and predicates for the use in
semantics definition of sequence, binding and special expressions are defined.
• monitors ⊂ N × RTEvent × Component × Environment∗ × Environment∗
× State∗ × State∗⊥
monitors(i, 〚rte〛, 〚c〛, e, e’, s, s’) ⇔
( eqMode(s(i), “running”) ∨ eqMode(s(i), “ready”) ) ∧ 〚c〛(e(i))(e’(i), s(i), s’(i)) ∧
∃ oe ∈ ObEvent: equals(rte, store(〚name(rte)〛)(e(i))) ∧
IF entryEvent(oe, c) THEN
data(c, s(i), s’(i)) ∧
( preconditions(c, e(i), e’(i), s(i), s’(i), “compromised”) ⇒
equals(s(i+1), s(i)) ∧ equals(s’(i+1), s(i+1)) ∧
setFlag(inState(s’(i+1)), “compromised”) ) ∨
( preconditions(c, e(i), e’(i), s(i), s’(i), “normal”)
⇒ setMode(s(i), “running”) ∧
LET cseq = components(c) IN
equals(s(i+1), s’(i)) ∧ equals(e(i+1), e’(i)) ∧
∀ c1 ∈ cseq, rte1 ∈ RTEvent:
arrives(rte1 , s(i+1)) ∧
monitor(i+1, rte1 , c1 , e(i+1), e’(i+1), s(i+1), s’(i+1))
END )
ELSE IF exitEvent(oe, c) THEN
data(c, s(i), s’(i)) ∧ eqMode(inState(s’(i)), “completed”) ∧
postconditions(c, e(i), e’(i), s(i), s’(i), “normal”)
⇒ equals(s(i+1), s’(i)) ∧ equals(e(i+1), e’(i)) ∧
setMode(inState(s’(i+1), “completed”)
ELSE IF allowableEvent(oe, c) THEN
equals(s(i+1), s’(i)) ∧ equals(e(i+1), e’(i))
ELSE
equals(s(i+1), s(i)) ∧ equals(s’(i+1), s(i+1)) ∧
setFlag(inState(s’(i+1)), “compromised”)
END
The predicate “monitors” captures the core semantics of the monitor
which is defined as a relation on
– number of observation i with respect to iteration of a component,
– an observation (runtime event) rte,
19
– corresponding component c under observation,
– a sequence of pre-environments e,
– a sequence of post-environments e′ ,
– a sequence of pre-states s and
– a sequence of post-states s′ .
The predicate monitors is defined such that, at any arbitrary observation if the current execution state s(i) of component c is “ready” or
“running” and behavior of component c has been evaluated and there
is a prediction oe which is semantically equal to an observation rte
and any of the following can happen:
– either the prediction respectively observation is an entry event of
the component c, then it waits until the complete data for the
component c arrives, if so, then
∗ either preconditions of “normal” behavior of the component
hold; if so then, the subnetwork of the component is initiated and the components in the subnetwork are monitored
iteratively with the corresponding arrival of the observation
∗ or preconditions of “compromised” behavior of the component hold, in this case the state is marked to “compromised”
and returns
– or the observation is an exit event and after the completion of
data arrival the postconditions hold and the resulting state is
marked as “completed”
– or the observation is an allowable event and just continues the
execution
– or the observation is an unexpected event (or any of the above
does not hold), then the state is marked as “compromised” and
returns.
The predicate monitors is used later in the semantics of the Execution
Monitor.
• entryEvent ⊂ ObEvent × Component: returns true only if the
given event is in a set of entry events of the given component.
• exitEvent ⊂ ObEvent × Component: returns true only if the
given event is in a set of exit events of the given component.
• allowableEvent ⊂ ObEvent × Component: returns true only if
the given event is in a set of allowable events of the given component.
20
• data ⊂ Component × State × State⊥ : returns true only if all
the data for the given component is received by transforming a given
pre-state (former) into a corresponding given post-state (latter).
• arrives ⊂ RTEvent × State: returns true only if the given runtime
event (observation) arrives in a given state.
• preconditions ⊂ Component × Environment × Environment
× State × State⊥ × Mode
returns true only if all the preconditions of the given component hold
in a given pair of pre- and post-environments, a pair of pre- and poststates and in a given mode.
• postconditions ⊂ Component × Environment × Environment
× State × State⊥ × Mode
returns true only if all the postconditions of the given component
hold in a given pair of pre- and post-environments, a pair of pre- and
post-states and in a given mode.
• startup ⊂ State × Target System: returns true only if the given
state is an initial state of the execution of the given target system.
• isTop ⊂ Component × (Environment → BehRelation): returns
true only if the given component is a top level component of the given
semantics of a System Architectural Model.
• enableDiagnosis: Environment → P(State × Value): results in
a given recovered state and a boolean value (true, if recovered safely,
false otherwise) from a given environment.
• respectsOrder ⊂ Identifier Sequence × Identifier Sequence
returns true only if the identifiers in the latter sequence has the same
order as the identifiers in the former sequence.
• buildEnv ⊂ Environment × List* × List* × List* × List*
→ Environment
builds the resulting environment by updating the given environment
with given sequences of values of
– resources
– resource mappings
– model mappings and
– vulnerabilities
respectively.
21
3.4
Definition of Valuation Functions
In this section we give the definition of the formal semantics of the interesting
syntactic domains (and associated domains) of the specification language,
e.g. system architectural model, Register mode, behavioral model and split
model. The semantics of other domains of the specification language are
very simple and can be easily rehearsed.
3.4.1
System Architectural Model
The System Architectural Model is give by the syntactic domain SAM such
that
• the Register model (syntactic domain RegModSeq) is defined at the top
of the System Architectural Model which gives the registered events
to be monitored at runtime
• followed by
– a hierarchical structural behavior (syntactic domain StrModSeq)
of components,
– a normal respectively compromised behavior (syntactic domain
BehModSeq) and
– corresponding split behaviors (syntactic domain SplModSeq) occurring in any of the structural behavior of the components.
For further details on the syntax of the model, please see Appendix A.
Semantically, an overall (system architectural) model holds (true) in
a given environment e such that it produces a new environment e′ and a
post-state e′ when executed in a pre-state s as defined below.
〚SAM〛(e)(e’, s, s’) ⇔
∀ e1 , e2 , e3 ∈ Environment, s1 , s2 , s3 ∈ State:
〚RegModSeq〛(e)(e1 , s, inState⊥ (s1 )) ∧ 〚StrModSeq〛(e1 )(e2 , s1 , inState⊥ (s2 )) ∧
〚BehModSeq〛(e2 ) (e3 , s2 , inState⊥ (s3 )) ∧ 〚SpltModSeq〛(e3 )(e’, s3 , s’)
In detail, the semantics of the System Architectural Model SAM holds in
a given environment e resulting in an environment e′ by transforming a
pre-state s into post-state s′ and
• the evaluation of the registered events in a given environment e results
in environment e1 transforming a pre-state s into a post-state s1 and
in principle
– the structural behavior of components hold in environment e1
(with some auxiliary transformations) and
22
– the functional behavior of components hold in environment e2
(with some auxiliary transformations) and finally
– the split behavior of components hold in e3 resulting in given
environment s′ and transforming a pre-state s3 into a given poststate s′ .
In the following, first we define the semantics of unit elements RegMod,
StrMod, BehMod and SplModSeq and then define corresponding sequence
domains RegModSeq, StrModSeq, BehModSeq and SplModSeq respectively.
3.4.2
Register Model
The syntactic domain (RegMod) defines a registered event as follows:
RegMod ::= register-event ’EvntName JavClaName JavMetName ’(JavParamSeq)
[ :static ObjName ]
[ :output-type JavParam ]
[ :bypass ObjNameStr ]
[ :EvntName ObjName]
Though the domain represents language independent event registration, in
this document we focus only on the the Java based target system. The
syntactic phrase RegMod states that a registered event can be represented
by a name (EvntName) whose source is a Java method (JavMetName) with
parameters (JavParmSeq) of corresponding class (JavClaName). The other
sub-clauses introduce further characterization of the method, e.g. the clause
:output-type represents the return type of the method.
A monitoring machinery of Architectural Differencer of the middleware
AWDRAT is based on these registered events.
In the following we define the semantics of a registered event such that
the evaluation of a registered event in a given environment e results in an
environment e′ transforming a pre-state s into a post-state s′ .
〚RegMod〛(e)(e’, s, s’) ⇔
∀ ej ∈ JTypeEnvironment, sj , sj ’ ∈ JState:
typeCheck(JavClaName)(ej )(sj , s′j ) ∧ equals(s, sj ) ∧ equals(e, ej )
∧ (∃ p ∈ JProcedure: equals(p(valseq, val), store(sj ’)(〚JavMetName〛(ej )))
∧ equals(valseq, store(sj ’)(〚JavParamSeq〛(ej )))
∧ equals(val, store(sj ’)(〚JavParam〛(ej ))))
∧ isStatic(...) ∧ byPass(...) ∧ otherEvents(...)
∧ e’ = push(e, EvntName)
∧ LET ev(valseq, val, s, s’) ∈ Event IN
s’ = update(s, 〚EvntName〛(e’), ev)
END
23
In detail, semantically, the Java class (JavClaName) is well-defined respectively type checked in an arbitrary environment ej transforming an arbitrary
pre-state sj into a corresponding arbitrary post-state sj ’ while ej and sj are
semantically equivalent to s and e respectively and
• there is some Java procedure p(valseq, val) which we get by evaluating
“JavMetName” with given environment ej such that the sequence of
input values valseq equals the evaluation of “JavParamSeq” in given
environment ej and the return value val of procedure p equals the
evaluation of “JavParam” in environment ej and
• finally we get e′ by pushing “EvntName” in given environment e and
• s′ is produces by updating the value ev for an identifier “EvntName”
in the given pre-state s.
3.4.3
Register Model Sequence
The syntax of the syntactic domain RegModSeq is defined as follows:
RegModSeq := EMPTY | (RegMod) RegModSeq
Semantically, when an EMPTY sub-phrase is evaluated in a given environment e then simply the resulting environment e′ equals e and a post-state
s′ equals the given pre-state s as defined below:
Case: EMPTY
〚EMPTY〛(e)(e’, s, s’) ⇔ e’ = e ∧ s’ = inState⊥ (s)
While in the second alternate of the domain “RegModSeq”, first semantics of the phrase “RegMod” in a given environment e produce an environment e′′ transforming a pre-state s into a post-state s′′ , then the evaluation
of the phrase “RegModSeq” in environment e′′ results in a given environment
e′ and transforms the pre-state s′′ into a given post-state s′ . The semantics
of the second alternative is formalized as follows:
Case: (RegMod) RegModSeq
〚(RegMod) RegModSeq〛(e)(e’, s, s’) ⇔
∀ e” ∈ Environment, s” ∈ State:
〚RegMod〛(e)(e”, s, inState⊥ (s”)) ∧ 〚RegModSeq〛(e”)(e’, s”, s’)
3.4.4
Structural Model
The structural behavior of the system is defined by the syntactic phrase
“StrMod” which represents a corresponding hierarchical model of the components. The syntax for the overall structural behavior of the component
24
“CompName” is defined by the syntactic phrase “StrMod” where different
clauses define three logical parts of the behavior as follows:
1. signals specify global control behavior of the component, e.g.
• the clauses :entry-events and :exit-events models the entry
and exit events of the component respectively and
• the other allowable events (while execution of the component) are
modeled with the clause :allowable-events
2. signature of the component consists of
• the sequences of objects for the clauses :inputs and :outputs
respectively
3. body of the component is modeled as a sub-network which involves different components as represented by the :components clause. These
components are connected through various nodes and links as follows:
(a) the control flows :controlflows which further have corresponding
splits :splits and joins :joins and
(b) the propagation of data among the components (via control flows)
is represented by the clause :dataflows.
(c) while the execution of the body, various computing resources :resources (each with a name, its type and its probabilities of being
in normal and hacked modes respectively) are involved which further requires
(d) the resource mappings :resource-mappings (where each resource
is mapped to a component that uses it) in addition to
(e) the model mappings :model-mappings (where the conditional
probability between the compromises and misbehaviors for each
of the component is given) and
(f) the vulnerabilities :vulnerabilities such that each resource is
mapped to a corresponding (possible) vulnerability (which is assumed to be defined as the part of an attack plan that is beyond
the scope of this document).
The syntactic domain of for the structural behavioral model (StrMod)
is defined as follows:
StrMod ::= define-ensemble CompName
:entry-events :auto | (EvntSeq1 )
:exit-events (EvntSeq2 )
:allowable-events (EvntSeq3 )
:inputs (ObjNameSeq1 )
25
:outputs (ObjNameSeq2 )
:components (CompSeq)
:controlflows (CtrlFlowSeq)
:splits (SpltCFSeq)
:joins (JoinCFSeq)
:dataflows (DataFlowSeq)
:resources (ResSeq)
:resource-mapping (ResMapSeq)
:model-mappings (ModMapSeq)
:vulnerabilities (VulnrabltySeq)
The semantics of the structural behavioral model in a given environment
e results in an environment e′ transforming a pre-state s into a post-state s′
as defined below:
〚StrMod〛(e)(e’, s, s’) ⇔
∀ e, e1 , e2 , e3 , e4 , e5 , e6 , e7 , e8 ∈ Environment, s, s1 , s2 , s3 , s4 , s5 , s6 , s7 , s8 ∈ State,
oeseq, oeseq1 , aeseq ∈ ObsEvent*, anameseq, enameseq, enameseq1 ∈ EvntNameSeq:
( eqFlag(s, “running”) ∧
〚EvntSeq3 〛(e)(e1 , s, inState⊥ (s1 ), enameseq, oeseq) ∧
∀ ename ∈ enameseq:
∃ se ∈ Event, rte ∈ RTEvent, oe ∈ oeseq:
IF equals(se, oe) THEN
LET rte = store(s1 )(ename) IN
IF equals(rte, se) THEN
true
ELSE
s1 = enableDiagnosis(e1 )(s1 , inBool(true))
END
END
ELSE
s1 = enableDiagnosis(e1 )(s1 , inBool(true))
END )
∨
( eqFlag(s, “running”) ∨ eqFlag(s, “ready”) ∧
〚EvntSeq1 〛(e)(e1 , s, inState⊥ (s1 ), enameseq, oeseq) ∧
∀ ename ∈ enameseq, oe ∈ oeseq:
∃ se ∈ Event, rte ∈ RTEvent:
equals(se, store(s1 )(ename)) ∧ equals(se[1], oe[1]) ∧
LET rte = store(s1 )(ename) IN
IF equals(rte[5][1], “entry” ) THEN
equals(rte[1], se[1])
ELSE equals(rte[2], se[2])
END
26
END ∧
∀ inseq ∈ Value∗ , c ∈ Component:
〚ObjNameSeq1 〛(e1 )(inState⊥ (s1 ), inseq) ∧ 〚CompName〛(e1 )(inValue(c)) ∧
IF equals(c[2][1], inseq) THEN
eqMode(s1 , “normal”)
ELSE
s1 = enableDiagnosis(e1 )(s1 , inBool(true))
END ∧
IF equals(c[3][1], inseq) THEN
eqMode(s1 , “compromised”) ∧ s1 = enableDiagnosis(e1 )(s1 , inBool(true))
ELSE true
END )
⇒ eqFlag(s1 , “running”) ∧
∀ compseq ∈ Component∗ : 〚CompSeq〛(e2 )(e3 , s2 , inState⊥ (s3 ), compseq) ∧
∀ rmseq, crmapseq, cpmapseq, vbltyseq ∈ List∗ :
〚ResSeq〛(e3 )(s3 , inState⊥ (s4 ), rmseq) ∧
〚ResMapSeq〛(e3 )(s3 , inState⊥ (s4 ), crmapseq) ∧
〚ModMapSeq〛(e3 )(s3 , inState⊥ (s4 ), cpmapseq) ∧
〚VulnrabltySeq〛(e3 )(s3 , inState⊥ (s4 ), vbltyseq) ∧
e4 = buildEnv(e3 , rmseq, crmapseq, cpmapseq, vbltyseq) ∧
〚CtrlFlowSeq〛(e4 )(e5 , s4 , inState⊥ (s5 )) ∧
〚SpltCFSeq〛(e5 )(e6 , s5 , inState⊥ (s6 )) ∧
〚JoinCFSeq〛(e6 )(e7 , s6 , inState⊥ (s7 )) ∧
〚DataFlowSeq〛(e7 )(e8 , s7 , inState⊥ (s8 )) ∧
〚EvntSeq2 〛(e8 )(e9 , s8 , s’, enameseq1 , oeseq1 ) ∧
∀ ename ∈ enameseq, oe ∈ oeseq:
∃ se ∈ Event, rte ∈ RTEvent:
equals(se, store(inState(s’))(ename)) ∧ equals(se[1], oe[1]) ∧
LET rte = store(inState(s’))(ename) IN
IF equals(rte[5][1], “entry” ) THEN
equals(rte[1], se[1])
ELSE equals(rte[2], se[2])
END
END
⇒
∀ outseq ∈ Value∗ , c ∈ Component:
〚ObjNameSeq2 〛(e9 )(s’, outseq) ∧ 〚CompName〛(e9 )(inValue(c)) ∧
( ( IF equals(c[2][2], outseq) THEN
eqMode(inState(s’), “normal”)
ELSE
s’ = enableDiagnosis(e9 )(inState(s’), inBool(true))
END ) ∨
( IF equals(c[3][2], outseq) THEN
eqMode(inState(s’), “compromised”) ∧
27
s’ = enableDiagnosis(e9 )(inState(s’), inBool(true))
END ) ) ∧
eqMode(inState(s’), “normal”) ∧
eqFlag(inState(s’), “completed”) ∧
LET sbeh = <inseq, outseq, s, s’>, nbeh = c[2], cbeh = c[3] IN
e’ = push(e9 , store(inState(s’))(〚CompName〛(e9 ))
, c(sbeh, nbeh, cbeh, s, s’))
END
In general, the semantics is defined as a big logical implication, where
the premise is a disjunction of two formulas as explained below:
1. either the current state s of the component is “running” and it receives
allowable events “EvntSeq3 ” and for every event oe in the allowable
event sequence oeseq there is a corresponding equivalent registered
event se for which we receive an equivalent runtime event rte such
that rte is one of the under observation (legal) event se, if not then
the runtime event is a result of the misbehavior of the component
so the diagnosis component of AWDRAT is activated by calling “enableDiagnosis(...)” which successfully recovers the compromised state
s1
2. or the current state s of the component is either “running” or “ready”
and it receives the entry events “EvntSeq1 ” (evaluating to oeseq) and
for every event oe in the sequence of entry events oeseq there is a
corresponding registered event se and the received runtime event rte
(equals se depending on its type “entry” or “exit”) is the monitored
event and
• if the sequence of input values inseq satisfies the pre-conditions
of the “normal” behavior (c[2][1])) of the component (c) then the
resulting state s1 is in “normal” mode
• otherwise (when pre-conditions are not satisfied) then the diagnosis component is activated which recovers the compromised state
s1 and
• if the sequence of input values inseq satisfies the pre-conditions of
the already “compromised” behavior (c[3][1]) of the component
(c) then the resulting state s1 is “compromised” state and we
restore it by enabling diagnostic engine
• otherwise the system is safe true to start executing component
respectively body/sub-network.
Semantically, if any of the above two holds then the
• the current state s1 is “running” and the components of the subnetwork evaluate to compseq and
28
• a new environment e4 is constructed based on the evaluation of the
resources, resource mappings, model mappings and vulnerabilities of
the sub-network to rmseq, crmapseq, cpmapseq, and vbltyseq respectively (such that all the trust model of the components is known before
the actual execution of the body starts) in which
• the execution blocks are evaluated (such that the evaluation of the
control flows their respective splits and joins and associated data flows
results in an environment e8 and a post-state s8 ) to complete the
executional behavior and
• once all the sub-network is executed (recursively), then the receiving
exit events (EvntSeq2 ) evaluate to oeseq1 and if for every event oe in
oeseq1 there is an equivalent registered event se and a runtime event
rte then
– either the sequence of output values outseq satisfies the postconditions of the “normal” behavior (c[2][2]) of the component (c)
then the post-state s′ is in “normal” mode otherwise the diagnosis
component restores the post-state s′ and
– or the sequence of output values outseq satisfies the post-conditions
of the misbehavior (c[3][2]) of the component (c) then the poststate must be in “compromised” mode and corresponding diagnosis component is enabled to recover the state back and
• the given and transformed final post-state s′ must be in “normal”
mode with “completed” flag and
• finally the resulting environment e′ is build with the evaluated behavior
of the component of the current component.
3.4.5
Structural Model Sequence
The syntactic domain StrModSeq is:
StrModSeq := EMPTY | (StrMod) StrModSeq
Case: EMPTY
〚EMPTY〛(e)(e’, s, s’) ⇔ e’ = e ∧ s’ = inState⊥ (s)
Case: (StrMod) StrModSeq
〚(StrMod) StrModSeq〛(e)(e’, s, s’) ⇔
∀ e” ∈ Environment, s” ∈ State:
〚StrMod〛(e)(e”, s, inState⊥ (s”)) ∧ 〚StrModSeq〛(e”)(e’, s”, s’)
29
The semantics of the domain “StrModSeq” of structural model sequence are
similar to the semantics of the domain “RegModSeq” as discussed above in
the corresponding section. Similarly, the semantics of the syntactic domains
of “BehModSeq” and “SplModSeq” can be exercised which are discussed in
the corresponding sections later in this document.
3.4.6
Behavioral Model
The behavioral model represents the functional behavior of a component,
which can be either “normal” or known “compromised” one. The functional
behavior of the component “CompName” consists of the following elements:
1. the inputs of the component as given by the clause :inputs,
2. the outputs of the component as represented by the corresponding
clause :outputs,
3. the allowable events :allowable-events represents the auxiliary communication of the component,
4. the pre-conditions of the component are specified in the clause :prerequisites while
5. the corresponding post-conditions are specified by the :postconditions clause.
Note that the “compromised” behavior is used to model already known
misbehaviors of the component (e.g. some attack) and needs corresponding
diagnosis which in this case is already known.
The syntactic domain “BehMod” for the behavioral model is defined as
follows:
BehMod ::= defbehavior-model (CompName normal | compromised)
:inputs (ObjNameSeq1 )
:outputs (ObjNameSeq2 )
:allowable-events (EvntSeq)
:prerequisites (BehCondSeq1 )
:postconditions (BehCondSeq2 )
Semantically, normal and compromised behavioral models results in modifying the corresponding elements of the environment value “Component”
as defined below:
〚BehMod〛(e)(e’, s, s’) ⇔
∀ e1 ∈ Environment, nseq ∈ EvntNameSeq, eseq ∈ ObsEvent*, inseq, outseq ∈ Value∗ :
〚ObjNameSeq1 〛(e)(inState⊥ (s), inseq) ∧ 〚BehCondSeq1 〛(e) (inState⊥ (s)) ∧
〚EvntSeq〛(e) (e1 , s, s’, nseq, eseq)
30
〚ObjNameSeq2 〛(e1 )(s’, outseq) ∧ 〚BehCondSeq2 〛(e1 ) (s’) ∧
∃ c ∈ Component: 〚CompName〛(e1 )(inValue(c)) ∧
IF eqMode(inState⊥ (s’), “normal”) THEN
LET sbeh = c[1], nbeh = <inseq, outseq, s, s’>, cbeh = c[3] IN
e’ = push(e1 , store(inState(s’))(〚CompName〛(e1 ))
, c(sbeh, nbeh, cbeh, s, s’))
END
ELSE
LET sbeh = c[1], nbeh = c[2], cbeh = <inseq, outseq, s, s’> IN
e’ = push(e1 , store(inState(s’))(〚CompName〛(e1 ))
, c(sbeh, nbeh, cbeh, s, s’))
END
END
In detail, if the semantics of of syntactic domain “BehMod” holds in a
given environment e resulting in environment e′ and transforming a pre-state
s into corresponding post-state s′ then
• the inputs “ObjNameSeq1 ” evaluates to a sequence of values inseq in
a given environment e and a given state s which satisfies the corresponding pre-conditions “BehCondSeq1 ” in the same e and s and
• the allowable events happens whose evaluation results in new environment e1 and given post-state s′ with some auxiliary sequences nseq
and eseq and
• the outputs “ObjNameSeq2 ” evaluates to a sequence of values outseq
in an environment e1 and given post-state s′ which satisfies the corresponding post-conditions “BehCondSeq2 ” in the same environment
e1 and state s′ and the given environment e′ can be constructed such
that
– if the post-state is “normal” then e′ is an update to the normal
behavior “nbeh” of the component “CompName” in environment
e1
– otherwise e′ is an update to the compromised behavior “cbeh” of
the component.
In the construction of the environment e′ the rest of the semantics of the
component do not change as represented in the corresponding LET-IN constructs.
3.4.7
Behavioral Model Sequence
The syntactic domain BehModSeq is:
BehModSeq := EMPTY | (BehMod) BehModSeq
31
Case: EMPTY
〚EMPTY〛(e)(e’, s, s’) ⇔ e’ = e ∧ s’ = inState⊥ (s)
Case: (BehMod) BehModSeq
〚(BehMod) BehModSeq〛(e)(e’, s, s’) ⇔
∀ e” ∈ Environment, s” ∈ State:
〚BehMod〛(e)(e”, s, inState⊥ (s”)) ∧ 〚BehModSeq〛(e”)(e’, s”, s’)
3.4.8
Split Model
Though the splits of control flows are declared in the “StrBeh” domain
but their corresponding definitions are given with the help of the domain
“SplMod” which consists of its
• name “SpltModName”,
• required sequence of parameters “SpltParamSeq” which are used by
the various branches of the split as defined in
• the split condition branches “SpltCondSeq”.
The syntax of the domain “SplMod” is given as follow:
SplMod ::= defsplit SpltModName? (SpltParamSeq) SpltCondSeq)
If the semantics of the split model “SplMod” in a given environment e
results in environment e′ and transforms a pre-state s into post-state s′ then
• first the parameters are evaluated in a given environment e which
results in an environment e1 and sequence of values vseq transforming
a given pre-state s into post-state s1 and
• the split conditions “SpltCondSeq” hold in environment e1 producing
environment e2 and given post-state s′ and finally
• given environment e′ is a result of a push operation on environment
e2 updating the value of the split “SpltModName” with the one constructed by the computed values vseq.
The semantics of the split behavior is formalized as follows:
〚SplMod〛(e)(e’, s, s’) ⇔
∀ e1 , e2 ∈ Environment, s1 ∈ State, vseq ∈ Value∗ :
〚SpltParamSeq〛(e)(e1 , s, inState⊥ (s1 ), vseq) ∧
〚SpltCondSeq〛(e1 ) (e2 , s1 , s’) ∧
LET s ∈ Split IN
e’ = push(e2 , store(inState(s’))(〚SpltModName〛(e2 )), s(vseq))
END
32
3.4.9
Split Model Sequence
The syntactic domain SplModSeq is:
SplModSeq := EMPTY | (SplMod) SplModSeq
Case: EMPTY
〚EMPTY〛(e)(e’, s, s’) ⇔ e’ = e ∧ s’ = inState⊥ (s)
Case: (SplMod) SplModSeq
〚(SplMod) SplModSeq〛(e)(e’, s, s’) ⇔
∀ e” ∈ Environment, s” ∈ State:
〚SplMod〛(e)(e”, s, inState⊥ (s”)) ∧ 〚SplModSeq〛(e”)(e’, s”, s’)
3.4.10
Attack Model
The attack model represents the different types of known/hypothetical attack, their corresponding probabilities and the respective vulnerabilities
causing the attack types. The attack model “AtkModName” has:
1. types of attack and their conditional probabilities as specified by the
clause :attack-types and
2. mapping between the types of attack and vulnerabilities as described
by the corresponding clause :vulnerability-mapping.
Additionally, the attack model is extended by the rules which map conditional probabilities of the attacks and vulnerabilities. The attack rule
“AtkRulName” has
1. a sequence of attack conditions which describe the attack situation as
specified by the clause if
2. and the attack consequences which map the probabilities of attacks
and vulnerabilities; the maps are represented by the clause then.
Note that the attack models can be used in the following ways:
• the models are already known attacks and thus already know the corresponding diagnosis
• or the models can be hypothetical attacks which can be used to generate rigorous monitors for the system.
The syntactic domain “AtkMod” for the attack model is defined as follows:
33
AtkMod ::= define-attack-model AtkModName
:attack-types (AtkTypeSeq)
:vulnerability-mapping (AtkVulnrabltyMapSeq)
While the syntactic domain “AtkRule” for defining attack rules is defined
as follows:
AtkRule ::= defrule AtkRulName (:forward)
if AtkCondSeq
then AtkConsSeq
Semantically, an attack model results in the environment value “AtkModel” as defined below:
〚AtkMod〛(e)(e’, s, s’) ⇔
∀ s” ∈ State, aseq, aseq’, vnseq ∈ ISeq, apseq ∈ Value∗ :
〚AtkTypeSeq〛(e)(s, inState⊥ (s”), aseq, apseq) ∧
〚AtkVulnrabltyMapSeq〛(e) (s”, s’, aseq’, vnseq) ∧ respectsOrder(aseq, aseq’) ∧
LET amod ∈ AtkModel IN
e’ = push(e, store(inState(s’))(〚AtkModName〛(e)), amod(aseq, apseq, vnseq)))
END
In detail, the semantics of the syntactic domain “AtkMod” updates the
environment e with a attack semantic value amod such that
• in a given environment e and state s, the evaluation of “AtkTypeSeq” results in a post-state s′′ , a sequence of attack types aseq and a
sequence of values (conditional probabilities) apseq and
• in a given environment e and state s, the evaluation of “AtkVulnrabltyMapSeq” results in post-state s′ , a sequence of attack types
aseq ′ and a sequence of vulnerabilities vnseq and
• the environment e′ is an update of environment e with the semantic
value amod which is a triple of
1. a sequence of attack types,
2. a sequence of corresponding probabilities and
3. a sequence of vulnerabilities causing the attack types, respectively.
However, if the semantics of the syntactic domain “AtkRule” holds in
an environment e, then
• there is some resource r such that (as given in “AtkCondSeq” respective “AtkCond”)
1. the resource name is “?resource-name” and
34
2. the resource type is “ResType” and
3. if the resource has been compromised by an attack “AtkTypeName”, then
• the resource r (and its associated component c) has behavior as specified by the evaluation of consequences “AtkCodSeq” in an environment
e and state s.
Formally, the semantics of the syntactic domain “AtkRule” is defined as:
〚AtkRule〛(e)(e’, s, s’) ⇔
∃ r ∈ Resource, c ∈ Component: 〚AtkCondSeq〛(e)(s, s’, r, c) ∧
〚AtkConsSeq〛(e) (s, s’, r, c) ∧ e’ = e
3.4.11
Attack Model Sequence
The syntactic domain AtkModSeq is:
AtkModSeq := EMPTY | (AtkMod) AtkModSeq
Case: EMPTY
〚EMPTY〛(e)(e’, s, s’) ⇔ e’ = e ∧ s’ = inState⊥ (s)
Case: (AtkMod) AtkModSeq
〚(AtkMod) AtkModSeq〛(e)(e’, s, s’) ⇔
∀ e” ∈ Environment, s” ∈ State:
〚AtkMod〛(e)(e”, s, inState⊥ (s”)) ∧ 〚AtkModSeq〛(e”)(e’, s”, s’)
4
Execution Monitor
In principle, Architectural Differencer synthesizes both the wrappers and
the execution monitor where the wrappers traces the execution of the target
system by creating an event stream (these traces are also called observations); while the role of an execution monitor is to interpret the stream
against the system (Architectural Model) specification (the execution of the
specification is also called predictions) by detecting inconsistencies between
observations and the predictions, if there are any.
We have already discussed the formal syntax and semantics of the predictions in the previous sections, now we first give the formal syntax of the
observations in this section and the corresponding formal semantics in the
following section.
35
4.1
Observation Model
Each runtime event (observation) consists of
• a name “EvntName”,
• its type, i.e. entry or exit,
• depending on the type of event
– either sequence of event parameters (if entry event)
– or a parameter representing return value of the event (if exit
event)
• a numeric value “Numeral” representing an operating system level
process id, which can be used later to get more information about the
event to detect any system level threats and other technical dependencies and
• a time “TimeStamp” of the event which later can be used to detect
inconsistencies in the sequence of events.
The syntax of the runtime event is defined by the syntactic domain
“Obsrv” as follow:
Obsrv := EvntName entry | exit
EvntParamSeq
Numeral
TimeStamp
If the semantics of an observation “Obsrv” in a given environment e
results in environment e′ and transforms a pre-state s into post-state s′
then
• first the parameters are evaluated in a given environment e which
results in an environment e1 and sequence of values pseq transforming
a given pre-state s into post-state s1 and
• evaluation of the numeric value “Numeral” results in a value n in
environment e1 and state s1 and
• also time stamp “TimeStamp” evaluates to a value t in environment
e1 and state s1 and finally
– if the observation is “entry” event the resulting environment e′
is a result of a push operation on environment e2 updating the
value of the observation “EvntName” with the semantic value
of the observation, i.e. of type “RTEvent” which is constructed
with the help of computed input values pseq, process id n and
time value t and
36
– if the observation is “exit” event the resulting environment e′ is a
result of a push operation on environment e2 updating the value
of the observation “EvntName” with the semantic value of the
observation, i.e. of type “RTEvent” which is constructed with
the help of computed output values pseq (sequence with a single
value), process id n and time value t.
The semantics of the observation is formalized as follows:
〚Obsrv〛(e)(e’, s, s’) ⇔
∀ e1 , e2 ∈ Environment, s1 ∈ State, pseq ∈ Value∗ , n, t ∈ Value:
〚EvntParamSeq〛(e)(e1 , s, inState⊥ (s’), pseq) ∧
〚Numeral〛(e1 ) (inState(s’), n) ∧ 〚TimeStamp〛(e1 ) (inState(s’), t) ∧
LET rte ∈ RTEvent IN
IF isEntry(Obsrv) THEN
e’ = push(e2 , store(inState(s’))(〚EvntName〛(e2 ))
, rte(pseq, EMPTY, s, s’, <“entry”, t, n>))
ELSE
e’ = push(e2 , store(inState(s’))(〚EvntName〛(e2 ))
, rte(EMPTY, pseq, s, s’, <“exit”, t, n>))
END
END
4.1.1
Observations
The event respectively observation stream is a sequence of observations,
which is modeled by corresponding syntactic domain ObsrvSeq as follows:
ObsrvSeq := EMPTY | (Obsrv) ObsrvSeq
The semantics of the observation sequence are similar to the other syntactic sequences discussed earlier in this document.
Case: EMPTY
〚EMPTY〛(e)(e’, s, s’) ⇔ e’ = e ∧ s’ = inState⊥ (s)
Case: (Obsrv) ObsrvSeq
〚(Obsrv) ObsrvSeq〛(e)(e’, s, s’) ⇔
∀ e” ∈ Environment, s” ∈ State:
〚Obsrv〛(e)(e”, s, inState⊥ (s”)) ∧ 〚ObsrvSeq〛(e”)(e’, s”, s’)
5
Semantics of the Execution Monitor
Though the technical details of the operation of the execution monitor are
discussed in [5], in the following we give their informal semantics.
37
We presume that a reasonable fine grained level behavior of the target
system is specified in the corresponding System Architectural Model. When
the target system starts execution, an initial “startup” event is generated
and dispatched to the top level component (module) of the system which
transforms the execution state of the component into “running” mode. The
component instantiates its subnetwork (of components, if there is one) and
also propagates the data along its data links by enabling the corresponding
control links (if involved). When the data arrives on the input port of
the component, the execution monitor checks if it is complete; if so, the
execution monitor checks the preconditions of the component for the data
and if they succeed, it transform the state of the component into “ready”
mode. In case, any of the preconditions fails, it enables diagnosis engine.
After the above startup of the target system, the execution monitor starts
monitoring the arrival of every observation (runtime event) as follows:
1. If the event is a “method entry”, then the execution monitor checks if
this is one of the “entry events” of the corresponding component in the
“ready” state; if so, then after receiving the data and the respective
preconditions are checked; if they succeed, then the data is applied on
the input port of the component and the mode of the execution state
is changed to “running”.
2. If the event is a “method exit”, then the execution monitor checks
if this one of the “exit events” of the component in the “running”
state; if so, it changes its state into “completed” mode and collects
the data from the output port of the component and checks for the
corresponding postconditions. Should the checks fail, the execution
monitor enables the diagnosis engine.
3. If the event is one of the “allowable events” of the component, it
continues execution and finally
4. if the event is an unexpected event, i.e. it is neither an “entry event”,
nor an “exit event” and also not in “allowable events”, then the execution monitor starts diagnosis.
Based on the above behavioral description of the execution monitor, we
have formalized the corresponding semantics of the execution monitor as
follows:
∀ app ∈ Target System, sam ∈ System Architectural Model, c ∈ Component,
s, s’ ∈ State, t, t’ ∈ States , d, d’ ∈ Environments , e, e’ ∈ Environment, rte ∈ RTEvent:
〚sam〛(d)(d’, t, t’) ∧ 〚app〛(e)(e’, s, s’) ∧ startup(s, app) ∧ isTop(c, 〚app〛(e)(e’, s, s’)) ∧
setMode(s, “running”) ∧ arrives(rte, s) ∧ equals(t, s) ∧ equals(d, e)
⇒
∀ p, p’ ∈ Environment∗ , m, n ∈ State∗⊥ :
38
equals(m(0), s) ∧ equals(p(0), e)
⇒
∃ k ∈ N, p, p’ ∈ Environment∗ , m, n ∈ State∗⊥ :
∀ i ∈ Nk : monitors(i, rte, c, p, p’, m, n) ∧
( eqMode(n(k), “completed”) ∧ eqFlag(n(k), “normal”) ∧
equals(s’, n(k))
∨
eqFlag(n(k), “compromised”)
⇒
enableDiagnosis(p’(k))(n(k), inBool(true)) ∧ equals(s’, n(k)) )
In detail, given a target system “app” and its specification“sam” and
their semantices are defined such that their corresponding pre-states are
equivalent. Furthermore, if the application starts “startup(...)”, and an
arbitrary c is a top-level component “isTop(...)”, then the current state of the
component is marked as “running” and when an observation “rte” arrives in
this state, then the monitor starts monitoring the event stream/sequence and
thus, here, we have formalized the corresponding semantics of the monitor by
the two sequences of pre- and post-states [3] and their respective sequences
of the pre- and post-environments. Both the former and later sequences
are constructed from their corresponding pre- and post objects. The arrival
and monitoring of the ith observation (event) transforms state pre(i) into
state post(i + 1) from which the state pre(i + 1) is constructed and the
same repeats for the construction of the corresponding environments. No
event can be accepted in an Error state and the corresponding monitoring
terminates either when the application has terminated with “normal” mode
or when there is some misbehavior is detected as indicated by the respective
“compromised” state. This semantics is formalized with the help of predicate
“monitor”, for details please see Subsection 3.3.
Finally, when there are sequences of states and environments for which
the predicate “monitor” holds, then either the given post-state s′ is equal
to the “monitor”ed post-state “n(k)” which is in “completed” mode and
has a “normal” flag or post-state “n(k)” is “compromised” and in this case
diagnosis is enabled which successfully transforms the compromised state
into a normal state which results in the given post-state s′ .
6
Conclusions and Future Work
In this report, we gave the formal definition of the syntax and semantices of
the System Architectural Model and the Execution Monitor of AWDRAT.
These definitions help to understand internal behavior of the corresponding
components on one hand, and also serves as a formal basis for ADWRAT to
extend the current system on the other hand. Based on this formalism, we
39
are currently working on the formal reliability (soundness) analysis of the
Execution Monitor of AWDRAT.
In future, we plan to extend AWDRAT such that a target system behavior is specified using Abstract State Machine (ASM) [1] based formalism
which then will automatically translate into a semantically equivalent System Architectural Model. This will allow to already check the inconsistencies
in the system behavior with existing ASM supported tools [2].
Acknowledgment
The authors cordially thank Adam Chilpala for his valuable and constructive
remarks and suggestions.
40
References
[1] E. Borger and Robert F. Stark. Abstract State Machines: A Method
for High-Level System Design and Analysis. Springer-Verlag New York,
Inc., Secaucus, NJ, USA, 2003.
[2] Jean-Baptiste Jeannin, Guido de Caso, Juan Chen, Yuri Gurevich,
Prasad Naldurg, and Nikhil Swamy. Dkal*: Constructing executable
specifications of authorization protocols. Technical Report MSR-TR2013-19, March 2013.
[3] Muhammad Taimoor Khan. On the Formal Semantics of MiniMaple and
Its Specification Language. In Proceedings of the 2012 10th International
Conference on Frontiers of Information Technology, FIT ’12, pages 169–
174, Washington, DC, USA, 2012. IEEE Computer Society.
[4] Schmidt, David A. Denotational Semantics: a methodology for language
development. William C. Brown Publishers, Dubuque, IA, USA, 1986.
[5] Howard Shrobe, Robert Laddaga, Bob Balzer, Neil Goldman, Dave Wile,
Marcelo Tallis, Tim Hollebeek, and Alexander Egyed. AWDRAT: A Cognitive Middleware System for Information Survivability’. In Proceedings
of the 18th Conference on Innovative Applications of Artificial Intelligence - Volume 2, IAAI’06, pages 1836–1843. AAAI Press, 2006.
[6] Shrobe, Howard E. Dependency Directed Reasoning for Complex Program Understanding. Technical report, 1979.
41
Appendices
Appendix A gives the formal abstract syntax (language grammar) for the
specification language “system architectural model” of AWDRAT.
A
A.1
Formal Syntax of System Architectural Model
Declaration of Syntactic Domains
/* top level syntactic domains */
SAM ∈ System Architectural Model
RegModSeq ∈ Register Model Sequence
StrModSeq ∈ Structural Model Sequence
BehModSeq ∈ Behavioral Model Sequence
SplModSeq ∈ Split Model Sequence
AtkModSeq ∈ Attack Model Sequence
AtkRuleSeq ∈ Attack Rule Sequence
/* top level syntactic sub-domains */
RegMod ∈ Register Model
StrMod ∈ Structural Model
BehMod ∈ Behavioral Model
SplMod ∈ Split Model
AtkMod ∈ Attack Model
AtkRule ∈ Attack Rule
/* event related syntactic domains */
Evnt ∈ Event
EvntSeq ∈ Event Sequence
EvntName ∈ Event Name
EvntNameSeq ∈ Event Name Sequence
EvntParamSeq ∈ Event Parameter Sequence
/* java related syntactic domains */
JavClaName ∈ Java Class Name
JavMetName ∈ Java Method Name
JavParam ∈ Java Parameter
JavParamSeq ∈ Java Parameter Sequence
JavParamName ∈ Java Parameter Name
JavParamType ∈ Java Parameter Type
/* object related syntactic domains */
ObjName ∈ Object Name
ObjtNameStr ∈ Object Name String
ObjNameSeq ∈ Object Name Sequence
42
ObjType ∈ Object Type
ObjComp ∈ Object Component
ObjCompSeq ∈ Object Component Sequence
/* behavioral condition, parameter and situation related syntactic domains
*/
BehCond ∈ Behavioral Condition
BehCondSeq ∈ Behavioral Condition Sequence
BehCondMode ∈ Behavioral Condition Mode
BehParam ∈ Behavioral Parameter
BehParamSeq ∈ Behavioral Parameter Sequence
BehSit ∈ Behavioral Situation
/* branch related syntactic domains */
BrnchName ∈ Branch Name
BrnchNameSeq ∈ Branch Name Sequence
BrnchCond ∈ Branch Condition
/* component related syntactic domains */
Comp ∈ Component
CompSeq ∈ Component Sequence
CompName ∈ Component Name
CompType ∈ Component Type
/* control flow related syntactic domains */
CtrlFlow ∈ Control Flow
CtrlFlowSeq ∈ Control Flow Sequence
/* function related syntactic domains */
FuncName ∈ Function Name
FuncParam ∈ Function Parameter
FuncParamSeq ∈ Function Parameter Sequence
/* split related syntactic domains */
SpltCF ∈ Split
SpltCFSeq ∈ Split Sequence
SpltName ∈ Split Name
SpltModName ∈ Split Model Name
SpltParamSeq ∈ Split Parameter Sequence
SpltCond ∈ Split Condition
SpltCondSeq ∈ Split Condition Sequence
/* join related syntactic domains */
JoinCF ∈ Join
JoinCFSeq ∈ Join Sequence
JoinName ∈ Join Name
43
JoinParamSeq ∈ Join Parameter Sequence
/* data flow related syntactic domains */
DataFlow ∈ Data Flow
DataFlowSeq ∈ Data Flow Sequence
/* resource related syntactic domains */
Res ∈ Resource
ResSeq ∈ Resource Sequence
ResName ∈ Resource Name
ResType ∈ Resource Type
ResMap ∈ Resource Mapping
ResMapSeq ∈ Resource Mapping Sequence
ResModMap ∈ Resource Model Mapping
/* model mapping syntactic domains */
ModMap ∈ Model Mapping
ModMapSeq ∈ Model Mapping Sequence
/* vulnerability related syntactic domains */
Vulnrablty ∈ Vulnerability
VulnrabltyName ∈ Vulnerability Name
VulnrabltySeq ∈ Vulnerability Sequence
/* attack related syntactic domains */
AtkType ∈ Attack Type
AtkTypeSeq ∈ Attack Type Sequence
AtkModName ∈ Attack Model Name
AtkCond ∈ Attack Condition
AtkCondSeq ∈ Attack Condition Sequence
AtkCons ∈ Attack Consequence
AtkConsSeq ∈ Attack Consequence Sequence
AtkTypeName ∈ Attack Type Name
AtkRulName ∈ Attack Rule Name
AtkVulnrabltyMap ∈ Attack Vulnerability Mapping
AtkVulnrabltyMapSeq ∈ Attack Vulnerability Mapping Sequence
/* other syntactic domains */
MembName ∈ Member Name
ParamName ∈ Parameter Name
DSCond ∈ Data Structure Condition
ISeq ∈ Identifier Sequence
44
A.2
Grammar
Based on the declarations of various syntactic domains, in this section we
discuss the grammar rules for the domains.
/* top level syntactic domains */
SAM ::= RegModSeq StrModSeq BehModSeq SplModSeq
RegModSeq ::= EMPTY
| (RegMod) RegModSeq
StrModSeq ::= EMPTY
| (StrMod) StrModSeq
BehModSeq ::= EMPTY
| (BehMod) BehModSeq
SplModSeq ::= EMPTY
| (SplMod) SplModSeq
AtkModSeq ::= EMPTY
| (AtkMod) AtkModSeq
AtkRuleSeq ::= EMPTY
| (AtkRule) AtkRuleSeq
/* top level syntactic sub-domains */
RegMod ::= register-event ’EvntName JavClaName JavMetName ’(JavParamSeq)
[ :static ObjName ]
[ :output-type JavParam ]
[ :bypass ObjNameStr ]
[ :EvntName ObjName]
StrMod ::= define-ensemble CompName
:entry-events :auto | (EvntSeq)
:exit-events (EvntSeq)
:allowable-events (EvntSeq)
:inputs (ObjNameSeq)
:outputs (ObjNameSeq)
:components (CompSeq)
:controlflows (CtrlFlowSeq)
:splits (SpltCFSeq)
:joins (JoinCFSeq)
:dataflows (DataFlowSeq)
:resources (ResSeq)
:resource-mapping (ResMapSeq)
:model-mappings (ModMapSeq)
:vulnerabilities (VulnrabltySeq)
BehMod ::= defbehavior-model (CompName normal | compromised)
:inputs (ObjNameSeq)
:outputs (ObjNameSeq)
:allowable-events (EvntSeq)
45
:prerequisites (BehCondSeq)
:postconditions (BehCondSeq)
SplMod ::= defsplit SpltModName? (SpltParamSeq) SpltCondSeq)
AtkMod ::= define-attack-model AtkModName
:attack-types (AtkTypeSeq)
:vulnerability-mapping (AtkVulnrabltyMapSeq)
AtkRule ::= defrule AtkRulName (:forward)
if AtkCondSeq
then AtkConsSeq
/* event related syntactic domains */
Evnt ::= EvntName | (EvntName [entry | exit] (EvntParamSeq))
EvntSeq ::= EMPTY
| Evnt EvntSeq
EvntParam ::= I Iseq
| nil I Iseq
| I ISeq nil
EvntParamSeq ::= EMPTY
| EvntParam EvntParamSeq
/* java related syntactic domains */
JavClaName ::= ”I”
JavMetName ::= ”ID” | ”<I>“
JavParam ::= (JavParamType JavParamName)
JavParamSeq ::= EMPTY
| JavParam JavaParamSeq
JavParamName ::= ”I“
JavParamType ::= ”ID“ | ”ID[]”
/* object related syntactic domains */
ObjComp ::= (ObjName CompName?)
ObjCompSeq ::= EMPTY
| ObjComp ObjCompSeq
ObjNameStr ::= ObjName | ”ObjName“
/* behavioral condition related syntactic domains */
BehCond ::= [ DSCond ObjName ObjType BehCondMode ]
| [ and BehCond ]
| [ or BehCond ]
| [ not BehCond ]
| [ SpecFuncName BehParamSeq BehSit ]
BehCondSeq ::= EMPTY
| BehCond BehCondSeq
BehCondMode ::= EMPTY | good
BehParam ::= ?ObjName | (MembName ?ObjName)
46
BehParamSeq ::= EMPTY
| BehParam BehParamSeq
BehSit ::= ?before-CompName | ?after-CompName
/* branch condition syntactic domain */
BrnchCond ::= FuncName FuncParamSeq
/* component related syntactic domains */
Comp ::= (CompName :type CompType :models (normal [compromised]))
CompSeq ::= EMPTY
| Comp CompSeq
/* control flow related syntactic domains */
CtrlFlow ::= (before | after CompName[?-BrnchName])
CtrlFlowSeq ::= EMPTY
| CtrlFlow CtrlFlowSeq
/* function related syntactic domains */
FuncParam ::= ?ParamName | ’ParamName | ParamName? | not (ParamName)
FuncParamSeq ::= EMPTY
| FuncParam FuncParamSeq
/* split of control flow related syntactic domains */
SpltCF ::= (SpltName? SpltModName? [(SpltParamSeq)] (BrnchNameSeq))
SpltCFSeq ::= EMPTY
| SpltCF SpltCFSeq
SpltCondSeq ::= EMPTY
| SpltCond SplitCondSeq
SpltCond ::= (BrnchName (BrnchCond))
/* join of control flow related syntactic domains */
JoinCF ::= (JoinName? [(JoinParamSeq)] (BrnchNameSeq))
JoinCFSeq ::= EMPTY
| JoinCF JoinCFSeq
/* data flow related syntactic domains */
DataFlow ::= (ObjCompSeq)
DataFlowSeq ::= EMPTY
| DataFlow DataFlowSeq
/* resource related syntactic domains */
Res ::= (ResName ResType [(normal | hacked FVal)]+)
ResSeq ::= EMPTY
| Res ResSeq
ResType ::= File | Port | Mem
47
ResMap ::= (CompName ResName)
ResMapSeq ::= EMPTY
| ResMap ResMapSeq
ResModMap ::= ResName normal | hacked FVal
| ((ResName normal | hacked)) FVal
/* model mapping related syntactic domains */
ModMap ::= (CompName noromal | compromised ResModMap)
ModMapSeq ::= EMPTY
| ModMap ModMapSeq
/* vulnerability related syntactic domains */
Vulnrablty ::= (ResName VulnrabltyName)
VulnrabltySeq ::= EMPTY
| Vulnrablty VulnrabltySeq
/* attack related syntactic domains */
AtkType ::= (AtkTypeName FVal)
AtkTypeSeq ::= EMPTY
| AtkType AtkTypeSeq
AtkCond ::= [resource ?ensemble ?ResName ?Res]
[resource-type-of ?Res ResType]
[resource-might-have-been-attacked ?Res AtkTypeName]
| [ and AtkCond ]
| [ or AtkCond ]
| [ not AtkCond ]
AtkCondSeq ::= EMPTY
| AtkCond AtkCondSeq
AtkCons ::= [attack-implies-compromised-mode AtkTypeName ?Res
normal | compromised FVal]
| [ and AtkCons ]
| [ or AtkCons ]
| [ not AtkCons ]
AtkConsSeq ::= EMPTY
| AtkCons AtkConsSeq
AtkVulnrabltyMap ::= (VulnrabltyName AtkTypeName)
AtkVulnrabltyMapSeq ::= EMPTY
| AtkVulnrabltyMap AtkVulnrabltyMapSeq
/* syntactic domains of various names and types */
CompName, CompType, FuncName, ObjName, ObjType,
EvntName, SpltName, SpltModName ::= I
JoinName, ResName, BrnchName, VulnrabltyName, SpecFuncName
, ParamName, MembName ::= I
AtkModName, AtkTypeName, AtkRulName ::= I
48
/* syntactic domains of various sequences */
ObjNameSeq, SpltParamSeq, JoinParamSeq, BrnchNameSeq ::= ISeq
/* other syntactic domains */
DSCond ::= EMPTY | dscs
ISeq :: = EMPTY
| I ISeq
I ::= any valid LISP system name
FVal ::= a sequence of decimal digits prefixed by a period (valid float value)
B
An Example of a System Architectural Model
In this section, we give the syntax of an example System Architectural Model
of MAF editor system which is discussed in detail in [5]. In the following,
we give a brief detail on how to read the example, i.e. maf-editor is the
top level component of the application whose structural behavior is specified
at first. Every sentence of the specification is self-explanatory. In principle,
the behavior of every component in the subnetwork of a parent component
has to be specified separately with two corresponding parts, e.g. a component maf-startup (which is in the subnetwork of top-level component as
mentioned in :components clause) has
1. structural behavior as specified by clause
define-ensemble maf-startup
2. normal behavior as specified by the clause
defbehavior-model (maf-startup norml)
3. and a corresponding compromised behavior is specified by the clause
defbehavior-model (maf-startup compromised)
The former part corresponds to the control level of the specification while
the latter two corresponds to the behavioral level of the specification of the
component.
Furthermore, as explained in Section 2, the split behavior of the component maf-create-events is further specified with the corresponding clauses,
e.g.
defsplit maf-more-events?
Also, any pre/postcondition that is followed by the dscs specifies the data
structure consistency property.
49
B.1
MAF Editor Model
(define-ensemble maf-editor
:entry-events :auto
:inputs ()
:outputs (the-model)
:components ((startup :type maf-startup :models (normal compromised))
(create-model :type maf-create-model :models (normal compromised))
(create-events :type maf-create-events :models (normal compromised))
(save :type maf-save :models (normal compromised)))
:controlflows ((before maf-editor before startup)
(after startup before create-model))
:dataflows ((the-model create-model the-model create-events)
(the-model create-events the-model save)
(the-model save the-model maf-save-model))
:resources ((imagery image-file (normal .7) (hacked .3))
(code-files loadable-files (normal .8) (hacked .2)))
:resource-mappings ((startup imagery)
(create-model code-files)
(create-events code-files)
(save-model code-files))
:model-mappings ((startup normal ((imagery normal)) .99)
(startup compromised ((imagery normal)) .01)
(startup normal ((imagery hacked)) .9)
(startup compromised ((imagery hacked)) .1)
(create-model
(create-model
(create-model
(create-model
normal ((code-files normal)) .99)
compromised ((code-files normal)) .01)
normal ((code-files hacked)) .9)
compromised ((code-files hacked)) .1)
(create-events normal ((code-files normal)) .99)
(create-events compromised ((code-files normal)) .01)
(create-events normal ((code-files hacked)) .9)
(create-events compromised ((code-files hacked)) .1)
(save normal ((code-files normal)) .99)
(save compromised ((code-files normal)) .001)
50
(save normal ((code-files hacked)) .01)
(save compromised ((code-files hacked)) .999))
:vulnerabilities ((imagery reads-complex-imagery)
(code-files loads-code)
))
(define-ensemble maf-startup
:entry-events (startup)
:exit-events (startup)
:allowable-events (post-validate create-client-frame
center-action load-image)
:inputs ()
:outputs ())
(defbehavior-model (maf-startup normal)
:inputs ()
:outputs ()
:prerequisites ()
:post-conditions ())
(defbehavior-model (maf-startup compromised)
:inputs ()
:outputs ()
:prerequisites ()
:post-conditions ())
;;; Need defbehaviors for each of these even if its empty
(define-ensemble maf-create-model
:entry-events (create-mission-action-action-performed)
:exit-events (mission-builder-submit)
:allowable-events (create-mission-builder-with-client-panel
create-mission-builder
create-mission-builder-with-hash-table
mission-builder-submit
(set-initial-info exit (the-model nil))
create-mission-action-action-performed
retrieve-info
create-mission-action-action-performed
(set-initial-info entry)
)
:inputs ()
:outputs (the-model))
51
(defbehavior-model (maf-create-model normal)
:inputs ()
:outputs (the-model)
:prerequisites ()
:post-conditions ([dscs ?the-model mission-builder good])
)
(defbehavior-model (maf-create-model compromised)
:inputs ()
:outputs (the-model)
:prerequisites ()
:post-conditions ([not [dscs ?the-model mission-builder good]])
)
(define-ensemble maf-create-events
:entry-events :auto
:exit-events ()
:allowable-events ()
:inputs (the-model)
:outputs (the-model)
:components ((get-next-cmd :type maf-get-next-cmd :models (normal))
(get-event-info :type maf-get-event-info :models (normal compromised))
(add-event-to-model :type maf-add-event-to-model :models
(normal compromised))
(get-leg :type maf-get-leg :models (normal compromised))
(get-movement :type maf-get-movement :models (normal compromised))
(get-sortie :type maf-get-sortie :models (normal compromised))
(add-additional-info-to-model :type maf-add-additional-info :models
(normal compromised))
(continue :type maf-create-events :models (normal compromised)))
:dataflows ((the-model maf-create-events the-model join-exit-exit)
(the-model maf-create-events the-model add-event-to-model)
(the-cmd get-next-cmd cmd more-events?)
(the-event get-event-info the-event add-event-to-model)
(the-model add-event-to-model the-model join-events-non-take-off)
(the-event get-event-info event takeoff?)
(the-leg get-leg the-leg add-additional-info-to-model)
(lms-event-counter get-leg event-number add-additional-info-to-model)
(the-movement get-movement the-movement add-additional-info-to-model)
(the-sortie get-sortie the-sortie add-additional-info-to-model)
(the-model add-event-to-model the-model add-additional-info-to-model)
(the-model add-additional-info-to-model the-model join-events-take-off)
52
(the-model join-events the-model continue)
(the-model continue the-model join-exit-recur)
(the-model join-exit the-model maf-create-events)
)
:controlflows ((after more-events?-build-event before add-event-to-model)
(after more-events?-exit before join-exit-exit)
(after takeoff?-get-additional-info before get-leg)
(after takeoff?-get-additional-info before get-movement)
(after takeoff?-get-additional-info before get-sortie)
(after takeoff?-exit before join-events-non-take-off))
:splits ((more-events? maf-more-events? (cmd) (build-event exit))
(takeoff? maf-takeoff? (event) (get-additional-info exit)))
:joins ((join-events (the-model) (take-off non-take-off))
(join-exit (the-model) (recur exit)))
:resources ((code-files loadable-files (normal .8) (hacked .2)))
:resource-mappings ((get-event-info code-files)
(add-event-to-model code-files)
(get-leg code-files)
(get-movement code-files)
(get-sortie code-files)
(add-additional-info-to-model code-files)
(continue code-files))
:model-mappings ((get-event-info normal code-files normal .99)
(get-event-info compromised code-files normal .01)
(get-event-info normal code-files hacked .9)
(get-event-info compromised code-files hacked .1)
(add-event-to-model
(add-event-to-model
(add-event-to-model
(add-event-to-model
(get-leg
(get-leg
(get-leg
(get-leg
normal code-files normal .99)
compromised code-files normal .01)
normal code-files hacked .9)
compromised code-files hacked .1)
normal code-files normal .99)
compromised code-files normal .001)
normal code-files hacked .01)
compromised code-files hacked .999)
(get-movement normal code-files normal .99)
53
(get-movement compromised code-files normal .001)
(get-movement normal code-files hacked .01)
(get-movement compromised code-files hacked .999)
(get-sortie
(get-sortie
(get-sortie
(get-sortie
normal code-files normal .99)
compromised code-files normal .001)
normal code-files hacked .01)
compromised code-files hacked .999)
(add-additional-info-to-model
(add-additional-info-to-model
(add-additional-info-to-model
(add-additional-info-to-model
(continue
(continue
(continue
(continue
normal code-files normal .99)
compromised code-files normal .001)
normal code-files hacked .01)
compromised code-files hacked .999)
normal code-files normal .99)
compromised code-files normal .001)
normal code-files hacked .01)
compromised code-files hacked .999))
:vulnerabilities ((code-files loads-code))
)
(defbehavior-model (maf-create-events normal)
:inputs (the-model)
:outputs (the-model)
:prerequisites ([dscs ?the-model mission-builder good])
:post-conditions ([dscs ?the-model mission-builder good])
)
(defbehavior-model (maf-create-events compromised)
:inputs (the-model)
:outputs (the-model)
:prerequisites ([dscs ?the-model mission-builder good])
:post-conditions ([not [dscs ?the-model mission-builder good]])
)
(define-ensemble maf-get-next-cmd
:entry-events (next-cmd)
:exit-events ((next-cmd exit (the-cmd)))
:inputs ()
:outputs (the-cmd))
(defbehavior-model (maf-get-next-cmd normal)
:inputs ()
54
:outputs (the-cmd)
:prerequisites ()
:post-conditions ())
(define-ensemble maf-get-event-info
:entry-events (create-mission-event-point)
:allowable-events (set-current-point
(create-mission-event-point exit)
create-mission-event-object
meo-set-information
mpl-action-performed
close-form
add-new-event-internal)
:exit-events ((got-event-info exit (the-event)))
:inputs ()
:outputs (the-event))
(defbehavior-model (maf-get-event-info normal)
:inputs ()
:outputs (the-event)
:prerequisites ()
:post-conditions ([dscs ?the-event event good]))
(defbehavior-model (maf-get-event-info compromised)
:inputs ()
:outputs (the-event)
:prerequisites ()
:post-conditions ([not [dscs ?the-event event good]]))
(define-ensemble maf-add-event-to-model
:entry-events (update-msn-evt)
:allowable-events
((update-msn-evt exit (mb event-number event))
add-new-event-internal
create-new-additional-mission-info-panel
)
:exit-events (mpl-action-performed)
:inputs (the-event the-model)
:outputs (the-model event-number))
(defbehavior-model (maf-add-event-to-model normal)
:inputs (the-event the-model)
:outputs (the-model event-number)
:prerequisites ([dscs ?the-event event good]
55
[dscs ?the-model mission-builder good])
:post-conditions
([add-to-map (events ?the-model)?event-number ?the-event
?before-maf-add-event-to-model]
[dscs ?the-model mission-builder good]))
(defbehavior-model (maf-add-event-to-model compromised)
:inputs (the-event the-model)
:outputs (the-model event-number)
:prerequisites ([not [dscs ?the-event event good]]
[not [dscs ?the-model mission-builder good]])
:post-conditions
([dscs ?the-model mission-builder good]))
(define-ensemble maf-get-leg
:entry-events (retrieve-leg)
:exit-events ((retrieve-leg exit (nil the-leg lms-event-counter)))
:allowable-events (create-mission-leg-object mlo-set-information)
:inputs ()
:outputs (the-leg lms-event-counter))
(defbehavior-model (maf-get-leg normal)
:inputs ()
:outputs (the-leg lms-event-counter)
:prerequisites ()
:post-conditions ([dscs ?the-leg leg good]))
(defbehavior-model (maf-get-leg compromised)
:inputs ()
:outputs (the-leg lms-event-counter)
:prerequisites ()
:post-conditions ([not [dscs ?the-leg leg good]]))
(define-ensemble maf-get-movement
:entry-events (retrieve-movement)
:exit-events ((retrieve-movement exit (nil the-movement)))
:allowable-events
(create-mission-movement-object mmo-set-information)
:inputs ()
:outputs (the-movement))
(defbehavior-model (maf-get-movement normal)
:inputs ()
:outputs (the-movement)
56
:prerequisites ()
:post-conditions ([dscs ?the-movement movement good]))
(defbehavior-model (maf-get-movement compromised)
:inputs ()
:outputs (the-movement)
:prerequisites ()
:post-conditions ([not [dscs ?the-movement movement good]]))
(define-ensemble maf-get-sortie
:entry-events (retrieve-sortie)
:exit-events ((retrieve-sortie exit (nil the-sortie)))
:allowable-events
(create-mission-sortie-object mso-set-information)
:inputs ()
:outputs (the-sortie))
(defbehavior-model (maf-get-sortie normal)
:inputs ()
:outputs (the-sortie)
:prerequisites ()
:post-conditions ([dscs ?the-sortie sortie good]))
(defbehavior-model (maf-get-sortie compromised)
:inputs ()
:outputs (the-sortie)
:prerequisites ()
:post-conditions ([not [dscs ?the-sortie sortie good]]))
(define-ensemble maf-add-additional-info
:entry-events ((retrieve-sortie exit))
:exit-events (Mission-builder-add-info)
:inputs (the-model the-leg the-movement the-sortie event-number)
:outputs (the-model))
(defbehavior-model (maf-add-additional-info normal)
:inputs (the-model the-leg the-movement the-sortie event-number)
:outputs (the-model)
:prerequisites ([dscs ?the-leg leg good]
[dscs ?the-movement movement good]
[dscs ?the-sortie sortie good]
[dscs ?the-model mission-builder good])
:post-conditions ([add-to-map (legs ?the-model) ?event-number ?the-leg
?before-maf-add-additional-info]
57
[add-to-map (sorties ?the-model) ?event-number ?the-sortie
?before-maf-add-additional-info]
[add-to-map (movements ?the-model) ?event-number ?the-movement
?before-maf-add-additional-info]
[dscs ?the-model mission-builder good]))
(defbehavior-model (maf-add-additional-info compromised)
:inputs (the-model the-leg the-movement the-sortie event-number)
:outputs (the-model)
:prerequisites ([dscs ?the-leg leg good]
[dscs ?the-movement movement good]
[dscs ?the-sortie sortie good]
[dscs ?the-model mission-builder good])
:post-conditions ([not [dscs ?the-model mission-builder good]]))
(defsplit maf-more-events? (cmd)
(build-event (equal ?cmd ’new-event))
(exit (equal ?cmd ’save-mission)))
(defsplit maf-takeoff? (event)
(get-additional-info (take-off-event? ?event))
(exit (not (take-off-event? ?event))))
(define-ensemble maf-save
:inputs (the-model)
:outputs ())
(defbehavior-model (maf-save normal)
:inputs (the-model)
:outputs ()
:prerequisites ([dscs ?the-model mission-builder good])
:post-conditions ([dscs ?the-model mission-builder good]))
(defbehavior-model (maf-save compromised)
:inputs (the-model)
:outputs ()
:prerequisites ([dscs ?the-model mission-builder good])
:post-conditions ([not [dscs ?the-model mission-builder good]]))
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;;
;;; attack models
;;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
58
(define-attack-model maf-attacks
:attack-types ((hacked-image-file-attack .3) (hacked-code-file-attack .5))
:vulnerability-mapping ((reads-complex-imagery hacked-image-file-attack)
(loads-code hacked-code-file-attack)))
;;; rules mapping conditional probabilities of vulnerability and attacks
(defrule bad-image-file-takeover (:forward)
if [and [resource ?ensemble ?resource-name ?resource]
[resource-type-of ?resource image-file]
[resource-might-have-been-attacked ?resource hacked-image-file-attack]]
then [and [attack-implies-compromised-mode hacked-image-file-attack
?resource hacked .9 ]
[attack-implies-compromised-mode hacked-image-file-attack
?resource normal .1 ]])
(defrule bad-image-file-takeover-2 (:forward)
if [and [resource ?ensemble ?resource-name ?resource]
[resource-type-of ?resource code-memory-image]
[resource-might-have-been-attacked ?resource hacked-image-file-attack]]
then [and [attack-implies-compromised-mode hacked-image-file-attack
?resource hacked .9 ]
[attack-implies-compromised-mode hacked-image-file-attack
?resource normal .1 ]])
(defrule hacked-code-file-takeover (:forward)
if [and [resource ?ensemble ?resource-name ?resource]
[resource-type-of ?resource loadable-files]
[resource-might-have-been-attacked ?resource hacked-code-file-attack]]
then [and [attack-implies-compromised-mode hacked-code-file-attack
?resource hacked .9 ]
[attack-implies-compromised-mode hacked-code-file-attack
?resource normal .1 ]])
(defrule hacked-code-file-takeover-2 (:forward)
if [and [resource ?ensemble ?resource-name ?resource]
[resource-type-of ?resource loadable-files]
[resource-might-have-been-attacked ?resource hacked-code-file-attack]]
then [and [attack-implies-compromised-mode hacked-code-file-attack
?resource hacked .9 ]
[attack-implies-compromised-mode hacked-code-file-attack
?resource normal .1 ]])
59
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;;;;
;;;;;
Hacked Code file attacks
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(defrule bad-code-file-takeover (:forward)
if [and [resource ?ensemble ?resource-name ?resource]
[resource-type-of ?resource code-file]
[resource-might-have-been-attacked ?resource hacked-code-file-attack]]
then [and [attack-implies-compromised-mode hacked-code-file-attack
?resource hacked .9 ]
[attack-implies-compromised-mode hacked-code-file-attack
?resource normal .1 ]])
(defrule bad-code-file-takeover-2 (:forward)
if [and [resource ?ensemble ?resource-name ?resource]
[resource-type-of ?resource code-memory-image]
[resource-might-have-been-attacked ?resource hacked-code-file-attack]]
then [and [attack-implies-compromised-mode hacked-code-file-attack
?resource hacked .9 ]
[attack-implies-compromised-mode hacked-code-file-attack
?resource normal .1 ]])
60
| 6cs.PL
|
Noname manuscript No.
(will be inserted by the editor)
Discovering More Precise Process Models from Event
Logs by Filtering Out Chaotic Activities
Niek Tax · Natalia Sidorova · Wil M. P.
van der Aalst
arXiv:1711.01287v1 [cs.DB] 3 Nov 2017
Received: date / Accepted: date
Abstract Process Discovery is concerned with the automatic generation of a
process model that describes a business process from execution data of that
business process. Real life event logs can contain chaotic activities. These activities
are independent of the state of the process and can, therefore, happen at rather
arbitrary points in time. We show that the presence of such chaotic activities
in an event log heavily impacts the quality of the process models that can be
discovered with process discovery techniques. The current modus operandi for
filtering activities from event logs is to simply filter out infrequent activities. We
show that frequency-based filtering of activities does not solve the problems that
are caused by chaotic activities. Moreover, we propose a novel technique to filter
out chaotic activities from event logs. We evaluate this technique on a collection
of seventeen real-life event logs that originate from both the business process
management domain and the smart home environment domain. As demonstrated,
the developed activity filtering methods enable the discovery of process models
that are more behaviorally specific compared to process models that are discovered
using standard frequency-based filtering.
Keywords Information Systems · Business Process Intelligence · Process Mining ·
Knowledge Discovery
1 Introduction
Process Mining [1] is a scientific discipline that bridges the gap between process
analytics and data analysis and focuses on the analysis of event data logged during
the execution of a business process. Events contain information on what was done,
by whom, for whom, where, when, etc. Such event data is often readily available
from information systems such as ERP, CRM, or BPM systems. Process discovery,
which plays a prominent role in process mining, is the task of automatically
Department of Mathematics and Computer Science
Eindhoven University of Technology
P.O. Box 513, 5600MB Eindhoven, The Netherlands
Email: E-mail: {n.tax,n.sidorova,w.m.p.v.d.aalst}@tue.nl
2
Niek Tax et al.
generating a process model that accurately describes a business process based on
such event data. Many process discovery techniques have been developed over the
last decade (e.g. [7, 15, 16, 17, 20, 33, 47]), producing process models in various
forms, such as Petri nets [29], process trees [7], and BPMN models [30].
Figure 1b shows an example process model from [1] that describes a compensation request process. The process model consists of eight process steps (called
activities): (A) register request, (B) examine thoroughly, (C) examine casually, (D)
check ticket, (E) decide, (F) re-initiate request, (G) pay compensation, and (H)
reject request. Figure 1a shows a small example event log consisting of six execution
trails of the process model. The Inductive Miner [20] process discovery algorithm
provides the guarantee that it can re-discover the process model from an event log
given that all pairs of activities that can directly follow each other in the process
are present in the event log, i.e., the log is directly-follows complete. Since the log
in Figure 1a is directly-follows complete, applying the Inductive Miner to this log
results in the process model in Figure 1b, which generated the log.
However, the presence of activities that can occur spontaneously at any point in
the process execution, which we will call chaotic activities, substantially impacts the
quality of the resulting process models obtained with process discovery techniques.
Figure 2a contains the event log obtained from the one in Figure 1a by adding
activity (X) the customer calls at random points, since customers can call the call
center multiple times at any point in time during the execution of the process.
Figure 2b shows the resulting process model discovered by the Inductive Miner
[20] from the event log of Figure 2a. The process model discovered from the “clean”
example log without activity X (Figure 1b) was very simple, interpretable, and
accurate with respect to the behavior allowed in the process. In contrast, the
process model discovered from the log containing X (Figure 2b) is very complex,
hard to interpret, and it overgeneralizes by allowing for too much behavior that
is not possible in the process. We consider X to be a so-called chaotic activity
because it does not have a clear position in the process model and it complicates
the discovery of the rest of the process. The reason for the decline in the quality
of process models discovered from logs with chaotic activities is that the directly
follows relations, which many process discovery algorithms operate on, are affected
by chaotic activities. Examples of such process discovery algorithms include the
Inductive Miner [19], the Heuristics Miner [45], and Fodina [4]. In a sequence of
activities h. . . , A, C, . . . i, where A was directly followed by C, the addition of a
chaotic activity X can turn the sequence into h. . . , A, X, C, . . . i, thereby obfuscating
the directly-follows relation between activities A and C.
Event sequences
hA,B,D,E,Hi
hA,D,C,E,Gi
hA,C,D,E,F,B,D,E,Gi
hA,D,B,E,Hi
hA,C,D,E,F,D,C,E,F,C,D,E,Hi
hA,C,D,E,Gi
B
G
A
C
D
E
H
F
(b)
(a)
Fig. 1 (a) Event log with A=register request, B=examine thoroughly, C=examine casually,
D=check ticket, E=decide, F=re-initiate request, G=pay compensation, H=reject request, and
(b) the Petri net mined from this log with the Inductive Miner [20].
Title Suppressed Due to Excessive Length
3
D
Event sequences
hX,A,B,D,E,Hi
hA,D,C,E,Gi
hA,X,C,D,E,F,B,D,X,E,Gi
hA,D,B,E,Hi
hA,C,D,E,F,D,C,E,F,C,D,E,X,Hi
hA,C,X,D,E,Gi
E
X
B
F
A
G
C
H
(a)
(b)
Fig. 2 (a) The event log from Figure 1a with an added chaotic activity X, and (b) the Petri
net mined from this log with the Inductive Miner [20].
In this paper, we show that existing approaches do not solve the problem of
chaotic activities and we present a technique to handle the issue. This paper is
structured as follows: in Section 2 we introduce basic concepts used throughout
the paper. In Section 3 we propose an approach to filter out chaotic activities.
In Section 4 we evaluate our technique using synthetic data where we artificially
insert chaotic activities and check whether the filtering techniques can filter out
the inserted chaotic activities. Additionally, Section 4 proposes a methodology
to evaluate activity filtering techniques in a real-life setting where there is no
ground truth knowledge on which activities are truly chaotic, and motivates this
methodology by showing that its results are consistent with the synthetic evaluation
on the synthetic datasets. In Section 5 the results on a collection of seventeen
real-life event logs are discussed. In Section 6 we discuss how the activity filtering
techniques can be used in a toggle-based approach for human-in-the-loop process
discovery. In Section 7 we discuss related techniques in the domains of process
discovery and the filtering of event logs. Section 8 concludes this paper and discusses
several directions for future work.
2 Preliminaries
In this section, we introduce concepts and notation throughout this paper.
X = {a1 , a2 , . . . , an } denotes a finite set. P(X) denotes the power set of X,
i.e., the set of all possible subsets of X. X\Y denotes the set of elements that
are in set X but not in set Y , e.g., {a, b, c}\{a, c}={b}. X ∗ denotes the set of all
sequences over a set X and σ = ha1 , a2 , . . . , an i denotes a sequence of length n,
with σ(i) = ai and hi the empty sequence. σX is the projection of σ on X, e.g.
ha, b, c, a, b, ci{a,c} = ha, c, a, ci. σ1 · σ2 denotes the concatenation of sequences σ1
and σ2 , e.g., ha, b, ci · hd, ei = ha, b, c, d, ei.
A partial function f ∈X9Y with domain dom(f ) can be lifted to sequences
over X using the following recursive definition: (1) f (hi) = hi; (2) for any σ∈X ∗
and x ∈ X:
4
Niek Tax et al.
f (σ · hxi) =
f (σ)
if x∈dom(f
/
),
f (σ) · hf (x)i if x∈dom(f ).
A multiset (or bag) over X is a function B : X→N which we write as
w2
wn
+
1
[aw
1 , a2 , . . . , an ], where for 1≤i≤n we have ai ∈X and wi ∈N . The set of all
multisets over X is denoted B(X).
In the context of process mining, we assume the set of all process activities Σ
to be given. Event logs consist of sequences of events where each event represents
a process activity.
Definition 1 (Event, Trace, and Event Log) An event e in an event log is
the occurrence of an activity e∈Σ. We call a (non-empty) sequence of events σ∈Σ +
a trace. An event log L∈B(Σ + ) is a multiset of traces.
L=[ha, b, ci2 , hb, a, ci3 ] is an example event log over process activities Σ = {a, b, c},
consisting of 2 occurrences of trace ha, b, ci and three occurrences of trace hb, a, ci.
Activities(L) denotes the set of process activities Σ that occur in L, e.g., Activities(L) =
{a, b, c}. #(a, L) denotes the number of occurrences of activity a in log L, e.g.,
#(a, L) = 5.
A process model notation that is frequently used in the area of process mining
is the Petri net. Petri nets can be automatically transformed into process model
notations that are commonly used in business environments, such as BPMN
and BPEL [23]. A Petri net is a directed bipartite graph consisting of places
(depicted as circles) and transitions (depicted as rectangles), connected by arcs. A
transition describes an activity, while places represent the enabling conditions of
transitions. Labels of transitions indicate the type of activity that they represent.
Unlabeled transitions (τ -transitions) represent invisible transitions (depicted as
gray rectangles), which are only used for routing purposes and are not recorded in
the event log.
Definition 2 (Labeled Petri net) A labeled Petri net N = hP, T, F, `i is a tuple
where P is a finite set of places, T is a finite set of transitions such that P ∩T =∅,
F ⊆(P ×T )∪(T ×P ) is a set of directed arcs, called the flow relation, and `:T 9Σ is
a partial labeling function that assigns a label to a transition, or leaves it unlabeled
(the τ -transitions).
We write •n and n• for the input and output nodes of n ∈ P ∪ T (according
to F ). A state of a Petri net is defined by its marking m∈B(P ) being a multiset
of places. A marking is graphically denoted by putting m(p) tokens on each place
p∈P . State changes occur through transition firings. A transition t is enabled (can
fire) in a given marking m if each input place p∈•t contains at least one token.
Once t fires, one token is removed from each input place p∈•t and one token is
added to each output place p0 ∈t•, leading to a new marking m0 =m− •t + t•.
A firing of a transition t leading from marking m to marking m0 is denoted as
t
step m−→m0 . Steps are lifted to sequences of firing enabled transitions, written
γ
0
m−→m and γ∈T ∗ is a firing sequence.
Defining an initial and a set of final markings allows defining the language
accepted by a Petri net as a set of finite sequences of activities.
Definition 3 (Accepting Petri Net) An accepting Petri net is a triplet APN =
(N, m0 , MF ), where N is a labeled Petri net, m0 ∈B(P ) is its initial marking, and
Title Suppressed Due to Excessive Length
5
MF ⊆B(P ) is its set of possible final markings. A sequence σ∈Σ ∗ is a trace of
γ
an accepting Petri net APN if there exists a firing sequence m0 −→mf such that
∗
mf ∈MF , γ∈T and `(γ)=σ.
In the Petri nets that are shown in this paper, places that belong to the initial
marking contain a token and places belonging to a final marking contain a bottom
right label fi with i a final marking identifier or are simply marked as
in case of
a single final marking.
The language L(APN ) is the set of all its traces, i.e., L(APN ) = {l(γ)|
γ
γ∈T ∗ ∧∃mf ∈M F m0 −→mf }, which can be of infinite size when APN contains
loops. While we define the language for accepting Petri nets, in theory, L(M ) can
be defined for any process model M with formal semantics. We denote the universe
of process models as M. For each M ∈M, L(M ) ⊆ Σ + is defined.
A process discovery method is a function PD : B(Σ + ) → M that provides a
process model for a given event log. The goal is to discover a process model that is
a good description of the process from which the event log was obtained, i.e., it
should allow for all the behavior that was observed in the event log (called fitness)
while it should not allow for too much behavior that was not seen in the event log
(called precision). For an event log L, L̃={σ∈Σ + |L(σ)>0} is the trace set of L.
For example, for log L=[ha, b, ci2 , hb, a, ci3 ], L̃={ha, b, cihb, a, ci}. For an event log
L and a process model M , we say that L is fitting on M if L̃⊆L(M ). Precision is
related to the behavior that is allowed by a model M that was not observed in the
event log L, i.e., L(M )\L̃.
3 Information-Theoretic Approaches to Activity Filtering
We consider a chaotic activity to be an activity that can occur at any point in
the process and that thereby complicates the discovery of the rest of the process
by obfuscating the directly-follows relations of the event log. In this section, we
propose a technique to detect chaotic activities in event logs and to filter them out
from those event logs.
We extend the function #(a, L) to the function #(σ, L) to count the number
of occurrence
P of a sequence σ, in L:
#(σ, L)= σ0 ∈L |{0≤i≤|σ 0 |−|σ| ∀1≤j≤|σ| σ 0 (i+j)=σ(j)}|.
The directly-follows ratio, denoted dfr (a, b, L), represents the ratio of the events
of activity a that are directly followed by an event of activity b in event log L, i.e.,
dfr (a, b, L)= #(ha,bi,L)
#(a,L) .
Likewise, the directly-precedes ratio, denoted dpr (a, b, L), represents the ratio
of the events of activity a that are directly preceded by an event of activity b in
event log L, i.e., dpr (a, b, L)= #(hb,ai,L)
#(a,L) .
Lc contains the traces of event log L appended with an artificial end event that
we represent with c. For each σ = he1 , e2 , . . . , en i in log L, log Lc contains a trace
σ c = he1 , e2 , . . . , en , ci. Likewise, Lb contains the traces of event log L prepended
with an artificial start event b, i.e., for each σ = he1 , e2 , . . . , en i in log L, log Lb
contains a trace σ b = hb, e1 , . . . , en i. The artificial start and end events allow us
to define the ratio of start events of an activity, e.g., dfr (a, c, Lc ) and dpr (a, b, Lb )
6
Niek Tax et al.
represent the ratio of events of activity a that respectively occur at the end of a
trace and at the beginning of a trace.
Assuming an arbitrary but consistent order over the set of process activities Activities(L), dfr (a, L) represents the vector of values dfr (a, b, Lc ) for all
b∈Activities(L) ∪ {c} and dpr (a, L) represents the vector of values dpr (a, b, Lb )
for all b ∈ Activities(L) ∪ {b}. From a probabilistic point of view, we can regard
dfr (a, L) and dpr (a, L) as the empirical estimates of the categorical distributions
over respectively the activities directly prior to a and directly after a, where the
empirical estimates are based on #(a, L) trials.
3.1 Direct Entropy-based Activity Filtering
We define the entropy of an activity in an event log L based on its directly-follows
ratio vector and the directly-precedes ratio vector by using the
Pusual definition of
function for the categorical probability distribution: H(X) = − x∈X x log2 (x). We
define the entropy of activity a ∈ Activities(L) in log L as: H(a, L) = H(dfr (a, L))+
H(dpr (a, L)). In case there are zero probability values in the directly follows or
directly precedes vectors, i.e., 0 ∈ dfr (a, L) ∨ 0 ∈ dpr (a, L), then the value of the
corresponding summand 0 log2 (0) is taken as 0, which is consistent with the limit
lim p log2 (p) = 0.
p→0+
For example, let event log L = [ha, b, c, xi10 , ha, b, x, ci10 , ha, x, b, ci10 ], then
20
10
, 0, 30
, 0i, using the arbitrary but consistent ordering ha, b, c, x, ci,
dfr (a, L) = h0, 30
indicating that 20 out of 30 events of activity a are followed by b and 10 out of 30
by x. Likewise dpr (a, L)=h0, 0, 0, 0, 1i, using the arbitrary but consistent ordering
ha, b, c, x, bi, indicating that all events of activity a are preceded by b. This leads
to H(dfr (a, L)) = 0.918, H(dpr (a, L)) = 0, and H(a, L) = 0.918. Furthermore,
H(b, L) = 1.837, H(c, L) = 1.837, and H(x, L) = 3.170, showing that activity x has
the highest entropy of the probability distributions for preceding and succeeding
activities. We conjecture that activities that are chaotic and behave randomly to a
high degree have high values of H(a, L).
Algorithm 1 An activity filtering approach based on entropy.
Input: event log L
Output: list of event logs Q
Initialisation :
0
1: L = L
0
2: Q = hL i
Main Procedure:
0
3: while |Activities(L )| > 2 do
4:
acts = Activities(L0 )
5:
a0 = arg maxa∈acts H(a, L0 )
6:
L0 = L0 acts\{a0 }
7:
Q = Q · hL0 i
8: end while
9: return Q
Title Suppressed Due to Excessive Length
7
Algorithm 1 describes a greedy approach to iteratively filter the most randomly
behaving (chaotic) activity from the event log. The algorithm takes an event log L
as input and produces a list of event logs, such that the first element of the list
contains a version of L with one activity filtered out, and each following element of
the list has one additional activity filtered out compared to the previous element.
In the example event log L, Algorithm 1 starts by filtering out activity x,
followed by activity b or c. The algorithm stops when there are two activities left in
the event log. The reason not to filter any more activities past this point is closely
related to the aim of process discovery: uncovering relations between activities.
From an event log with less than two activities no relations between activities can
be discovered.
3.2 The Entropy of Infrequent Activities and Laplace Smoothing
We defined entropy of the activities in an event log L is based on the directly-follows
ratios dfr and the directly-precedes ratios dpr of the activities in L. The empirical
estimates of the categorical distributions dfr (a, L) and dpr (a, L) become unreliable
for small values of #(a, L). In the extreme case, when #(a, L)=1, dfr (a, L) assigns
an estimate of 1 to the activity that the single activity a in L happens to be
preceded by and contains a probability of 0 for the other activities. Likewise, when
#(a, L)=1, dpr (a, L) assigns value 1 to one activity and value 0 to all others.
Therefore, #(a, L)=1 leads to H(dfr (a, L))=0 and H(dfr (a, L))=0. This shows an
undesirable consequence of Algorithm 1, infrequent activities are unlikely to be
filtered out. In the extreme case, the activities that occur only once, which are the
last in line activities to be filtered out. This effect is undesired, as very infrequent
activities should not be the primary focus of the process model discovered from an
event log.
We aim to mitigate this effect by applying Laplace smoothing [48] to the
empirical estimate of the categorical distributions over the preceding and succeeding
activities. Therefore, we define a smoothed version of the directly-follows and
α + #(ha,bi,L)
directly-precedes ratios, dfr s (a, b, L)= α(|Activities(L)|+1)+#(a,L)
, with smoothing
parameter α∈R≥0 . The value of dfr s (a, b, L) will always be between the empirical
1
estimate dfr (a, b, L) and the uniform probability |Activities(L)|+1
, depending on
s
the value α. Similar to dfr and dpr , dfr (a, L) represents the vector of values
dfr s (a, b, Lc ) for all b∈Activities(L) ∪ {c} and dpr s (a, L) represents the vector
of values dpr s (a, b, Lb ) for all b ∈ Activities(L) ∪ {b}. From a Bayesian point
of view, Laplace smoothing corresponds to the expected value of the posterior
distribution that consists of the categorical distribution given by dfr (a, L) and a
Dirichlet distributed prior that assigns equal probability to each of the possible
number of next activities |Activities(L)| + 1 (including c). Parameter α indicates
the weight that is assigned to the prior belief w.r.t. the evidence that is found in
the data. An alternative definition of the entropy of log L, based on the smoothed
distributions over the preceding and succeeding activities, is as follows: H s (a, L) =
H(dfr s (a, L)) + H(dpr s (a, L)). The smoothed direct entropy-based activity filter is
similar to Algorithm 1, where function H in line 5 of the algorithm is replaced by
H s . Function H(a, L) starts from the assumption that an activity is non-chaotic
unless we see sufficient evidence in the data for it’s chaoticness, function H s (a, L)
8
Niek Tax et al.
Algorithm 2 An indirect activity filtering approach based on entropy.
Input: event log L
Output: list of event logs Q
Initialisation :
0
1: L = L
0
2: Q = hL i
Main Procedure:
0
3: while |Activities(L )| > 2 do
4:
acts = Activities(L0 )
5:
a0 = arg mina∈acts H(L0 acts\{a} )
6:
L0 = L0 acts\{a0 }
7:
Q = Q · hL0 i
8: end while
9: return Q
in contrast starts from the assumption that is is chaotic, unless we see evidence
sufficient evidence in the data for it’s non-chaoticness.
Categorical distribution dfr (a, L) consists of |Activities(L)| + 1, therefore, the
maximum entropy of an activity decreases as more activities get filtered out of
the event log. The keep the values of H s (a, L) comparable between iterations of
the filtering algorithm, we propose to gradually increase the weight of the prior by
1
setting weight parameter α to |Activities(L)|
.
3.3 Indirect Entropy-based Activity Filtering
An alternative approach to the method proposed in Algorithm 1 is to filter out
activities such that the other activities in the log become less chaotic. We define
the total entropy of an
Pevent log L as the sum of the entropies of the activities in
the log, i.e., H(L) = a∈Activities(L) H(a, L).
Algorithm 2 describes a greedy approach that iteratively filters out the activity
that results in the lowest total log entropy. We call this approach the indirect
entropy-based activity filter, as opposed to the direct entropy-based activity filter
(Algorithm 1), which selects the to-be-filtered activity directly based on the activity
entropy, instead of based on the total log entropy after removal.
3.4 An Indirect Entropy-based Activity Filter with Laplace Smoothing
Just like the direct entropy-based activity filter, the indirect entropy-based activity
filter is sensitive to infrequent activities. To deal with this problem, the ideas of
the indirect entropy-based activity filtering method and Laplace smoothing can be
combined, using the following definition for smoothed log entropy:
P
H s (L) = a∈Activities(L) H s (a, L).
The algorithm for indirect entropy-based activity filtering with Laplace smoothing is identical to Algorithm 2, in which function H in line 5 is replaced by function
H s.
Title Suppressed Due to Excessive Length
(2)
9
Log with Frequent
Randomly-Positioned
Activities
insert k frequent
randomly-positioned activities
(3)
Filtered Event Log
filter activities
k = 1, 2, . . .
Process Model
Synthetic
(1) generate Event Log
(2)
insert k infrequent
Log
randomly-positioned activities
with Infrequent
Randomly-Positioned
Activities
(3)
Filtered Event Log
filter activities
k = 1, 2, . . .
(2)
insert k
uniform
randomly-positioned activities
(4)
Log with Uniform
Randomly-Positioned
Activities
(3)
Filtered Event Log
filter activities
k = 1, 2, . . .
compare
Fig. 3 An overview of the proposed evaluation methodology on synthetic data.
4 Evaluation using Synthetic Data
In this section we evaluate the activity filtering techniques using synthetic data.
Figure 3 gives an overview of the evaluation methodology. First, as step (1), we
generate a synthetic event log from a process model such that we know that
all activities of this model are non-chaotic. We take well-known process models
introduced by Maruster et al. [27], which respectively consist of 12 and 22 activities
and are commonly referred to as the Maruster A12, A22 models. The Maruster
A12 and A22 models are shown respectively in Figures 4a and 5a. We generated
25 traces by simulation from Maruster A12 to form log LA12 and generated 400
traces from Maruster A22 to form log LA22 . Then, in step (2), we artificially
insert activities that we position at random positions in the log. Since we chose the
positions in the log of those activities randomly, we assume those activities to be
chaotic. We vary the number (k) of randomly-positioned activities that we insert,
to assess how well the chaotic activity filtering techniques are able to deal with
different numbers of randomly-positioned activities in the event log. Furthermore,
we vary the frequency of the randomly-positioned activities that we insert, where
we distinguish between three types of randomly-positioned activities:
Frequent randomly-positioned activities the number of events inserted for
all k randomly-positioned activities is maxa∈Activities(L) #(a, L).
Infrequent randomly-positioned activities the number of events inserted for
all k randomly-positioned activities is mina∈Activities(L) #(a, L).
Uniform randomly-positioned activities for each of the k inserted randomlypositioned activities the frequency is chosen at randomly from a uniform probability distribution with minimum value mina∈Activities(L) #(a, L) and maximum
value maxa∈Activities(L) #(a, L).
In step (3) we filter out all the inserted randomly-positioned activities from the
event log, by removing activities one-by-one using the activity filtering approaches,
until all k artificially inserted activities have been removed again. We then count how
many of the activities that were originally in the process model we also removed
during this procedure (step (4)). Using this approach, we compare the direct
entropy-based activity filtering approach (with and without Laplace smoothing)
10
Niek Tax et al.
(a)
(b)
(c)
Fig. 4 (a) The synthetic process model Maruster A12, from which we generate an event log L,
consisting of 25 traces, from which the process model can be rediscovered with the Inductive
Miner [19], (b) the process model discovered by the Inductive Mining when we insert one
uniform randomly-positioned activity X to LA12 , and (c) the process model discovered by the
Inductive Miner after inserting a second randomly-positioned activity Y to LA12 .
with the indirect entropy-based activity filtering approach (with and without
Laplace smoothing). Furthermore, we compare those activity filtering techniques
with activity filtering techniques that are based on the frequency of activities,
such as filtering out the activities starting from the least frequent activity (leastfrequent-first), or starting from the most frequent activity (most-frequent-first).
Frequency-based activity filtering techniques are the current default approach for
filtering activities from event logs.
The original process models A12 and A22 can be rediscovered from generated
event logs LA12 and LA22 with the Inductive Miner [19] when there are no added
randomly-positioned activities. Figure 4b shows the process model discovered by
the Inductive Miner [19] after inserting one uniform randomly-positioned activity,
Title Suppressed Due to Excessive Length
11
(a)
(b)
Fig. 5 (a) The synthetic process model Maruster A22, from which we generate an event
log LA22 , consisting of 400 traces, from which the process model is re-discoverable with the
Inductive Miner [19], and (b) the process model discovered by the Inductive Miner after inserting
a uniform randomly-positioned activity X to LA22 .
activity X, into LA12 . The insertion of activity X causes the Inductive Miner to
create a model that overgeneralizes the behavior of the event log, as indicated by
many silent transitions in the process model that allow activities to be skipped.
Adding a second uniform randomly-positioned activity Y to LA12 results in the
Inductive Miner discovering a process model (shown in Figure 4c) that overgeneralizes even further, allowing for almost all sequences over the set of activities. Figure
5b shows the process model discovered by the Inductive Miner after inserting two
uniform randomly-paced activities (X and Y ) into LA22 . The addition of X and
Y has the effect that activity C is no longer positioned at the correct place in the
process model, but it is instead put in parallel to the whole process, making the
process model overly general, as it wrongly allows for activity C to occur before A
and B, or after D, E, F , and G. Figures 4b, 4c and 5b further motivate the need
for filtering out chaotic activities.
12
Niek Tax et al.
Table 1 The number of incorrectly filtered activities per filtering approach on LA12 and LA22
with k added Uniform (U) / Frequent (F) /Infrequent (I) chaotic activities.
Approach
Direct
1
Direct (α= |A|
)
Indirect
1
Indirect (α= |A|
)
Least-frequent-first
Most-frequent-first
1
Maruster A12 (Number of inserted randomly-positioned activities →)
4
8
16
32
64
128
U
F
I
U
F
I
U
F
I
U
F
I
U
F
I
U
F
I
U
F
I
U
F
I
0
0
0
0
9
11
0
0
0
0
12
0
0
0
0
0
0
12
0
0
0
0
11
3
0
0
0
0
12
0
0
0
0
0
0
12
0
0
0
0
6
7
0
0
0
0
12
0
0
0
0
0
0
12
0
0
0
0
11
10
0
0
0
0
12
0
0
0
1
1
0
12
0
0
1
1
11
12
0
0
0
0
12
0
0
0
1
1
0
12
0
0
1
1
12
12
0
0
0
0
12
0
12
0
1
1
0
12
4
4
2
2
12
12
0
0
0
0
12
0
12
6
1
1
0
12
10
6
3
2
12
12
1
2
1
1
12
0
12
12
6
10
0
12
2
Maruster A22 (Number of inserted randomly-positioned activities →)
4
8
16
32
64
U
F
I
U
F
I
U
F
I
U
F
I
U
F
I
U
F
I
U
F
I
U
F
I
0
0
0
0
16
7
0
0
0
0
22
0
0
0
0
0
0
22
0
0
0
0
17
8
0
0
0
0
22
0
0
0
0
0
0
22
0
0
0
0
6
19
0
0
0
0
22
0
0
0
0
0
0
22
0
0
0
0
21
17
0
0
0
0
22
0
1
1
1
1
0
22
0
0
0
0
19
19
0
0
0
0
22
0
0
0
1
1
0
22
0
0
1
1
22
22
0
0
0
0
22
0
0
0
1
1
0
22
0
0
1
0
22
22
0
0
0
0
22
0
0
0
1
1
0
22
0
0
1
1
22
22
0
0
0
0
22
0
5
5
1
1
0
22
Approach
Direct
1
Direct (α= |A|
)
Indirect
1
Indirect (α= |A|
)
Least-frequent-first
Most-frequent-first
2
1
128
Frequent randomly-positioned activities will impact the quality of process models discovered with process discovery to a higher degree than infrequent randomlypositioned activities. Each randomly-positioned activity that is inserted at a random
position in the event log is placed in-between two existing events in that log (or at the
start or end of the trace). By inserting randomly-positioned activity X in-between
two events of activities A and C respectively, the directly-follows relation between
activities A and C gets weakened. Therefore, the impact of randomly-positioned
activity X is proportional to its frequency #(X, L).
4.1 Results
Table 1 reports the number of activities that were originally part of the synthetic
process models A12 and A22 that were wrongly filtered out from LA12 and LA22
as an effect of removing all inserted randomly-positioned activities from these
logs. If this number is 12 for Maruster A12 or 22 for Maruster A22 this indicates
that all activities of the original process model needed to be filtered out before
the activity filtering technique was able to remove all inserted chaotic activities.
The results show that the direct filtering approach can perfectly distinguish between actual activities from the process and artificial chaotic activities for up
to 32 uniform randomly-positioned activities inserted activities to LA12 , up to
64 frequent randomly-positioned activities, and up to 16 infrequent randomlypositioned activities. Infrequent randomly-positioned activities are the hardest type
of randomly-positioned activities to correctly filter out, as their infrequency can
have the effect that the probability distributions over their surrounding activities
1
can by chance have low entropy. Using Laplace smoothing with α = |Activities(L)|
mitigates this effect, but does not completely solve it: the number of incorrectly
removed activities drops from 12 to 0 as an effect of Laplace smoothing for 32
added randomly-positioned activities, and from 12 to 6 for 64 added randomlypositioned activities. The indirect activity filter starts making errors already at
lower numbers of added randomly-positioned activities than the direct activity
filter; however, it is more stable to errors for higher numbers of added randomlypositioned activities, i.e., fewer activities get incorrectly removed for 64 and 128
Title Suppressed Due to Excessive Length
13
added randomly-positioned activities. In contrast to direct activity filtering, Laplace
smoothing does not seem to reduce the number of wrongly removed activities for
indirect activity filtering. In fact, surprisingly, the number of incorrectly removed
activities even increased from 6 to 10 as an effect of using Laplace smoothing for 128
infrequent randomly-positioned activities added to LA12 . The direct and indirect
filtering approaches, both with and without Laplace smoothing, outperform the
currently widely used approach of filtering out infrequent activities from the event
log (least-frequent-first filtering). Furthermore, a second frequency-based activity
filtering technique is included in the evaluation in which the most-frequent activities
are removed from the event log (most-frequent first filtering). Both Frequency-based
filtering approaches are not able to filter out the randomly-positioned activities
inserted to LA12 and LA22 , even for small numbers of added randomly-positioned
activities.
4.2 An Evaluation Methodology for Event Data without Ground Truth
Information
In a real-life data evaluation that we perform in the following section, there is
no ground truth knowledge on which activities of the process are chaotic. This
motivates a more indirect evaluation in which we evaluate the quality of the process
model discovered from the event log after filtering out activities with the proposed
activity filtering techniques. In this section we propose a methodology for evaluation
of activity filtering techniques by assessing the quality of discovered process models,
we apply this evaluation methodology to the Maruster A12 and Maruster A22
event logs, and we discuss the agreement between the findings of Table 1 and the
quality of the discovered process models.
There are several ways to quantify the quality of a process model for an event
log. Ideally, a process model M should allow for all behavior that was observed in
the event log L, i.e., L̃ \ L(M ) should be as small as possible, preferably empty.
The fitness quality dimension covers this. Furthermore, model M should not allow
for too much additional behavior that was not seen in the event log, i.e., L(M ) \ L̃
should be as small as possible. This aspect is called precision. For each process
model that we discovered, we measure fitness and precision with respect to the
filtered log. Fitness is measured using the alignment-based fitness measure [3] and
we measure precision using negative event precision [44]. Based on the fitness and
precision results we also calculate F-score [11], i.e., the harmonic mean between
fitness and precision.
Precision is likely to increase by filtering out one or more activities from an
event log independently of which activities are removed from the log, as a result of
two factors. First, precision measures express L(M ) \ L̃ in terms of the number
of activities that are enabled at certain points in the process, w.r.t. the number
of activities seen that were actually observed at these points in the process. With
the log and model containing fewer activities after filtering, the number of enabled
activities is likely to decrease as well. Secondly, activity filtering leads to log L0 that
contains less behavior than original log L (i.e., L̃0 is smaller than L̃), this makes it
easier for process discovery methods to discover a process model with less behavior.
These two factors make precision values between event logs with different numbers
of activities filtered out incomparable. The degree to which the behavior of filtered
14
Niek Tax et al.
1.00
Maruster A12 − 0 added
Maruster A12 − 1 added
Maruster A12 − 2 added
0.75
0.50
0.25
0.00
0.00
F−score
1.00
0.25
0.50
0.75
0.00
0.25
Maruster A12 − 4 added
0.50
0.75
0.00
Maruster A12 − 8 added
0.25
0.50
0.75
Maruster A12 − 16 added
0.75
0.50
0.25
0.00
0.00
1.00
0.25
0.50
0.75
0.00
0.25
0.50
0.75
0.00
0.25
Maruster A12 − 32 added
0.50
0.75
Filter
0.75
Direct
0.50
Direct (a=1/|A|)
Indirect
0.25
0.00
0.00
Indirect (a=1/|A|)
0.25
0.50
0.75
Least−frequent−first
Minimum % of activities explained
Fig. 6 F-score on the log generated from the Maruster A12 model with inserted artificial
chaotic activities.
Maruster A12 − 0 added
Maruster A12 − 1 added
5.0
0.5
Nondeterminism
0.0
0.00
6
2.5
0.25
0.50
0.75
0.25
0.50
0.75
Maruster A12 − 8 added
16
12
8
2
25
20
15
10
5
0
0.00
0.0
0.00
Maruster A12 − 4 added
4
0
0.00
Maruster A12 − 2 added
7.5
1.0
4
0.25
0.50
0.75
0
0.00
0.25
0.50
0.75
4
3
2
1
0
0.00
25
20
15
10
5
0
0.00
0.25
0.50
0.75
Maruster A12 − 16 added
0.25
0.50
0.75
Filter
Maruster A12 − 32 added
Direct
Direct (a=1/|A|)
Indirect
Indirect (a=1/|A|)
0.25
0.50
0.75
Least−frequent−first
Minimum % of activities explained
Fig. 7 Nondeterminism on the log generated from the Maruster A12 model with inserted
artificial chaotic activities.
log L0 decreases w.r.t. an unfiltered log L depends on the activities that are filtered
out: when very chaotic activities are filtered from L the behavior decreases much
more than when very structured activities are filtered from L. One effect of this
is that too much behavior in a process model affects the precision of that model
more for the log from which the non-chaotic activities are filtered out than for the
log from which the chaotic activities are filtered out.
To measure the behavior allowed by the process model independent of which
activities are filtered from the event log is to determine the average number of
enabled activities when replaying the traces of the log on the model. To deal
with traces of the event log that do not fit the behavior of the process model,
we calculate alignments [3] between log and model. Alignments are a function
Γ m : M × Σ + → B(P )+ that map each trace from the event log to a sequence
of markings hm0 , . . . , mf i that are reached to replay that trace on the model,
Title Suppressed Due to Excessive Length
15
with m0 the initial marking and mf ∈MF , such that for each two consecutive
markings hmi , mi+1 i there exists a transition t ∈ T such that mi+1 = mi − •t + t•.
Furthermore, alignments also provide a function Γ t : M × Σ + → T + that provides
the sequence of transitions ht0 , . . . tn i that matches the changes in the sequence of
markings, i.e., m1 = m0 − •t0 + t0 •, etc. For each trace σ ∈ Σ + that fits a process
model N ∈ M the alignment l(Γ t (N, σ)) = σ. For unfitting traces σ ∈ Σ + , the
alignment is such that l(Γ t (N, σ)) is as close as possible to l according to some cost
function. We refer to Adriansyah et al. [3] for a more exhaustive introduction of
alignments. Let Γ t denote the sequence consisting of only the visible transitions in
Γ t , and let Γ m correspondingly denote the sequence of markings prior to each firing
of a visible transition. Given a marking m ∈ B(P ) we define the nondeterminism
of that marking to be the number of reachable visible transitions that can be fired
γ
as first next visible transition from m, i.e., nondeterminism(m) = |{a∈Σ|m −→
mi ∧ t∈γ ∧ l(t) = a ∧ ∀γi ∈γ γi ∈dom(l) =⇒ γi =t}|. We define the nondeterminism
of a model N ∈ M given a trace σ ∈ Σ + as the average nondeterminism of the
markings Γ m (N, σ) and define the nondeterminism for a model N and a log L as
the average nondeterminism over the traces of L.
Figure 6 shows the F-scores measured for different percentages of activities
filtered out from the Maruster LA12 log with different numbers of uniform chaotic
activities added. Note that the line stops when further removal of activities does
not lead to further improvement in F-score. Note that on the original event log with
0 chaotic activities added the F-score on the original log is already 1.0, resulting
in no lines being drawn. With one chaotic activity added, the least-frequent-first
filter needs to remove 75% of the activities before it ends up with F-score 1, which
can be explained by the fact that 9 out of 12 non-chaotic needed to be removed
in order with the least-frequent-first filter to remove all uniform chaotic activities,
as shown in Table 1. All entropy-based activity filtering techniques remove the
chaotic activity in the first filtering step, immediately leading to an F-score of 1.0.
Up until 8 added chaotic activities there is no difference between the entropy-based
activity filtering techniques in terms of F-score of the resulting process models,
which is consistent with the fact that all these filtering techniques were found to
filter without errors for these number of inserted chaotic activities in Table 1. For
16 and 32 activities, the direct filtering methods outperform the indirect filtering
methods, consistent with the fact that the indirect approach made one filtering error
according to the ground truth for these numbers of added chaotic activities. Note
that the least-frequent-first filter is outperformed by the entropy-based filtering
methods in terms of F-score of the discovered models, as would be expected given
the filtering results according to the ground truth.
Figure 7 shows the results in terms of nondeterminism measured for different
percentages of activities filtered out from the Maruster LA12 log with various
numbers of uniform chaotic activities added. The results show very clearly that
when filtering out a number of activities that is identical to the number of added
chaotic activities (this corresponds to 92% for one added activity, 86% for two
added activities, 75% for 4 added activities, 60% for 8 added activities, 43% for 16
added activities, and 27% for 32 added activities), the nondeterminism reaches a
value of 1.5, which is the nondeterminism value of the model discovered from the
original log without added chaotic activities. The least-frequent-first filter, however,
leads to process models where many activities are enabled on average, therefore
16
Niek Tax et al.
Table 2 An overview of the event logs used in the experiments
Name
BPI’12 [41]
BPI’12 resource 10939 [39]
Environmental permit [6]
SEPSIS [26]
Traffic Fine [10]
Bruno [5]
CHAD 1600010 [28]
MIT A [36]
MIT B [36]
Ordonez A [31]
van Kasteren [18]
Cook hh102 labour [9]
Cook hh102 weekend [9]
Cook hh104 labour [9]
Cook hh104 weekend [9]
Cook hh110 labour [9]
Cook hh110 weekend [9]
Category
Business
Business
Business
Business
Business
Human behavior
Human behavior
Human behavior
Human behavior
Human behavior
Human behavior
Human behavior
Human behavior
Human behavior
Human behavior
Human behavior
Human behavior
# traces
# events
# activities
13087
49
1434
1050
150370
57
26
16
17
15
23
18
18
43
18
21
6
164506
1682
8577
15214
561470
553
238
2772
1962
409
220
576
210
2100
864
695
184
23
14
27
16
11
14
10
27
20
12
7
18
18
19
19
17
14
overgeneralizing the process behavior, as an effect of filtering out nonchaotic
activities instead of the added chaotic activities.
5 Evaluation using Real Life Data
For the experiments on real-life event logs we do not artificially insert chaotic
activities to event logs, but instead filter directly on the activities that are present
in these logs. Whether these logs contain chaotic activities that impact process
discovery results is not known upfront. Therefore, we apply different activity
filtering techniques to these logs and use them to filter out a varying number of
activities, after which we assess the quality of the process model that is discovered
from these filtered logs. Table 2 gives an overview of the real-life event logs that we
use in the experiment. In total, we include five event logs from the business domain.
Furthermore, we include twelve event logs that contain events of human behavior,
recorded in smart home environments or through wearable devices. Mining process
model descriptions of daily life is a novel application of process mining that has
recently gained popularity [12, 22, 35, 40, 38]. Furthermore, human behavior event
data are often challenging for process discovery because of the presence of highly
chaotic activities, like going to the toilet. We perform the experiments with activity
filtering techniques on real-life data with RapidProM [2], which is an extension
that adds process mining capabilities to the RapidMiner platform for repeatable
scientific workflows.
For each event log, we apply seven different activity filtering techniques for comparison: 1) direct entropy filter without Laplace smoothing, 2) direct entropy filter
1
with Laplace smoothing (α= |Activities(L)|
), 3) indirect entropy filter without Laplace
1
smoothing, 4) indirect entropy filter with Laplace smoothing (α= |Activities(L)|
), 5)
least-frequent-first filtering, 6) most-frequent-first filtering, 7) filtering the activities
from the log in a random order. Recall that the activity filtering procedure stops
at the point where all but two activities are filtered from the event log because
process models that contain just one activity do not communicate any information
Title Suppressed Due to Excessive Length
BPI '12
BPI '12 resource 10939
17
Environmental permit
SEPSIS
Traffic Fine Management
1.00
0.75
IM
0.50
Filter
F−score
0.25
Direct
Direct (a=1/|A|)
0.00
1.00
Indirect
Indirect (a=1/|A|)
Least−frequent−first
0.75
IMf 20
0.50
0.25
0.00
0.00
0.25
0.50
0.75
0.00
0.25
0.50
0.75
0.00
0.25
0.50
0.75
0.00
0.25
0.50
0.75
0.00
0.25
0.50
0.75
Minimum % of activities explained
Fig. 8 F-score on business logs dependent on the minimum share of activities remaining.
regarding the relations between activities. For each event log and for each activity
filtering approach we discover a process model after each filtering step (i.e., after
each removal of an activity). The process discovery step is performed with two
process discovery approaches: the Inductive Miner [19], and the Inductive Miner
infrequent (20%) [20].
5.0.1 Results on Business Process Event Logs
Figure 8 shows the F-score of the process models discovered with the Inductive
Miner [19] and the Inductive Miner with infrequent behavior filtering [20] (20%
filtering) on the five business event logs for different percentages of activities filtered
out and different activity filtering techniques. The figure shows an increasing trend
in F-score for all event logs when more activities are filtered from the event log.
Furthermore, the line for the least-frequent-first filtering approach is below the lines
of the entropy-based filtering techniques for most of the percentages of activities
removed on most event logs, which shows that entropy-based filtering enables the
discovery of models with higher F-score compared to simply filtering out infrequent
activities. There are a few exceptions where filtering out infrequent activities
outperforms the entropy-based techniques, e.g., the Inductive Miner on the BPI ’12
resource 10939 event log (around 40% of activities explained) and the traffic fines
event log (around 55% of activities explained). It differs between event logs which
of the entropy-based techniques performs best: for the environmental permit log the
indirect filter without Laplace smoothing almost dominates the other techniques
while for the SEPSIS log the direct filter without Laplace smoothing outperforms
the other techniques. Generally, it seems that the use of Laplace smoothing harms
F-score, as most parts of the lines of indirect filtering with Laplace smoothing are
below the lines of the indirect approach without Laplace smoothing, and similar for
the direct approach with and without Laplace smoothing. However, the detrimental
effect of Laplace smoothing does not seem to be large, and in some cases, the usage
of Laplace smoothing in filtering increases the F-score of the discovered models.
Figure 9 shows the nondeterminism of the process models as a function of the
minimum percentage of activities. The green dashed line indicates the nondeterminism of the flower model, i.e., the process model that allows for all behavior over the
18
Niek Tax et al.
BPI '12
BPI '12 resource 10939
Environmental permit
SEPSIS
Traffic Fine Management
15
10
IM
Filter
Nondeterminism
5
Direct
Direct (a=1/|A|)
0
Indirect
Indirect (a=1/|A|)
6
Least−frequent−first
Flower model
IMf 20
4
2
0
0.00
0.25
0.50
0.75
0.00
0.25
0.50
0.75
0.00
0.25
0.50
0.75
0.00
0.25
0.50
0.75
0.00
0.25
0.50
0.75
Minimum % of activities explained
Fig. 9 Nondeterminism on business behavior logs dependent on the minimum share of the
activities remaining.
1.00
BPI '12
BPI '12 resource 10939
Environmental permit
SEPSIS
Traffic Fine Management
0.75
Filter
Direct (a=1/|A|)
0.50
IMf 20
Fitness
Direct
Indirect
Indirect (a=1/|A|)
Least−frequent−first
Most−frequent−first
Random
0.25
0.00
Fig. 10 Fitness on business logs with least 75% of the activities remaining.
activities. The lines stop when further removal of activities does not lead to further
improvement of nondeterminism. It is clear that the filtering mechanism of the
Inductive Miner helps to discover process models that are more behaviorally constrained, as the nondeterminism values are lower for the Inductive Miner infrequent
20% compared to the Inductive Miner without filtering. However, the results show
even when already using the 20% frequency filter of the Inductive Miner infrequent,
the chaotic activity filter can lead to an additional reduction of nondeterminism.
Furthermore, the results on the environmental permit log and the SEPSIS log show
that filtering several chaotic activities from the event log also enables the discovery
of a model with low nondeterminism using the Inductive Miner without filtering.
Which of the activity filtering approaches works best seems to be dependent on
the event log: the indirect entropy-based filter leads to the models with the lowest
nondeterminism on the traffic fine event log, the environmental permit event log,
while the direct entropy-based filter works better for some percentages of remaining
activities for the SEPSIS log and the BPI ’12 resource 10939 log.
Title Suppressed Due to Excessive Length
BPI '12
BPI '12 resource 10939
Environmental permit
19
SEPSIS
Traffic Fine Management
0.5
0.4
IM
0.3
0.2
Precision
0.1
0.0
0.5
0.4
IMf 20
0.3
0.2
0.1
0.0
Fig. 11 Precision on business logs with least 75% of the activities remaining.
Figures 10 and 11 show the fitness and precision values for the business process
event logs at the filtering step that leads to the highest F-score while describing at
least 75% of the activities of the original log. In addition to the filtering techniques
shown in Figure 8 it also shows the frequency-based activity filter where the most
frequent activities are filtered out first, and a random baseline is shown which
iteratively picks a random activity from the event log to filter out. The error bar for
the random activity filter indicates one standard error of the mean (SEM) based on
eight repetitions of applying the filter. The black dotted horizontal lines indicate
the fitness and precision values of the process models discovered from the original
event log without filtering any activities. Note that the fitness values are only shown
for the Inductive Miner infrequent 20% [20] because the Inductive Miner without
infrequent behavior filter [19] provides the formal guarantee that the fitness of the
discovered model is 1. Figure 10 shows that generally, the differences in fitness
between the models discovered from the filtered logs are very minor, and very close
to the fitness of the unfiltered log (i.e., the dotted line). Figure 11, however, shows
that the entropy-based filtering approaches outperform filtering out activities based
on frequency and filtering out random activities from the event log. The F-scores of
the discovered process models is determined mostly by the precision of the models
because the activity filtering impacts precision more than it impacts fitness. One
exception is the BPI’12 resource 10939 log [39], where the fitness decreases to
below 0.75 as a result of applying one of the two frequency-based filters, while the
precision increase as an effect of applying the filter is only minor.
5.0.2 Results on Human Behavior Event Logs
Figure 12 shows the maximum F-score for different human behavior event logs
as a function of the minimum percentage of activities that are remaining in the
log. Again, the general pattern is that the F-score of the discovered process model
decreases when the minimum percentage of events explained increases, as the
process discovery task gets easier for smaller numbers of activities. The figure
shows that filtering infrequent activities from the event log is dominated in terms
20
Niek Tax et al.
Bruno
CHAD 1600010
MIT A
MIT B
Ordonez A
van Kasteren
1.00
0.75
IM
0.50
Filter
F−score
0.25
Direct
Direct (a=1/|A|)
0.00
1.00
Indirect
Indirect (a=1/|A|)
Least−frequent−first
0.75
IMf 20
0.50
0.25
0.00
0.00 0.25 0.50 0.75
0.0 0.2 0.4 0.6 0.8 0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.0 0.2 0.4 0.6 0.8
Minimum % of activities explained
Fig. 12 F-score on human behavior logs dependent on the minimum share of activities.
Bruno
CHAD 1600010
MIT A
MIT B
Ordonez A
van Kasteren
15
Nondeterminism
IM
10
Filter
5
Direct
Direct (a=1/|A|)
0
Indirect
Indirect (a=1/|A|)
Least−frequent−first
10
Flower model
IMf 20
5
0
0.00 0.25 0.50 0.75
0.0 0.2 0.4 0.6 0.8 0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.0 0.2 0.4 0.6 0.8
Minimum % of activities explained
Fig. 13 Nondeterminism on human behavior logs dependent on the minimum share of the
activities remaining.
of F-score by the entropy-based filtering techniques. Like on the business process
event logs, there are mixed results on which of the four configurations of the
entropy-based filtering technique leads to the highest F-score: on the CHAD event
log the indirect activity filter outperforms the direct activity filter when using the
Inductive Miner infrequent 20%; however, the direct activity filter leads to higher
F-score for the Inductive Miner when filtering more than 50% of the activities.
Figure 13 shows the nondeterminism results for the human behavior event logs.
It is noticeable that the nondeterminism values of the process models that are
discovered when filtering very few activities are much closer to the flower model
compared to what we have seen before for the business process event logs. This is
caused by human behavior event logs having much more variability in behavior
compared to execution data from business processes, resulting in a much harder
process discovery task. After filtering several chaotic activities, the nondeterminism
drops significantly to ranges comparable to nondeterminism values seen for logs
from the business process domain. This shows that the problem of chaotic activities
is much more prominent in human behavior event logs than in business process event
logs. The entropy-based activity filtering approaches lead to more deterministic
Title Suppressed Due to Excessive Length
Bruno
CHAD
MIT A
21
MIT B
Ordonez A
van Kasteren
0.8
0.6
IM
0.4
Filter
Direct
Precision
0.2
Direct (a=1/|A|)
0.0
Indirect
0.8
Least−frequent−first
Indirect (a=1/|A|)
Most−frequent−first
0.6
IMf 20
0.4
Random
0.2
0.0
Fig. 14 Precision on human behavior logs with at least 50% of the activities.
process models compared to filtering out infrequent activities. Two clear examples
of this are the MIT B log and the Ordonez A log, on which filtering out infrequent
activities after several filtering steps results in a flower model (i.e., nondeterminism
is identical to that of the flower model), while entropy-based activity filters enable
the discovery of a model with nondeterminism close to one (i.e., very close to a
sequential model) while at the same time keeping 75% of the activities in the event
log.
Figure 14 shows the precision values for the human behavior logs for the
filtering step that leads to the highest F-score while describing at least 50% of the
activities of the original log. Similarly to what we have seen in the nondeterminism
graph, removing random activities from the log and removing infrequent activities
from the log results in smaller precision increases compared to the entropy-based
activity filters. Furthermore, it is noticeable that removing frequent activities from
the log works quite well to improve the precision of models discovered from the
human behavior application domain. The reason for this is that some of the chaotic
activities that are present in many of those event logs, including going to the toilet
and getting a drink, also happen to be frequent. On the van Kasteren event log
the indirect activity filter with Laplace smoothing leads to the largest increase in
precision when mining a model with at least 50% of the activities (from 0.324 to
0.732 with the Inductive Miner infrequent 20%).
Table 3 shows in which order activities are filtered from the van Kasteren event
log by 1) the indirect entropy-based activity filter with Laplace smoothing and
2) the least-frequent-first filter. It shows that the entropy-based filter filters use
toilet as the first activity, which from domain knowledge we know to be a chaotic
activity, as people generally just go to the toilet whenever they need to, regardless
of which other activities they have just performed. For the infrequent activity filter
use toilet would be the last choice of the activities to filter out, because it is the
most frequent activity in the van Kasteren event log.
Figures 15a and 15b show the corresponding process models discovered with
the Inductive Miner infrequent 20% from the logs filtered with the indirect activity
filter with Laplace smoothing and the infrequent activity filter respectively. The
22
Niek Tax et al.
Table 3 Left: the order in which activities are filtered using the direct activity filter with
1
Laplace smoothing (α = |Activities(L)|
) on the van Kasteren log. Right: the order in which the
activities are filtered using the least-frequent-first filter.
Order
1
2
3
4
5
6
7
Filtered activity
(indirect entropy-based filter
with Laplace smoothing)
Filtered activity
(least-frequent-first filter)
Use toilet
Get drink
Leave house
Take shower
Go to bed
Prepare breakfast
Prepare dinner
Prepare dinner
Get drink
Prepare breakfast
Take shower
Go to bed
Leave house
Use toilet
(a)
(b)
Fig. 15 (a) The model discovered with Inductive Miner infrequent 20% on the Van Kasteren
log after filtering all but four activities with the indirect approach with Laplace smoothing,
and (b) the model discovered from the same log with the same miner when filtering all but
four activities when filtering out the least frequent activities.
process model discovered after filtering three activities with the Indirect entropybased activity filter with Laplace smoothing is very specific on the behavior that it
described: after going to bed, either the logging ends, or prepare breakfast occurs
next, followed by taking a shower. After taking a shower, there is a possibility to
either go to bed again or to prepare dinner before going to bed. The process model
discovered after filtering three activities with the infrequent activity filter allows
for many more traces: it starts with go to bed followed by use toilet, after which
any of the activities go to bed, take shower, and leave house can occur as next event
or the logging can end. Furthermore, the activities leave house and take shower
can occur in any order, and take shower can also be skipped.
Figure 16 shows the results on F-score for the human behavior event logs by
Cook et al. [9]. The results on the Cook event logs are in-line with the results on
the human behavior event logs, however, on these event logs, it is even more clear
Title Suppressed Due to Excessive Length
hh102 labour
hh102 weekend
hh104 labour
23
hh104 weekend
hh110 labour
hh110 weekend
1.00
0.75
IM
0.50
Filter
F−score
0.25
Direct
Direct (a=1/|A|)
0.00
1.00
Indirect
Indirect (a=1/|A|)
Least−frequent−first
0.75
IMf 20
0.50
0.25
0.00
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
Minimum % of activities explained
Fig. 16 F-score on cook’s human behavior logs dependent on the minimum share of the
activities remaining.
hh102 labour
hh102 weekend
hh104 labour
hh104 weekend
hh110 labour
hh110 weekend
16
12
IM
Nondeterminism
8
Filter
4
Direct
Direct (a=1/|A|)
0
Indirect
12.5
Indirect (a=1/|A|)
Least−frequent−first
10.0
Flower model
IMf 20
7.5
5.0
2.5
0.0
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
0.00 0.25 0.50 0.75
Minimum % of activities explained
Fig. 17 Nondeterminism on cook’s human behavior logs dependent on the minimum share of
the activities remaining.
that filtering out infrequent activities leads to suboptimal process models in terms
of F-score. Which of the filtering approaches results in the optimal process model
in terms of F-score is very dependent on the event log and the minimum number
of activities to be remained after filtering: each of the four configurations of the
entropy-based filtering approach is optimal for at least one combination of log and
minimum percentage of activities explained.
Figure 17 shows the results in terms of nondeterminism for the same event
logs. Filtering infrequent activities at high percentages of activities explained has
much lower nondeterminism compared to the flower model, while further left on
the graph, after filtering out more activities, the nondeterminism of filtering out
infrequent activities gets closer to the flower model. This shows that filtering out
infrequent activities can even be harmful to the quality of the obtained process
discovery result. The nondeterminism values obtained with the four configurations
of the entropy-based filtering approach are generally close to each other, where the
optimal configuration is dependent on the log and the number of filtered activities.
24
Niek Tax et al.
Average Winning Number
4
Filter
3
Direct
Direct (a=1/|A|)
Indirect
Indirect (a=1/|A|)
2
Least−frequent−first
Most−frequent−first
Random
1
0
0.00
0.25
0.50
0.75
Minimum % of activities explained
Fig. 18 The average winning number for the seven activity filtering techniques dependent
on the minimum ratio of activities explained, averaged over the 17 event logs used in the
experiment.
Table 4 Kendall τb rank correlation between five activity filtering methods, mean and standard
deviation over the 17 event logs.
Direct
1
Direct (α= |A|
)
Indirect
1
Indirect (α= |A|
)
Least-frequent-first
Direct
1
Direct (α= |A|
)
Indirect
1
Indirect (α= |A|
)
Least-frequent-first
1.0
0.2956
0.0829
0.1408
0.0504
0.2956
1.0
0.0698
0.0536
0.1454
0.0829
0.0698
1.0
0.6852
-0.0275
0.1408
0.0536
0.6852
1.0
-0.0392
0.0504
0.1454
-0.0275
-0.0392
1.0
5.0.3 Aggregated Analysis Over All Event Logs
We have observed in Figures 9, 13, and 17 that the entropy-based activity filtering
techniques perform differently on different datasets and for different numbers of
activities filtered. To evaluate the overall performance of activity filtering techniques,
we use the number of other filtering techniques that it can beat over all the seventeen
event logs of Table 2. This metric, known as winning number, is commonly used
for evaluation in the Information Retrieval (IR) field [32, 37]. Formally, winning
number is defined as
P
P7
x
x
Wix = 17
j=1
k=1 1{Ni (j)<Nk (j)}
where j is the index of an event log, i and k are indices of activity filtering
techniques, Nix (j) is the performance of the i-th algorithm on the j-th event log
in terms of nondeterminism where each least x% of activities are explained and
1{Nix (j)<Nkx (j)} is the indicator function
(
1, if Nix (j) < Nkx (j),
1{Nix (j)<Nkx (j)} =
0, otherwise.
x
Wx
We define W i = 17i as the average number of other activity filtering techniques
that are outperformed by filtering technique i at the point where at least x% of
activities are explained.
x
Figure 18 shows the average winning number W i for different values of x and
for the seven different activity filtering techniques. We observe that for higher ratios
Title Suppressed Due to Excessive Length
25
Table 5 Number of event logs for which we can reject the null hypothesis that the orderings
of activities returned by activity filters are uncorrelated, according to the tau test.
Direct
1
Direct (α= |A|
)
Indirect
1
Indirect (α= |A|
)
Least-frequent-first
Direct
1
Direct (α= |A|
)
Indirect
1
Indirect (α= |A|
)
Least-frequent-first
17
5
1
2
0
5
17
1
1
3
1
1
17
17
3
2
1
17
17
3
0
3
3
3
17
of activities explained the differences between filtering techniques are smaller than
for lower numbers of activities explained. Intuitively this can be explained by the
fact that for lower ratios of activities explained more activities have been filtered
out from the log. Therefore the effect of the filtering techniques is more clearly
visible. The figure shows that, up until +-74% of activities explained, the indirect
entropy-based activity filtering technique leads to the most deterministic process
models averaged over all event logs included in the experiment, where it outperforms
between 4 and 4.5 other filtering techniques. Between +-75% and +- 87.5% the
indirect entropy-based activity filtering technique with Laplace smoothing results
in the highest average winning number, although the difference with the indirect
entropy-based filtering technique seems negligible. Filtering out random activities
from the event log outperforms none of the 6 other activities filtering techniques for
the most of the graph, indicating that frequency-based filtering clearly outperforms
filtering random activities.
To investigate to what degree the order in which activities are removed from the
logs differs between the activity filtering techniques we calculate Kendall’s tau (τb )
rank correlation for each log between the activity filtering techniques in a pairwise
way. Table 4 shows the rank correlation values found between the activity filters,
averaged over the 17 event logs. The indirect activity filter with Laplace smoothing
and the indirect activity filter without Laplace smoothing generate orderings over
the activities of a log that are strongly correlated. Between the direct activity filter
without Laplace smoothing and the direct activity filter without Laplace smoothing
there is only a weak correlation. All the other activity filtering techniques are
uncorrelated or very weakly correlated. Using the Kendall τb statistic, we apply a
tau test for each pair of activity filtering techniques on each event log to test the
null hypothesis that the two orderings in which activities are filtered by the two
activity filtering techniques are uncorrelated, using a significance level α = 0.05.
For each pair of activity filtering techniques Table 5 shows the number of
event logs for which the null hypothesis was rejected, i.e., the number of event
logs for which the order in which activities are filtered is statistically correlated.
The indirect activity filters with and without Laplace smoothing create correlated
orderings of activities for all seventeen event logs. For all other pairs of activity
filtering techniques the orderings in which activities are filtered are only correlated
with for low numbers of event logs.
26
Niek Tax et al.
Fig. 19 A mockup of the proposed way to use the activity filters in an interactive setting.
6 Entropy-based Toggles for Process Discovery
In the previous section we have shown that all four configurations of the entropybased activity filtering technique lead to more deterministic process models compared to simply filtering out infrequent activities. However, the differences in
determinism of the process models that are discovered after applying any of the
four configurations are small and dependent on the event log to which they are
applied. Furthermore, all four configurations of the activity filtering technique
simply impose an ordering over the activities, but do not specify at which step the
filtering should be stopped. Additionally, the proposed filtering technique ignores
the semantics of activities: activities that are chaotic may still be relevant for the
process. Leaving them out of the process model to discover will harm the usefulness
of the discovered process model.
To address the three issues we propose to use the filtering technique as a sorting
technique over the activities in combination with toggles that interactively allow the
process analyst to “disable” (filter out) or “enable” activities, and then rediscover
and visualize the process model according to the new settings. This approach is
similar to the Inductive Visual Miner [21], an interactive implementation of the
Inductive Miner [20] algorithm which allows the process analyst to filter the event
log interactively using a slider-based approach. The Inductive visual miner contains
two sliders: with one slider activities can be filtered using the least-frequent-first
filter, where the user can control how many activities are filtered out by moving
the slider up and down. We propose to replace this slider with a sorted list of
activities and toggles, as this allows the process analyst to override the ordering of
the activities that is determined by the activity filtering technique with domain
knowledge. Figure 19 shows a mockup of the proposed way to use the activity
filter. Activities are by default sorted using the chaotic activity filter, showing the
entropy to indicate the assessed degree of chaoticness of each activity. Based on
this information, the process analyst can choose to rely on the filtering technique
and filter out the top of the list or to override this list with domain knowledge.
Furthermore, other activity filtering techniques, such as the least-frequent-first
filter, can be included as an additional column on which the activities of the process
Title Suppressed Due to Excessive Length
27
can be sorted. This allows the process analyst to control how many activities, and
which activities, are filtered out of the process model, and thereby also empowers
the user to prevent the removal of semantically important activities that should
not be removed. Furthermore, this approach allows the process analyst to explore
himself which of the filtering techniques leads to the most useful process model
from the event log that he is analyzing.
7 Related Work
Real life events logs often contain all sorts of data quality issues [34], include
incorrectly logged events, events that are logged in the wrong order, and events
that took place without being logged. Instances of such data quality issues are
often referred to as noise. Many event log filtering techniques have been proposed
to address the problem of noise. Existing filtering techniques in the process mining
field can be classified into four categories: 1) event filtering techniques, 2) process
discovery techniques that have an integrated filtering mechanism build in, 3) trace
filtering techniques, and 4) activity filtering techniques. We use these categories to
discuss and structure related work.
7.1 Event filtering
Conforti et al. [8] recently proposed a technique to filter out outlier events from an
event log. The technique starts by building a prefix automaton of the event log,
which is minimal in terms of the number of arcs in the automaton, using an Integer
Linear Programming (ILP) solver. Infrequent arcs are removed from the minimal
prefix automaton, and finally, the events belonging to removed arcs are filtered out
from the event log.
Lu et al. [25] advocate the use of event mappings [24] to distinguish between
events that are part of the mainstream behavior of a process and outlier events.
Event mappings compute similar behavior and dissimilar behavior between each
two executions of the process as a mapping: the similar behavior is formed by all
pairs of events that are mapped to each other, whereas events that are not mapped
are dissimilar behavior.
Fani Sani et al. [13] proposes the use of sequential pattern mining techniques
to distinguish between events that are part of the mainstream behavior and outlier
events.
All three of the event filtering techniques listed above aim filter out outlier
events from the event log, while keeping the mainstream behavior. Event filtering
techniques model the frequently occurring contexts of activities and filter out the
contexts of activities that occur infrequently in the log. For example, consider an
activity B such that 98% of its occurrences are in context h. . . , A, B, C, . . . i, with
the remaining 2% of the events of activity B are in context h. . . , D, B, E, . . . i, then
the B events that occur between D and E will be filtered out by event filtering
techniques. Note that our filtering technique is orthogonal to event filtering: it would
consider activity B to be nonchaotic and would not filter out anything. However,
when a log L contains a chaotic activity X, then event filtering techniques are not
able to remove all events of this chaotic activity. One of the contexts of X will
28
Niek Tax et al.
by chance be more frequent than other contexts, i.e., for some activity A, it will
hold that ∀B ∈ Activities(L) : #(hA, Xi, L) > #(hB, Xi, L), even though hA, Xi
might only be slightly more frequent. This will result in X events after a B being
removed, while the X events after an A remain in the log. Applying a process
discovery technique to this filtered log will then result in a process model where
activity X is misleadingly positioned after activity A, while in fact X can happen
anywhere in the process. The activity filtering technique presented in this paper
will instead detect that activity X is chaotic, and completely remove it from the
event log, preventing the misleading effect of event filtering.
7.2 Process Discovery Techniques with Integrated Filtering
Several process discovery algorithms offer integrated filtering mechanisms as part
of the approach. The Inductive Miner (IM) [19] is a process discovery algorithm
which first discovers a directly-follows graph from the event logs, where activities
are connected that directly follow each other in the log, from which in a second
step a process model is discovered. The directly-follows relations are affected by
the presence of a chaotic activity X: sequence h. . . , A, X, C, . . . i leads to false
directly-follows relations between A and X and between X and C, while the
directly-follows relation between A and C is obfuscated by X. The Inductive Miner
infrequent (IMf) [20] is an extension of the IM where infrequent directly-follows
relations are filtered out from the set of directly-follows relations that are used
to generate to process models. The filtering mechanism of IMf can help to filter
out the directly-follows relations between A and X and between X and C, but it
does not help to recover the obfuscated directly-follows relation between A and
C. Instead, the activity filtering technique presented in this paper filters out the
chaotic activity X, leading to sequence h. . . , A, X, C, . . . i being transformed into
h. . . , A, C, . . . i, thereby recovering the directly follows relation between A and C.
The Heuristics Miner [45] and the Fodina algorithm [4], in addition to the
directly-follows relation, defines an eventually-follows relation between activities
and allows the process analyst to filter out infrequent directly-follows and eventually
follows relations. Two activities A and B are in an eventually-follows relation when
A is eventually followed by B, before the next appearance of A or B. The eventuallyfollows relation, unlike the directly-follows relation, is not impacted by the presence
of chaotic activities. The Heuristic Miner [45] and Fodina [4] both include filtering
methods for the directly-follows and eventually-follows relations that are similar in
nature to the filtering mechanism that is used in the Inductive Miner infrequent
[20]. However, the use of sequential orderings and parallel constructs in the mining
approaches of the Heuristic Miner [11] and Fodina [4] is based on the directlyfollows relations only, with the eventually follows relations being used for the
mining of long-term dependencies. Furthermore, in contrast to the Inductive Miner,
the process models discovered with the Heuristic Miner [45] or Fodina [4] can be
unsound, i.e., the can contain deadlocks.
The ILP-miner [46] is a process discovery algorithm where a set of behavioral
constraints over activities is discovered for each prefix (called the prefix-closure)
of the event log, based on which a process model is discovered that satisfies these
constraints using Integer Linear Programming (ILP). Van Zelst et al. [47] proposed
a filtering technique for the ILP-miner where the prefix closure of the event log
Title Suppressed Due to Excessive Length
29
is filtered prior to solving the ILP problem by removing infrequently observed
prefixes. It is easy to see that a chaotic activity X affect the prefix-closure that
is discovered from the event log: given log consisting of two traces hA, X, Ci and
hX, A, Ci, activity X causes the prefixes closures of the two traces to have no
overlap in states, while without activity X the two traces are identical. This makes
the filtering method of the prefix-closure proposed by Van Zelst et al. [47] less
effective, as frequent prefixes randomly get distributed over several infrequent
prefixes when chaotic activities are present. Instead, the chaotic activity filtering
technique presented in this paper would remove chaotic activity X, leading to
traces hA, X, Ci and hX, A, Ci becoming identical after filtering, therefore leading
to a simpler process model while still describing the behavior of the event log
accurately.
The Fuzzy Miner [16] is a process discovery algorithm that aims at mining
models from flexible processes, and it discovers a process model without formal
semantics. The Fuzzy Miner discovers this graph by extracting the eventually
follows relation from the event log, which is not affected by chaotic activities.
Similar to the Heuristics Miner [45] and Fodina [4] the Fuzzy Miner allows to filter
out infrequent eventually-follows relations between activities. In practice, the lack
of formal semantics of the Fuzzy Miner models hinders the usability of the models,
as the models are not precise on what behavior is allowed in the process under
analysis.
7.3 Trace filtering
Ghionna et al. [14] proposed a technique to identify outlier traces from the event
log that consists of two steps: 1) mining frequent patterns from the event log, and
2) applying MCL clustering [43] on the traces, where the similarity measure for
traces is defined on the number of patterns that jointly characterize the execution
of the traces. Traces that are not assigned to a cluster by the MCL clustering
algorithm are considered to be outlier traces and are filtered from the event log.
It is easy to see that trace filtering techniques address a fundamentally different
problem than chaotic activity filtering: in the event log shown in Figure 2b there
are only two traces that do not contain an instance of chaotic activity X, therefore,
even if a trace filtering technique would be able to perfectly filter out traces that
contain a chaotic event, the number of remaining traces will become too small to
mine a fitting and precise process model when the chaotic activity is frequent.
7.4 Activity filtering
The modus operandi for filtering activities is to simply filter out infrequent activities
from the event log. The plugin ’Filter Log using Simple Heuristics’ in the ProM
process mining toolkit [42] offers tool support for this type of filtering. The Inductive
Visual Miner [21] is an interactive process discovery tool that implements the
Inductive Miner [20] process discovery algorithm in an interactive way: the process
analyst can filter the event log using sliders and is then shown the process model
that is discovered from this filtered log. One of the available sliders in the Inductive
Visual Miner offers the same frequency-based activity filtering functionality. The
30
Niek Tax et al.
working assumption behind filtering out infrequent activities is that when there
are just a few occurrences of an activity, there is probably not enough evidence
to establish their relation to other activities to model their behavior. However,
as we have shown in this paper, for frequent but chaotic activities, while they
are frequent enough to establish their relation to other activities, complicate the
process discovery task by lowering directly-follows counts between other activities
in the event log. The activity filtering technique presented in this paper is able to
filter out chaotic activities, thereby reconstructing the directly-follows relations
between the non-chaotic activities of the event log, at the expense of losing the
chaotic activities.
8 Conclusion & Future Work
In this paper, we have shown the possible detrimental effect of the presence of
chaotic activities in event logs on the quality of process models produced by process
discovery techniques. We have shown through synthetic experiments that frequencybased techniques for filtering activities from event logs, which is currently the modus
operandi for activity filtering in the process mining field, do not necessarily handle
chaotic activities well. As shown, chaotic activities can be frequent or infrequent.
We have proposed four novel techniques for filtering chaotic from event logs, which
find their roots in information theory and Bayesian statistics. Through experiments
on seventeen real-life datasets, we have shown that all four proposed activity
filtering techniques outperform frequency-based filtering on real data. The indirect
entropy-based activity filter has been found to be the best performing activity
filter overall averaged over all datasets used in the experiments; however, the
performance of the four proposed activity filtering techniques is highly dependent
on the characteristics of the event log.
Because the performance of the filtering techniques was found to be logdependent, we propose the use the activity filtering techniques in a slider-based
approach where the user can filter activities interactively and directly see the
process model discovered from the filtered event log. Ultimately, only the user can
decide which activities to include. In future work, we aim to construct a hybrid
activity filtering technique that combines the four techniques proposed in this
paper by using supervised learning techniques from the data mining field to predict
the effect of removing a particular activity.
References
1. van der Aalst WMP (2016) Process mining: data science in action. Springer
2. van der Aalst WMP, Bolt A, van Zelst SJ (2017) RapidProM: Mine your
processes and not just your data. In: Hofmann M, Klinkenberg R (eds) RapidMiner: Data Mining Use Cases and Business Analytics Applications, Chapman
& Hall/CRC Data Mining and Knowledge Discovery Series, p To Appear.
3. Adriansyah A, van Dongen BF, van der Aalst WMP (2011) Conformance
checking using cost-based fitness analysis. In: Proceedings of the 15 IEEE
International Enterprise Distributed Object Computing Conference (EDOC),
IEEE, pp 55–64
Title Suppressed Due to Excessive Length
31
4. vanden Broucke SKLM, De Weerdt J (2017) Fodina: a robust and flexible
heuristic process discovery technique. Decision Support Systems
5. Bruno B, Mastrogiovanni F, Sgorbissa A, Vernazza T, Zaccaria R (2013)
Analysis of human behavior recognition algorithms based on acceleration
data. In: Proceedings of the IEEE International Conference on Robotics and
Automation, IEEE, pp 1602–1607
6. Buijs JCAM (2014) Receipt phase of an environmental permit application process (WABO), CoSeLoG project. doi:10.4121/uuid:
a07386a5-7be3-4367-9535-70bc9e77dbe6
7. Buijs JCAM, van Dongen BF, van der Aalst WMP (2012) A genetic algorithm
for discovering process trees. In: Proceedings of the 2012 IEEE Congress on
Evolutionary Computation, IEEE, pp 1–8
8. Conforti R, La Rosa M, ter Hofstede AHM (2017) Filtering out infrequent
behavior from business process event logs. IEEE Transactions on Knowledge
and Data Engineering 29(2):300–314
9. Cook DJ, Crandall AS, Thomas BL, Krishnan NC (2013) CASAS: A smart
home in a box. Computer 46(7):62–69
10. De Leoni M, Mannhardt F (2015) Road traffic fine management process.
doi:10.4121/uuid:270fd440-1057-4fb9-89a9-b699b47990f5
11. De Weerdt J, De Backer M, Vanthienen J, Baesens B (2011) A robust Fmeasure for evaluating discovered process models. In: Proceedings of the IEEE
Symposium on Computational Intelligence and Data Mining (CIDM), IEEE,
pp 148–155
12. Dimaggio M, Leotta F, Mecella M, Sora D (2016) Process-based habit mining: Experiments and techniques. In: Proceedings of the International IEEE
Conference on Ubiquitous Intelligence & Computing, IEEE, pp 145–152
13. Fani Sani M, van Zelst SJ, van der Aalst WMP (2017) Improving process
discovery results by filtering outliers using conditional behavioural probabilities.
In: Proceedings of the International Workshop on Business Process Intelligence,
Springer
14. Ghionna L, Greco G, Guzzo A, Pontieri L (2008) Outlier detection techniques
for process mining applications. In: International Symposium on Methodologies
for Intelligent Systems, Springer, pp 150–159
15. Goedertier S, Martens D, Vanthienen J, Baesens B (2009) Robust process
discovery with artificial negative events. Journal of Machine Learning Research
10(Jun):1305–1340
16. Günther CW, van der Aalst WMP (2007) Fuzzy mining–adaptive process
simplification based on multi-perspective metrics. In: International Conference
on Business Process Management, Springer, pp 328–343
17. Herbst J (2000) A machine learning approach to workflow management. In:
European Conference on Machine Learning, Springer, pp 183–194
18. van Kasteren T, Noulas A, Englebienne G, Kröse B (2008) Accurate activity recognition in a home setting. In: Proceedings of the 10th International
Conference on Ubiquitous Computing, ACM, pp 1–9
19. Leemans SJJ, Fahland D, van der Aalst WMP (2013) Discovering blockstructured process models from event logs - a constructive approach. In: International Conference on Applications and Theory of Petri Nets and Concurrency,
Springer, pp 311–329
32
Niek Tax et al.
20. Leemans SJJ, Fahland D, van der Aalst WMP (2013) Discovering blockstructured process models from event logs containing infrequent behaviour. In:
International Conference on Business Process Management, Springer, pp 66–78
21. Leemans SJJ, Fahland D, van der Aalst WMP (2014) Process and deviation
exploration with inductive visual miner. In: Proceedings of the BPM Demo
Track, CEUR-WS.org, vol 1295, p 46
22. Leotta F, Mecella M, Mendling J (2015) Applying process mining to smart
spaces: Perspectives and research challenges. In: International Conference on
Advanced Information Systems Engineering, Springer, pp 298–304
23. Lohmann N, Verbeek E, Dijkman R (2009) Petri net transformations for
business processes–a survey. In: Transactions on petri nets and other models
of concurrency II, Springer, pp 46–63
24. Lu X, Fahland D, van der Aalst WMP (2014) Conformance checking based on
partially ordered event data. In: International Conference on Business Process
Management, Springer, pp 75–88
25. Lu X, Fahland D, van den Biggelaar FJHM, van der Aalst WMP (2015) Detecting deviating behaviors without models. In: Proceedings of the International
Workshop on Business Process Intelligence, Springer, pp 126–139
26. Mannhardt F (2016) Sepsis cases - event log. doi:10.4121/uuid:
915d2bfb-7e84-49ad-a286-dc35f063a460
27. Maruster L, Weijters AJMM, Aalst WMPvd, Bosch Avd (2006) A rule-based
approach for process discovery: Dealing with noise and imbalance in process
logs. Data Mining & Knowledge Discovery 13(1):67–87
28. McCurdy T, Glen G, Smith L, Lakkadi Y (2000) The national exposure research laboratory’s consolidated human activity database. Journal of Exposure
Analysis and Environmental Epidemiology 10(6):566–578
29. Murata T (1989) Petri nets: Properties, analysis and applications. Proceedings
of the IEEE 77(4):541–580
30. Object Management Group (2011) Notation (BPMN) version 2.0. OMG Specification
31. Ordónez FJ, de Toledo P, Sanchis A (2013) Activity recognition using hybrid
generative/discriminative models on home environments using binary sensors.
Sensors 13(5):5460–5477
32. Qin T, Liu TY, Xu J, Li H (2010) LETOR: A benchmark collection for
research on learning to rank for information retrieval. Information Retrieval
13(4):346–374
33. Solé M, Carmona J (2013) Region-based foldings in process discovery. IEEE
Transactions on Knowledge and Data Engineering 25(1):192–205
34. Suriadi S, Andrews R, ter Hofstede AHM, Wynn MT (2017) Event log imperfection patterns for process mining: Towards a systematic approach to cleaning
event logs. Information Systems 64:132–150
35. Sztyler T, Völker J, Carmona Vargas J, Meier O, Stuckenschmidt H (2015)
Discovery of personal processes from labeled sensor data: An application of
process mining to personalized health care. In: Proceedings of the International
Workshop on Algorithms & Theories for the Analysis of Event Data, CEURWS.org, pp 31–46
36. Tapia EM, Intille SS, Larson K (2004) Activity recognition in the home
using simple and ubiquitous sensors. In: International Conference on Pervasive
Computing, Springer, pp 158–175
Title Suppressed Due to Excessive Length
33
37. Tax N, Bockting S, Hiemstra D (2015) A cross-benchmark comparison of 87
learning to rank methods. Information Processing & Management 51(6):757–
772
38. Tax N, Sidorova N, Haakma R, van der Aalst WMP (2016) Event abstraction
for process mining using supervised learning techniques. In: Proceedings of the
SAI Intelligent Systems Conference, Springer
39. Tax N, Sidorova N, Haakma R, van der Aalst WMP (2016) Mining local process
models. Journal of Innovation in Digital Ecosystems 3(2):183–196
40. Tax N, Sidorova N, Haakma R, van der Aalst WMP (2017) Mining process
model descriptions of daily life through event abstraction. In: Intelligent Systems
and Applications, Springer, p To appear.
41. Van Dongen B (2012) BPI challenge 2012. doi:10.4121/uuid:
3926db30-f712-4394-aebc-75976070e91f
42. Van Dongen BF, de Medeiros AKA, Verbeek HMW, Weijters AJMM, Van
Der Aalst WMP (2005) The ProM framework: A new era in process mining
tool support. In: International Conference on Application and Theory of Petri
Nets, Springer, pp 444–454
43. Van Dongen S (2008) Graph clustering via a discrete uncoupling process. SIAM
Journal on Matrix Analysis and Applications 30(1):121–141
44. Vanden Broucke SKLM, De Weerdt J, Vanthienen J, Baesens B (2013) Determining process model precision and generalization with weighted artificial
negative events. IEEE Transactions on Knowledge and Data Engineering
45. Weijters AJMM, Ribeiro JTS (2011) Flexible heuristics miner (FHM). In:
Proceedings of the IEEE Symposium on Computational Intelligence and Data
Mining (CIDM), IEEE, pp 310–317
46. van der Werf JMEM, van Dongen BF, Hurkens CAJ, Serebrenik A (2009)
Process discovery using integer linear programming. Fundamenta Informaticae
94(3):387–412
47. van Zelst SJ, van Dongen BF, van der Aalst WMP (2015) Avoiding overfitting in ILP-based process discovery. In: International Conference on Business
Process Management, Springer International Publishing, pp 163–171
48. Zhai C, Lafferty J (2004) A study of smoothing methods for language models
applied to information retrieval. ACM Transactions on Information Systems
22(2):179–214
| 2cs.AI
|
arXiv:1802.01880v1 [cs.CV] 6 Feb 2018
Learning Image Representations by Completing Damaged Jigsaw Puzzles
Dahun Kim
KAIST
Donghyeon Cho
KAIST
Donggeun Yoo
KAIST
In So Kweon
KAIST
[email protected]
[email protected]
[email protected]
[email protected]
Abstract
In this paper, we explore methods of complicating selfsupervised tasks for representation learning. That is, we
do severe damage to data and encourage a network to recover them. First, we complicate each of three powerful
self-supervised task candidates: jigsaw puzzle, inpainting,
and colorization. In addition, we introduce a novel complicated self-supervised task called “Completing damaged
jigsaw puzzles” which is puzzles with one piece missing and
the other pieces without color. We train a convolutional
neural network not only to solve the puzzles, but also generate the missing content and colorize the puzzles. The recovery of the aforementioned damage pushes the network
to obtain robust and general-purpose representations. We
demonstrate that complicating the self-supervised tasks improves their original versions and that our final task learns
more robust and transferable representations compared to
the previous methods, as well as the simple combination of
our candidate tasks. Our approach achieves state-of-the-art
performance in transfer learning on PASCAL classification
and semantic segmentation.
1. Introduction
The goal of representation learning is to learn robust and
general-purpose visual features. Typically, the amount of
labeled data decreases as the extent of annotation increases.
The networks trained on limited amount of labeled data are
easily overfitted and have poor representation ability. Representation learning is used to avoid this problem by pretraining visual features on large-scale data before training
on target tasks.
Conventional yet still popular method to learn such features is to pre-train image classification [11, 20, 33, 34] on
millions of human-labeled data such as ImageNet [32]. It
provides powerful representations and image priors when
the target task and data are similar. However, the dependency on human supervision of this traditional method limits its scalability and adaptability to dissimilar target tasks
and domains(e.g. depth prediction).
(a)
(b)
Figure 1. Learning image representations by completing damaged jigsaw puzzles. We sample 3-by-3 patches from an image
and create damaged jigsaw puzzles. (a) is the puzzles after shuffling the patches, removing one patch, and decolorizing. We push
a network to recover the original arrangement, the missing patch,
and the color of the puzzles. (b) shows the outputs; while the
pixel-level predictions are in ab channels, we visualize with their
original L channels for the benefit of the reader.
Many researches have been conducted to minimize human supervision in computer vision. For example, weaklysupervised learning [10, 15–17, 27] has been proposed to
learn object localization using weak image-level annotations rather than bounding boxes or pixel-level annotations.
In the same vein, recent representation learning has also
been improved to minimize human supervision. The emerging family of such methods is self-supervised learning; It
manufactures a supervised task and labels from raw images,
so that unlimited amount of labeled data can be used. A
considerable number of such methods [4–6, 21, 22, 25, 26,
30,36–39] have been proposed in last few years. They often
train a network to infer geometrical configuration [4,25], recover missing pixels [30] or channels [21,38,39] of images.
The features learned by these methods have been successfully transferred to different target tasks, such as classification, detection, and semantic segmentation, and resulted in
promising performances.
The common intuition of these approaches is that a network obtains useful representations of scenes and objects
while struggling to solve a challenge task that requires highlevel reasoning. Based on this idea, we propose a concept
of complicating a self-supervised task where we raise the
difficulty of the task. More specifically, we design more
difficult versions of jigsaw puzzle, inpainting, and colorization tasks. We investigate the effectiveness of our approach
by transferring the learned features on PASCAL VOC classification, detection, and segmentation tasks [7, 8]. In order to further the idea, we design a task called “Completing
damaged jigsaw puzzle”, which is puzzles with one piece
missing and the other pieces without color. Then, jigsaw
puzzle, inpainting, and colorization tasks are jointly optimized. The network learned in this way preserves better
feature representations for classification, detection and semantic segmentation.
In summary, our main contributions are as follows:
• We propose an approach of making self-supervised
tasks more challenging for representation learning.
• We design a problem of completing damaged jigsaw
puzzles where three different self-supervised tasks are
complicated and incorporated simultaneously.
• We show that the representations learned by our approach achieve state-of-the-art performances on PASCAL classification and semantic segmentation [7, 8]
when transferred on AlexNet, compared to existing
self-supervised learning methods.
2. Related works
A considerable number of unsupervised learning approaches have been studied to learn image representations
without relying on human-annotation. The most fundamental example is the autoencoder [35], which is a generative
model that reconstructs the input data, aiming to extract the
data representation. Since then, various generative models
rooted in the autoencoder have been proposed. For example, DCGAN [31] and variational auto-encoders [3] have
been proposed for further photorealistic reconstruction and
feature learning.
Our study falls into self-supervised learning which has
emerged as a new stream of unsupervised learning. This
technique manufactures supervision signal from the raw visual data and achieves promising results in learning discriminative features. Recent methods commonly use images [4, 6, 21, 25, 26, 30, 38, 39], and often video [12,
24, 29, 36], or other sensory data such as egomotion and
sound [1, 2, 13, 28].
Different supervision signals encourage the network
to pay attention to different characteristics in images.
Thus, the virtues of the learned representations also differ
across the self-supervised tasks. Recent methods on selfsupervised feature learning can be broadly categorized according to the type of knowledge preferred in the training:
spatial configuration, context, and cross-channel relations.
Spatial Configuration.
The methods that operate on
the spatial dimension of images usually extract the patches
from the image and learn the network to infer spatial relations between them. Doersch et al. [4] proposed a problem with 3-by-3 puzzles, where the network sees one of the
outer patches, and predicts its relative position to the center patch. Noroozi and Favaro [25] learn image representations by solving the jigsaw puzzle with the 3-by-3 patches
which imposes a challenging task of estimating what permutation has been used in shuffling. The learned features
well capture the geometrical configuration of the objects as
mentioned in [4].
Image Context. A contextual autoencoder was proposed
by Pathak et al. [30] in order to drive representation learning. The supervisory signal comes from inpainting task
where the network is encouraged to recover dropped part
of the image from the surrounding pixels. Also, Isola et
al. [12] exploited a co-occurance cues as a self-supervision
where the network takes two isolated patches and predict
whether or not they were taken from nearby locations in an
image. These methods allow the network to learn contextual
relations between part of an image and the rest or between
each object parts/instances in an image.
Cross-Channel Relations.
The methods that manipulate the images in channel domain have also been proposed.
Typically, they remove one subset of the image channels,
and train the network to recover it from the remaining channel(s). Zhang et al. [38] and Larsson et al. [21] obtain selfsupervision from the task of colorization where the network
predicts ab channels given L channel. Zhang et al. [39] took
a one step further by learning colorization together with the
inverse mapping from ab channels to L channel.
Combining Multiple Self-supervised Tasks. The aforementioned methods are essentially relying on a single supervisory signal. Recently, representation learning by multiple supervisory signals has also emerged. Zhang et al. [39]
proposed a bidirectional cross-channel prediction to aggregate complementary image representations. They propose
a network split into to two groups, and each subnetworks
are trained separately. Wang et al. [37] exploited two selfsupervised approaches to unify different types of invariance appearing in the two approaches. Doersch and Zisserman [5] combine multiple self-supervised tasks to create a
single universal representation. However, each of the methods have limitations. In [39], splitting the network reduces
the number of parameters by half which might limit the feature transferability. Also, [37] trains two tasks in sequential order. That is, the training on ranking video frames [36]
comes only after the training on estimating relative position [4] finishes. Lastly, the involved tasks in [5] operate
(a)
(b)
(c)
(d)
(e)
(f)
(g)
Figure 2. Illustrations of complicating self-supervised tasks. (a) Conventional 2×2 jigsaw puzzles. (b) Complicated 2×2 jigsaw puzzles;
each patch’s L or ab channel is dropped. (c) Complicated 2×2 jigsaw puzzles; one of the patches is completely dropped. (d) Conventional
inpainting. (e) Complicated inpainting; it outputs in ab channels from an input in only L channel. (f) Conventional colorization. (g)
Complicated colorization; only one-quarter of the entire image is given for colorization.
on very different inputs, which hinders simultaneous training of all tasks and requires special handlings.
Our study shares the goal with [37,39] and [5] where we
want to learn representations that have all-round capability
in every downstream task. However, our approach differs in
the strategy; We squeeze a network to solve more complicated tasks, and in the same vain, our final method combines
the complicated tasks and trains them simultaneously.
3. Approach
A number of recent self-supervised learning methods
commonly operate via damage-and-recover mechanisms.
In other words, the networks are supervised to recover intentionally damaged image data. For the purpose of representation learning, the damages are designed so that the
recovery requires the high-level understanding of the objects and the scene. During training, the representations
that are necessary for the recovery are learned, resulting in
task/damage-specific features. For example, the spatial configuration is damaged in jigsaw puzzles, so the learned representations are focused on the configuration and geometry
of objects. Similarly, the representations learned from inpainting and colorization preferably encode contextual and
cross-channel relations as analyzed in [30] and [21], respectively.
Motivated by the mechanism above, we design a strategy where we drive the network to recover even more severe damage. More specifically, we do further damage to
the data in jigsaw puzzle, inpainting, and colorization to
make them more challenging as illustrated in Fig. 2. The
methods of complicating each of the tasks are explained in
Sec. 3.1. Furthermore, in order to maximize the effectiveness of our approach, we incorporate those three tasks in a
single problem, “Completing damaged jigsaw puzzles”, as
detailed in Sec. 3.2.
3.1. Complicating Each Self-supervised tasks
In this section, we briefly review each of jigsaw puzzle, inpainting and colorization, and explain the methods
of complicating them. Considering that different damages
teach different lessons, we do additional damage to the data
domains that have remained intact in the original task. The
effectiveness of the complicated versions is quantitatively
evaluated in Sec. 5.1.
Jigsaw Puzzle. With 2-by-2 puzzles, let us define S a sequence of puzzle patches X 1 -X 4 shuffled by a permutation
P . The spatial configuration of objects is intermixed by the
permutation.
Accordingly, we consider two additional types of damage that make jigsaw puzzles more difficult. First, we do
damage in the channel-wise domain, where half of the puzzles have only the L channel and the other half, ab channels, as shown in Fig. 2-(b). Successfully solving the puzzles requires not only the knowledge on spatial configuration, but also the understanding of the cross-channel relations between L and ab channels. Second, we damage the
image context by removing one piece from a complete set
of puzzles, as shown in Fig. 2-(c). In practice, a piece is
discarded with a probability of 0.4 and the missing contents
are replaced with Gaussian noise. Doing well on this task
may require extra understanding on the full context without
seeing the missing area.
As in [25], we train an AlexNet-based network to learn
a mapping P̂ = f jig (S) to a probability distribution over
24(that is, 4!) possible permutations P̂ ∈ [0, 1]24 with loss,
Ljig = −
X
P log(P̂ ).
(1)
fc10
fc9
conv1
~
conv7
*
L_jig.
*
conv8
conv8
conv8
conv1
~
conv7
*
conv10
conv9
conv1
~
conv7
conv9
L_inp.
L_col.
: Shuffle
: Concatenate
: Arrange
*
: Shared weights
Figure 3. The architecture for “Completing damaged jigsaw puzzles”. It is a 9-tower siamese network. The shared tower(colored in
gray) consists of AlexNet conv1-7 layers. note f c6-7 are converted into equivalent conv6-7 layers for the pixel-level outputs. The task
branches for jigsaw puzzle, inpainting, and colorization are marked in blue, red, and orange, respectively. The learned shared tower is used
for transfer learning on downstream tasks.
Inpainting. Inpainting is a problem of restoring lost regions of an image. In the field of representation learning, a
small patch Xp is removed from the image X, and remaining parts Xr are used for inferring the removed patch Xp .
It is formulated as
X̂p = f inp (Xr ).
(2)
By solving this problem, the network learns contextual information of Xr and between Xr and Xp .
In order to do damage of a different flavor, we discard
a subset of image channels. Unlike the original inpainting
where all channels are given (Fig. 2-(d)), our complicated
inpainting requires generation of ab channels of the missing
region from the surrounding pixels in L channel (Fig. 2-(e)).
While struggling to solve this problem, the network learns
cross-channel relations as well as the contextual information. We use Euclidean distance between the prediction and
the ground truth as a loss as proposed in [30] as
Linp = Xˆpab − f inp (XrL )
2
,
2
(3)
where superscripts L and ab denote the input’s L and ab
channels, respectively.
Colorization. Colorization and other cross-channel prediction tasks [21, 38, 39] discard and recover a subset of image channels to learn cross-channel relations.
Additional damage for more difficult colorization takes
place in the context domain. We encourage the network
to see only part of images, and colorize in the absence of
the full context. Specifically, we feed the network with the
L channel of only one patch out of the 2×2 puzzles, and
push it to colorize the patch as shown in Fig. 3.1-(g). The
colorization becomes more difficult since only one-quarter
of the entire image is available.
As in [38], the network learns a mapping Xˆab =
fcol (X L ) to a probability distribution over possible colors
Xˆab ∈ [0, 1]313 , where the ab are quantized to 313 values.
We train the network with the 313-way classification loss
as,
Lcol = −
X
v(X ab log(Xˆab )),
(4)
where v(·) denotes a color-class rebalancing term.
3.2. Completing Damaged Jigsaw Puzzles
In order to further develop our idea, we design our final
problem, “Completing damaged jigsaw puzzles”, by involving all the damages and recoveries mentioned above. As its
name indicates, this problem requires the simultaneous recovery of the following damages: (1) shuffling the image
patches, (2) discarding one patch, (3) dropping ab channels
in all the patches. During training, the network is encouraged to arrange the puzzles, recover the missing context,
and colorize the patches. In practice, recovering the missing patch is defined as generating ab channels of the missing
region from the surrounding pixels in L channel.
Recent self-supervised learning methods that use multiple self-supervised tasks either assign separate features to
each tasks [39], train each tasks in sequential order [37],
or jointly train the tasks [5]. We share with them the goal
of learning a single set of well-rounded representations.
However, our approach complicates each involved tasks to
fuel the damage-and-recover, whereas the previous methods adopt the original form of existing self-supervised tasks.
More specifically, our final problem involves a jigsaw puzzle with one piece missing, inpainting across channels, and
colorization with a narrower view, which are more complicated than their predecessors. Also, each tasks are intermingled in a way that some tasks share the knowledge. That is,
the understanding of cross-channel relation supports both
the colorization and the inpainting, and the contextual information is shared across all tasks. As a result, the network learns to effectively integrate and propagate the different knowledge on the spatial configuration, image context, and cross-channel relations into the final representations. Finally, all our involved tasks share the input space: a
set of damaged puzzles. This makes our approach immune
to the risk in [5, 39] that use different inputs for each tasks,
where the network might task-specifically encode the representations depending on the type of inputs, as stated in [5].
In practice, our method operate on 3×3 puzzles rather
than 2×2, for more discriminative representations.
Architecture and Losses. Our architecture is shown in
Fig. 3. It is a 9-tower siamese network as in [25]. The
shared tower follows the standard AlexNet [20] to provide a
fair comparison with recent self-supervised learning methods [4, 6, 21, 25, 26, 30, 36–39]. The task branches of the
jigsaw puzzle, inpainting, and colorization are rooted to the
shared tower, and colored in blue, red, and orange, respectively.
In the jigsaw branch, 9 sets of the common features
(conv7 features) pass through a fully-connected layer,
f c8(blue), and are concatenated, then fed into two more
fully-connected layers up to f c10(blue), resulting in a 1000long vector. We use the same Ljig as Eq. (1). In the inpainting branch, the 9 features go through a 1×1 convolutional
layer. This time, we arrange the features before concatenating them as we know what permutation has been used in the
inputs. After two more 1×1 convolutions(conv9, conv10,
red), the features have a volume of 7 × 7 × 313, where 313
denotes the number of quantized color values as in [38].
Note that we use a classification loss rather than Eq. (3) as,
X
Lcls
v(Xpab log(Xˆpab )),
(5)
inp = −
where Xˆpab denotes the predicted chromaticity values of the
missing puzzle. Each of the 9 features is fed into the colorization branch, resulting in 9 branches. Each branch is an
equivalent form of the network in [38] which has two more
1 × 1 convolutions(conv8, conv9, orange) after the shared
tower, resulting in features of 7 × 7 × 313. Our colorization
loss is a sum of the 9 losses of Eq. (7) as,
Lcol
9 X
X
=−
(
v(Xiab log(Xˆiab ))),
three losses as,
Lf inal = Ljig + αLcls
inp + β Lcol ,
where α and β are weighting parameters.
Simple Combination. We also consider combining the
original forms of self-supervised tasks, conceptually following [5]. We jointly train original versions of the three
tasks: jigsaw puzzles, inpainting, and colorization. Although the types of involved tasks are different to [5], we
provide a self-comparison on the effectiveness of our approach and the simple combination in Sec. 5.3.
4. Training
We train our proposed network on 1.3M images from the
training set of ImageNet without annotations. We resize the
input images to 312×312 pixels, and extracted patches of
140×140 and 85×85, in 2-by-2 and 3-by-3 puzzles, respectively. We use caffe [14] for implementation. The network
is trained by ADAM optimizer [18] for 350K iterations with
batch size of 64 on a machine with a GTX 1080-Ti GPU and
an intel i7 3.4GHz CPU. The learning rate is set to 10−3 ,
and is dropped by a factor of 0.1 every 100K iterations. We
use α, β = 0.01 for the experiment in Sec. 3.2. Inpainting and colorization of Sec. 3.1 follow the protocol of their
original papers [30, 38], respectively.
5. Results and Discussions
In this section, we provide both quantitative and qualitative evaluations and discussions of our self-supervised
learning approach. Further transfer learning results on new
tasks(e.g. depth prediction) and with deeper network(e.g.
vgg [33]) are presented in our supplementary material.
5.1. Fine-tuning on PASCAL
In this section we evaluate the effectiveness of both the
“Complicating each self-supervised tasks” in Sec. 3.1 and
our final task, “Completing damaged jigsaw puzzles” in
Sec. 3.2. To do this, we transfer the learned representations
to a standard AlexNet [20] and rescale the weights via [19].
We test on some or all of the PASCAL tasks, using VOC
2007 [7] for classification and detection, VOC 2012 [8] for
segmentation; these are standard benchmarks for representation learning.
5.1.1
(6)
i=1
where Xi denotes ith of the input patches. Finally, our loss
for “Completing damaged jigsaw puzzles” is the sum of the
(7)
Complicating each self-supervised task
In Sec. 3.1, we explore the idea of complicating the jigsaw
puzzle, inpainting, and colorization to benefit representation learning. We evaluate the effectiveness of each complications by comparing the performances before and after
Method
Jigsaw( Sec. 3.1)
Jigsaw( Sec. 3.1)
Jigsaw( Sec. 3.1)
Inpainting [30]
Inpainting( Sec. 3.1)
Colorization [38]
Colorization( Sec. 3.1)
Complication
None
L-or-ab dropped
A piece removed
None
Cross-Channel
None
Narrow view
Class.
64.7
65.5
65.3
56.5
57.7
65.9
66.7
Segm.
34.9
35.7
35.7
29.7
30.2
35.7
36.8
Table 1. Effectiveness of complicating self-supervised tasks on
PASCAL. Classification is evaluated on PASCAL VOC 2007
with testing frameworks from [19], using mean average precision(mAP) as a performance measure. Segmentation is evaluated
on PASCAL VOC 2012 with testing framework from [23], which
reports mean intersection over union(mIU).
the complications in downstream tasks: classification and
semantic segmentation.
The results are shown in Table. 1. In all cases, the complicated self-supervised tasks consistently achieve higher
scores than their predecessors both in classification and segmentation. These results indicate that the capacity of the
network was still above the difficulty of the existing selfsupervised tasks, and that indeed, useful representations can
be extracted more via solving more difficult tasks.
5.1.2
Completing Damaged Jigsaw Puzzles
We evaluate how beneficial is our final self-supervised task,
“Completing damaged jigsaw puzzles”, in learning representations. We transfer the learned weights from the shared
tower Fig. 3 on classification, detection, and semantic segmentation. As shown in Table. 2, our method outperforms
all the previous methods in classification and segmentation,
and achieves the second best performance in the detection
task, even though the network has been exposed only on
grayscale images during pretraining. We also summarize
the comparison on classification and segmentation tasks in
Fig. 4 which indicates that our approach learns more robust
and general-purpose representations in comparison to each
of the involved tasks and all the conventional methods.
5.2. Linear Classification on ImageNet
We test the task-generality of our learned representations
on large-scale representation learning benchmarks. As proposed in [38], we freeze each layer of our learned features from conv1 to conv5, and initialize the subsequent
unfrozen layers with random values. Then, we train linear
classifiers on top of each layer on labeled ImageNet [32]
dataset.
The result is shown in Table. 3. ImageNet-pretrained
AlexNet shows the best performance and is the upper bound
in this comparison. Since our network only learns from L
channel, conv1 features suffer lack of input information,
resulting in slightly lower score compared to other meth-
Method
ImageNet [20]
Random
RelativePosition [4]
Jigsaw [25]
Ego-motion [36]
Adversarial [6]
Inpainting [30]
Colorization [38]
Split-Brain [39]
ColorProxy [21]
WatchingObjectMove [29]
Counting [26]
CDJP
Class.
79.9
53.3
65.3
67.6
54.2
58.6
56.5
65.9
67.1
65.9
61.0
67.7
69.2
Det.
56.8
43.4
51.1
53.2
43.9
46.2
44.5
46.9
46.7
52.2
51.4
52.4
Segm.
48.0
19.8
37.6
34.9
29.7
35.6
36.0
38.4
36.6
39.3
Table 2. Evaluation of transfer learning on PASCAL. Classification and detection are evaluated on PASCAL VOC 2007 with
testing frameworks from [23] and [9], respectively. Both tasks are
evaluated using mean average precision(mAP) as a performance
measure. Segmentation is evaluated on PASCAL VOC 2012 with
testing framework from [23], which reports mean intersection
over union(mIU).
Method
ImageNet [20]
Random
RelativePosition [4]
Jigsaw [25]
Adversarial [6]
Inpainting [30]
Colorization [38]
Split-Brain [39]
Counting [26]
CDJP
conv1
19.3
11.6
16.2
18.2
14.1
17.7
12.5
17.7
18.0
14.5
conv2
36.3
17.1
23.3
28.8
20.7
24.5
24.5
29.3
30.6
27.2
conv3
44.2
16.9
30.2
34.0
21.0
31.0
30.4
35.4
34.3
32.8
conv4
48.3
16.3
31.7
33.9
19.8
29.9
31.5
35.2
32.5
34.3
conv5
50.5
14.1
29.6
27.1
15.5
28.0
30.3
32.8
25.7
32.9
Table 3. Linear classification on ImageNet. We train linear classifiers on top of each layer of the learned feature representations.
We use publicly available testing code from [38] and report top1 accuracy of AlexNet on ImageNet 1000-way classification. The
learned weights between conv1 and the displayed layer are frozen.
ods. However, it overcomes this handicap immediately
from conv2 layer, and achieves competitive performances
in higher layers. Finally, conv4 and conv5 features achieve
the second best and state-of-the-art performances, respectively.
As shown in [25], the last layers of the pretrained network tend to be task-specific, while the first layers are
general-purpose. In our proposed architecture(Fig. 3), this
transition from general-purpose to task-specific is delayed
and left to the task branches. Since the last features of the
shared tower must support all three different, they should remain as general as possible, rather than get biased to either
of the tasks. Also, the network can hardly assign separate
features to each tasks since the features required by the tasks
often overlap, thus it has to integrate and hold the different
features up to the last layers.
Combination
Class.
Segm.
Jig.
66.6
36.8
Jig.+Inp.
67.4
37.9
Jig.+Col.
68.4
38.6
Jig.+Inp.+Col./simple
68.0
38.1
Jig.+Inp.+Col.(CDJP)
69.2
39.3
Table 4. Comparing different combinations of self-supervised
tasks on PASCAL. we evaluate different combinations of selfsupervised tasks on PASCAL classification and segmentation in
the same setting as in Table. 1. We make different combinations using our architecture (Fig. 3) with or without certain task
branches; this may have caused slight performance differences
from the original task. We also report the result of simple combinations where the original versions of each tasks are jointly
trained.
5.3. Comparing Combinations of Self-supervised
tasks
In order to show the impact of each task, we evaluate different combinations on PASCAL classification and semantic segmentation tasks. We experiment with the same architecture Fig. 3, but with or without certain task branches to
make different combinations. In addition, as we mentioned
in Sec. 3.2, we provide the result of the simple combination of the tasks in their original form, which conceptually
follows [5].
The results are shown in Table. 4. We set the jigsaw
puzzle as our starting point and add different tasks to it. We
can see that the performances increase every time the tasks
are combined. Our final method which combines all three
tasks obtains the best scores and improves our jigsaw puzzle
by 2.6% and 2.5%scores both in classification and semantic
segmentation tasks. The simple combination of the original
versions slightly improves their single-task baselines [25,
30, 38] in both test tasks, but not better than our Jig.+Col.
and Jig.+Inp.+Col.(CDJP) methods.
5.4. Nearest Neighbor Search
The pretrained networks recognize the semantic similarity of data by their own standards. We qualitatively evaluate
the validity of this reasoning of the networks by performing
‘nearest neighbor search’ which has been proposed in [4]
and further used in [26, 37]. In this experiment we compare AlexNets [20] pretrained by different methods: jigsaw
puzzle [25], inpainting [30], colorization [38], ours, and ImageNet classification [20]. We perform retrieval on f c6 (the
feature before the concatenation) for jigsaw puzzle, conv5
(the last layer of the encoder) for inpainting, and conv7/f c7
features for the remaining methods.
Single-task Baselines. As in figure 5, the learned representations in each methods show distinct characteristics.
For example, the jigsaw puzzle representations retrieve ob-
Figure 4. Summarization of performances of different selfsupervised learning methods and combinations. We compare
the state-of-the-art methods(Table. 2), our final method (CDJP),
and each involved tasks in our final method and the simple combination( Table. 4). The involved tasks, their original versions,
the simple combination, the other existing methods, and our final
method are marked in orange, green, gray, blue and red, respectively. Note that Jig. is what we reproduced in our architecture.
jects with the same pose and shape. Even in the blurred image, it retrieves objects with similar silhouettes. In inpainting, objects that would co-occur or share the similar background are retrieved, such as things to ride for horse and
caregivers for baby. The features learned by colorization
is often color-specific, and retrieves babies wearing pink
clothes for baby, and sometimes false samples with bluegreen color for bottle. Also, blurred objects a retrieved for
the blurred image. Such color-sensitivity sometimes misrepresent semantics, e.g. a brown chair back is retrieved for
horse image.
Similarity to ImageNet Classification Pretraining. Note
that we consider pretraining on ImageNet classification as
our gold standard in this qualitative evaluation. Our approach integrates the characteristics of the single-task baselines, yet mostly complement and overcome the aforementioned sensitivities. First, our approach is more invariant to
pose/viewpoint variations compared to jigsaw puzzle baseline, and represents horses and babys in different pose and
viewpoint as semantically nearby, which is also the case
in ‘ImageNet’ model. Furthermore, our representations are
more robust in intra-class color variations, and retrieves objects with various colors according to horse, baby, and bottle query images, which also raise our model closer to our
gold standard. Our model also adopts the virtues of the
single-task baselines. To illustrate, for blurred object, as
in colorization, our model retrieves images that are semantically ambiguous. We can see the same tendency in the
Jig.
Inp.
Col.
Ours
INet
Jig.
Inp.
Col.
Ours
INet
Figure 5. Nearest Neighbor Search. We perform image retrieval on the object instances cropped from the PASCAL VOC 2012 [8] trainval
dataset. The query images are in red boxes. Down from the top rows are the retrieval results of jigsaw puzzle, inpainting, colorization, our
method, and ImageNet classification, respectively.
‘ImageNet’ model, where it may consider the query image to be vague, and retrieves also blurred objects in different categories. Finally, our model adopts a reasonable
understanding on the image context, which enabled the retrieval of co-occurable objects, e.g., person with horse and
parent with baby. Interestingly, we observe that the ‘ImageNet’ retrieves images where person and horse; caregiver
and baby appear together, similarly to ours. These results
can be viewed as one reason that our approach can propagate the high-level semantics through our model, and raise
its robustness and task generality of our representations.
ing. Furthermore, we design “Completing damaged jigsaw puzzles” as a more complicated and complex problem
for self-supervised representation learning. While learning
to recover and colorize original image content simultaneously, rich and general-purpose visual features are encoded
into the network. Experiments contain transfer learning on
PASCAL VOC classification, detection and segmentation,
ImageNet linear classification as well as nearest neighbor
search. All of the results clearly show that the features
learned by our method generalize well across different highlevel visual tasks.
6. Conclusions
Acknowledgements This research is supported by the
Study on representation learning for object recognition
funded by the Samsung Electronics Co., Ltd (Samsung Research)
In this paper, we study complicating self-supervised
tasks for representation learning. We propose complicated versions of jigsaw puzzles, inpainting and colorization and show their effectiveness on representation learn-
References
[1] P. Agrawal, J. Carreira, and J. Malik. Learning to see by
moving. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13,
2015, 2015.
[2] R. Arandjelovic and A. Zisserman. Ambient sound provides
supervision for visual learning. In Proc. of Int’l Conf. on
Computer Vision (ICCV), 2017.
[3] M. W. Diederik P Kingma. Auto-encoding variational bayes.
In Proc. of Int’l Conf. on Learning Representations (ICLR),
2014.
[4] C. Doersch, A. Gupta, and A. A. Efros. Unsupervised visual
representation learning by context prediction. In Proc. of
Int’l Conf. on Computer Vision (ICCV), 2015.
[5] C. Doersch and A. Zisserman. Multi-task self-supervised
visual learning. In Proc. of Int’l Conf. on Computer Vision
(ICCV), 2017.
[6] J. Donahue, P. Krähenbühl, and T. Darrelln. Adversarial feature learning. In Proc. of Int’l Conf. on Learning Representations (ICLR), 2017.
[7] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn,
and A. Zisserman. The PASCAL Visual Object Classes
Challenge 2007 (VOC2007) Results. http://www.pascalnetwork.org/challenges/VOC/voc2007/workshop/index.html.
[8] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn,
and A. Zisserman. The PASCAL Visual Object Classes
Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html.
[9] R. Girshick. Fast r-cnn. In Proc. of Int’l Conf. on Computer
Vision (ICCV), 2015.
[10] H.Bilen and A. Vedaldi. Weakly supervised deep detection
networks. In Proc. of Computer Vision and Pattern Recognition (CVPR), 2016.
[11] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proc. of Computer Vision and
Pattern Recognition (CVPR), 2016.
[12] P. Isola, D. Zoran, D. Krishnan, and E. H. Adelson. Learning visual groups from co-occurrences in space and time. In
ICLR Workshop, 2015.
[13] D. Jayaraman and K. Grauman. Learning image representations tied to egomotion from unlabeled video. Int’l Journal
of Computer Vision (IJCV), 125(1-3):136–161, 2017.
[14] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint
arXiv:1408.5093, 2014.
[15] Z. Jie, Y. Wei, X. Jin, J. Feng, and W. Liu. Deep self-taught
learning for weakly supervised object localization. In Proc.
of Computer Vision and Pattern Recognition (CVPR), 2017.
[16] V. Kantorov, M. Oquab, M. Cho, and I. Laptev. Contextlocnet: Context-aware deep network models for weakly supervised localization. In Proc. of European Conf. on Computer
Vision (ECCV), 2016.
[17] D. Kim, D. Cho, and D. Yoo. Two-phase learning for weakly
supervised object localization. In Proc. of Int’l Conf. on
Computer Vision (ICCV), 2017.
[18] D. P. Kingma and J. Ba. Adam: A method for stochastic
optimization. In Proc. of Int’l Conf. on Learning Representations (ICLR), 2015.
[19] P. Krähenbühl, C. Doersch, J. Donahue, and T. Darrell. Datadependent initializations of convolutional neural networks.
In Proc. of Int’l Conf. on Learning Representations (ICLR),
2016.
[20] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Proc.
of Neural Information Processing Systems (NIPS), 2012.
[21] G. Larsson, M. Maire, and G. Shakhnarovich. Colorization
as a proxy task for visual understanding. In Proc. of Computer Vision and Pattern Recognition (CVPR), 2017.
[22] H.-Y. Lee, J.-B. Huang, M. Singh, and M.-H. Yang. Unsupervised representation learning by sorting sequences. In
Proc. of Int’l Conf. on Computer Vision (ICCV), 2017.
[23] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional
networks for semantic segmentation. In Proc. of Computer
Vision and Pattern Recognition (CVPR), 2015.
[24] I. Misra, C. L. Zitnick, and M. Hebert. Shuffle and learn:
Unsupervised learning using temporal order verification. In
Proc. of European Conf. on Computer Vision (ECCV), 2016.
[25] M. Noroozi and P. Favaro. Unsupervised visual representation learning by context prediction. In Proc. of European
Conf. on Computer Vision (ECCV), 2016.
[26] M. Noroozi, H. Pirsiavash, and P. Favaro. Representation
learning by learning to count. In Proc. of Int’l Conf. on Computer Vision (ICCV), 2017.
[27] M. Oquab, L. Bottou, I. Laptev, and J. Sivic. Is object localization for free? - weakly-supervised learning with convolutional neural networks. In Proc. of Computer Vision and
Pattern Recognition (CVPR), 2015.
[28] A. Owens, J. Wu, J. H. McDermott, W. T. Freeman, and
A. Torralba. Look, listen and learn. In Proc. of European
Conf. on Computer Vision (ECCV), 2016.
[29] D. Pathak, R. B. Girshick, P. Dollár, T. Darrell, and B. Hariharan. Learning features by watching objects move. In Proc.
of Computer Vision and Pattern Recognition (CVPR), 2017.
[30] D. Pathak, P. Krähenbühl, J. Donahue, T. Darrell, and A. A.
Efros. Context encoders: Feature learning by inpainting. In
Proc. of Computer Vision and Pattern Recognition (CVPR),
2016.
[31] A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. In Proc. of Int’l Conf. on Learning Representations (ICLR), 2016.
[32] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual
Recognition Challenge. Int’l Journal of Computer Vision
(IJCV), 115(3):211–252, 2015.
[33] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR,
abs/1409.1556, 2014.
[34] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. In Proc. of Computer Vision and Pattern Recognition (CVPR), 2015.
[35] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A.
Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising
criterion. J. Mach. Learn. Res., 11:3371–3408, 2010.
[36] X. Wang and A. Gupta. Unsupervised learning of visual representations using videos. In Proc. of Int’l Conf. on Computer Vision (ICCV), 2015.
[37] X. Wang, K. He, and A. Gupta. Transitive invariance for
self-supervised visual representation learning. In Proc. of
Int’l Conf. on Computer Vision (ICCV), 2017.
[38] R. Zhang, P. Isola, and A. A. Efros. Colorful image colorization. In Proc. of European Conf. on Computer Vision
(ECCV), 2016.
[39] R. Zhang, P. Isola, and A. A. Efros. Split-brain autoencoders:
Unsupervised learning by cross-channel prediction. In Proc.
of Computer Vision and Pattern Recognition (CVPR), 2017.
| 1cs.CV
|
1
1
PREPRINT VERSION
Matrix
MultiMatrix Factorization
Factorization for
for Nonparametric
Nonparametric MultiSource
Properties
SourceLocalization
Localization Exploiting
Exploiting Unimodal
Unimodal Properties
arXiv:1711.07457v1 [cs.IT] 20 Nov 2017
Junting
JuntingChen
Chenand
and Urbashi
Urbashi Mitra
Mitra
Ming
Hsieh
Department
of
Electrical
Engineering,
California
Ming Hsieh Department of Electrical Engineering, University
University of
of Southern
Southern California
Los
Angeles,
CA
90089
USA,
email:{juntingc,
ubli}@usc.edu
Los Angeles, CA 90089 USA, email:{juntingc, ubli}@usc.edu
Abstract—Herein,
Abstract—Herein,the
theproblem
problemofofsimultaneous
simultaneous localization
localization
of ofmultiple
multiplesources
sourcesgiven
givena anumber
number ofof energy
energy samples
samples atat
different
differentlocations
locationsis isexamined.
examined.The
Thestrategies
strategiesdo
donot
not require
require
knowledge
of of
thethe
signal
knowledge
signalpropagation
propagationmodels,
models,nor
nordo
dothey
theyexploit
exploit
thethespatial
spatialsignatures
signaturesofofthe
thesource.
source.AAnonparametric
nonparametric source
source
localization
framework
localization
frameworkbased
basedonona amatrix
matrixobservation
observationmodel
modelisis
developed.
It It
is is
shown
developed.
shownthat
thatthe
thesource
sourcelocation
locationcan
canbe
beestimated
estimated
byby
localizing
localizingthethepeaks
peaksofofa apair
pairofoflocation
locationsignature
signaturevectors
vectors
extracted
extractedfrom
fromthetheincomplete
incompleteenergy
energyobservation
observation matrix.
matrix. AA
robustpeak
peaklocalization
localizationalgorithm
algorithmisisdeveloped
developedand
and shown
shown toto
robust
decrease
source
localizationmean
meansquared
squarederror
error(MSE)
(MSE)faster
faster
decrease
thethe
source
localization
1.51.5
than
O(1/M
) with
samples.ToToextract
extractthe
thesource
sourcesignature
signature
than
O(1/M
) with
MMsamples.
vectors
from
a matrixwith
withmixed
mixedenergy
energyfrom
frommultiple
multiplesources,
sources,
vectors
from
a matrix
a unimodal-constrainedmatrix
matrixfactorization
factorization(UMF)
(UMF)problem
problemisis
a unimodal-constrained
formulated,
andtwo
tworotation
rotationtechniques
techniquesare
aredeveloped
developedtotosolve
solve
formulated,
and
UMF
efficiently.
Our
numericalexperiments
experimentsdemonstrate
demonstratethat
that
thethe
UMF
efficiently.
Our
numerical
proposed
schemeachieves
achievessimilar
similarperformance
performanceasasthe
thekernel
kernel
thethe
proposed
scheme
regression
baseline
usingonly
only1/5
1/5energy
energymeasurement
measurementsamples
samples
regression
baseline
using
detecting
singlesource,
source,and
andthe
theperformance
performancegain
gainisismore
more
in in
detecting
a asingle
significant
in
the
cases
of
detecting
multiple
sources.
significant in the cases of detecting multiple sources.
IndexTerms—Source
Terms—Sourcelocalization,
localization, unimodal,
unimodal, sparse
sparse signal
signal
Index
processing,
matrixcompletion,
completion,nonparametric
nonparametricestimation
estimation
processing,
matrix
I NTRODUCTION
I. I.I NTRODUCTION
Source
localizationis isimportant
importantininmany
manydomains,
domains,such
suchas
as
Source
localization
salvage,
exploration,
tactical
surveillance,
and
hazard
findsalvage, exploration, tactical surveillance, and hazard finding.However,
However,ininmany
manyapplication
applicationscenarios,
scenarios, itit isis difficult
difficult
ing.
to obtain the correct propagation parameters of the source
to obtain the correct propagation parameters of the source
signal for localization. For example, in underwater localization
signal for localization. For example, in underwater localization
with acoustic signals, the signal propagation depends on the
with acoustic signals, the signal propagation depends on the
water temperature, pressure, and salinity, which are locationwater temperature, pressure, and salinity, which are locationdependent. In gas source localization, the gas diffusion chardependent. In gas source localization, the gas diffusion characteristics depends on the chemical type and the atmospheric
acteristics depends on the chemical type and the atmospheric
conditions. Therefore, model-based parametric localization
conditions.
Therefore,
model-based
localization
methods [1]–[6]
may not
be reliableparametric
in application
scenarios
methods
[1]–[6]
may
not
be
reliable
in
application
scenarios
with a temporal and spatial varying nature.
with Model-free
a temporal positioning
and spatialschemes,
varying nature.
such as connectivity based
Model-free
positioning
schemes,
as connectivity
based
localizations and weighted centroidsuch
localizations
(WCL),
have
localizations
and
weighted
centroid
localizations
(WCL),
have
attracted a lot of interest due to their simplicity in impleattracted
a lot
interest
due tototheir
simplicity
in implementation
andof the
robustness
variations
of propagation
mentation
and
the
robustness
to
variations
of
propagation
properties [7]–[12]. However, connectivity based techniques
properties
[7]–[12].
However,
connectivity
based
techniques
[7], [8] can
only provide
coarse
localization
results
and the
[7],performance
[8] can only
provide
coarse
localization
results
and
the
of WCL [9], [10] highly depends on the
choice
performance
of
WCL
[9],
[10]
highly
depends
on
the
choice
of parameters and the propagation environments. Recently,
of machine
parameters
and the
propagation
Recently,
learning
techniques,
such environments.
as kernel regression
and
machine learning techniques, such as kernel regression and
This research has been funded in part by one or more of the following
grants:
ONR N00014-15-1-2550,
CCF-1718560,
This
research
has been funded inNSF
part CNS-1213128,
by one or moreNSF
of the
following
NSF ONR
CCF-1410009,
NSF CPS-1446901,
and AFOSR FA9550-12-1-0215.
grants:
N00014-15-1-2550,
NSF CNS-1213128,
NSF CCF-1718560,
NSF CCF-1410009, NSF CPS-1446901, and AFOSR FA9550-12-1-0215.
Figure 1. Left: Energy measurements at different locations (colored-squares)
Figure 1. Left: Energy measurements at different locations (colored-squares)
to localize two sources (red stars). Right: The underlying energy field appears
to localize two sources (red stars). Right: The underlying energy field appears
as one peak when the two sources are close to each other.
as one peak when the two sources are close to each other.
support vector
vector machines
machines [13]–[15]
[13]–[15] have
support
have also
also been
been explored
explored
for
localization.
However,
these
methods
usually
for localization. However, these methods usually require
require aa
separate training
training phase
phase which
which may
may not
separate
not be
be available
available in
in practice.
practice.
This
paper
studies
nonparametric
methods
for
This paper studies nonparametric methods for localizing
localizing
several sources
sources based
based on
on aa few
several
few energy
energy measurements
measurements atat
different
sensing
locations
as
illustrated
in
Fig.
different sensing locations as illustrated in Fig. 1.
1. Our
Ourprevious
previous
works
studied
the
single
source
case
in
[11],
[16],
works studied the single source case in [11], [16], where
where aa
trust region was developed for targeting the source and a multitrust region was developed for targeting the source and a multistep exploration-exploitation strategy was developed for active
step exploration-exploitation strategy was developed for active
search using an underwater robot. The results were extended
search using an underwater robot. The results were extended
to the cases of two or more sources by exploiting novel
to the cases of two or more sources by exploiting novel
coordinate system rotation techniques [12], [17]. However,
coordinate system rotation techniques [12], [17]. However,
these works were based on a decomposable assumption on the
these works were based on a decomposable assumption on the
matrix formed by the energy measurements in a discretized 2D
matrix
formed by the energy measurements in a discretized 2D
area.
area.
In this paper, we find that the decomposable assumption
In this paper, we find that the decomposable assumption
is not necessary. Specifically, we show that the source loiscation
not necessary.
Specifically, we show that the source loin 2D can be found by localizing the peaks of a
cation
in
2D
can
be found
by localizing
the peaks
of aa
pair of location signature
vectors
that are obtained
from
pair
of
location
signature
vectors
that
are
obtained
from
sparsely observed energy matrix based on measurements ina
sparsely
observed
energy
measurements
a discretized
2D area,
andmatrix
such a based
result on
holds
universally asin
along
discretized
2D
area,
and
such
a
result
holds
universally
as
as the signal energy decays in all directions
away from
long
as
the
signal
energy
decays
in
all
directions
away
from
the source location. Therefore, the method is robust in a lot of
the
source propagation
location. Therefore,
the method
is robust
a lot of
unknown
environments.
In addition,
weinquantify
unknown
propagation
environments.
In
addition,
we
quantify
the performance degradation when the actual energy matrix is
the
degradation
when
the actual
energystrategies
matrix is
notperformance
ideally rank-1.
Based on this
model,
we develop
not
ideally rank-1.
Based
on this
model, we develop
strategies
to localize
multiple
sources
by overcoming
three additional
tochallenges:
localize multiple
by peak
overcoming
threealgorithms
additional
First, wesources
formulate
localization
challenges:
First, we vectors
formulate
peak localization
algorithms
for noise corrupted
extracted
from a sparse
energy
for
noise corrupted
a sparse
energy
observation
matrix. vectors
Second, extracted
we designfrom
matrix
factorization
observation
matrix.
we design
matrix
factorization
techniques to
extractSecond,
the source
signature
vectors
from the
techniques to extract the source signature vectors from the
PREPRINT VERSION
2
matrix that collects mixed energy from multiple sources.
Note that the singular value decomposition (SVD) framework
in our preliminary work [12], [16] does not work herein
since the signature vectors may not be orthogonal. Third, we
develop optimization techniques for the optimal rotation of
the coordinate system, which determines how to arrange the
measurements into the horizontal and vertical entries of the
observation matrix.
To summarize, the following contributions are made herein:
• We propose an energy matrix observation model and
prove that the matrix has a unimodal property, where the
left and right dominant singular vectors, corresponding to
each source, are unimodal with their peaks representing
the source locations.
• We develop localization algorithms that exploit the unimodal and symmetric property. We prove that the location
MSE decreases faster than O(1/M 1.5 ) in the single
source case.
• For the case of multiple sources, we develop optimization
techniques to establish an optimal coordinate system
with an rotation angle adaptive to the unknown source
topology.
• We show that the optimal rotation can be found by
maximizing the normalized dominant singular value of
the observation matrix in the two source case.
• For the case of arbitrary number of sources, we solve
a UMF problem and show that the convergence can be
enhanced by rotating the coordinate system to minimize
the normalized dominant singular value.
The rest of the paper is organized as follows. Section II
establishes the matrix observation model and develops the
unimodal property of source signature vectors. Section III
studies the case of single source and develops the MSE
localization performance. Section IV and V develop rotation
optimization techniques for establishing the coordinate system
in the two source and arbitrary number of sources cases.
Numerical results are presented in Section VI and conclusion
is given in Section VII.
Notation: Vectors are written as bold italic letters x and
matrices as bold capital italic letters X . Random variables,
random vectors, and random matrices are written as x, bold
letters x, and bold capital letters X, respectively. For a matrix
X, Xij denotes the entry in the ith row and jth column of X.
For a vector x, xi denotes the ith entry of x. For a set X , |X |
denote the Cardinality of X , i.e., the number of elements in
X . The notation o(x) means limx→0 o(x)/x → 0, and O(x)
means lim supx→0 O(x)/x < ∞.
II. S YSTEM M ODEL
A. Signal Model
Assume that there are K sources in an area with radius L/4.
The location of source k is denoted by s k ∈ R2 . The sources
continuously emit signals that form an aggregated energy field,
which can be measured at M different sensing locations that
are distributed uniformly and randomly in the target area A
with radius L/2. Let d(z , s) = kz − sk2 be the distance
between the sensing location at z ∈ R2 and a source location
s. The energy measurement h(m) at the mth sensing location
z (m) is given by
h(m) =
K
X
αk h(d(z (m) , s k )) + n(m)
(1)
k=1
where αk is the transmit power of source k and n(m) is
the additive noise of the mth measurement with zero mean
and variance σn2 . The function h(d) is a non-negative strictly
decreasing function of the distance d from the source. In
addition, we assume that h(d) is Lipschitz continuous and
square-integrable.
Without loss of generality (w.l.o.g.), assume
RR
that R2 h(d(z, s))2 dz = 1. Note that neither the source
power αk nor the function h(d) are known.
B. Non-parametric Matrix Observation Model
1) Matrix Model of the Energy Field: We discretize the
L × L area containing the target area, A, into N
√1 × N2 grid
points. W.l.o.g., assume that N1 = N2 = N ≥ M , and the
grid points are equally spaced.
Let H (k) ∈ RN ×N be the discretized energy field matrix
of source k, i.e., the (i, j)th element of H (k) is given by
(k)
Hij =
L
αk h(d(c i,j , s k ))
N
(2)
where c i,j ∈ R2 is the center location of the (i, j)th grid and
L
is for normalization purposes:
the factor N
X L 2
2
h(d(c i,j , s k ))2
H (k) F = α2k
N
i,j
ZZ
L
2
≈ α2k (3)
= α2k
h(d(z, s k ))2 dz + o
N
A
where we have assumed that h(d(z,
s)) has a negliRR
h(d(z,
s k ))2 dz ≈
gible
tail
outside
of
A,
i.e.,
2
R
RR
2
1
h(d(z, s k )) dz.
A
P
Consider the SVD of H (k) = i σk,i u k,i v Tk,i , where σk,i
denote the ith largest singular value of H (k) . In particular,
denote uk = uk,1 and vk = vk,1 . Then, we have the following
model for the K source energy field.
Definition 1 (Signature vector and signature matrix). The
signature matrix for all K sources is defined as
H ,
K
X
k=1
H (k) =
K
X
k=1
σk,1 uk vkT +
K X
N
X
σk,i u k,i v Tk,i (4)
k=1 i=2
where the vectors uk and vk , are called the signature vectors
of source k.
Definition 2 (Unimodal). A vector v ∈ RN is unimodal if the
following is satisfied:
0 ≤v1 ≤ v2 ≤ · · · ≤ vs
vs ≥ vs+1 ≥ · · · ≥ vN ≥ 0
(5)
(6)
for some integer 1 ≤ s ≤ N , where vi is the ith entry of v.
1 Note that the proposed algorithm does not require such an assumption.
This approximation is just for the ease of discussion.
PREPRINT VERSION
3
We show that the signature vectors uk and vk are unimodal
and symmetric about the source location.
Theorem 1 (Unimodal Signature Vector). The signature vectors uk and vk are unimodal. In addition, suppose that source
k locates inside the (m, n) th grid. Then the peaks of uk and
vk locate at the mth entry of uk and the nth entry of vk ,
respectively.
Proof: See Appendix A.
Theorem 1 confirms the unimodal property of the signature
vectors extracted from the rank one approximation of any
energy field. Such a property holds universally as long as
the propagated energy is a strictly decreasing function of the
distance from the source. Note that the result in Theorem 1
is not trivial, as such a unimodal property does not hold for
other singular vectors uk,i , vk,i , for i = 2, 3, . . . , N .
Consider a Cartesian coordinate system C, where the xaxis corresponds to a row of the grid centers and the y-axis
corresponds to a column. Denote the vector of x coordinates
of a row ci,1 , ci,2 , . . . , ci,N as cX = [cX,1 , cX,2 , . . . , cX,N ]
for all rows i, and the vector of y coordinates of a column
cN,j , cN −1,j , . . . , c1,j as cY = [cY,1 , cY,2 , . . . , cY,N ] for all
columns j.
Proposition 1 (Symmetric Signature Vector). For each source
k, the signature vectors can be approximated by unimodal and
(k)
symmetric vectors ůk and v̊k , such that Hij −αk ůk,i v̊k,j ≤
K
L
h
√
, where Kh is a bounded constant such that |h(d1 ) −
2N
h(d2 )| ≤ Kh |d1 − d2 | for all d1 ,p
d2 ≥ 0. In addition, ůk
L/N w(cY,i + δ2 − sk,2 )
and v̊k can be
obtained
as
ů
=
k,i
p
and v̊k,j = L/N w(cX,j + δ1 − sk,1 ), where [sk,1 , sk,2 ] =
sk are the source coordinates, [δ1 , δ2 ] = sk − cm,n is the
distance from the source sk to the nearest grid center cm,n ,
and w(x) is a non-negative, unimodal, and symmetric function,
d
i.e., w(x) ≥ 0, w(x) = w(−x), and dx
w(x) < 0 for x > 0.
Proof: See Appendix B.
The results in Theorem 1 and Proposition 1 suggest a
decoupled source localization strategy for the x coordinate and
the y coordinate, respectively: the source location sk can be
found from localizing the peaks of one-dimension data arrays
uk and vk extracted as the dominant singular vectors of H (k) .
Note that, by contrast, existing nonparametric methods (e.g.,
weighted centroid and kernel regressions) require iterative
search in two-dimensional (or, in general, 2K-dimensional)
source location space.
Such insight leads to the following observation model.
2) Matrix Observation Model: Let H be the N × N
observation matrix that contains the M energy measurements.
Specifically, the (i, j)th entry of H is given by
Hij =
L (m)
h
N
(7)
if the mth energy measurement is taken inside the (i, j)th grid.
Note that the measurement location z (m) may not be at the
center ci,j of the grid.
There are two strategies for constructing the observation
√
matrix H. In a conservative construction, we choose N = M
and the matrix is fully observed.2 We may extract the signature
vectors directly
√ from H. In an aggressive construction, we
choose N > M and the matrix H is partially observed.
C. Problem Formulation
Denote by UsN the cone specified by the unimodal
conS
N
straints (5) – (6) for a fixed s. Denote U N = N
s=1 Us as
the non-negative unimodal cone, and U N ×K as the set of
N × K real matrices where all the columns are in U N . Let
U , V ∈ RN ×K be the matrices that each contains K pairs of
signature vectors {uk , vk } to be determined. Let W ∈ RN ×N
be an indicator matrix that describes the sampling strategy,
where Wij = 1, if (i, j) ∈ Ω, and Wij = 0, otherwise, where
Ω denotes the set of entries that are assigned values based on
(7), |Ω| = M .
Based on the unimodal property developed in Theorem 1,
we are interested in extracting from H the vectors that are
unimodal. Specifically, the source signature vectors can be
extracted by solving the following UMF problem with missing
values:
2
P1 : minimize
W H − UV T F
(8)
U ,V
subject to
U ∈ U N ×K , V ∈ U N ×K
(9)
where denotes the Hadamard product, i.e., W H is an
N × N matrix computed entry-by-entry with W H ij =
Wij Hij .
Note that the above sparse matrix factorization problem
is formulated under a specific coordinate system C. If we
rotate the coordinate system θ degrees from a reference system
C0 , the resulting rotated coordinate system is denoted by Cθ .
Then the problem P1 is to find Uθ and Vθ from observation
matrices Hθ and Wθ in the coordinate system Cθ .
In general, P1 is difficult to solve because the objective
function is non-convex and the constraints are also nonconvex. However, as to be discussed in the remaining part
of this paper, there are some θ that make P1 easier to solve.
In the special cases of K = 1, 2, one can even find unique
solutions to a relaxed version of P1 for some coordinate
system Cθ .
III. S PECIAL C ASE I: S INGLE S OURCE
In the single source case, the eigenstructure of the signature
matrix H is invariant under any rotation of the coordinate
system Cθ . We show that the factorization problem P1 can
be easily solved by a matrix completion problem followed by
SVD. Therefore, the remaining challenge is to estimate the
peaks of the signature vectors and to study the localization
performance to justify the estimation strategy.
A. Solution via Matrix Completion
Without the unimodal constraints (9), P1 is a classical
rank-K matrix completion problem. It has been shown in
the sparse signal processing literature that, under some mild
2 For easy elaboration, assume that the M sensing locations distribute over
M distinct grids.
PREPRINT VERSION
4
regularization conditions on the low rank matrix H (e.g.,
strong incoherence property and small rank property of H,
see [18], [19]), the matrix H can be recovered, with a high
probability, from the sparse and noisy observation W H.
Specifically, the noisy recovery of H can be obtained as a
solution, X̂, to the following convex optimization problem
[16], [19]:
P2 :
minimize kXk∗
X
X
Xij − Hij
subject to
(10)
2
(i,j)∈Ω
≤ ǫ2
where kXk∗ denotes the nuclear norm of X (i.e., the sum
of the singular values of X), and ǫ2 is a small parameter
(depending on the observation noise [19]) for the tolerance of
the observation noise in H.
Note that under exact recovery X̂ = H in the K = 1 case,
the signature vectors can be extracted from the SVD of the
rank-1 matrix X̂ = α̂1 û1 v̂T1 . The unimodal constraints (9) are
then automatically satisfied. As a result, an efficient solution
to the sparse matrix factorization problem P1 can be obtained
as the dominant singular vectors of X̂ as the solution to P2.
B. Peak Localization Exploiting Symmetry
We first establish the property of the symmetric function
w(x) specified in Proposition 1. The autocorrelation of w(x)
is given by
Z ∞
τ (t) =
w(x)w(x − t)dx
(11)
−∞
which can be shown to be monotonic for t > 0.
Lemma 1 (Monotone property of τ (t)). The autocorrelation
function τ (t) is non-negative and symmetric. In addition, τ (t)
is strictly decreasing in t > 0.
Proof: The result can be easily derived using the unimodal and symmetric property of w(x). The details are
omitted here due to page limit.
From Lemma 1, τ (t) is maximized as t = 0. As a result, the
non-negative, unimodal, and symmetric function w(x − s1,1 )
from Proposition 1 has the following autocorrelation function
Z ∞
w(x − s1,1 )w(−x + t − s1,1 )dx
(12)
−∞
Z ∞
=
w(x − s1,1 )w(x − t + s1,1 )dx
(13)
Z−∞
∞
=
w(z)w(z + 2s1,1 − t)dz
(14)
−∞
= τ (t − 2s1,1 )
which is maximized at t = 2s1,1 , where the first equality (13)
is due to symmetry w(x) = w(−x), and the second equality
(14) is from the change of variable z = x − s1,1 .
Let v̂1 be the dominant right singular of X̂, the solution to
P2. From Proposition 1, v̂1 is a noisy discretization of w(x−
s1,1 ). As an estimate of the location s1,1 , the x coordinate of
the source, can be given by
1
(15)
ŝ1,1 (v̂1 ) = argmax R(t; v̂1 )
2 t∈R
where
R(t; v̂) =
Z
∞
v̂(x)v̂(−x + t)dx
(16)
−∞
is the reflected correlation function given v̂ ∈ RN , in which,
v̂(x) is a continuous (nonparametric) regression function based
on v̂. For example, v̂(x) can be obtained by v(x) = v̂i if
x = cX,i , i = 1, 2, . . . , N , and by linear interpolation between
v̂i and v̂i+1 if cX,i < x < cX,i+1 , where v̂i is the ith entry of
v̂.3
The estimator (15), as motivated by the form of the integral
(12), computes the correlation of the data from both sides of
the point the function v̂(x) is symmetric about. As a result,
the peak location estimator from v̂ uses all N entries for
estimation rather than merely a local subset of the entries
around the peak location.
The location estimate, ŝ1,2 , from û1 can be obtained in a
similar fashion.
C. Squared Error Bound
Let e(x) = v̂(x) − w(x − s1,1 ) be the error between
the non-parametric regression function v̂(x) obtained from
interpolating v̂1 and the function w(x− s1,1 ) from Proposition
1. Define κ = (σ1,1 − σ1,2 )/α1 as the normalized difference
between the first dominant singular value and the second
dominant singular value of the energy filed matrix H (1) in
(4) for a single source. We have the following theorems to
characterize the squared estimation error kŝ1 − s1 k22 of the
source location s1 .
We first consider the conservative
construction of the ob√
servation matrix, where N = M and all the entries of H
are observed, and therefore the signature vectors û1 and v̂1
are directly extracted as the dominant left and right singular
vectors of H.
Theorem 2 (Squared Error Bound under the Conservative
Construction). Suppose that there exists 0 < RCe < ∞,
∞
such that the interpolation error e(x) satisfies R −∞ w(x −
∞
s1,1 ) e(−x + t) − e(−x + t′ ) dx ≤ Ce −∞ w(x −
′
s1,1 )e(x)dx for all t, t ≥ 0 and N → ∞. Then, for M = N 2
under asymptotically large N ,
8L2 σn2
2Ce 4Kh2 L4
(17)
+
kŝ1 − s1 k22 ≤ ′′
2
3
|τ (0)|κ
N
N α21
with probability 1, where τ ′′ (0) is the second order derivative
of τ (t) evaluated at t = 0.
Proof: See Appendix C.
Remark 1. The terms Kh (Lipschitz continuity parameter)
and −τ ′′ (0) capture the sharpness of p
the energy field. For a
2
Gaussian source energy field, h(d) = 2γ/πe−γd , one can
′′
calculate that −τ (0) = γ.
We now evaluate
the case of the aggressive construction,
√
where N > M and the matrix H is partially observed. The
3 Other nonparametric smoothing methods can be applied, such as knearest neighbor (KNN) regression and kernel regression [13], [14]. However,
determining the best method to obtain v̂(x) is beyond the scope of this paper.
PREPRINT VERSION
5
signature vectors û1 and v̂1 are extracted from X̂, a solution
to the matrix completion problem P2 based on H.
Theorem 3 (Squared Error Bound under the Aggressive
Construction). Suppose that the measurement noise n(m) is
bounded by |n(m) | < σ̄n and the sampling error of H is
P
2
bounded as (i,j) Hij −Hij ≤ ǫ2 , where ǫ is the parameter
used in P2. Assume that the matrix dimension N = N (M ) is
chosen as the largest integer such that M ≥ βCN (log N )2 for
some constant C and β. Then, under the condition of Theorem
2, for asymptotically large N ,
512Ce L2 Kh2 L2
kŝ1 − s1 k22 ≤ ′′
+
|τ (0)|κ2 2N 2
√
2Kh L σ̄n
σ̄ 2
+ n2 (18)
N
α1 α1
with high probability.
Proof: See Appendix C.
We draw the following observations from the analytical
results in Theorems 2 and 3.
Efficiency of the Signature Matrix Model H: The parameter
0 < κ ≤ 1 captures how precisely the outer product u1 v1T of
the signature vectors may approximate the energy field matrix
H for a single source. In the ideal case where H is rank1, we have κ = 1 leading to a low MSE. For most practical
propagation models we have tested (e.g., propagations of radio
signals over the air, acoustic signals in the water, etc.), κ is
close to 1.
Performance Advantage of the Aggressive Construction
Strategy: It can be verified that in the high signal-to-noise ratio
(SNR) case α1 /σn2 , α1 /σ̄n2 ≫ 1, as the number of samples M
increases, the squared error bound (18), which exploits matrix
completion techniques under partial sampling, decreases at a
higher rate in terms of M than the result in (17) under full
sampling. Therefore, with a proper choice of the parameters
β and C, the aggressive construction strategy can achieve the
same localization performance using fewer measurements.
Sparsity and Noise Suppression Tradeoff: While the aggressive construction strategy achieves higher squared error
decay rate in terms of the number of samples M under high
SNR , α1 /σ̄n2 ≫ 1, it is less tolerant of measurement
noise as observed from the last terms in (17) and (18),
respectively. Specifically, in the low SNR case, the squared
error bound of the aggressive construction strategy scales as
1
1
), whereas, it scales as O( N SNR
) for the conservative
O( SNR
construction strategy.
Performance Scaling Law: As a performance benchmark,
for a naive scheme that estimates the source location directly
from the position of the measurement sample that observes
the highest power, the localization squared error decreases as
O(1/M
√ ), whereas, even for the conservative construction case
N = M of the proposed scheme, the squared error decreases
as O(1/M 3/2 ) in high SNR case α1 /σn2 ≫ 1, order-wise
faster than the naive scheme. These results then confirm that
by exploiting the unimodal and symmetry properties as well
as sparse signal processing techniques, the proposed algorithm
will significantly improve the localization resolution.
IV. S PECIAL C ASE II: T WO S OURCES
In the case of two sources, the SVD may not extract
the desired signature vectors from H in (4), because u1
and u2 are not necessarily orthogonal. However, it turns out
that by choosing an appropriate coordinate system C, the
UMF problem P1 can be trivially solved (under some mild
conditions). In this section, we propose rotation techniques to
select the optimal coordinate system for source separation and
localization.
A. Optimal Rotation of the Coordinate System
We fix the origin at the center of the target area and rotate
the coordinate system such that the two sources are aligned,
w.l.o.g., on the y axis. Thus, the source locations satisfy s1,1 =
s2,1 . Correspondingly, the signature vectors v1 = v2 , since
they are discretized from w(x − s1,1 ) according to Proposition
1. This
approximately yields a rank-1 model H = α1 u1 +
α2 u2 v1T by ignoring the minor components in (4).
As a result, an algorithm for the two source case can be
designed as follows. First, extract the vectors α̂1 û1 + α̂2 û2 and
v̂1 by solving the matrix completion problem P2 followed by
the SVD as developed in Section III. Second, obtain û1 and
û2 from the composite vector α̂1 û1 + α̂2 û2 . If the signature
vectors in (4) are unimodal, then the solutions obtained will
also satisfy the unimodal constraint (9).
The remaining challenge is to find the optimal rotation based
on the M measurement samples as the source topology is not
known.
Denote H(θ) as the observation matrix constructed in coordinate system Cθ with θ degrees of rotation to reference
coordinate system C. The desired rotation θ can be obtained
as
σ12 (H(θ))
(19)
ρ(θ)
,
P3 : maximize
P
N
2
θ∈[0, π
2]
k=1 σk (H(θ))
where σk (H) is defined as the kth largest singular value of
X̂(H), the solution to the matrix completion problem P2
based on H.
Note that ρ(θ) ≤ 1 for all θ ∈ [0, π2 ]. In addition, ρ(θ∗ ) = 1,
when H(θ) becomes a rank-1 matrix where the sources are
aligned with one of the axes.
The optimization problem P3 is, in general, non-convex.
An exhaustive search for the solution θ∗ is computationally
expensive, since for each θ, problem P2 might need be solved
followed by the SVD to obtain the singular value profile of
H(θ). However, it can be shown that ρ(θ) has a nice property
that enables efficient optimization.
Let H(θ) be the signature matrix defined in (4) under the
coordinate system Cθ . If H(θ) can be perfectly recovered from
H(θ), i.e., X̂(H(θ)) = H(θ), we can show a locally unimodal
property of ρ(θ).
Theorem 4 (Property of ρ(θ)). Assume that the two sources
have equal transmission power α1 = α2 . In addition, suppose
that λk (H(θ)) = λk (H(θ)) for k = 1, 2. Then, ρ(θ) is
periodic, i.e., ρ(θ) = ρ(θ + π2 ). In addition, ρ(θ) is strictly
increasing over (θ∗ − π4 , θ∗ ) and strictly decreasing over
PREPRINT VERSION
6
(θ∗ , θ∗ + π4 ), if the energy field satisfies
′
′
s · τ (t) > t · τ (s)
for all 0 < s < t, where τ ′ (t) ,
maximizer of ρ(θ).
d
dt τ (t)
(20)
and θ∗ is the
V. A RBITRARY N UMBER
Proof: See Appendix D.
The result in Theorem 4 confirms that the function ρ(θ) has
a unique local maximum within a π2 -window under a mild condition, in the ideal case of perfect recovery X̂(H(θ)) = H(θ).
The property motivates a simple bisection search algorithm to
efficiently search for the globally optimal solution, θ∗ , to (19).
Note that condition (20) can be satisfied by a variety of
energy fields. For example, for Laplacian field h(x, y) =
γe−γ|x|−γ|y|, we have the autocorrelation function τ (t) =
(1 + γt)e−γt , and its p
derivative τ ′ (t) = −γ 2 te−γt ; for
2
Gaussian field h(d) = 2γ/πe−γd , where d2 = x2 + y 2 ,
2
2
we have τ (t) = e−γt /2 , and τ ′ (t) = −γte−γt /2 . In both
cases, condition (20) is satisfied.
B. Source Separation
Suppose that in the coordinate system Cθ , the sources are
aligned with the y axis. Then, the dominant left and right
singular vectors of H(θ) are proportional to α1 u1 + α2 u2
and v1 , respectively. Denote the dominant left and right
singular vectors of X̂(H(θ)) (solution to P2) as û1 and v̂1 ,
respectively. Using a technique similar to that in Section III,
we can estimate the x coordinate from
1
(21)
ŝ1,1 = ŝ2,1 = argmax R(t; v̂1 ).
2 t∈R
To find the y coordinates, if the two sources have equal
transmission power α1 = α2 , the combined signature vector
α1 u1 + α2 u2 is symmetric about the mid-point c = (s1,2 +
s2,2 )/2, which can be estimated as
1
ĉ = argmax R(t; û1 ).
2 t∈R
(22)
As the two sources have equal distance r to the mid-point
(s1,1 , c), the variable r can be estimated as
r̂ = argmax Q(r; û1 , v̂1 )
(23)
r≥0
where
Q(r; û1 , v̂1 ) ,
1
2
Z
∞
−∞
that is observed would be unimodal with only one peak in
û1 . As a comparison, the proposed procedure (21) – (23) can
resolve such a limitation.
û(x) v̂(x − ĉ − r) + v̂(x − ĉ + r) dx
in which, û(x) is a non-parametric regression function from
the interpolation of û1 and v̂(x) is from interpolating v̂1 .
Using calculations similar to (12) – (14), it can be shown that
Q(r; u1 + u2 , v1 ) is maximized at r∗ = (s1,2 − s2,2 )/2 if the
interpolation is perfect, i.e., û(x) = w(x − s1,2 ) + w(x − s2,2 )
and v̂(x − s1,1 ) = w(x − s1,1 ).
As a result, the estimated source locations are given by
ŝ1 = (ŝ1,1 , ĉ + r̂) and ŝ2 = (ŝ2,1 , ĉ − r̂).
As a benchmark, consider a naive scheme that estimates s1,2
and s2,2 by finding the peaks of û1 . However, such a naive
strategy cannot work for close spread sources, because the
effective energy field function α1 h(d(z, s1 )) + α2 h(d(z, s2 ))
OF
S OURCES
In the case of an arbitrary number of sources, we first
study a general algorithm framework to solve P1. We then
discuss efficient approximations for fast implementation of the
algorithm. Finally, an optimization of the coordinate system Cθ
is studied to enhance the convergence of the algorithm.
A. The Gradient Projection
2
Let f (U , V ) = W H − U V T F . With some algebra
and matrix calculus, it can be shown that the gradients of f
are
∂
f = −2(W H)V + 2(W (U V T ))V
∂U
∂
f = −2(W T HT )U + 2(W T (V U T ))U
∂V
and the iteration of the projected gradient algorithm can be
computed as
n
o
∂
U (t + 1) = PU U (t) − µt
f (U (t), V (t))
(24)
∂U
o
n
∂
f (U (t + 1), V (t))
(25)
V (t + 1) = PU V (t) − νt
∂V
where PU {·} is a projection operator to project any N × K
matrix onto the unimodal cone U N ×K , and the step size µt and
νt are chosen to ensure the decrease of the objective function
f (for example, using a back-pressure rule).
B. Efficient Unimodal Projection
The projection PU {X} onto the unimodal cone is formally
defined as the solution that minimizes kX − Y kF over
Y ∈ U N ×K . Due to the property of the Frobenius norm,
the projection can be computed column-by-column.
While it is not straight-forward to efficiently project onto the
convex set UsN (specified by constraints (5) – (6)) as it may
seem to be, it is relatively easier to compute the projection
onto an isotonic cone, where an isotonic sequence is defined
as a non-increasing (or non-decreasing) sequence. Recently, a
fast algorithm for exact isotonic projection was developed in
[20], which finds the solution within N − 1 steps.
With such a tool, a fast approximate algorithm to compute
PU {X} can be described as follows.
Fast approximate unimodal projection:
1) For each s, compute the isotonic projection for
xk , the kth column of X, to form an ascend(1)
ing branch y1 , y2 , . . . , ys and a descending branch
(2)
ys , ys+1 , . . . , yN , respectively, using the exact isotonic
projection algorithm in [20].
(1) (2)
(s)
2) Let ys
=
max{ys , ys }. Then, x̃k
:=
(y1 , y2, . . . , ys , . . . , yN ) is an approximate unimodal
projection for xk with the kth entry taking the
maximum value.
PREPRINT VERSION
7
Table I
E VALUATION OF UNIMODAL PROJECTION ALGORITHMS , ǫ2 = 10−10
Vector Size
Brute-force
Proposed
E{e2 }
P e>ǫ
N = 10
2.28 sec
6.41 ms
1.5 × 10−3
0.078
N = 20
5.13 sec
20.5 ms
1.8 × 10−10
0.053
3) Repeat the previous steps to compute a series of projections for s = 1, 2, . . . , N .
(s)
(s)
4) Choose the solution x̃k that minimizes kx̃k − xk k
over s = 1, 2, . . . , N .
where Step 2) is an approximation because the way in which
ys is determined is not optimal.4 Since Step 1) has complexity
at most N , the overall complexity is at most N 2 .
In Table I, we compare the performance of the proposed
fast projection with a brute-force projection method over independent and identically distributed (i.i.d.) standard Gaussian
vectors x ∼ N (0, IN ×N ). In the brute-force method, the vector x is projected to the unimodal cone by solving N convex
problems: x̃(s) = arg min ky − xk2 subject to (5) – (6), for
s = 1, 2, . . . , N , and pick x̃∗ = x̃(s) to minimize kx̃(s) − xk2 .
The projection error is defined as e = kx̃− x̃∗ k2 /kx̃∗ k2 , where
x̃ is the solution of the proposed method. It is shown that the
proposed fast approximate projection is more than 200 times
faster, but has only marginal performance loss.
Proposition 2 (Partial convergence). Assume perfect sampling
H = H and H in (4) has rank K. Suppose that the algorithm
initialization X(0) is close enough to the optimal solution X̂
to P1. Then the following holds
T
d
(27)
E(Xe ) ≤ −2λK (V̂ V̂)kUe k2F + o kUe k2F
dt
for kVe kF = o kUe kF , where λK (A) denotes the smallest
eigenvalue of A. Moreover,
T
d
(28)
E(Xe ) ≤ −2λK (Û Û)kVe k2F + o kVe k2F
dt
for kUe kF = o kVe kF .
Proof: See Appendix E.
Proposition 2 shows that the rate of convergence depends
T
T
on the eigenvalues of V̂ V̂ and Û Û, where V̂ and Û carry the
location signatures of the source. Specifically, if the sources
are aligned with either the x axis or the y axis, then either
Û or V̂ tends to have identical columns, which leads to rank
T
T
deficiency of matrices Û Û or V̂ V̂, corresponding to small
eigenvalues λK and hence slow convergence.
Propositions 2 provides useful intuition for algorithm design; they show that gradient type algorithms work better when
sources are well-separated in both axes.
D. Rotation for Convergence Improvement
C. Local Convergence Analysis
Specifically, we frame the analysis according to the following two points. First, it observed that the globally optimal
solution X̂ = (Û, V̂) is in the interior of the unimodal cone
U N ×K × U N ×K , i.e., the unimodal constraints are not active
when the algorithm iterate approaches the neighborhood of X̂.
As a result, problem (8) becomes an unconstrained one in the
neighborhood of X̂.
Second, note that the function is bi-convex. Therefore, we
can study partial convergence, where the convergence of the
variable U is analyzed while fixing the other variable V to
be in the neighborhood of V̂. Thus, such analysis serves as a
qualitative prediction on the convergence performance.
Denote g(X) = [ ∂f /∂U )T ∂f /∂V )T ]T as the gradient
function of f (X), where X = (U , V ). Suppose X(0) is
sufficiently close to X̂, such that the unimodal constraints are
not active. As a continuous counter-part to the discrete iteration
(24) – (25), the continuous algorithm trajectory X(t) can be
given as
d
X(t) = −g(X(t)).
(26)
dt
Let E(Xe (t)) = 12 kXe (t)k2F be the normed error function
for the convergence error Xe (t) , X(t)− X̂. Let Ue = U − Û
and Ve = V − V̂ with the time index t dropped for notational
brevity. The following result suggests that if either Ue or Ve
is much smaller than the other variable, then the algorithm
trajectory X(t) converges exponentially to X̂.
4 In fact, optimizing the value at the sth turning point may involve
complexity scaling with N .
As it is preferred to analyze the case where the sources are
well-separated in both axes of the coordinate system, we may
need to establish a coordinate system with the optimal rotation
for the desired source topology. However, the challenge is that
we have no prior knowledge of the source locations.
Recall that H(θ) denotes the observation matrix constructed
in coordinate system Cθ with θ degrees of rotation with respect
to the reference coordinate system C. Similar to P3, the
desired rotation θ can be obtained as
P4 :
σ12 (H(θ))
ρ(θ)
,
minimize
P
N
2
θ∈[0, π
2]
k=1 σk (H(θ))
(29)
where σk (H) is defined as the kth largest singular value of
X̂(H), the solution to the matrix completion problem P2
based on H.
While problem P3 is to align the sources with one of the
axes, problem P4 tries to avoid alignments with any axes.
Fig. 2 demonstrates the performance of the UMF with
optimal coordinate system rotation under noise-free sampling
σn2 = 0,pwhere L = 1 and the energy field is given by
2γ/π exp(−γd2 ) with λ = 20. The observation
h(d) =
matrices constructed with dimension N satisfy N 2 /2 ≈ M .
There are two key observations: (i) the coordination system
rotation does improve the convergence as demonstrated by the
comparison between scheme “Rotated UMF”, which solves
P1 in the optimal coordinate system Cθ∗ with θ∗ solved
from P4, and scheme “Simple UMF”, which solves P1
in a fixed coordinate system C. (ii) UMF performs better
in the recovery of sparse unimodal structures as compared
to conventional sparse matrix completion methods, scheme
“Complete UMF”, which first solves the matrix completion
8
8
PREPRINT VERSION
8
Root
MSE
[km]
Root
MSE
[km]
MSE
RootRoot
MSE
modeled as (1 + d1.5 A(f )d )−1 , where Thorp’s formula [23]
−1
modeled
as (1
+
d1.5 A(f
)d )10
, where
formula
is used to
arrive
log
A(f
) =Thorp’s
0.11f 2/(1
+ f 2[23]
)+
Simple UMF
1.5 at 10
d −1
modeled
as
(1
+
d
A(f
)
)
,
where
Thorp’s
formula
[23]
2
2
2
2
−4
2
Complete UMF
is
used
to
arrive
at
10
log
A(f
)
=
0.11f
/(1
+
f
)
+
44f
/(4100
+
f
)
+
2.75
×
10
f
+
0.003
dB/km.
The
10
Simple
2
RotatedUMF
UMF
is used2 to arrive at2(m)
10 log10 A(f
) −4
= 0.11f
/(1 + f 2 ) +
2
Complete UMF
44f
/(4100
+
f
)
+
2.75
×
10
f
+
0.003
dB/km.
The
ambient
noise
n
is
modeled
as
a
zero
mean,
Gaussian
0.2
Rotated UMF
44f 2 /(4100 + f 2 ) +
2.75 × 10−4 f 2 + 0.003 dB/km.
The
2
ambient
noise(m)n(m)
isnormalized
modeled as
a zeroσmean,
Gaussian
random
variable
with
0.2
n /P = −34 dB,
ambient noise n
is modeled
as a variance
zero mean,
2 Gaussian
random
normalized
variance
−34 dB,
wherevariable
Pvariable
= 1 with
is with
thenormalized
total
transmission
The=sensor
has
2 σn /P
random
variance σpower.
n /P = −34 dB,
where
P
=
1
is
the
total
transmission
power.
The
sensor
has
a
receive
window
of
4
seconds
from
the
detection
of
the
first
where P = 1 is the total transmission power. The sensor has
apath.
receive
window
4 seconds
thethe
detection
thematrix
first
The
parameter
N
for constructing
observation
a receive
window
of 4ofseconds
fromfrom
the detection
of theoffirst
2matrix
path.
The
parameter
N
for
constructing
the
observation
H
is
chosen
as
the
largest
integer
satisfying
N
(log
N
)
≤
M.
path. The parameter N for constructing the observation matrix
0.1
2
H
is
chosen
as
the
largest
integer
satisfying
N
(log
)
≤
M
.
2 Ncase
The
proposed
algorithms
for
the
single
source
((15)
0.1
H is chosen as the largest integer satisfying N (log N ) ≤ M .
The
proposed
algorithms
for
the
single
source
case
((15)
in
Section
III),
two
source
((21)
–
(23)
in
Section
IV),
and
The proposed algorithms for the single source case ((15)
two
source
((21)
–(Section
(23)
in V)
Section
IV),
an Section
arbitrary
number
of((21)
sources
are compared
in in
Section
III),III),
two
source
– (23)
in Section
IV),
and and
an
arbitrary
number
of
sources
(Section
V)
are
compared
two number
baselineofschemes.
Baseline V)
1, are
Naive
scheme for
an with
arbitrary
sources (Section
compared
with
baseline
schemes.
1, Naive
scheme
with
twotwo
baseline
schemes.
Baseline
Naive
scheme
for for
single
source:
the
location
ofBaseline
the 1,sensor
that
observes
the
0
single
source:
the
location
of
the
sensor
that
observes
the
single
source:
the
location
of
the
sensor
that
observes
the
120
20
40
60
80
140
100
highest
energy
is
identified
as
the
source
location.
Baseline
0
120
20
40
60 Number80of sensors,
140
100M
highest
energy
is identified
as the
source
location.
Baselineis
highest
energy
is
identified
as the
source
location.
Baseline
2, Weighted
centroid
localization
[10]:
the location
estimate
Number of sensors, M
2,
Weighted
centroid
localization
the location
estimate
2, Weighted
centroid
localization
[10]:[10]:
the location
estimate
is is
updated
by
Figure 2. Localization performance under noise-free sampling in a two source updated
by by
updated
!
(m) (m)
Gaussian
field. performance
Figure
2. energy
Localization
performanceunder
undernoise-free
noise-free
sampling
a two
source
Figure 2.
Localization
sampling
in aintwo
source
̺(m)
z
P ! m∈R(ŝ(n))
(m) (m)
Gaussian
energy field.
field.
!
Gaussian energy
̺
z
ŝ(n + 1) =m∈R(ŝ(n))
̺
z (m)
m∈R(ŝ(n))
(m)
̺
ŝ(nŝ(n
+ 1)
+=1) =P ! m∈R(ŝ(n))
̺(m)̺(m)
m∈R(ŝ(n))
m∈R(ŝ(n))
1
(m)
(m)
until convergence, where
=
Weighted centroid
1
(m)̺(m)
(m) h(m) are the weights
until
convergence,
where
̺
=
h
are are
the weights
until
convergence,
where
̺
=
h
the are
weights
Naive
scheme
Weighted centroid
and R(x) specifies a set of measurements
that
taken
Kernelscheme
method
andand
R(x)
specifies
a
set
of
measurements
that
are taken
Naive
R(x)
specifies
a
set
of
measurements
that
are
taken
Proposed
within a radius r from location x, in which, the
raKernel
method
within
a radius
r from
location
x, in
which,
the raDecay rate
Proposed
within
ais radius
r L/8
from
location
x, as
in towhich,
the M/4
radius
r
at
least
or
as
large
include
Decay rate
dius
r is
at at
leastleast
L/8L/8
or as
largelarge
as toas include
M/4M/4
dius
r is
or
to include
measurements.
Baseline
3, as
Kernel
regression:
The
algomeasurements.
Baseline
3,
Kernel
regression:
The
algomeasurements.
Baseline
3, {α
Kernel
regression:
The
algorithm
chooses
parameters
,
ẑ
}
and
λ
to
minimize
k"
rithm
parameters
{αk , ẑkk} and
λ to minimize
" chooses
! chooses
!
Kparameters(m)
{α
λ totheminimize
P rithm
k ,2;ẑλ)
k}
"h(m)P−K!
"22 and
0.1
(m)
(m)
α
B(z
,
ẑ
based
measure"
"
!
k
k
k=1
m " (m)
− k=1K
αk B(z ,(m)
ẑk ; λ) based
on theonmeasure" based
m h
0.1
on the
measureh
−
α
B(z
,
ẑ
;
λ)
(m)
(m)
k
k
m (m)
k=1
(m)
ment
{h
,
z
},
where
two
classes
of
kernel
functions
ment {h (m)
, z (m)
}, where two classes of kernel functions are are
ment
{h , zGaussian
}, where
two B
classes
kernel
are−
considered,
kernel
ẑkof;=λ)
= functions
exp(−λ$z
G
considered, Gaussian kernel
BG (z,
ẑ(z,
exp(−λkz
−
k ; λ)
considered,
Gaussian
kernel
B
(z,
ẑ
;
λ)
=
exp(−λ$z
−−
2
) and
Laplacian
kernel
BGLẑ(z,
ẑkk ;=λ)exp(−λkz
= exp(−λ$z
ẑk kẑ2k)$2and
Laplacian
kernel
BL (z,
−
k ; λ)
ẑ
$
)
and
Laplacian
kernel
B
(z,
ẑ
;
λ)
=
exp(−λ$z
−
2
L norm.
k Cross-validation
in which,
$1 denotes
Cross-validation
ẑk kẑ21kk),$21in), which,
k · k$1 ·denotes
the Lthe
1 L1 norm.
ẑisk $used
· $1 best
denotes
thefunction,
L1 norm.
Cross-validation
1 ),
towhich,
choose
the
kernel
and
the set
data
is used
to in
choose
the$best
kernel
function,
and the
data
is set is
is
used
to
choose
the
best
kernel
function,
and
the
data
set
is
partitioned
to
70%
for
parameter
training
and
30%
for
MSE
partitioned
to
70%
for
parameter
training
and
30%
for
MSE
0.01
partitioned
to
70%
for
parameter
training
and
30%
for
MSE
10
100
1000
performance
validation.
The The
parameters
ẑk give
0.01
performance
validation.
parameters
ẑk the
givelocation
the location
10
100
1000
performance
validation.
The1parameters
ẑkcannot
give
the
location
Number of samples, M
estimates.
Note
that that
Baselines
and
2 cannot
differentiate
estimates.
Note
Baselines
1
and
2
differentiate
Number of samples, M
estimates.
Note
that
Baselines
1
and
2
cannot
differentiate
multiple
sources
when
theythey
appear
as one.
Therefore,
these these
multiple
sources
when
appear
as one.
Therefore,
Figure
3. MSE
MSE of
of the
thesource
sourcelocation
locationversus
versusthethe
number
sensors,
sources
when
they
appear
as
one.
Therefore,
these
Figure 3.
number
of of
sensors,
M ,M
in, in twomultiple
baselines
are
evaluated
in
the
single
source
case
only.
two
baselines
are
evaluated
in
the
single
source
case
only.
Figure
3. source
MSE of
the source location versus the number of sensors, M , in
the
case.
the single
single source
case.
two
baselines
are
evaluated
in
the
single
source
case
only.
Fig.Fig.
3 shows
the the
rooted
meanmean
squared
error error
(RMSE)
of
the single source case.
3 shows
rooted
squared
(RMSE)
of
Fig.
3
shows
the
rooted
mean
squared
error
(RMSE)
of
thethe
source
location
versus
the
number
of
sensors,
M
,
in
source location versus the number of sensors, M , in
source
location
the number
of outperforms
sensors, M , in
thethe
single
source
case. versus
The proposed
scheme
, and
then the single source case. The proposed scheme outperforms
problem
P2 in
in the
the optimal
optimalcoordinate
coordinatesystem
systemCθC∗ ,θ∗and
then
problem P2
6
naive
scheme
andcase.
the The
weighted
centroid
scheme.
It
single
source
proposed
scheme
outperforms
∗ , and then thethe
problem
P2for
the optimal
coordinate
system
Cθthe
solves P1
vector
extraction.
AsAsthe
UMF
is is the naive scheme and the weighted centroid scheme.66 It
solves
P1
forinsignature
signature
vector
extraction.
UMF
also
the kernel
method
from moderate
large It
theoutperforms
naive scheme
and the
weighted
centroid toscheme.
solves
P1
for
signature
vector
extraction.
As
the
UMF
is
the kernel method from moderate to large
aware of the
“Rotated
UMF”
aware
the unimodal
unimodalstructure,
structure,the
theproposed
proposed
“Rotated
UMF” also outperforms
of sensors. Inthe
thekernel
case ofmethod
more than
500moderate
sensors, the
also outperforms
from
to large
aware
of the the
unimodal structure,
the proposed
5 5 “Rotated UMF” number
outperforms
number
of
sensors.
In the case of more than 500 sensors, the
outperforms
the “Complete
“CompleteUMF”
UMF”scheme.
scheme.
performance
limited In
by the
the case
ambient
noise.than
The 500
dashed
line the
number ofissensors.
of more
sensors,
outperforms the “Complete UMF” scheme.5
performance
is limited
by
the ambient
noise.
The
dashed
line
fitsperformance
the performance
of theby
proposed
scheme
in the
Mdashed
= 10 line
is limited
the ambient
noise.
The
VI. N UMERICAL R ESULTS
fits
the
performance
of
the
proposed
scheme
in
the
M
=
10
to fits
500the
region
using logof
= a log
It gets
VI. N UMERICAL R ESULTS
10 RMSE
10 M +inb.the
performance
the proposed
scheme
M = 10
VI.
N
UMERICAL
R
ESULTS
to −0.88
500 region
using logMSE
= a as
logO(1/M
b. ).It gets
We consider the source and sensor deployment model in a =
10 RMSE
10 M +1.76
and hence
decreases
500 region
using the
log10
RMSE
= a log
M + b. It
gets
1.76
We consider
the
source
and
sensor
deployment
model in to
10 as
a
=
−0.88
and
hence
the
MSE
decreases
O(1/M
Section
II
with
L
=
5
km
in
an
underwater
environment.
Note
that
Theorem
2
and
3
show
that
the
MSE
of
the
proposed
We consider the source and sensor deployment model in a = −0.88 and hence the MSE decreases as O(1/M 1.76 ).).
Section
II with
L = 5 kmtransmit
in an underwater
environment.
1.5MSE of the proposed
Noteshould
that Theorem
and 3 showO(1/M
that the
The sources
simultaneously
signals at f environment.
= 5 kHz. scheme
decrease22faster
) (fullofsampling
Section
II with
L = 5 km in an underwater
that Theorem
and 3 than
show that the MSE
1.5 the proposed
The
sources
simultaneously
transmit
signals
at
f modeled
= 5 kHz. Note
scheme
should
decrease
faster
than
O(1/M
(full sampling
The
propagation
of
the
signal
from
each
source
is
case),
which
is
confirmed
by
our
numerical
results.
The sources simultaneously transmit signals at f = 5 kHz. scheme should decrease faster than O(1/M 1.5 ))(full
sampling
The
propagation
of the signal
from
each
source
is modeled Fig.
using
N
=
15
discrete
paths,
where
the
path
inter-arrival
case),
which
is
confirmed
by
our
numerical
results.
p
4
shows
the
RMSE
of
the
source
locations
versus
The propagation of the signal from each source is modeled case), which is confirmed by our numerical results. the
using
N
=
15
discrete
paths,
where
the
path
inter-arrival
p − τp are exponentially distributed with mean 100
times τN
Fig.
showsM
the
of the
locations
versus
p+1
of 4sensors,
, inRMSE
two source
andsource
three source
cases.
In the
using
p = 15 discrete paths, where the path inter-arrival number
Fig. 4 shows the RMSE of the source locations versus the
times
τp+1
−path
τp are
exponentially
distributed
with mean
100both
ms,
and
the
amplitudes
are
Rayleigh
distributed
with
number
of
sensors,
M
,
in
two
source
and
three
source
cases.
In
cases,
the
proposed
methods
significantly
outperform
the
times τp+1 − τp are
exponentially
distributed with mean 100 number of sensors, M , in two source and three source cases. In
p −τ1 )
ms,
andscaled
the path
amplitudes
are respect
Rayleigh
distributed
with kernel
power
as e−ϕ(τ
with
to the
first path,
methods.
Note
that
the
proposed
schemes
only
require
both cases, the proposed methods significantly outperform the
ms, and the−1path −ϕ(τ
amplitudes
are Rayleigh distributed with both
cases, the proposed methods significantly outperform the
p −τ1 )
power
e [22].
the first path,
ϕ = 2scaled
sec as[21],
Thewith
pathrespect
energytoattenuation
is
kernel methods. Note that the proposed schemes only require
power scaled−1as e−ϕ(τp−τ1 ) with respect to the first path, 6kernel
methods.
Note
that experiments
the proposed
schemes
require
from our
numerical
that the
weights ̺only
and the
ϕ 5 = 2 sec −1 [21], [22]. The path energy attenuation
is It is observed
ϕ However,
= 2 sec
[22].
we also [21],
note that
in theThe
case path
of low energy
SNR, i.e., attenuation
σn2 ≫ 1, P2 is window
6 R in the weighted centroid scheme should depend on M and h(d).
It is
observed
our
numerical
experiments
thatnon-trivial.
the weights ̺ and the
6 It
However,
thefrom
best our
weights
and window
are highly
can5 serve as an efficient de-noising step to help faster convergence 2of UMF.
ischoosing
observed
numerical
experiments
that
the weights
andh(d).
the
we also note that in the case of low SNR, i.e., σ2n ≫ 1, P2
window
R in thefrom
weighted
centroid scheme
should
depend
on M ̺and
5However,
P2
weefficient
also note
that in the
of low
SNR,
i.e., σn ≫of1,UMF.
window
in the weighted
shouldare
depend
M and h(d).
can However,
serve as an
de-noising
stepcase
to help
faster
convergence
However,Rchoosing
the bestcentroid
weights scheme
and window
highlyonnon-trivial.
can serve as an efficient de-noising step to help faster convergence of UMF.
However, choosing the best weights and window are highly non-trivial.
99
1
1
0.8
0.8
Root MSE [km]
Root MSE [km]
PREPRINT VERSION
0.6
0.6
0.4
0.4
0.2 KernelKernel
method,
method,
K = 3K = 3
Proposed,
Proposed,
K = 3K = 3
method,
KernelKernel
method,
K = 2K = 2
Proposed,
Proposed,
K = 2K = 2
0
0
10
100 100
10
Number of sensors, M
Number of sensors, M
9
0.2
10001000
Figure
5.
Localizing
the
three
peaks
from
the
elevation
data
North
Jutland,
Figure
5. Localizing
Localizing
thethree
three
peaks
from
the
elevation
of of
North
Jutland,
Figure
5.
the
peaks
from
thevisual
elevation
datadata
of North
Jutland,
Denmark.
The
result
matches
with
the
expectation.
Denmark.
The
result
matches
with
the
visual
expectation.
Denmark. The result matches with the visual expectation.
Figure 4.
4. MSE ofofthe
thesource
sourcelocation
locationversus
versusthethe
number
of
sensors,
, in
number
sensors,
M, ,M
FigureFigure
4. MSEMSE
of the source
location versus
the number
ofofsensors,
M
inin
two source
source and
and three
threesource
sourcecases.
cases.
two source and three source cases.
coordinate
system.
Our
numerical
experiments
demonstrate
coordinate
system.
Our
numerical
experiments
demonstrate
coordinate
system.
Our
numerical
experiments
demonstrate
that
the
proposed
scheme
achieves
similar
performance
the
that the
similar
performance
as the
that
the proposed
proposedscheme
schemeachieves
achieves
similar
performance
asasthe
baselines
using
no more
measurement
samples.
h(d)
is nonusing
no more
thanthan
1/5 1/5
measurement
samples.
the generic
generic property
propertythat
thatthe
thesource
sourceenergy
energyfield
field
h(d)
is non- baselines
the generic property that the source energy field h(d) is non- baselines using no more than 1/5 measurement samples.
negative and
d to
thethe
source.
negative
andstrictly
strictlydecreasing
decreasingininthe
thedistance
distance
d to
source.
negative
and strictly
decreasing inalso
the distancesuch
d toproperty,
the source.
A PPENDIX
A A
Although
the
thethe
Although
the kernel
kernelfunctions
functions alsocapture
capture such
property,
A PPENDIX
Although
the
kernel
functions
also
capture
such
property,
the
AOF
PPENDIX
A 1
P
ROOF
T
HEOREM
kernel
methods
suffer
from
parameter
estimation
error
when
kernel methods suffer from parameter estimation error when
P ROOF OF T HEOREM
1
kernel
methods
suffer
from
parameter
estimation
error
when
P
ROOF
OF
T
HEOREM
1
fitting the
As As
the the
signature
vectors
u
and
v
correspond
to thetodomfitting
the data
data toto inaccurate
inaccurateparametric
parametricmodels.
models.
k
k
signature vectors uk and vk(k)correspond
the domfitting the
data to inaccurate
parametric of
models.
Fig.
method
As singular
the signature
vectors
ukvk,1
andofvkHcorrespond
to theon
domvectors
u
and
,
we
(k)can focus
Fig. 55 demonstrates
demonstratesthe
therobustness
robustness ofthetheproposed
proposed
methodinant
k,1
inant
singular
vectors
u
and
v
of
H
,
we
can
focus on
k,1
k,1
(k)
Fig.
5localizing
demonstrates
the
robustness
of theThe
proposed
method
in
several
peaks
on
real
data.
dataset
contains
only
one
source
and
drop
the
source
index
k
here
for
brevity.
singular vectors uk,1 and vk,1 of H , P
we can focus on
in localizing several peaks on real data. The dataset containsinant
only one source and drop
the source index
kN
here
for brevity.
T
T
in localizing
several
peaks
on real
data. The
contains
altitude data
(colored
pixels)
measured
on the
networks only
!
fromand
(4), drop
we write
H = αuv
+k There
λiNubrevity.
one
source
the
source
index
for
i vi
altitude
data
(colored
pixels)
measured
ondataset
theroad
road
networksSpecifically,
i=2
!
Specifically,
from
(4),
we
write
H
=
αuv
+
λi uiTviT
N
i=2u
in North
Jutland, Denmark
[24], where
red
pixels
represent Specifically,
T
altitude
data
(colored
pixels)
measured
on
the
road
networks
(for
the
kth
source,
where
k
=
1),
in
which
u
and
fromsource,
(4), wewhere
write kH== 1),
αuvin +
ui i viu
in North Jutland, Denmark [24], where red pixels represent (for the kth
whichi=2uλiand
i
high altitude locations.
an altitude
is selected are the left singular
in North
DenmarkSuch
[24],
reddataset
pixels
vectors
ofk H,
v1),andinviwhich
are theu right
source,
where
=
and
ui
high Jutland,
altitude locations.
Such where
an altitude
datasetrepresent
is selected(forarethethekth
left
singular
vectors
of
H,
v
and
v
are
the
right
toaltitude
emulatelocations.
noise-corrupted
measurements
of an energy
field singular vectors, and α is the largest singular value,i α > λi ,
high to
Such an
altitude dataset
selected
singularand
vectors
of largest
H, v singular
and vi are
theα right
emulate noise-corrupted
measurements
of anis energy
fieldare the leftvectors,
α is the
value,
> λi ,
generated
by
K
=
3
hidden
sources.
We
downsample
the i =singular
2,
3,
.
.
.
,
N
.
to emulate
noise-corrupted
measurements
of
an
energy
field
singular
vectors,
and
α
is
the
largest
singular
value,
α
>
λi ,
generated
by
K
=
3
hidden
sources.
We
downsample
the
T
i
=
2,
3,
.
.
.
,
N
.
datasetby
over
equallysources.
spaced grid
points
and form
R = H H. Then, the (i, j)th entry of R is given
generated
K 40
=××340
hidden
We
downsample
thea ai =Let
T
dataset
over
40
40
equally
spaced
grid
points
and
form
2,
3,
.
.
.
,
N
.
T
R = H
H. Then,
j)th
entry of
of H.
R isIn given
sparse elevation matrix. We find the peaks (black triangles) by RLet
, where
hi isthe
the(i,ith
column
ij = hi hT
dataset
overelevation
40 × 40matrix.
equally spaced
gridpeaks
points(black
and form
a
sparse
the
triangles)
Rij ==HhjTH.
the
(i,
entry
of R of
is given
by solving
P1 to extractWe
the find
signature
vectors followed
by theLet
by
R
hj , Then,
where
hithat
is j)th
the columns
ith column
following
lemma,
we
show
the
of
R
areH. In
i
T
sparse
elevation
matrix.
We find
the
peaks vectors
(black triangles)
by
solving
P1
to
extract
the
signature
followed
by
h
,
where
h
is
the
ith
column
of
by
R
=
h
peak localization using (15). The result matches with a visual unimodal.
i jlemma, we ishow that the columns ofH.
theijfollowing
R In
are
by solving
P1
to
extract
the
signature
vectors
followed
by
7visual
peak
localization
using
(15).
The
result
matches
with
a
the
following
lemma,
we
show
that
the
columns
of
R
are
determination of possible locations of the hidden sources.
unimodal.
peakdetermination
localization using
(15). The
result of
matches
with sources.
a visual7 unimodal.
Lemma 2. Suppose that the source locates at the (m, n)th
of possible
locations
the hidden
thatforthe
source
locates
(m, n)th
gridLemma
centered2.at Suppose
cm,n . Then,
each
column
of R, at
thethe
entries
determination of possible locations of the hidden sources.7
VII. C ONCLUSIONS
Lemma
2. Suppose
that
the
source
atofthe
(m,
n)th
Rijgrid
arecentered
increasing,
Ri+1,j
if ilocates
<
n, and
theythe
areentries
at cR
.<
Then,
for, each
column
R,
ij
m,n
VII.
C ONCLUSIONS
This paper developed
nonparametric
algorithms for localiz- grid
at>cm,n
. Then,
for
R,and
the they
entries
decreasing,
Ri+1,j
, ij
if i<
≥R
n.each
Rcentered
increasing,
R
, if i <ofn,
are
ij
ij are R
i+1,jcolumn
VII.based
C ONCLUSIONS
ingThis
multiple
on a moderate
number offorenergy
are increasing,
R
<, R
, if i < n, and they are
papersources
developed
nonparametric
algorithms
localiz-Rijdecreasing,
R
>
R
if
i
≥
n.
ij
i+1,j
ij
i+1,j
Proof: Since the source location s is inside the (m, n)th
measurements
from different
locations.
A matrix
observation
This
paper
developed
nonparametric
algorithms
for
localizRij >
Ri+1,j
, have
if i ≥d(c
n. , s) > d(c
ing
multiple
sources
based on
a moderate
number
of energydecreasing,
grid
centered
cm,n
, wesource
, s)(m,
≥ n)th
p,i
p,i+1the
Proof:atSince
the
location
s is inside
model was
proposed
andon
proven
to have number
unimodal
syming multiple
sources
based
a moderate
energy
measurements
from
different
locations.
A matrixofand
observation
d(c
,
s),
for
i
<
n
and
all
p
=
1,
2,
.
.
.
,
N
.
Similarly,
p,n
Proof:
Sinceatthe
inside
thep,i+1
(m,,n)th
grid
centered
c source
, we location
have d(csp,iis, s)
> d(c
s) ≥
metric properties
for every
under
mildobservation
conditions.
measurements
different
locations.
matrix
model
wasfrom
proposed
and energy
provenfield
to A
have
unimodal
and sym- d(cp,i , s) < d(cp,i+1 ,m,n
i ≥
np and
all2,
Recall
that
<s),
n for
and
alld(c
=, s)
1,
.,N
. Similarly,
gridd(c
centered
at cim,n
, we
have
>p.. .d(c
, s) ≥
A
nonparametric
source
localization
algorithm
exploiting
the
p,n , s), for
p,i
p,i+1
metric
energy
fieldunimodal
under mild
conditions.
model
was properties
proposed for
andevery
proven
to have
and
sym- Hij = h(d(ci,j , s)) and h(d) is a non-negative decreasing
d(c
, s)for
< id(c
, s), all
for pi ≥
Recall that
, s),
<p,i+1
n and
= n1, and
2, . . all
. , Np.. Similarly,
unimodal and symmetric property was developed and shown d(c
p,n p,i
A properties
nonparametric
sourceenergy
localization
algorithm
exploiting
the function.
metric
for every
field under
mild conditions.
We
thus have
1.5
to decrease the localization MSE faster than O(1/M ) using d(c
H
=
h(d(c
,
s))
and
h(d)
is
a
non-negative
decreasing
,
s)
<
d(c
,
s),
for
i
≥
n
and
all
p.
Recall
that
i,j
p,iij
p,i+1
unimodal andsource
symmetric
propertyalgorithm
was developed
and shown
A nonparametric
localization
exploiting
the
N
X
M sensors in the single source case. In the two source
case,
function.
We
thus
have
h(d(c , s)) and h(d) is a non-negative decreasing
1.5
to decrease
the localization
MSEwas
faster
than O(1/M shown
) usingHij =
unimodal
and symmetric
property
Rij = hTi i,j
hj =
Hpi Hpj
the source
locations can
be found
bydeveloped
choosing and
the
optimal function.
N
We thus
have "
1.5
M
sensors
in
the
single
source
case.
In
the
two
source
case,
p=1
to decrease
the
localization
MSE
faster
than
O(1/M
)
using
T
rotation of the coordinate system such that the normalized
h
=
R
=
h
N
j
ij
i
N
the
source
locations
canofbe
found
choosing
the case,
optimal
" Hpi Hpj
M sensors
in singular
the
single
source
case.
Inby
the
two issource
X
dominant
value
the
sample
matrix
maximized.
T
Rij = hi hj<= p=1
pi H
HH
Hpjpj = hTi+1 hj = Ri+1,j
rotation
of of
thearbitrary
coordinate
system
such
that
normalized
p,i+1
the source
locations
can benumber
found
by sources,
choosing
the
optimal
In the case
of
wethe
localize
the
N
"
p=1
p=1
dominant
value
of problem
the such
sample
matrix
is maximized.
rotation
of the
coordinate
system
the normalized
sources
bysingular
solving
a UMF
inthat
an
optimally
rotated
Hp,i+1 Hpj = hTi+1 hj =TRi+1,j
<N
In
the
case
of
arbitrary
number
of
sources,
we
localize
the
dominant singular value of the sample matrix is maximized. for i < n. Similarly,"
we
can
show that TRij = hi hj >
p=1
7 In this experiment, we failed to successfully implement the kernel re< i≥
H
sources
by arbitrary
solving anumber
UMF problem
in anweoptimally
rotated
hTi+1 hj = Ri+1,j , for
n.p,i+1 Hpj = hi+1 hj = Ri+1,j
In the
case
of
of
sources,
localize
the
gression method to localize the peaks. There are too many local optima of
p=1
for i the
< condition
n. Similarly,
we can
thatR R
hTi hj >
Under
of Lemma
2, ifshow
we raise
toijthe=power
sources
by solving
a UMF
problem
in an the
optimally
rotated
the7 nonlinear
regression
problem,
and furthermore,
local optima
are not
T
q
In
this
experiment,
we
failed
to
successfully
implement
the
kernel
reRi+1,j
forwe
iare
≥can
n.
of h
q, the
of, R
unimodal,
with R
their =
peaks
necessarily consistent with the peak locations.
<hjcolumns
n.= Similarly,
show that
hTi hatj >
ij
gression method to localize the peaks. There are too many local optima offor ii+1
7 In this experiment, we failed to successfully implement the kernel reUnder
the
condition
of
Lemma
2,
if
we
raise
R
to
the power
T
the nonlinear regression problem, and furthermore, the local optima are nothi+1 hj = Ri+1,j , for i ≥ n.
q
gression
method consistent
to localizewith
the the
peaks.
are too many local optima of
of
q,
the
columns
of
R
are
unimodal,
with
their
necessarily
peakThere
locations.
Under the condition of Lemma 2, if we raise R to thepeaks
powerat
the nonlinear regression problem, and furthermore, the local optima are not
of q, the columns of Rq are unimodal, with their peaks at
necessarily consistent with the peak locations.
PREPRINT VERSION
10
the nth entry. Specifically, define R(q) , Rq /tr Rq . We
show, in the following lemma, that the columns of R(q) are
unimodal.
(q)
(q)
Lemma 3. Let Rij be the (i, j)th entry of R(q) . Then, Rij <
(q)
(q)
(q)
Ri+1,j , if i < n, and Rij > Ri+1,j , if i ≥ n.
Proof: First, it can be easily verified that the result
holds for q = 1 according to Lemma 2. Then, suppose
(q+1)
that
=
the result holds
for some q ≥ 1. Note that R
tr R(q) R(q) R/tr R(q+1) and that R(q) is symmetric. We
have
N
tr Rq X (q)
(q+1)
= q+1
Rij
Rpi Rpj
tr R
p=1
N
tr Rq X (q)
(q+1)
<
Rpj = Ri+1,j
R
tr Rq+1 p=1 p,i+1
(q+1)
(q+1)
> Ri+1,j , for
for i < n. Similarly, we can show that Rij
i ≥ n. Therefore, by deduction, the result holds for all q ≥ 1.
Let R∞ , limq→∞ Rq /tr Rq . It can be shown that the
limit exists and equals R∞
= vv T . To see this, we can easily
PN
T
q
2q
T
compute
+ i=2 λ2q
i vi vi and the normalization
Rq = α 2qvv P
N
2q
term tr R = (α + i=2 λi ), and note that (λi /α)2q → 0
as q → ∞, for i = 2, 3, . . . , N , leading to R∞ being rank-1.
On the other hand, from Lemma 3, each column of R∞ is
unimodal. Note that the ith column of R∞ can be written as
vi v, where vi is the ith entry of v. We therefore confirm that
v is unimodal, with its nth entry being the peak.
Similarly, by constructing Q = HH T, we can also show
that u is unimodal with its mth entry being the peak.
P ROOF
A PPENDIX B
OF P ROPOSITION 1
Consider a slight shift of the grids c̊i,j = ci,j + [δ1 , δ2 ],
and denote c̊X = cX + δ1 and c̊X = cY + δ2 . As a result, the
source locates exactly at the center of the (m, n)th grid, i.e.,
sk = c̊m,n .
Let H̊ (k) be the signature matrix measured at the shifted
grids {c̊i,j }. Then, we have
Z 1
L
(k)
(k)
h(dij ) − h(dij ) +
h′ (dij + tδ̊ij )dt
Hij − H̊ij =
N
0
Kh L
≤
δ̊ij
N
Kh L2
≤√
2N 2
where dij = kci,j − sk k, δ̊ij = kc̊i,j − sk k − dij , and h′ (d) =
limt↓0 1t [h(d + t) − h(d)] denotes the right derivative of h(d).
Consider the largest sub-block [H̊ (k) ]J of H̊ (k) that includes rows m − J to m + J, and columns n − J to n + J for
the largest possible integer J. The sub-matrix [H̊ (k) ]J then
has dimension (2J + 1) × (2J + 1). In addition, the submatrix is symmetric about the (J + 1, J + 1)th entry, where
the jthe row and (2J + 2 − j)th row are identical, and the
jth column and the (2J + 2 − j)th column are identical.
This is because the girds c̊m−i,n−j , c̊m+i,n−j , c̊m−i,n+j , and
c̊m+i,n+j have equal distance to the source sk = c̊m,n , and
(k)
thus the corresponding entries H̊pq are identical.
As a result, the rank-1 sub-matrix [H̊ (k) ]Jm ×Jn has symmetric singular vectors that satisfy [ůk ]Jm = [v̊k ]Jn , and
the singular vectors are symmetric about the (J + 1)th entry,
where Jm and Jn are some index sets that contain 2J + 1
elements, respectively. Note that vectors [ůk ]Jm and [v̊k ]Jn
are sub-blocks of the singular vectors ůk and v̊k of the rank1 full matrix H̊ (k) , respectively. We thus know that ůk and
v̊k are symmetric vectors about the mth entry and nth entry,
respectively. Moreover, ůk and v̊k are unimodal by Theorem
1.
Therefore, w.l.o.g., assume m < n ≤ N/2. We can
theoretically construct
a continuous function w(x) that takes
p
N/Lv̊k,j
value w(x) =
p for x = c̊X,j − sk,1 and j =
1, 2, . . . , n − 1, w(x) = N/Lůk,i for x = c̊Y,i − sk,2 and
i = m, m + 1, . . . , N , and by smooth interpolation elsewhere
such that w(x) is non-negative, unimodal, and symmetric
about x = 0.
A PPENDIX C
P ROOF OF T HEOREM 2
We first compute the peak localization error bound given
signature vector perturbations.
Let v̂1 be the dominant right singular vectors of H, the
observation
matrix in the case of conservative construction,
√
√
N = M . In the case of aggressive construction, N > M ,
let v̂1 be the dominant right singular vector of X̂, the solution
to P2. Denote e1 = v̂1 − v1 .
Recall that the signature vector v1 can be approximately
expressed as a discretization of w(x − s1,1 ) from Proposition
1, and in addition, e(x) = v̂(x) − w(x − s1,1 ) is the error of
the continuous non-parametric regression v̂(x) obtained from
interpolating v̂1 . We denote w1 (x) = w(x − s1,1 ) for the ease
of elaboration. From (16), it follows that
R(t; v̂1 )
Z ∞
w1 (x) + e(x) w1 (−x + t) + e(−x + t) dx
=
Z−∞
Z ∞
∞
=
w1 (x)w1 (−x + t)dx +
w1 (x)e(−x + t)dx
−∞
−∞
Z ∞
Z ∞
e(x)w1 (−x + t)dx +
e(x)e(−x + t)dx
−∞
−∞
Z ∞
≈ τ (t − 2s1,1 ) +
w1 (x)e(−x + t)dx
−∞
Z −∞
e(−y + t)w1 (y)(−dy)
(30)
+
+∞
Z ∞
w1 (x)e(−x + t)dx
(31)
= τ (t − 2s1,1 ) + 2
−∞
where theR approximation (30) drops the second or∞
der
R ∞ term −∞ e(x)e(−x + t)dx and uses the fact that
−∞ w1 (x)w1 (−x + t)dx = τ (t − 2s1,1 ) from the integral
(12).
PREPRINT VERSION
11
As t = 2ŝ1,1 maximizes R(t; v̂1 ) in (31), we have
Z ∞
τ (2ŝ1,1 − 2s1,1 ) + 2
w1 (x)e(−x + 2ŝ1,1 )dx
−∞
Z ∞
≥ τ (2s1,1 − 2s1,1 ) + 2
w1 (x)e(−x + 2s1,1 )dx
−∞
Z ∞
= τ (0) + 2
w1 (x)e(−x + 2s1,1 )dx
−∞
where τ (0) =
R∞
w (x)2 dx ≈
−∞ 1
P
2
j v1,j = 1. As a result,
τ (0) − τ (2ŝ1,1 − 2s1,1 )
Z ∞
h
i
w1 (x) e(−x + 2ŝ1,1 ) − e(−x + 2s1,1 ) dx
≤2
−∞
Z ∞
≤ 2Ce
w1 (x)e(x)dx
−∞
X
v1,j e1,j
≈ 2Ce
=
j
2Ce v1T e1
.
In addition, using the Taylor’s expansion of τ (t) at t = 0,
we have
τ (2ŝ1,1 − 2s1,1 )
1
2
≈ τ (0) + τ (0)′ × 2 ŝ1,1 − s1,1 + τ (0)′′ × 4 ŝ1,1 − s1,1
2
≥ τ (0) − 2Ce v1T e1
where, it can be verified that the first order derivative τ (0)′ = 0
and the second order derivative τ ′′ (0) < 0 by the definition
of τ (t) in (11). As a result,
ŝ1,1 − s1,1
2
≤
Ce
v T e1
−τ ′′ (0) 1
(32)
for asymptotically large N (since all the approximations
become accurate for large N ).
2
Similar result can be derived for ŝ1,2 − s1,2 by analyzing
the perturbation of u1 , and the result is statistically identical
to (32). Therefore, the squared error bound is given by (as
τ ′′ (0) < 0)
2Ce
v T e1 .
kŝ1 − s1 k22 ≤ ′′
|τ (0)| 1
We now derive the signature perturbation v1T e1 by discussing two cases.
A. The Case of Conservative Construction N =
√
M
Consider that the sensing location z (m) is inside the grid
centered at ci,j . Then, according to the matrix observation
model (1), (2), and (7), we have
2
2
L2
Hij − Hij = 2 h(m) − α1 h(dij )
N
Z
L2 1 ′
n(m) 2
= α21 2
h (dij + tδ (m) )dt +
N
α1
0
n(m)
n(m) 2
α21 L2 2
Ξ + 2Ξ
+
=
(33)
N2
α1
α1
where dij = kci,j − s1 k2 , δ (m) = kz (m) − s1 k2 − dij , h′ (d) =
limt↓0 1t [h(t + d) − h(d)] is the right derivative of h(d), and
R1
Ξ , 0 h′ (dij + tδ (m) )dt.
Note that by the √
Lipschitz continuity of h(d), we have Ξ ≤
Kh δ (m) ≤ Kh L/( 2N ). Therefore,
n
2 o α21 Kh2 L4
α2 L2 σ 2
E Hij − Hij
≤
+ 1 2 n2 , ω 2
(34)
4
2N
N α1
Using (34), we now derive an upper bound of v1T e1 .
Note that v1T e1 < 0 because 1 = kv̂1 k22 = kv1 + e1 k22 =
1 + 2v1T e1 + keT1 e1 k22 , and hence, 2v1T e1 = −keT1 e1 k22 < 0.
Then, the singular vector perturbation can be obtained as
q
2
sin ∠(v1 , v̂1 ) = 1 − v1T (v1 + e1 )
q
= −2v1T e1 + (v1T e1 )2
q
≥ 2 v1T e1
(35)
where | · | denotes the absolute value operator.
Let E = H − H. Using the singular vector perturbation
results in [25], we know that
2σ(E)
σ1 − σ2
(36)
σ(E)2 < σ(Ē)2 ≈ 4ω 2 N
(37)
sin ∠(v1 , v̂1 ) ≤
where σ(E) is the spectral norm of E, and σ1 and σ2 are the
first and second dominant singular values of H. Note that we
have σ1 − σ2 = κα1 .
We now compute σ(E). From the uniform random sampling
strategy for H, the elements of E are independent random
variables with zero mean and variance strictly smaller than ω 2
in (34). Let Ē be an N × N matrix whose elements are i.i.d.
random variables with zero mean, variance ω 2 and bounded
fourth order moment. Then we know that the spectral norm
σ(E) < σ(Ē) almost surely for asymptotically large N . On the
other hand, using random matrix theory, it has been shown
in [26, Theorem 2.1] that ω√1 N σ(Ē) → 2 almost surely, as
N → ∞. As a result, we have
almost surely.
Combining the results in (35) – (37), we have
1 2σ(E) 2
8L2 σn2
4K 2 L4
v1T e1 ≤
≈ 2h 3 + 2
2 κα1
κ N
κ N α21
(38)
for asymptotically large N , almost surely.
B. The Case of Aggressive Construction N >
√
M
We first characterize the completed matrix X̂ as the solution
to P2.
Lemma 4 (Matrix completion with noise [19]). Suppose that
the parameter ǫ in P2 is chosen to satisfy ǫ ≥ kPΩ (H−H)kF.
In addition, assume the N is the largest integer chosen such
that M ≥ βCN (log N )2 for some constant C. Then, with
high probability,
s
(2 + p)N
ǫ + 2ǫ
(39)
η , kX̂ − HkF ≤ 4
p
PREPRINT VERSION
12
where p = M/N 2 .
In the case of α1 = α2 , it can be shown that the SVD of
H is given by
Using (33), we have
Hij − Hij
2
α2 L2 Kh2 L2
≤ 12
+
N
2N 2
Using the fact that
kPΩ (H −
H)k2F
=
X
(i,j)∈Ω
√
2Kh L σ̄n
σ̄ 2
+ n2 , ε̄2 .
N
α1
α1
2
|Hij − Hij | ≤ M ε̄
2
√
the choice of parameter ǫ = M ε̄, and M ≈ βCN (log N )2
for asymptotically large N , the bound (39) can be simplified
as
s
√
(2 + βCN (log N )2 /N 2 )N √
η≤4
M
ε̄
+
2
M ε̄
CN (log N )2 /N 2
r
32 N √
≈
M ε̄
βC log N
, η̄.
(40)
Let Ê be an N × N matrix, where the elements are i.i.d.
random variables with zero mean and variance ω̃ 2 = η̄ 2 /N 2 ,
and
bounded fourth order moment. As a result, we have
E kÊk2F ≥ kX̂−Hk2F with probability 1. Therefore, using the
results in [26, Theorem 2.1], the spectral norm of Ê converges
to ω̃√1 N σ(Ê) → 2 almost surely as N → ∞. As a result,
σ(X̂ − H)2 < σ(Ê)2 ≈ 4ω̃ 2 N
H = σ1 p1 q1T + σ2 p2 q2T
(41)
where σ1 (θ) = 12 α1 ku1 + u2 k2 kv1 + v2 k2 and σ2 (θ) =
1
2 α1 ku1 − u2 k2 kv1 − v2 k2 are the singular values, and
u1 + u2
,
ku1 + u2 k2
u1 − u2
p2 =
,
ku1 − u2 k2
p1 =
v 1 + v2
kv1 + v2 k2
v 1 − v2
q2 =
kv1 − v2 k2
q1 =
are the corresponding singular vectors. Here, all the components are functions of θ.
As a result,
σ2 (θ)2
σ1 (θ)2
h(w − wc )2 ih(w − ws )2 i
≈
h(w + wc )2 ih(w + ws )2 i
1 − hw · wc i 1 − hw · ws i
=
1 + hw · wc i 1 + hw · ws i
µ(θ) ,
(42)
where we have used the factthat h(w−wc )2 i = hw2 i+hwc2 i−
2hw · wc i = 2 1 − hw · wc i .
In addition, from properties of calculus, if f (x, θ) and
∂
∂θ f (x, θ) are continuous in θ, then
Z ∞
d
d
hf i =
f (x, θ)dx
dθ
dθ
Z ∞ −∞
D∂ E
∂
f (x, θ)dx =
f .
=
∂θ
−∞ ∂θ
almost surely.
Using the singular vector perturbation results in [25] and
Therefore, defining
following similar calculations in (35) – (38), we have
d
√
wc′ (x, θ) ,
w(x) x=x−D cos θ
2Kh L σ̄n σ̄n2
256L2 Kh2 L2
1 2σ(Ê) 2
dx
T
.
+
+
≈
v1 e 1 ≤
d
2 κα1
κ2
2N 2
N
α1 α21
ws′ (x, θ) ,
w(x) x=x−D sin θ
dx
we have
A PPENDIX D
P ROOF OF T HEOREM 4
d
∂
hw · wc i = hw ·
wc (x, θ)i = hw · wc′ iD sin θ
dθ
∂θ
W.l.o.g., assume that the two sources locate at s1 = (0, 0)
∂
d
and s2 = (D cos θ, D sin θ). Define wc (x, θ) = w(x−D cos θ)
hw · ws i = hw ·
ws (x, θ)i = −hw · ws′ iD cos θ.
dθ
∂θ
and ws (x, θ) = w(x − D sin θ), where w(x) is the unimodal
With some algebra, the derivative of µ(θ) can be obtained
and symmetric function
R ∞ defined in Proposition 1. In addition,
as
using (3), we have −∞ w(x)2 dx ≈ 1.
h
d
Using Proposition 1, we have the following approximation
µ(θ) = η D cos θhw · ws′ i 1 − hw · wc i2
Z ∞
dθ
2
i
w(x) + wc (x, θ) dx
ku1 + u2 k22 ≈
− D sin θhw · wc′ i 1 − hw · ws i2
−∞
h
i
, h(w + wc )2 i
= η − t · τ ′ (s) 1 − τ (t)2 + s · τ ′ (t) 1 − τ (s)2
−2
−2
where we have defined an integration operator h·i as
where η = 2 1 + hw · wc i
1 + hw · ws i
, t = D cos θ,
Z ∞
and s = D sin θ.
hf i ,
f (x, θ)dx
Note that 0 < s < t for 0 < θ < π . Applying condition
−∞
for a function f (x, θ). Note that the operator h·i is linear and
satisfies the additive property, i.e., haf i = ahf i and hf + gi =
hf i + hgi, for a constant a and a function g(x, θ).
Similarly, kv1 + v2 k22 ≈ h(w + ws )2 i, ku1 − u2 k22 ≈ h(w −
wc )2 i, and kv1 − v2 k22 ≈ h(w − ws )2 i.
4
(20), we have
h
i
d
µ(θ) > η · t · τ ′ (s) 1 − τ (s)2 − 1 − τ (t)2
dθ
= η · t · τ ′ (s) τ (t)2 − τ (s)2
>0
PREPRINT VERSION
13
since τ ′ (s) < 0 and τ (t) < τ (s) for 0 < s < t.
This confirms that µ(θ) is a strictly increasing function, and
hence ρ(θ) is a strictly decreasing function in θ ∈ (0, π4 ). The
result is thus proved.
P ROOF
A PPENDIX E
OF P ROPOSITION 2
For notation brevity, we drop the symbol t for the variables related to the continuous-time algorithm dynamic X(t)
wherever the meaning is clear.
Denote the Hessian function of f (X) along the direction
ξ ∈ RN ×N as
i
1h
g(X + γξ) − g(X) .
h(ξ, X) = lim
γ→0 γ
Then, a Taylor’s expansion of the gradient function g(X)
yields
g(X) = g(X̂) + h(sξ, X̂) + o(s)
where ξ = γ1 (X − X̂) and γ = kX − X̂kF . Therefore, as
g(X̂) = 0, it holds that g(X) ≈ h(Xe , X̂) for small s.
In addition, it holds that
n
o
T d
d
E(Xe (t)) = tr X(t) − X̂
X(t)
dt
dt
n
o
T
= −tr X(t) − X̂ g(X(t)) .
n
o
d
As a result, dt
E(Xe ) = −tr XTe h(XTe , X̂) + o(kXe k2F ).
T
Using (24) – (25) and the fact that H = ÛV̂ , it can be
shown that
o
1 n T
tr Xe h(Xe , X) = tr UTe W Ue V T V
2
T
T
T
U
+ V e W Ve U
+ tr UTe W U VTe V
+ tr VTe W T V UTe U .
Note that under H = H, we have W = 1N ×N . In addition,
if kVe kF = o kUe kF , i.e., kVe kF ≪ kUe kF , then
o
n
o
1 n T
tr Xe h(Xe , X) = tr UTe Ue V T V + o kUe k2F
2
o
n
= tr Ue V T V Ue + o kUe k2F
=
N
X
j=1
≥
(e)
N
X
j=1
(e)
(e)T
uj V T V uj
+ o kUe k2F
(e)
λK V T V kuj k2 + o kUe k2F
where uj is the jth row vector of the matrix Ue . As a result,
we have
T
d
E(Xe ) ≤ −2λK (V̂V̂ )kUe k2F + o kUe k2F
dt
proving (27).
If kUe kF = o kVe kF , the derivation to show (28) is similar.
R EFERENCES
[1] A. Beck, P. Stoica, and J. Li, “Exact and approximate solutions of source
localization problems,” IEEE Trans. Signal Process., vol. 56, no. 5, pp.
1770–1778, 2008.
[2] H.-D. Qi, N. Xiu, and X. Yuan, “A lagrangian dual approach to
the single-source localization problem,” IEEE Trans. Signal Process.,
vol. 61, no. 15, pp. 3815–3826, 2013.
[3] X. Sheng and Y.-H. Hu, “Maximum likelihood multiple-source localization using acoustic energy measurements with wireless sensor networks,”
IEEE Trans. Signal Process., vol. 53, no. 1, pp. 44–53, 2005.
[4] C. Meesookho, U. Mitra, and S. Narayanan, “On energy-based acoustic
source localization for sensor networks,” IEEE Trans. Signal Process.,
vol. 56, no. 1, pp. 365–377, 2008.
[5] Y. Liu, Y. H. Hu, and Q. Pan, “Distributed, robust acoustic source
localization in a wireless sensor network,” IEEE Trans. Signal Process.,
vol. 60, no. 8, pp. 4350–4359, 2012.
[6] C. Liu, Y. V. Zakharov, and T. Chen, “Broadband underwater localization
of multiple sources using basis pursuit de-noising,” IEEE Trans. Signal
Process., vol. 60, no. 4, pp. 1708–1717, 2012.
[7] H. Chen, Q. Shi, R. Tan, H. V. Poor, and K. Sezaki, “Mobile element
assisted cooperative localization for wireless sensor networks with
obstacles,” IEEE Trans. Wireless Commun., vol. 9, no. 3, 2010.
[8] T. He, C. Huang, B. M. Blum, J. A. Stankovic, and T. Abdelzaher,
“Range-free localization schemes for large scale sensor networks,” in
Proc. Int. Conf. Mobile Computing and Networking, 2003, pp. 81–95.
[9] J. Blumenthal, R. Grossmann, F. Golatowski, and D. Timmermann,
“Weighted centroid localization in Zigbee-based sensor networks,” in
Prof. IEEE Int. Symp. Intelligent Signal Process., 2007, pp. 1–6.
[10] J. Wang, P. Urriza, Y. Han, and D. Cabric, “Weighted centroid localization algorithm: theoretical analysis and distributed implementation,”
IEEE Trans. Wireless Commun., vol. 10, no. 10, pp. 3403–3413, 2011.
[11] S. Choudhary and U. Mitra, “Analysis of target detection via matrix
completion,” in Proc. IEEE Int. Conf. Acoustics, Speech, and Signal
Processing, 2015, pp. 3771–3775.
[12] J. Chen and U. Mitra, “Rotated eigenstructure analysis for source
localization without energy-decay models,” in Proc. Int. Conf. Digital
Signal Process., London, UK, Aug. 2017.
[13] R. Lefort, G. Real, and A. Drémeau, “Direct regressions for underwater
acoustic source localization in fluctuating oceans,” Applied Acoustics,
vol. 116, pp. 303–310, 2017.
[14] Y. Jin, W.-S. Soh, and W.-C. Wong, “Indoor localization with channel
impulse response based fingerprint and nonparametric regression,” IEEE
Trans. Wireless Commun., vol. 9, no. 3, pp. 1120–1127, 2010.
[15] W. Kim, J. Park, J. Yoo, H. J. Kim, and C. G. Park, “Target localization
using ensemble support vector regression in wireless sensor networks,”
IEEE Trans. on Cybernetics, vol. 43, no. 4, pp. 1189–1198, 2013.
[16] S. Choudhary, N. Kumar, S. Narayanan, and U. Mitra, “Active target
localization using low-rank matrix completion and unimodal regression,”
arXiv preprint arXiv:1601.07254, 2016.
[17] J. Chen and U. Mitra, “Underwater acoustic source localization using
unimodal-constrained matrix factorization,” in Proc. Asilomar Conf.
Signals, Systems and Computers, Pacific Grove, CA, USA, Nov. 2017.
[18] E. Candes and B. Recht, “Exact matrix completion via convex optimization,” Commun. of the ACM, vol. 55, no. 6, pp. 111–119, 2012.
[19] E. J. Candes and Y. Plan, “Matrix completion with noise,” Proceedings
of the IEEE, vol. 98, no. 6, pp. 925–936, 2010.
[20] A. Németh and S. Németh, “How to project onto an isotone projection
cone,” Linear Algebra and its Applications, vol. 433, no. 1, pp. 41–51,
2010.
[21] C. R. Berger, S. Zhou, J. C. Preisig, and P. Willett, “Sparse channel
estimation for multicarrier underwater acoustic communication: From
subspace methods to compressed sensing,” IEEE Trans. Signal Process.,
vol. 58, no. 3, pp. 1708–1721, 2010.
[22] S. Beygi and U. Mitra, “Multi-scale multi-lag channel estimation using
low rank approximation for OFDM,” IEEE Trans. Signal Process.,
vol. 63, no. 18, pp. 4744–4755, 2015.
[23] L. M. Brekhovskikh, Fundamentals of ocean acoustics.
Springer
Science & Business Media, 2003.
[24] M. Kaul, B. Yang, and C. S. Jensen, “Building accurate 3D spatial
networks to enable next generation intelligent transportation systems,”
in Proc. IEEE Int. Conf. Mobile Data Management, 2013, pp. 137–146.
[25] V. Vu, “Singular vectors under random perturbation,” Random Structures
& Algorithms, vol. 39, no. 4, pp. 526–538, 2011.
[26] M. Rudelson and R. Vershynin, “Non-asymptotic theory of random
matrices: extreme singular values,” arXiv preprint arXiv:1003.2990,
2010.
| 7cs.IT
|
An Improved Subsumption Testing Algorithm for the
Optimal-Size Sorting Network Problem
arXiv:1707.08725v1 [cs.DS] 27 Jul 2017
Cristian Frăsinaru and Mădălina Răschip
Faculty of Computer Science, ”Alexandru Ioan Cuza” University, Iaşi, Romania
Abstract
In this paper a new method for checking the subsumption relation for the optimalsize sorting network problem is described. The new approach is based on creating
a bipartite graph and modelling the subsumption test as the problem of enumerating all perfect matchings in this graph. Experiments showed significant improvements over the previous approaches when considering the number of subsumption
checks and the time needed to find optimal-size sorting networks. We were able to
generate all the complete sets of filters for comparator networks with 9 channels,
confirming that the 25-comparators sorting network is optimal. The running time
was reduced more than 10 times, compared to the state-of-the-art result described
in [6].
Keywords: Comparator networks. Optimal-size sorting networks. Subsumption.
1. Introduction
Sorting networks are a special class of sorting algorithms with an active research area since the 1950’s [10], [3], [2]. A sorting network is a comparison
Email address: {acf, mionita}@info.uaic.ro (Cristian Frăsinaru and Mădălina
Răschip)
Preprint submitted to Elsevier
July 28, 2017
network which for every input sequence produces a monotonically increasing output. Since the sequence of comparators does not depend on the input, the network
represents an oblivious sorting algorithm. Such networks are suitable in parallel implementations of sorting, being applied in graphics processing units [9] and
multiprocessor computers [3].
Over time, the research was focused on finding the optimal sorting networks
relative to their size or depth. When the size is considered, the network must
have a minimal number of comparators, while for the second objective a minimal
number of layers is required. In [1] a construction method for sorting network
of size O(nlogn) and depth O(logn) is given. This algorithm has good results in
theory but it is inefficient in practice because of the large constants hidden in the
big-O notation. On the other side, the simple algorithm from [3] which constructs
networks of depth O(log 2n) has good results for practical values of n.
Because optimal sorting networks for small number of inputs can be used
to construct efficient larger networks the research in the area focused in the last
years on finding such small networks. Optimal-size and optimal-depth networks
are known for n ≤ 8 [10]. In [12] the optimal-depth sorting networks were provided for n = 9 and n = 10. The results were extended for 11 ≤ n ≤ 16 in [4].
The approaches use search with pruning based on symmetries on the first layers.
The last results for parallel sorting networks are for 17 to 20 inputs and are given
in [8], [5]. On the other side, the paper [6] proved the optimality in size for the
case n = 9 and n = 10. The proof is based on exploiting symmetries in sorting networks and on encoding the problem as a satisfiability problem. The use of
powerful modern SAT solvers to generate optimal sorting networks is also investigated in [11]. Other recent results can be found in [7], where a revised technique
2
to generate, modulo symmetry, the set of saturated two-layer comparator networks
is given. Finding the minimum number of comparators for n > 10 is still an open
problem. In this paper, we consider the optimal-size sorting networks problem.
Heuristic approaches were also considered in literature, for example approaches
based on evolutionary algorithms [15] that are able to discover new minimal networks for up to 22 inputs, but these methods cannot prove their optimality.
One of the most important and expensive operation used in [6] is the subsumption testing. This paper presents a new better approach to implement this
operation based on matchings in bipartite graphs. The results show that the new
approach makes the problem more tractable by scaling it to larger inputs.
The paper is organized as follows. Section 2 describes the basic concepts
needed to define the optimal-size sorting-network problem and a new model of
the subsumption problem. Section 3 presents the problem of finding the minimalsize sorting network. Section 4 discusses the subsumption problem while Section
5 the subsumption testing. Section 6 presents the new way of subsumption testing
by enumerating all perfect matchings. Section 7 describes the experiments made
to evaluate the approach and presents the results.
2. Basic Concepts
A comparator network Cn,k with n channels (also called wires) and size k is a
sequence of comparators c1 = (i1 , j1 ); . . . ; ck = (ik ; jk ) where each comparator
ct specifies a pair of channels 1 ≤ it < jt ≤ n. We simply denote by Cn a
comparator network with n channels, whenever the size of the network is not
significant in a certain context.
Graphically, a comparator network may be represented as a Knuth diagram
3
[10]. A channel is depicted as a horizontal line and a comparator as a vertical
segment connecting two channels.
1
0
0
0
0
0
0
1
1
1
1
0
1
1
0
0
0
1
0
0
1
1
1
1
channel 1
channel 2
channel 3
channel 4
Figure 1: The sorting network C = (1, 2); (3, 4); (2, 4); (1, 3); (2, 3), having 4 channels and 5
comparators, operating on the input sequence 1010. The output sequence is 0011.
An input to a comparator network Cn may be any sequence of n objects taken
from a totally ordered set, for instance elements in Zn . Let x = (x1 , . . . , xn ) be
an input sequence. Each value xi is assigned to the channel i and it will ”traverse”
the comparator network from left to right. Whenever the values on two channels
reach a comparator c = (i, j) the following happens: if they are not in ascending
order the comparator permutes the values (xi , xj ), otherwise the values will pass
through the comparator unmodified. Therefore, the output of a comparator network is always a permutation of the input. If x is an input sequence, we denote by
C(x) the output sequence of the network C.
A comparator network is called a sorting network if its output is sorted ascending for every possible input.
The zero-one principle [10] states that if a comparator network Cn sorts correctly all 2n sequences of zero and one, then it is a sorting network. Hence, without loss of generality, from now on we consider only comparator networks with
binary input sequences. In order to increase readability, whenever we represent a
4
binary sequence we only write its bits; so 1010 is actually the sequence (1, 0, 1, 0).
The output set of a comparator network is outputs(C) = {C(x)|∀x ∈ {0, 1}n }.
Let x be a binary input sequence of length n. We make the following notations:
zeros(x) = {1 ≤ i ≤ n|xi = 0} and ones(x) = {1 ≤ i ≤ n|xi = 1}. The output
set of a comparator network Cn can be partitioned into n + 1 clusters, each cluster
containing sequences in outputs(C) having the same number of ones. We denote
by cluster(C, p) the cluster containing all sequences having p ones:
cluster(C, p) = {x ∈ outputs(C) | |ones(x)| = p}.
Consider the following simple network C = (1, 2); (3, 4). The output clusters
of C are: cluster(C, 0) = {0000}, cluster(C, 1) = {0001, 0100}, cluster(C, 2) =
{0011, 0101, 1100}, cluster(C, 3) = {0111, 1101}, cluster(C, 4) = {1111}.
The following proposition states some simple observations regarding the output set and its clusters.
Proposition 1. Let C be a comparator network having n channels.
(a) C is the empty network ⇔ |outputs(C)| = 2n .
(b) C is a sorting network ⇔ |outputs(C)| = n + 1 (each cluster contains
exactly one element).
(c) |cluster(C, p)| ≤
n
p
, 1 ≤ p ≤ n − 1.
(d) |cluster(C, 0)| = |cluster(C, n)| = 1.
We extend the zeros and ones notations to output clusters in the following manner. Let C be a comparator network. For all 0 ≤ p ≤ n we denote
S
S
zeros(C, p) = {zeros(x)|x ∈ cluster(C, p)} and ones(C, p) = {ones(x)|x ∈
cluster(C, p)}. These sets contain all the positions between 1 and n for which
5
there is at least one sequence in the cluster having a zero, respectively an one,
set at that position. Considering the clusters from the previous example, we
have: zeros(C, 0) = zeros(C, 1) = zeros(C, 2) = {1, 2, 3, 4}, zeros(C, 3) =
{1, 3}, zeros(C, 4) = ∅, ones(C, 0) = ∅, ones(C, 1) = {2, 4}, ones(C, 2) =
ones(C, 3) = ones(C, 4) = {1, 2, 3, 4}.
We introduce the following equivalent representation of the zeros and ones
sets, as a sequence of length n, where n is the number of channels of the network,
and elements taken from the set {0, 1}. Let Γ be a cluster:
• zeros(Γ) = (γ1 , . . . , γn ), where γi = 0 if i ∈ zeros(Γ), otherwise γi = 1,
• ones(Γ) = (γ1′ , . . . , γn′ ), where γi′ = 1 if i ∈ ones(Γ), otherwise γi′ = 0.
In order to increase readability, we will depict 1 values in zeros, respectively 0
values in ones with the symbol ∗. Considering again the previous example, we
have: zeros(C, 3) = (0 ∗ 0∗) and ones(C, 1) = (∗1 ∗ 1).
If C is a comparator network on n channels and 1 ≤ i < j ≤ n we denote
by C; (i, j) the concatenation of C and (i, j), i.e. the network that has all the
comparators of C and in addition a new comparator connecting channels i and j.
The concatenation of two networks C and C ′ having the same number of channels
is denoted by C; C ′ and it is defined as the sequence of all comparators in C and
C ′ , first the ones in C and then the ones in C ′ . In this context, C represents a
prefix of the network C; C ′ . Obviously, size(C; C ′ ) = size(C) + size(C ′ ).
Let π be a permutation on {1, . . . , n}. Applying π on a comparator network C = (i1 , j1 ); . . . ; (ik , jk ) will produce the generalized network π(C) =
(π(i1 ), π(j1 )); . . . ; (π(ik ), π(jk )). It is called generalized because it may contain
comparators (i, j) with i > j, which does not conform to the actual definition of
6
a standard comparator network. An important result in the context of analyzing
sorting networks (exercise 5.3.4.16 in [10]) states that a generalized sorting network can always be untangled such that the result is a standard sorting network of
the same size. The untangling algorithm is described in the previously mentioned
exercise. Two networks Ca and Cb are called equivalent if there is a permutation
π such that untangling π(Cb ) results in Ca .
Applying a permutation π on a binary sequence x = (x1 , . . . , xn ) will permute
the corresponding values: π(x) = (xπ(1) , . . . , xπ(n) ). Applying π on a set of
sequences S (either a cluster or the whole output set) will permute the values of
all the sequences in the set: π(S) = {π(x)|∀x ∈ S}. For example, consider the
permutation π = (4, 3, 2, 1) and the set of sequences S = {0011, 0101, 1100}.
Then, π(S) = {1100, 1010, 0011}
3. Optimal-size sorting networks
The optimal size problem regarding sorting networks is: ”Given a positive
integer n, what is the minimum number of comparators sn needed to create a
sorting network on n channels?”.
Since even the problem of verifying whether a comparator network is a sorting
network is known to be Co-N P complete [13], we cannot expect to design an
algorithm that will easily answer the optimal size problem. On the contrary.
In order to prove that sn ≤ k, for some k, it is enough to find a sorting network
of size k. On the other hand, to show that sn > k one should prove that no network
on n channels having at most k comparators is a sorting network.
Let Rkn denote the set of all comparator networks having n channels and k
comparators. The naive approach to identify the sorting networks is by generating
7
the whole set Rkn , starting with the empty network and adding all possible comparators. In order to find a sorting network on n channels of size k, one could
iterate through the set Rkn and inspect the output set of each network. According
to proposition 1 (b), if the size of the output is n + 1 then we have found a sorting
network. If no sorting network is found, we have established that sn > k.
Unfortunately, the size of Rkn grows rapidly since |Rkn | = (n(n − 1)/2)k and
constructing the whole set Rkn is impracticable even for small values of n and k.
We are actually interested in creating a set of networks Nkn that does not include all possible networks but contains only ”relevant” elements.
Definition 1. A complete set of filters [6] is a set Nkn of comparator networks on
n channels and of size k, satisfying the following properties:
(a) If sn = k then Nkn contains at least one sorting network of size k.
opt
n
(b) If k < sn = k ′ then ∃Cn,k
′ an optimal-size sorting network and ∃Cn,k ∈ Nk
such that C is a prefix of C opt .
Since the existence of Nkn is guaranteed by the fact that Rkn is actually a complete set of filters, we are interested in creating such a set that is small enough (can
be computed in a ”reasonable” amount of time).
4. Subsumption
In order to create a complete set of filters in [6] it is introduced the relation of
subsumption.
Definition 2. Let Ca and Cb be comparator networks on n channels. If there
exists a permutation π on {1, . . . , n} such that π(outputs(Ca )) ⊆ outputs(Cb )
8
we say that Ca subsumes Cb , and we write Ca Cb (or Ca ≤π Cb to indicate the
permutation).
For example, consider the networks Ca = (0, 1); (1, 2); (0, 3) and
Cb = (0, 1); (0, 2); (1, 3). Their output sets are:
ouputs(Ca) = {{0000}, {0001, 0010}, {0011, 0110}, {0111, 1011}, {1111}},
ouputs(Cb) = {{0000}, {0001, 0010}, {0011, 0101}, {0111, 1011}, {1111}}.
It is easy to verify that π = (0, 1, 3, 2) has the property that Ca ≤π Cb .
Proposition 2. Let Ca and Cb be comparator networks on n channels, having
|outputs(Ca)| = |outputs(Cb)|. Then, Ca Cb ⇔ Cb Ca .
Proof. Assume that Ca ≤π Cb ⇒ π(outputs(Ca)) ⊆ outputs(Cb) and since
|outputs(Ca)| = |outputs(Cb)| ⇒ π(outputs(Ca)) = outputs(Cb). That means
that π is actually mapping each sequence in outputs(Ca ) to a distinct sequence
in outputs(Cb). The inverse permutation π −1 is also a mapping, this time from
outputs(Cb) to outputs(Ca), implying that π −1 (outputs(Cb )) = outputs(Ca) ⇒
Cb ≤π−1 Ca .
The following result is the key to creating a complete set of filters:
Lemma 1. Let Ca and Cb be comparator networks on n channels, both having
the same size, and Ca Cb . Then, if there exists a sorting network Cb ; C of size
k, there also exists a sorting network Ca ; C ′ of size k.
The proof of the lemma is presented in [6] (Lemma 2) and [4] (Lemma 7).
The previous lemma ”suggests” that when creating the set of networks Rkn
using the naive approach, and having the goal of creating actually a complete set
9
of filters, we should not add two networks in this set if one of them subsumes the
other.
The algorithm to generate Nkn
Require: n, k ∈ Z+
Ensure: Returns Nkn , a complete set of filters
N0n = {Cn,0 } {Start with the empty network}
for all p = 1 . . . k do
n
Npn = ∅ {Generate Npn from Np−1
, adding all possible comparators}
n
do
for all C ∈ Np−1
for all i = 1 . . . n − 1, j = i + 1 . . . n do
if the comparator (i, j) is redundant then
continue
end if
C ∗ = C; (i, j) {Create a new network C ∗ }
if 6 ∃C ′ ∈ Npn such that C ′ C ∗ then
Npn = Npn ∪ C ∗
Remove from Npn all the networks C ′′ such that C ∗ C ′′ .
end if
end for
end for
end for
return Nkn
A comparator c is redundant relative to the network C if adding it at the end
of C does not modify the output set: outputs(C; c) = outputs(C). Testing if a
10
comparator c = (i, j) is redundant relative to a network C can be easily implemented by inspecting the values xi and xj in all the sequences x ∈ outputs(C).
If xi ≤ xj for all the sequences then c is redundant.
The key aspect in implementing the algorithm above is the test for subsumption.
5. Subsumption testing
Let Ca and Cb be comparator networks on n channels. According to definition
2, in order to check if Ca subsumes Cb we must find a permutation π on {1, . . . , n}
such that π(outputs(Ca)) ⊆ outputs(Cb). If no such permutation exists then Ca
does not subsume Cb .
In order to avoid iterating through all n! permutations, in [6] several results are
presented that identify situations when subsumption testing can be implemented
efficiently. We enumerate them as the tests ST1 to ST4 .
(ST1 ) Check the total size of the output
If |outputs(Ca)| > |outputs(Cb)| then Ca cannot subsume Cb .
(ST2 ) Check the size of corresponding clusters (Lemma 4 in [6])
If there exists 0 ≤ p ≤ n such that |cluster(Ca , p)| > |cluster(Cb , p)| then Ca
cannot subsume Cb . When applying a permutation π on a sequence in outputs(Ca ),
the number of bits set to 1 remains the same, only their positions change. So,
if π(outputs(Ca)) ⊆ outputs(Cb ) then ∀0 ≤ p ≤ n π(cluster(Ca ), p) ⊆
cluster(Cb , p), which implies that |cluster(Ca )| = |π(cluster(Ca), p)| ≤
|cluster(Cb, p)| for all 0 ≤ p ≤ n.
(ST3 ) Check the ones and zeros (Lemma 5 in [6])
Recall that zeros and ones represent the sets of positions that are set to 0, respec11
tively to 1. If there exists 0 ≤ p ≤ n such that |zeros(Ca , p)| > |zeros(Cb , p)| or
|ones(Ca , p)| > |ones(Cb , p)| then Ca cannot subsume Cb .
For example, consider the networks Ca = (0, 1); (2, 3); (1, 3); (0, 4); (0, 2) and
Cb = (0, 1); (2, 3); (0, 2); (2, 4); (0, 2). cluster(Ca , 2) = {0011, 00110, 01010},
cluster(Cb , 2) = {00011, 01001, 01010}, ones(Ca , 2) = {2, 3, 4, 5}, ones(Cb , 2) =
{2, 4, 5}, therefore Ca 6 Cb .
(ST4 ) Check all permutations (Lemma 6 in [6])
The final optimization presented in [6] states that if there exists a permutation π
such that π(outputs(Ca )) ⊆ outputs(Cb) then ∀0 ≤ p ≤ n zeros(π(Ca , p)) ⊆
zeros(Cb , p) and ones(π(Ca , p)) ⊆ ones(Cb , p). So, before checking the inclusion for the whole output sets, we should check the inclusion for the zeros and
ones sets, which is computationally cheaper.
The tests (ST1 ) to (ST3 ) are very easy to check and are highly effective in
reducing the search space. However, if none of them can be applied, we have
to enumerate the whole set of n! permutations, verify (ST4 ) and eventually the
definition of subsumption, for each one of them. In [6] the authors focused on
n = 9 which means verifying 362, 880 permutations for each subsumption test.
They were successful in creating all sets of complete filters Nk9 for k = 1, . . . , 25
and actually proved that s9 = 25. Using a powerful computer and running a
parallel implementation of the algorithm on 288 threads, the time necessary for
9
creating these sets was measured in days (more than five days only for N14
).
Moving from 9! to 10! = 3, 628, 800 or 11! = 39, 916, 800 does not seem
feasible. We have to take in consideration also the size of the complete filter sets,
9
for example |N14
| = 914, 444.
We present a new approach for testing subsumption, which greatly reduces the
12
number of permutations which must be taken into consideration. Instead of enumerating all permutations we will enumerate all perfect matchings in a bipartite
graph created for the networks Ca and Cb being tested.
6. Enumerating perfect matchings
Definition 3. Let Ca and Cb be comparator networks on n channels. The subsumption graph G(Ca , Cb ) is defined as the bipartite graph (A, B; E(G)) with
vertex set V (G) = A ∪ B, where A = B = {1, . . . , n} and the edge set E(G)
defined as follows. Any edge e ∈ E(G) is a 2-set e = {i, j} with i ∈ A and j ∈ B
(also written as e = ij) having the properties:
• i ∈ zeros(Ca , p) ⇒ j ∈ zeros(Cb , p), ∀0 ≤ p ≤ n;
• i ∈ ones(Ca , p) ⇒ j ∈ ones(Cb , p), ∀0 ≤ p ≤ n.
So, the edges of the subsumption graph G represent a relationship between
positions in the two output sets of Ca and Cb . An edge ij signifies that the position
i (regarding the sequences in outputs(Ca )) and the position j (regarding Cb ) are
”compatible”, meaning that a permutation π with the property π(outputs(Ca )) ⊆
outputs(Cb) might have the mapping i to j as a part of it.
As an example, consider the following zeros and ones sequences, corresponding to Ca = (0, 1); (2, 3); (1, 3); (1, 4) and Cb = (0, 1); (2, 3); (0, 3); (1, 4).
zeros(Ca ) = {00000,00000,000-0,000--,000--,-----},
zeros(Cb ) = {00000,00000,00000,000--,000--,-----},
ones(Ca ) = {-----,---11,1-111,11111,11111,11111},
ones(Cb ) = {-----,---11,-1111,11111,11111,11111}.
The subsumption graph G(Ca , Cb ) is pictured below:
13
1
2
3
4
5
1
2
3
4
5
Figure 2: The subsumption graph corresponding to the comparator networks
Ca = (0, 1); (2, 3); (1, 3); (1, 4) and Cb = (0, 1); (2, 3); (0, 3); (1, 4)
A matching M in the graph G is a set of independent edges (no two edges in
the matching share a common node). If ij ∈ M we say that i and j are saturated.
A perfect matching is a matching that saturates all vertices of the graph.
Lemma 2. Let Ca and Cb be comparator networks on n channels. If Ca ≤π Cb
then π represents a perfect matching in the subsumption graph G(Ca , Cb).
Proof. Suppose that Ca ≤π Cb , π(i) = j and ij 6∈ E(G). That means that
∃0 ≤ p ≤ n such that i ∈ zeros(Ca , p)∧j 6∈ zeros(Cb , p) or i ∈ ones(Ca , p)∧j 6∈
ones(Cb , p). We will asumme the first case. Let x a sequence in cluster(Ca , p)
such that x(i) = 0. Since π(outputs(Ca )) ⊆ outputs(Cb) ⇒ π(x) ∈ cluster(Cb , p).
But π(i) = j, therefore in cluster(Cb , p) there is the sequence π(x) having the bit
at position j equal to 0, contradiction.
The previous lemma leads to the following result:
Corollary 1. Let Ca and Cb be comparator networks on n channels. Then Ca
subsumes Cb if and only if there exists a perfect matching π in the subsumption
graph G(Ca , Cb).
The graph in figure 2 has only four perfect matchings: (2, 1, 3, 4, 5), (3, 1, 2, 4, 5),
(2, 1, 3, 5, 4), (3, 1, 2, 5, 4). So, when testing subsumption, instead of verifying
5! = 120 permutations it is enough to verify only 4 of them.
14
If two clusters are of the same size, then we can strengthen the previous
result even more. If there is a permutation π such that π(cluster(Ca , p)) =
cluster(Cb , p) then π −1 (cluster(Cb , p) = cluster(Ca , p). Using the same reasoning, when creating the subsumption graph C(Ga , Cb) we add the following
two condition when defining an edge ij:
• j ∈ zeros(Cb , p) ⇒ i ∈ zeros(Ca , p), ∀0 ≤ p ≤ n such that |cluster(Ca , p)| =
|cluster(Cb, p)|,
• j ∈ ones(Cb , p) ⇒ i ∈ ones(Ca , p), ∀0 ≤ p ≤ n such that |cluster(Ca , p)| =
|cluster(Cb, p)|.
In order to enumerate all perfect matchings in a bipartite graph, we have implemented the algorithm described in [14]. The algorithm starts with finding a
perfect matching in the subsumption graph G(Ca , Cb ). Taking into consideration
the small size of the bipartite graph, we have chosen the Ford-Fulkerson algorithm which is very simple and does not require elaborate data structures. Its time
complexity is O(n|E(G)|). If no perfect matching exists, then we have established that Ca does not subsume Cb . Otherwise, the algorithm presented in [14]
identifies all other perfect matchings, taking only O(n) time per matching.
7. Experimental results
We implemented both variants of subsumption testing:
• (1) enumerating all permutations and checking the inclusions described by
(ST4 ) before verifying the actual definition of subsumption;
• (2) verifying only the permutations that are actually perfect matchings in
the subsumption graph, according to Corollary 1.
15
We made some simple experiments on a regular computer (Intel i7-4700HQ
@2.40GHz), using 8 concurrent threads. The programming platform was Java SE
Development Kit 8.
Several suggestive results are presented in the table below:
(n, k)
|Nkn |
total
sub
perm1
time1
perm2
time2
(7, 9)
678
1, 223, 426
5, 144
26, 505, 101
2.88
33, 120
0.07
(7, 10)
510
878, 995
5, 728
25, 363, 033
2.82
24, 362
0.06
(8, 7)
648
980, 765
2, 939
105, 863, 506
13.67
49, 142
0.14
(8, 8)
2088
9, 117, 107
9, 381
738, 053, 686
94.50
283, 614
0.49
(8, 9)
5703
24, 511, 628
29, 104
4, 974, 612, 498
650.22
1, 303, 340
1.96
The columns of the table have the following significations:
• (n, k) - n is the number of channels, k is the number of comparators;
• |Nkn | - the size of the complete set of filters generated for the given n and k;
• total - the total number of subsumption checks;
• sub - the number of subsumptions that were identified;
• perm1 - how many permutations were checked, using the variant (1);
• time1 - the total time, measured in seconds, using the variant (1);
• perm2 - how many permutations were checked, using the variant (2);
• time2 - the total time, measured in seconds, using the variant (2);
As we can see from this results, using the variant (2) the number of permutations that were verified in order to establish subsumption is greatly reduced.
16
Despite the fact that it is necessary to create the subsumption graph and to iterate
through its set of perfect matchings, this leads to a much shorter time needed for
the overall generation of the complete set of filters.
This new approach enabled us to reproduce the state-of-the-art result concerning optimal-size sorting networks, described in [6]. Using an Intel Xeon E5-2670
@ 2.60GHz computer, with a total of 32 cores, we generated all the complete set
of filters for n = 9. The results are presented in the table below.
k
1
2
3
4
5
6
7
8
|Nk9 |
1
3
7
20
59
208
807
3415
time(s)
0
0
0
0
0
0
0
0
k
9
10
11
12
13
14
15
16
14343
55991
188730
490322
854638
914444
607164
274212
4
48
769
6688
25186
40896
24161
5511
17
18
19
20
21
22
23
24
25
94085
25786
5699
1107
250
73
27
8
1
610
36
2
0
0
0
0
0
0
|Nk9 |
time(s)
k
|Nk9 |
time(s)
9
In [6] the necessary time required to compute |N14
| using the generate-and-
prune approach was estimated at more than 5 days of computation on 288 threads.
Their tests were performed on a cluster with a total of 144 Intel E8400 cores
clocked at 3 GHz. In our experiments, the same set was created in only 11 hours,
which is actually a significant improvement.
8. Acknowledgments
We would like to thank Michael Codish for introducing us to this research
topic and Cornelius Croitoru for his valuable comments. Furthermore, we thank
17
Mihai Rotaru for providing us with the computational resources to run our experiments.
9. Conclusions
In this paper we have extended the work in [6], further investigating the relation of subsumption. In order to determine the minimal number of comparators
needed to sort any input of a given length, a systematic BFS-like algorithm generates incrementally complete sets of filters, that is sets of comparator networks
that have the potential to prefix an optimal-size sorting network. To make this
approach feasible it is essential to avoid adding into these sets networks that subsume one another. Testing the subsumption is an expensive operation, invoked
a huge number of times during the execution of the algorithm. We described a
new approach to implement this test, based on enumerating perfect matchings
in a bipartite graph, called the subsumption graph. Computer experiments have
shown significant improvements, greatly reducing the number of invocations and
the overall running time. The results show that, using appropriate hardware, it
might be possible to approach in this manner the optimal-size problem for sorting
networks with more than 10 channels.
References
[1] M. Ajtai, J. Komlós, and E. Szemerédi. An 0(n log n) sorting network. In
Proceedings of the Fifteenth Annual ACM Symposium on Theory of Computing, STOC ’83, pages 1–9. ACM, 1983.
[2] Sherenaz W. Al-Haj Baddar and Kenneth E. Batcher. Designing Sorting
Networks: A New Paradigm. Springer Science & Business Media, 2012.
18
[3] Kenneth E. Batcher. Sorting networks and their applications. In Proceedings
of the April 30–May 2, 1968, spring joint computer conference, pages 307–
314. ACM, 1968.
[4] Daniel Bundala and Jakub Závodnỳ. Optimal sorting networks. In International Conference on Language and Automata Theory and Applications,
pages 236–247. Springer, 2014.
[5] Michael Codish, Luı́s Cruz-Filipe, Thorsten Ehlers, Mike Müller, and Peter
Schneider-Kamp. Sorting networks: to the end and back again. Journal of
Computer and System Sciences, 2016.
[6] Michael Codish, Luı́s Cruz-Filipe, Michael Frank, and Peter SchneiderKamp. Twenty-five comparators is optimal when sorting nine inputs (and
twenty-nine for ten). In Tools with Artificial Intelligence (ICTAI), 2014 IEEE
26th International Conference on, pages 186–193. IEEE, 2014.
[7] Michael Codish, Luı́s Cruz-Filipe, and Peter Schneider-Kamp. The quest
for optimal sorting networks: Efficient generation of two-layer prefixes. In
Symbolic and Numeric Algorithms for Scientific Computing (SYNASC), 2014
16th International Symposium on, pages 359–366. IEEE, 2014.
[8] Thorsten Ehlers and Mike Müller. New bounds on optimal sorting networks.
In Conference on Computability in Europe, pages 167–176. Springer, 2015.
[9] Peter Kipfer and Rüdiger Westermann. Improved gpu sorting. GPU gems,
2:733–746, 2005.
[10] Donald E. Knuth. The Art of Computer Programming, Volume 3: Sorting
19
and Searching (2nd ed.). Addison Wesley Longman Publishing Co., Inc.,
Redwood City, CA, USA, 1998.
[11] Andreas Morgenstern and Klaus Schneider. Synthesis of parallel sorting
networks using sat solvers. In MBMV, pages 71–80, 2011.
[12] Ian Parberry. A computer-assisted optimal depth lower bound for nine-input
sorting networks. Mathematical systems theory, 24(1):101–116, 1991.
[13] Ian Parberry. On the computational complexity of optimal sorting network
verification. In Parle ’91 Parallel Architectures and Languages Europe:
Volume I: Parallel Architectures and Algorithms, pages 252–269. Springer,
1991.
[14] Takeaki Uno. Algorithms for enumerating all perfect, maximum and maximal matchings in bipartite graphs. In Proceedings of the 8th International
Symposium on Algorithms and Computation, ISAAC ’97, pages 92–101,
London, UK, UK, 1997. Springer-Verlag.
[15] Vinod K. Valsalam and Risto Miikkulainen. Using symmetry and evolutionary search to minimize sorting networks. Journal of Machine Learning
Research, 14:303–331, 2013.
20
| 8cs.DS
|
Chained Multi-stream Networks Exploiting Pose, Motion, and Appearance for
Action Classification and Detection
arXiv:1704.00616v2 [cs.CV] 26 May 2017
Mohammadreza Zolfaghari , Gabriel L. Oliveira, Nima Sedaghat, and Thomas Brox
University of Freiburg
Freiburg im Breisgau, Germany
{zolfagha,oliveira,nima,brox}@cs.uni-freiburg.de
Abstract
General human action recognition requires understanding of various visual cues. In this paper, we propose a network architecture that computes and integrates the most
important visual cues for action recognition: pose, motion, and the raw images. For the integration, we introduce a Markov chain model which adds cues successively.
The resulting approach is efficient and applicable to action
classification as well as to spatial and temporal action localization. The two contributions clearly improve the performance over respective baselines. The overall approach
achieves state-of-the-art action classification performance
on HMDB51, J-HMDB and NTU RGB+D datasets. Moreover, it yields state-of-the-art spatio-temporal action localization results on UCF101 and J-HMDB.
Figure 1: The chained multi-stream 3D-CNN sequentially
refines action class labels by analyzing motion and pose
cues. Pose is represented by human body parts detected by
a deep network. The spatio-temporal CNN can capture the
temporal dynamics of pose. Additional losses on YP ose and
YOF are used for training. The final output of the network
YRGB is provided at the end of the chain.
1. Introduction
Human action recognition is a complex task in computer
vision, due to the variety of possible actions is large and
there are multiple visual cues that play an important role. In
contrast to object recognition, action recognition involves
not only the detection of one or multiple persons, but also
the awareness of other objects, potentially involved in the
action, such as the pose of the person, and their motion.
Actions can span various time intervals, making good use
of videos and their temporal context is a prerequisite for
solving the task to its full extent [38, 37].
The success of convolutional networks in recognition has
also influenced action recognition. Due to the importance of
multiple visual cues, as shown by Jhuang et al. [12], multistream architectures have been most popular. This trend was
initiated by Simonyan and Zisserman [33], who proposed a
simple fusion of the action class scores obtained with two
separate convolutional networks, where one was trained on
raw images and the other on optical flow. The relative success of this strategy shows that deep networks for action
recognition cannot directly infer the relevant motion cues
from the raw images, although, in principle, the network
could learn to compute such cues.
In this paper, we propose a three-stream architecture that
also includes pose, see Figure 1. Existing approaches model
the temporal dynamics of human postures with hand-crafted
features. We rather propose to compute the position of human body parts with a fast convolutional network. Moreover, we use a network architecture with spatio-temporal
convolutions [37]. This combination can capture temporal
dynamics of body parts over time, which is valuable to improve action recognition performance, as we show in dedicated experiments. The pose network also yields the spatial
localization of the persons, which allows us to apply the
approach to spatial action localization in a straightforward
manner.
1
The second contribution is on the combination of the
multiple streams, as also illustrated in Figure 1. The combination is typically done by summation of scores, by a linear classifier, or by early or late concatenation of features
within the network. In this paper, we propose the integration of different modalities via a Markov chain, which leads
to a sequential refinement of action labels. We show that
such sequential refinement is beneficial over independent
training of streams. At the same time, the sequential chain
imposes an implicit regularization. This makes the architecture more robust to over-fitting – a major concern when
jointly training very large networks. Experiments on multiple benchmarks consistently show the benefit of the sequential refinement approach over alternative fusion strategies.
Since actions may span different temporal resolutions,
we analyze videos at multiple temporal scales. We demonstrate that combining multiple temporal granularity levels
improves the capability of recognizing different actions. In
contrast to some other state-of-the-art strategies to analyze
videos over longer time spans, e.g., temporal segmentation networks [43], the architecture still allows the temporal localization of actions by providing actionness scores of
frames using a sliding window over video. We demonstrate
this flexibility by applying the approach also to temporal
and spatio-temporal action detection. Compared to previous spatio-temporal action localization methods, which are
typically based on region proposals and action tubes, the
pose network in our approach directly provides an accurate
person localization at no additional computational costs.
Therefore, it consistently outperforms the previous methods
in terms of speed and mean average precision.
LSTM to classify video sequences. More recently, several
CNN based works presented efficient deep models for action recognition [6, 29, 37]. Tran et al. [37] employed a 3D
architecture to learn spatio-temporal features from videos.
Fusion of multiple modalities. Zisserman et al. [33]
proposed a two-stream CNN to capture the complementary
information from appearance and motion, each modality in
an independent stream. Feichtenhofer et al. [8] investigated
the optimal position within a convolution network in detail
to combine the separate streams. Park et al. [28] proposed
a gated fusion approach. In a similar spirit, Wang et al. [46]
presented an adaptive fusion approach, which uses two regularization terms to learn fusion weights. In addition to optical flow, some works made use of other modalities like
audio [46], warped flow [43], and object information [11]
to capture complementary information for video classification. In the present work, we introduce a new, flexible fusion technique for early or late fusion via a Markov chain
and show that it outperforms previous fusion methods.
Pose feature based methods. Temporal dynamics of
body parts over time provides strong information on the performing action. Thus, this information has been employed
for action recognition in several works [4, 19, 39]. Cheron
et al. [4] used pose information to extract high-level features from appearance and optical flow. They showed that
using pose information for video classification is highly effective. Wang et al. [39] used data mining techniques to
obtain a representation for each video and finally, by using a bag-of-words model to classify videos. In the present
work, we compute the human body layout efficiently with
a deep network and learn the relevant spatio-temporal pose
features within one of the streams of our action classification network.
2. Related work
Feature based approaches. Many traditional works in
the field of action recognition focused on designing features
to discriminate action classes [17, 40, 5, 16]. These features
were encoded with high order encodings, e.g., bag of words
(BoW) [35] or Fisher vector based encodings [31], to produce a global representation for video and to train a classifier on the action labels. Recent research showed that most
of these approaches are not only computationally expensive,
but they also fail on capturing context and high-level information.
CNN based approaches. Deep learning has enabled
the replacement of hand-crafted features by learned features, and the learning of whole tasks end-to-end. Several works employed deep architectures for video classification [24, 37, 41]. Thanks to their hiearchical feature representation, deep networks learn to capture localized features as well as context cues and can exploit high-level information from large scale video datasets. Baccouche et
al. [2] firstly used a 3D CNN to learn spatio-temporal features from video and in the next step they employed an
3. Inputs to the Network
We rely on three input cues: the raw RGB images, optical flow, and human pose in the form of human body part
segmentation. All inputs are provided as spatio-temporal
inputs covering multiple frames.
3.1. Optical Flow
We compute the optical flow with the method from
Zach et al. [48], which is a reliable variational method that
runs sufficiently fast. We convert the x-component and ycomponent of the optical flow to a 3 channel RGB image by
stacking components and magnitude of them [29]. The flow
and magnitude values in the image are multiplied by 16 and
quantized into the [0,255] interval [18, 29, 42, 43].
3.2. Body Part Segmentation
Encoder-decoder architectures with an up-convolutional
part have been used successfully for semantic segmentation
tasks [23, 22, 30, 3, 27], depth estimation [20] and optical
2
Figure 2: Human body part segmentation architecture. Convolutions are shown in green, pooling in blue, feature map dropout
in brown, up-convolutional layers in red and softmax in yellow.
flow estimation [7]. For this work, we make use of Fast-Net
[27], a network for human body part segmentation, which
will provide our action recognition network with body pose
information. Figure 2 illustrates the architecture of FastNet. The encoder part of the network is initialized with the
VGG network [34]. Skip connections from the encoder to
the decoder part ensure the reconstruction of details in the
output up to the original input resolution.
We trained the Fast-Net architecture on the J-HMDB
[12] and the MPII [1] action recognition datasets. J-HMDB
provides body part segmentation masks and joint locations,
while MPII provides only joint locations. To make body
part masks compatible across datasets, we apply the following methodology, which only requires annotation for the
joint locations. First, we derive a polygon for the torso from
the joint locations around that area. Secondly, we approximate the other parts by ellipses scaled consistently based on
the torso area and the distance between the respective joints;
see second column of Fig. 3. We convert the body part segmentation into a 3 channel RGB image, mapping each label
to a correspondent pre-defined RGB value.
To the best of our knowledge, we are the first who trained
a convolutional network on body part segmentation for the
purpose of action recognition. Figure 3 shows exemplary
results of the body part segmentation technique on J-HMDB
and MPII datasets. Clearly, the network provides good accuracy on part segmentation and is capable of handling images with multiple instances. The pose estimation network
has a resolution of 150×150 and runs at 33 fps.
Figure 3: Qualitative results on J-HMDB and MPII datasets
(task with 15 body parts). First column: Input image. Second column: Ground truth. Third column: Result predicted with Fast-Net. First two rows correspond to results
on J-HMDB and the last ones on MPII.
4. Action Recognition Network
optical flow stream, and finally apply a refinement by the
RGB stream.
We use the assumption that the class predictions are conditionally independent due to the different input modalities.
Consequently, the joint probability over all input streams
factorizes into the conditional probabilities over the separate input streams.
In a Markov chain, given a sequence of inputs X =
{X1 , X2 , ..., XS }, we wish to predict the output sequence
Y = {Y1 , Y2 , ..., YS } such that P (Y |X) is maximized. Due
4.1. Multi-stream Fusion with a Markov Chain
To integrate information from the different inputs we rely
on the model of a multi-stream architecture [33], i.e., each
input cue is fed to a separate convolutional network stream
that is trained on action classification. The innovation in our
approach is the way we combine these streams. In contrast
to the previous works, we combine features from the different streams sequentially. Starting with the human body part
stream, we refine the evidence for an action class with the
3
YPose
YOF
YRGB
FC2
FC
FC
FC
FC1
FC
FC
Ypred
h1
YPose
h2
At each subsequent stage s > 2, we obtain a refined prediction ys by combining the hidden state and the predictions
from the previous stage.
FC
YPose YOF
h3
Concatenated Features
FC
h2 = f ([h1 , 3DCNN(XOF ), (Y1 )])
3DCNN
3DCNN
FC
3DCNN
FC
3DCNN
FC
3DCNN
FC
3DCNN
FC
Pose
OF
RGB
Pose
OF
RGB
P (Y2 |X, Y<2 ) = softmax(Net2 (h2 ))
(4)
h3 = f ([h2 , 3DCNN(XRGB ), (Y1 , Y2 )])
P (Y3 |X, Y<3 ) = softmax(Net3 (h3 ))
In the proposed model, at each stage, the next prediction
is made conditioned on all previous predictions and the new
input. Therefore, when training the network, the prediction
of the output class label does not only depend on the input,
but also on the previous state. Thus, the network in that
stream will learn complementary features to refine the class
labels from the previous streams. With this chaining and
joint training, the information at the previous stages serve
as the present belief for the predictions at the current stage,
as shown in Figure 4-right. This sequential improvement
of the class label enables the combination of multiple cues
within a large network, while keeping the risk of over-fitting
low.
This is in contrast to the fusion approaches that combine features from different, independently trained streams.
In such a case, the different streams are not enforced to
learn complementary features. In the other extreme, approaches that train all streams jointly but not sequentially,
are more prone to over-fitting, because the network is very
large, and, in such case, lacks the regularization via the separate streams and their additional losses.
It should be expected that the ordering of the sequence
plays a role for the final performance. We compared different ordering options in our experiments and report them
in the following section. The ordering that starts with the
pose as input and ends with the RGB image yielded the best
results.
It is worth noting that the concept of sequential fusion
could be applied to any layer of the network. Here we
placed the fusion after the first fully-connected layer, but
the fusion could also be applied to the earlier convolutional
layers.
Figure 4: Baseline fusion architecture (left) and the proposed approach (right). In the chained architecture, there
is a separate loss function for each stream. The final class
label is obtained at the end of the chain (rightmost prediction).
to the Markov property, P (Y |X) can be decomposed:
P (Y |X) = P (Y1 |X)
S
Y
P (Ys |X, Y1 , . . . , Ys−1 ) (1)
s=2
For the state s ∈ {1, . . . , S}, we denote by hs the hidden
state of that stream. We use deep networks to model the
likelihood in (1):
hs = f ([hs−1 , 3DCNN(Xs ), (Y1 , . . . , Ys−1 )])
P (Ys |X, Y<s ) = softmax(Nets (hs )),
(2)
where f is a non-linearity unit (ReLU), hs−1 denotes the
hidden state from the previous stream, and ys is the prediction of stream s. For the 3DCNN(·), we use the convolutional part of the network presented in Figure 5 to encapsulate the information in the input modality, and Nets is the
fully connected part in Figure 5.
At each fusion stage, we concatenate the output of the
function 3DCNN(·) with the hidden state and the outputs from the previous stream and apply the non-linearity
f before feeding them to Nets . Finally, at the output part, we use Nets to predict action labels from hs .
With the softmax(·) function we convert these scores into
(pseudo-)probabilities.
Using the above notation, we consider input modalities
as X = {Xpose , XOF , XRGB }, and Xs = {xt }Tt=1 , where
xt is the t-th frame in Xs , and T is the total number of
frames in Xs . At the stage s = 1, by considering X1 =
Xpose we start with an initial hidden state and obtain an
initial prediction (see Figure 4-right):
h1 = 3DCNN(Xpose )
P (Y1 |X) = softmax(Net1 (h1 ))
4.2. Network Configuration
In all streams, we use the C3D architecture [37] as the
base architecture, which has 17.5M parameters. The network has 8 three-dimensional convolution layers with kernel size of 3×3×3 and stride 1, 5 three-dimensional pooling
layers with kernel size of 2×2×2 and stride 2 and two fully
connected layers followed by a softmax; see Figure 5. Each
stream is connected with the next stream via layer FC6; see
Figure 4-right. Each stream takes 16 frames as input.
(3)
4
Figure 5: Base architecture used in each stream of the action recognition network. The convolutional part is a 3DCNN
architecture. We define the remaining fully connected layers as N ets .
4.3. Training
poral windows across a video and 10 crop scores per clip.
Apart from averaging, we also tested a multi-resolution approach, which we call multi-granular (MG), where we
trained separate networks for three different temporal resolutions. These are assembled as (1) 16 consecutive frames,
(2) 16 frames from a temporal window of 32 frames by a
sample rate of 2, and (3) 16 frames sampled randomly from
the entire video. For the final score, we take the average
over the scores produced by these temporal resolution networks. This approach extends the temporal context that the
network can see, which can be useful for more complex actions with longer duration.
In case of temporal action detection, we localize the action in time by thresholding the score provided for each
frame. Clearly, the MG approach is not applicable here.
In addition to the action score, also the human body part
network helps in temporal localization: we do not detect an
action as long as no human is detected. More details on the
spatio-temporal action detection are provided in the experimental section and in the supplemental material.
The network weights are learned using mini-batch
stochastic gradient descent (SGD) with a momentum of
0.9 and weight decay of 5e−4 . We jointly optimize the
whole network without truncating gradients and update the
weights of each stream based on the full gradient including the contribution from the following stream. We initialize the learning rate with 1e−4 and decrease it by a factor
of 10 every 2k for J-HMDB, 20k for UCF101 and NTU,
and at multiple steps for HMDB51. The maximum number
of iterations was 20k for J-HMDB, 40k for HMDB51 and
60k for the UCF101 and NTU datasets. We initialize the
weights of all streams with an RGB network pre-trained on
the large-scale Sports-1M dataset [14].
We split each video into clips of 16 frames with an overlap of 8 frames and feed each clip individually into the network stream with size of 16 × 112 × 112. We apply corner
cropping as a form of data augmentation to the training data.
Corner cropping extracts regions from the corners and the
center of the image. It helps to prevent the network from
bias towards the center area of the input. Finally, we resize
these cropped regions to the size of 112 × 112. In each iteration, all streams take the same clip from the video with the
same augmentation but with different modalities as input.
We used Caffe [13] and an NVIDIA Titan X GPU to run
our experiments. The training time for the J-HMDB dataset
was ∼ 10 hours for the full network.
5. Experiments
5.1. Datasets
UCF-101 [36] contains more than 2 million frames in
more than 13, 000 videos, which are divided into 101 human action classes. The dataset is split into three folds
and each split contains about 8000 videos for training.
The UCF101 dataset also comes with a subset for spatiotemporal action detection.
HMDB51 [15] contains 6766 videos divided into 51 action classes, each with at least 101 samples. The evaluation
follows the same protocol used for UCF-101.
J-HMDB contains a subset of videos from the HMDB
dataset, for which it provides additional annotation, in particular optical flow and joint localization [12]. Thus, it is
well-suited for evaluating the contribution of optical flow,
body part segmentation, and the fusion of all cues via a
4.4. Temporal Processing of the Whole Video
At test time, we feed the architecture with a temporal
window of 16 frames. The stride over the video is 8. Each
set of inputs is randomly selected for cropping operations,
which are 4 corners and 1 center crop for the original image and their horizontal flipping counterpart. We extract
scores before the softmax normalization in the last stream
(Y RGB).
In case of action classification, the final score of a video
is calculated by taking the average of scores over all tem5
Streams
1
RGB+OF
3 w/o GT
3 with GT
Variant
RGB
OF
Pose
Pose (GT)
baseline
chained
chained+MG
baseline
chained
chained+MG
baseline
chained
UCF101
84.2%
79.6%
56.9%
87.1%
88.9%
89.1%
90.4%
91.3%
-
HMDB
53.3%
45.2%
36.0%
55.6%
61.7%
66.0%
57.5%
62.1%
71.1%
-
J-HMDB
60.8%
61.9%
45.5%
56.8%
62.7%
72.8%
70.2%
79.1%
72.0%
83.2%
Datasets
Methods
TS Fusion [8]
LTC [38]
Two-stream [33]
TSN [43]
CPD [26]
Multi-Granular [18]
M-fusion [28]
KVMF [49]
P-CNN [4]
Action tubes [9]
TS R-CNN [29]
MR-TS R-CNN [29]
Ours (chained)
Table 1: The value of different cues and their integration
for action recognition on the UCF101, HMDB51, and JHMDB datasets (split 1). Adding optical flow and pose
is always beneficial. Integration via the proposed Markov
chain clearly outperforms the baseline fusion approach. In
all cases, the accuracy achieved with estimated optical flow
and body parts almost reaches the upper bound performance
when providing ground truth values for those inputs.
UCF101
HMDB51
J-HMDB
92.5%
91.7%
88.0%
94.2%
92.3%
90.8%
89.1%
93.1%
91.1%
65.4%
64.8%
59.4%
69.4%
66.2%
63.6%
54.9%
63.3%
69.7%
61.1%
62.5%
70.5%
71.1%
76.1%
Table 2: Comparison to the state of the art on UCF101,
HMDB51, and J-HMDB datasets (over all three splits).
is lost due to erroneous optical flow and pose estimates. Surprisingly, the difference between the results is rather small,
showing that the network does not suffer much from imperfect estimates. This conclusion can be drawn independently
of the fusion method.
Finally, the temporal multi-granularity fusion (MG) further improves results. Especially on HMDB51, there is a
large benefit.
Markov chain. The dataset comprises 21 human actions.
The complete dataset has 928 clips and 31838 frames.
There are 3 folds for training and testing for this dataset.
The videos in J-HMDB are trimmed and come with bounding boxes. Thus, it can be used also as a benchmark for
spatial action localization.
NTU RGB+D is a recent action recognition dataset that
is quite large and provides depth and pose ground truth
[32]. It contains more than 56000 sequences and 4 million
frames. NTU provides 60 action classes and 3D coordinates
for 25 joints. Additionally, the high intra-class variations
make NTU one of the most challenging datasets.
5.2.1
Comparison with the state-of-the-art
Table 3 compares the proposed network to the state of the
art in action classificaation. In contrast to Table 1, the comparison does not show the direct influence of single contributions anymore, since this table compares whole systems that are based on quite different components. Many of
these systems also use other features extraction approaches,
such as improved dense trajectories (IDT), which generally
have a positive influence on the results, but also make the
system more complicated and harder to control. Our network outperforms the state of the art on J-HMDB, NTU,
and HMDB51. Also, on UCF101 dataset our approach is
on par with the current state of the art while it does not
rely on any additional hand-crafted features. In two stream
case (RGB+OF), if we replace the 3DCNN network by the
TSN approach [43], we obtain a classification accuracy of
94.05% on UCF101 (over 3 splits), which is the state of the
art also on this dataset. However, the TSN approach does
not allow for action detection anymore.
Finally, we ran the network on the recent NTU RGB+D
dataset, which is larger and more challenging than the previous datasets. The dataset is popular for the evaluation of
methods that are based on human body pose. Clearly, the result of our network, shown in Table ??, compares favorably
5.2. Action Classification
Table 1 shows that fusion with the sequential Markov
chain model outperforms the baseline fusion consistently
across all datasets. The baseline fusion is shown in Figure 4
and can be considered a strong baseline. It consists of fusing the multiple modalities through feature concatenation
followed by a set of fully connected layers. The network is
trained jointly.
Adding pose leads to a substantial improvement over the
two-stream version. This confirms that pose plays an important role as complementary modality for action recognition
tasks. Again, the Markov chain fusion is advantageous with
a large margin.
For the J-HMDB dataset, ground truth for optical flow
and pose is available and can be provided to the method.
While not being relevant in practice, running the recognition with this ground truth shows on how much performance
6
Methods
Fusion Location
FC7
FC6
Cross Subject %
Deep LSTM [32]
P-LSTM [32]
HOGˆ2 [25]
FTP DS [10]
ST-LSTM [21]
Ours (Pose)
Ours (RGB+OF+Pose - Baseline)
Ours (RGB+OF+Pose - Chained)
60.7%
62.93%
32.2%
60.23%
69.2%
67.8%
76.9%
80.8%
UCF101
89.8%
89.6%
Dataset
J-HMDB (RGB)
NTU RGB+D (Pose)
OPR
59.8%
86.8%
ORP
57.3%
86.2%
RPO
54.8%
84.3%
ROP
54.1%
84.7%
PRO
56.4%
85.1%
POR
60.0%
87.1%
Y Pose
55.7%
40.9%
47.1%
Y OF
83.0%
56.4%
65.3%
5.2.4
Effect of clip length
We analyzed the effect of the size of the temporal window
on the action recognition performance. Larger windows
clearly improve the accuracy on all datasets; see Table 7.
For the J-HMDB dataset (RGB modality) we use a temporal window ranging from 4 to 16 frames every 4 frames.
The highest accuracy is obtained with a 16 frames clip size.
Based on the J-HMDB minimum video size, 16 is the highest possible time frame to be explored. We also tested multiple temporal resolutions for the NTU dataset (pose modality). Again, we obtained the best results for the network
with the larger clip length as input.
The conducted experiments confirm that increasing the
length of the clip, we decrease the chance of getting unrelated parts of an action in a video. In addition, with longer
sequences, 3D convolutions can better exploit their ability
to capture abstract spatio-temporal features for recognizing
actions.
to the existing methods. As a result, the used pose estimation network is competitive with pose estimates using depth
images and that our way to integrate this information with
the raw images and optical flow is advantageous.
Ordering of modalities in the Markov chain.
Table 4 shows an analysis on how the order of the modalities
affects the final classification accuracy. Clearly, the ordering has an effect. The proposed ordering starting with the
pose and then adding the optical flow and the RGB images
performed best, but there are alternative orders that do not
perform much worse.
Table 5 quantifies the improvement in accuracy when
adding a modality. Clearly, each additional modality improves the results.
5.2.3
Accuracy
44.8%
49.6%
58.7%
60.8%
61.6%
67.8%
the outcome of the study by Feichtenhofer et al. [8], where
the last convolutional layer worked best.
Y RGB
90.4%
62.1%
79.1%
Table 5: Sequential improvement of classification accuracy
on UCF101, HMDB51 and J-HMDB datasets (Split1) by
adding modalities to the chained network.
5.2.2
Clip length
4
8
12
16
16
32
Table 7: Effect of the temporal window size. Using more
frames as input to the network consistently increases classification performance.
Table 4: Impact of chain order on the performance (clip
accuracy) on UCF101 and HMDB51 datasets (split1). ”O”
= Optical flow, ”P” = Pose and ”R” = RGB.
Dataset
UCF101
HMDB51
J-HMDB
J-HMDB
73.9%
79.1%
Table 6: Classification performance for different fusion
locations on UCF101, HMDB51 and J-HMDB datasets
(split1).
Table 3: Comparison to literature on the NTU RGB+D
benchmark.
Dataset
HMDB51
UCF101
HMDB51
61.3%
62.1%
5.3. Action Detection
To demonstrate the generality of our approach, we show
also results on action detection on UCF101 and J-HMDB.
Many of the top performing methods for action classification are not applicable to action detection, because they integrate information over time in a complex manner, are too
slow, or are unable to spatially localize the action.
This is different for our approach, which is efficient and
can be run in a sliding window manner over time and provides good spatial localization via the human body part segmentation. In order to create temporally consistent spatial
detections, we link action bounding boxes over time to pro-
Fusion location
In principle the chained fusion can be applied to any layer in
the network. We studied the effect of this choice. In contrast
to the large scale evaluation in Feichtenhofer et al. [8], we
tested only two locations: FC6 and FC7. Table 6 shows
a clear difference only on the J-HMDB dataset. There it
seems that an earlier fusion, at a level where the features
are not too abstract yet, is advantageous. This is similar to
7
3DCNN
Per-frame scores
RGB
OF
Pose
Chained
Ground truth
Prediction
Basketball
Basketball
Temporal localization
Pose
Spatial localization
Figure 6: Scheme for spatio-temporal action detection. The
chained network provides action class scores and body part
segmentations per frame. From these we compute action
tubes and their actionness scores; see the supplemental material for details.
Figure 7: Qualitative results on the action detection task.
The first two rows correspond to detections on UCF101,
the last ones on J-HMDB. Ground truth bounding boxes are
shown in green and detections in red. Our spatial localization is accurate and robust to unusual pose.
duce action tube [9]; see the supplemental material for details. We use the frame level action classification scores to
make predictions at the tube level. Figure 6 schematically
outlines the detection procedure.
We also present a set of qualitative action detection experiments for the UCF and J-HMDB datasets. Figure 7
shows several examples where we can robustly localize the
action, even when unusual pose, illumination, viewpoints
and motion blur are presented. Additional results exploring
failure cases are provided in supplementary material.
Following recent works on action detection [9, 44, 29],
we report video-AP. A detection is considered correct if the
intersection over union (IoU) with the ground-truth is above
a threshold δ and the action label is predicted correctly.
The IoU between two tubes is defined as the IoU over the
temporal domain, multiplied by the average of the IoU between boxes averaged over all overlapping frames. VideoAP measures the area under the precision-recall curve of the
action tube predictions.
Table 8 and Table 9 show the video mAP results on spatial and spatio-temporal action detection with different IoU
thresholds on J-HMDB and UCF101 (split1) datasets respectively. Although we did not optimize our approach for
action detection, we obtain state-of-the-art results on both
datasets. Moreover, the approach is fast: spatial detection
runs at a rate of 31 fps and spatio-temporal detection with
10 fps. Compared to the recent works [9, 45, 29, 47], our
detection framework has two desirable properties: (1) the
pose network directly provides a single detection box per
person, which causes a large speed-up; (2) the classification
J-HMDB
IoU threshold (δ)
0.1
0.2
0.3
0.4
0.5
Actionness [42]
56.4
ActionTubes [9]
53.3
Weinzaepfel et al. [44]
63.1 63.5 62.2 60.7
Peng et al. [29]
74.3
73.1
Ours
78.81 78.20 77.12 75.05 73.47
Table 8: Spatial action detection results (Video mAP) on
the J-HMDB dataset. Across all IoU thresholds, our model
outperforms the state of the art.
UCF101
IoU threshold (δ)
0.05
0.1
0.2
0.3
Weinzaepfel et al. [44] 54.28 51.68 46.77 37.82
Yu et al. [47]
42.80
Peng et al. [29]
54.46 50.39 42.27 32.70
Weinzaepfel et al. [45] 62.8
45.4
Ours
65.22 59.52 47.61 38.00
Table 9: Spatio-temporal action detection results (Video
mAP) on UCF101 dataset (split1). Across all IoU thresholds, our model outperforms the state of the art.
8
takes advantage of three modalities and the chained fusion,
which yields highly accurate per-frame scores.
[9]
[10]
6. Conclusion
We have proposed a network architecture that integrates
multiple cues sequentially via a Markov chain model. We
have shown that this sequential fusion clearly outperforms
other ways of fusion, because it can consider the mutual
dependencies of cues during training while avoiding overfitting due to very large network models. Our approach provides state-of-the-art performance on all four challenging
action classification datasets UCF101, HMDB51, J-HMDB
and NTU RGB+D while not using any additional handcrafted features. Moreover, we have demonstrated the value
of a reliable pose representation estimated via a fast convolutional network. Finally, we have shown that the approach
generalizes also to spatial and spatio-temporal action detection, where we obtained state-of-the-art results as well.
[11]
[12]
[13]
[14]
7. Acknowledgements
[15]
We acknowledge funding by the ERC Starting Grant
VideoLearn and the Freiburg Graduate School of Robotics.
References
[16]
[1] M. Andriluka, L. Pishchulin, P. Gehler, and B. Schiele. 2d
human pose estimation: New benchmark and state of the
art analysis. Conference on Computer Vision and Pattern
Recognition (CVPR), 2014. 3
[2] M. Baccouche, F. Mamalet, C. Wolf, C. Garcia, and
A. Baskurt. Sequential deep learning for human action
recognition. In Proceedings of the Second International
Conference on Human Behavior Unterstanding, HBU’11,
pages 29–39, 2011. 2
[3] V. Badrinarayanan, A. Kendall, and R. Cipolla. Segnet: A
deep convolutional encoder-decoder architecture for image
segmentation. arXiv preprint arXiv: 1511.00561, 2015. 2
[4] G. Chéron, I. Laptev, and C. Schmid. P-CNN: Pose-based
CNN Features for Action Recognition. In ICCV, 2015. 2, 6
[5] N. Dalal and B. Triggs. Histograms of oriented gradients for
human detection. In Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern
Recognition (CVPR’05) - Volume 1 - Volume 01, CVPR ’05,
pages 886–893, Washington, DC, USA, 2005. IEEE Computer Society. 2
[6] A. Diba, A. M. Pazandeh, and L. V. Gool. Efficient twostream motion and appearance 3d cnns for video classification. CoRR, abs/1608.08851, 2016. 2
[7] A. Dosovitskiy, P. Fischer, E. Ilg, P. Hausser, C. Hazrba,
V. Golkov, P. v.d. Smagt, D. Cremers, and T. Brox. Flownet:
Learning optical flow with convolutional networks. In IEEE
International Conference on Computer Vision (ICCV), Dec
2015. 3
[8] C. Feichtenhofer, A. Pinz, and A. Zisserman. Convolutional
two-stream network fusion for video action recognition. In
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
9
IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2, 6, 7
G. Gkioxari and J. Malik. Finding action tubes. 2015. 6, 8
J. F. Hu, W. shi Zheng, J. Lai, and J. Zhang. Jointly learning heterogeneous features for rgb-d activity recognition. In
Computer Vision and Pattern Recognition (CVPR) (In press),
2015. 7
M. Jain, J. C. van Gemert, and C. G. M. Snoek. What do
15,000 object categories tell us about classifying and localizing actions? In 2015 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), pages 46–55, June 2015. 2
H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black.
Towards understanding action recognition. In International
Conf. on Computer Vision (ICCV), pages 3192–3199, 2013.
1, 3, 6
Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint
arXiv:1408.5093, 2014. 5
A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar,
and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 5
H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre.
HMDB: a large video database for human motion recognition. In Proceedings of the International Conference on
Computer Vision (ICCV), 2011. 5
I. Laptev. On space-time interest points. Int. J. Comput.
Vision, 64(2-3):107–123, Sept. 2005. 2
Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradientbased learning applied to document recognition. In Proceedings of the IEEE, pages 2278–2324, 1998. 2
Q. Li, Z. Qiu, T. Yao, T. Mei, Y. Rui, and J. Luo. Action
recognition by learning deep multi-granular spatio-temporal
video representation. In Proceedings of the 2016 ACM on International Conference on Multimedia Retrieval, ICMR ’16,
pages 159–166, New York, NY, USA, 2016. ACM. 2, 6
I. Lillo, J. C. Niebles, and A. Soto. A hierarchical pose-based
approach to complex action understanding using dictionaries
of actionlets and motion poselets. CoRR, abs/1606.04992,
2016. 2
F. Liu, C. Shen, G. Lin, and I. D. Reid. Learning depth from
single monocular images using deep convolutional neural
fields. IEEE Trans. Pattern Anal. Mach. Intell., 38(10):2024–
2039, 2016. 2
J. Liu, A. Shahroudy, D. Xu, e. B. Wang, Gang”, J. Matas,
N. Sebe, and M. Welling. Spatio-Temporal LSTM with Trust
Gates for 3D Human Action Recognition, pages 816–833.
Springer International Publishing, 2016. 7
W. Liu, A. Rabinovich, and A. C. Berg. Parsenet: Looking
wider to see better. arXiv preprint arXiv: 1506.04579, 2015.
2
J. Long, E. Shelhamer, and T. Darrell. Fully convolutional
networks for semantic segmentation. CVPR, Nov. 2015. 2
B. Mahasseni and S. Todorovic. Regularizing long short term
memory with 3d human-skeleton sequences for action recognition. In The IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), June 2016. 2
[25] E. Ohn-Bar and M. M. Trivedi. Joint angles similarities and
hog2 for action recognition. In Proceedings of the 2013
IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPRW ’13, pages 465–470, Washington,
DC, USA, 2013. IEEE Computer Society. 7
[26] K. Ohnishi, M. Hidaka, and T. Harada. Improved dense trajectory with cross streams. In Proceedings of the 2016 ACM
on Multimedia Conference, MM ’16, pages 257–261, New
York, NY, USA, 2016. ACM. 6
[27] G. L. Oliveira, W. Burgard, and T. Brox. Efficient deep models for monocular road segmentation. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS),
2016. 2, 3
[28] E. Park, X. Han, T. L. Berg, and A. C. Berg. Combining
multiple sources of knowledge in deep cnns for action recognition. In WACV, pages 1–8. IEEE Computer Society, 2016.
2, 6
[29] X. Peng and C. Schmid. Multi-region two-stream R-CNN
for action detection. In ECCV 2016 - European Conference
on Computer Vision, Amsterdam, Netherlands, Oct. 2016. 2,
6, 8
[30] O. Ronneberger, P.Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention
(MICCAI), volume 9351 of LNCS, pages 234–241. Springer,
2015. 2
[31] J. Sanchez, F. Perronnin, T. E. J. Mensink, and J. Verbeek.
Image classification with the fisher vector: Theory and practice. International Journal of Computer Vision, 2013. 2
[32] A. Shahroudy, J. Liu, T.-T. Ng, and G. Wang. Ntu rgb+d:
A large scale dataset for 3d human activity analysis. In The
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 6, 7
[33] K. Simonyan and A. Zisserman. Two-stream convolutional
networks for action recognition in videos. In Advances in
Neural Information Processing Systems, 2014. 1, 2, 3, 6
[34] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. ICLR, 2015. 3
[35] J. Sivic and A. Zisserman. Video google: A text retrieval
approach to object matching in videos. In Proceedings of the
Ninth IEEE International Conference on Computer Vision Volume 2, ICCV ’03, pages 1470–, Washington, DC, USA,
2003. IEEE Computer Society. 2
[36] k. Soomro, A. Roshan Zamir, and M. Shah. UCF101: A
dataset of 101 human actions classes from videos in the wild.
In CRCV-TR-12-01, 2012. 5
[37] D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri.
Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the 2015 IEEE International
Conference on Computer Vision (ICCV), pages 4489–4497,
2015. 1, 2, 4
[38] G. Varol, I. Laptev, and C. Schmid. Long-term temporal
convolutions for action recognition. CoRR, abs/1604.04494,
2016. 1, 6
[39] C. Wang, Y. Wang, and A. L. Yuille. An approach to posebased action recognition. In Proceedings of the 2013 IEEE
Conference on Computer Vision and Pattern Recognition,
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
10
CVPR ’13, pages 915–922, Washington, DC, USA, 2013.
IEEE Computer Society. 2
H. Wang and C. Schmid. Action recognition with improved
trajectories. In Proceedings of the 2013 IEEE International
Conference on Computer Vision, ICCV ’13, pages 3551–
3558, Washington, DC, USA, 2013. IEEE Computer Society.
2
L. Wang, Y. Qiao, and X. Tang. Action recognition with
trajectory-pooled deep-convolutional descriptors. In CVPR,
pages 4305–4314, 2015. 2
L. Wang, Y. Qiao, X. Tang, and L. Van Gool. Actionness estimation using hybrid fully convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pages 2708–2717, 2016. 2, 8
L. Wang, Y. Xiong, Z. Wang, Y. Qiao, D. Lin, X. Tang, and
L. Val Gool. Temporal segment networks: Towards good
practices for deep action recognition. In ECCV, 2016. 2, 6
P. Weinzaepfel, Z. Harchaoui, and C. Schmid. Learning to
track for spatio-temporal action localization. In ICCV 2015
- IEEE International Conference on Computer Vision, pages
3164–3172, Santiago, Chile, Dec. 2015. IEEE. 8
P. Weinzaepfel, X. Martin, and C. Schmid. Towards weaklysupervised action localization. CoRR, abs/1605.05197,
2016. 8
Z. Wu, Y. Jiang, X. Wang, H. Ye, X. Xue, and J. Wang.
Fusing multi-stream deep networks for video classification.
CoRR, abs/1509.06086, 2015. 2
G. Yu and J. Yuan. Fast action proposals for human action
detection and search. In 2015 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), pages 1302–1311,
June 2015. 8
C. Zach, T. Pock, and H. Bischof. A duality based approach
for realtime tv-l1 optical flow. In Proceedings of the 29th
DAGM Conference on Pattern Recognition, pages 214–223,
Berlin, Heidelberg, 2007. Springer-Verlag. 2
W. Zhu, J. Hu, G. Sun, X. Cao, and Y. Qiao. A key volume mining deep framework for action recognition. In 2016
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1991–1999, June 2016. 6
| 9cs.NE
|
Approximated Computation of Belief Functions for
Robust Design Optimization
Massimiliano Vasile1 and Edmondo MInisci2
University of Strathclyde, G1 1XJ , Glasgow, UK
Quirien Wijnands3
ESA/ESTEC, 2200 AG, Noordwijk, The Netherlands
This paper presents some ideas to reduce the computational cost of evidence-based
robust design optimization. Evidence Theory crystallizes both the aleatory and epistemic
uncertainties in the design parameters, providing two quantitative measures, Belief and
Plausibility, of the credibility of the computed value of the design budgets. The paper
proposes some techniques to compute an approximation of Belief and Plausibility at a cost
that is a fraction of the one required for an accurate calculation of the two values. Some
simple test cases will show how the proposed techniques scale with the dimension of the
problem. Finally a simple example of spacecraft system design is presented.
Nomenclature
SA
ta
taq
Pl
A
A
AL
AML
ANTemp
aPCU
B
Bel
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
solar array solar aspect angle, rad
access time, s
acquisition time, s
solar array specific mass, kg/m2
amplifier case mass fraction
antenna specific mass, kg/m2
solar cell efficiency
battery efficiency
PCU efficiency
generic threshold
focal element
Faraday rotation, rad
cumulative Plausibility function
generic proposition
archive
atmospheric losses, dB
antenna misalignment loss, dB
antenna noise temperature, K
PCU mass coefficient, kg/W
onboard data volume, bits
cumulative Belief function
Bl,i
Cmin
c
Dant
D
DOD
d
=
=
=
=
=
=
=
generic box in U
battery capacity, Wh
speed of light, m/s
antenna diameter, m
design space
Depth Of Discharge
design parameter vector
SA
CMR
A
cell
b
PCU
f
1
Reader, Mechanical & Aerospace Department, 75 Montrose Street.
Lecturer, Mechanical & Aerospace Department, 75 Montrose Street.
3
Senior staff, European Space Research & Technology Centre Postbus 2992200 AG.
1
American Institute of Aeronautics and Astronautics
2
Ed
e
FL
FSL
f,g,h
fT
GAMP
Gr
HG
Id
IL
Ld
P0
Pd
PL
PLd
Pe
PSA
Rt
RA
RaL
rGS
SLAT
T
TAMP
Tdata
U
u
Xe
Xd
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
battery energy density, Wh/kg
elevation error, rad
feeder loss, dB
free space loss, dB
generic functions
carrier frequency, MHz
amplifier gain, dB
ground station gain, dB
ground station altitude, m
inherent degradation
implementation loss, dB
time degradation
generated power per unit area, W/m2
power in daylight, W
polarization mismatch loss, dB
required transmission power, W
power in eclipse, W
total required power, W
data rate, dB
rain absorption
rain absorption loss, dB
distance from the ground station, km
horn lateral surface, m2
amplifier type
amplifier noise, K
transmitted data, bits
uncertain space
uncertain parameter vector
power system efficiency in eclipse
power system efficiency in daylight
I. Introduction
I
N recent times, Evidence Theory has been proposed in place of Probability Theory for robust design of
engineering systems. Authors like Oberkampf et al.1 demonstrated the potentiality of Evidence Theory to model
both epistemic and aleatory uncertainties in the design of engineering systems. Similar examples can be found in the
work of Agarwal et al.2 or in the work of Bae et al.3, Fetz et al.4 He et al.5 and Mourela et al. 6, mainly with
applications to structural design. Denoeux proposed a technique to compute an inner and outer approximation of
Belief and Plausibility functions7. Helton et al.8,9,10 proposed a number of techniques to reduce the dimensionality of
problems treated with Evidence-based models. More recently Vasile11 and Croissard et al.12 provided some
examples of application of Evidence Theory to the robust optimal design of space systems and space trajectories.
The uncertainties in the design parameters of the main spacecraft subsystems were modeled using Evidence Theory.
The design process was then formulated as an Optimization Under Uncertainties (OUU) and the Belief function was
optimized (maximized) together with all the other criteria that define the optimality of the system.
With Evidence Theory, also know as Dempster-Shafer’s theory13, both aleatory and epistemic uncertainties, coming
from a poor or incomplete knowledge of the design parameters, can be correctly modeled. The values of uncertain or
vague design parameters can be expressed by means of intervals with associated basic belief assignment or bpa.
Each expert participating in the design, assigns an interval and a bpa according to their experience. Ultimately, all
the pieces of information associated to each interval are fused together to yield two cumulative values, Belief and
Plausibility, that express the confidence range in the optimal design point. In particular, the value of Belief expresses
the lower limit on the probability that the selected design point remains optimal (and feasible) even under
uncertainties. The benefits coming from the use of Evidence Theory are considerable but the computation of Belief
and Plausibility requires running a number of optimizations that grows exponentially with the number of dimensions
and becomes intractable even for problems of moderate size.
2
American Institute of Aeronautics and Astronautics
This paper presents some ideas on how to reduce the computational cost to obtain an approximation of Belief
and Plausibility cumulative functions in space system engineering. Some of the techniques presented in this paper
are not problem dependent others exploit the partial decomposability of space system engineering design problems.
The paper starts with a brief introduction to Evidence Theory and its use in the context of robust design
optimization. It then presents some techniques to compute an optimal design solution under uncertainty when
Evidence Theory is used for uncertainty quantification. A few ideas are then proposed to reduce the computational
cost and their effectiveness is experimentally proven on some scalable analytic functions. The preliminary robust
design of an integrated power and telecommunication system of a satellite is then used to illustrate the application of
Evidence-based Robust Design Optimization to the design and margin quantification of space systems. A final
section introduces a problem decomposition technique that looks promising to solve large scale space system
engineering problems in polynomial time.
II EVIDENCE-BASED ROBUST DESIGN OPTIMISATION
Evidence Theory, developed by Shafer13, belongs to the class of imprecise probability theories conceived to
adequately treat both epistemic and aleatory uncertainty when no information of probability distributions is
available. The theory does not require additional assumptions when the available information is poor or incomplete
and provides a nice framework to incorporate multiple pieces of evidence in support to a statement. In most current
engineering design applications of Evidence Theory, domain experts are expected to express their belief on the value
of an uncertain parameter u being within a certain set of intervals. Each interval can be considered as an elementary
proposition, and all the intervals form the so-called frame of discernment , which is a set of mutually exclusive
elementary propositions. The frame of discernment can be viewed as the counterpart of the finite sample space in
probability theory. The power set of is U=2 or the set of all the subsets of (the uncertain space in the
following) The level of confidence an expert has in an element of U is quantified using the Basic Probability
Assignment (bpa) m() that satisfies the axioms:
m( ) 0, U 2 ;
m( ) 0, U 2 ;
m() 0;
(1)
m( ) 1
U
Note that the bpa is actually a belief in the values of rather than an actual probability. An element of U that has
a non-zero bpa is named a focal element . When more than one parameter is uncertain, the focal elements are the
result of the Cartesian product of all the intervals associated to each uncertain parameter. The bpa of a given focal
element is then the product of the bpa of each interval. All the pieces of evidence completely in support of a given
proposition form the cumulative belief function Bel while all the pieces of evidence partially in support of a given
proposition from the cumulative plausibility function Pl. In mathematical terms the two functions are defined as
follows:
Bel ( A)
m( )
i A
Pl ( A)
i A0
i
(2)
m(i )
where A is the proposition about which the Belief and Plausibility need to be evaluated. For example, the
proposition can be expressed as:
A u U | f (u)
(3)
where f is the system process and the threshold is the value of a design budget (e.g. the mass). Thus, focal
elements intercepting the set A but not included in A are considered in Pl but not in Bel. It is important to note that
the set A can be disconnected or present holes, likewise the focal elements can be disconnected or partially
overlapping.
A. Robust Design Formulation
3
American Institute of Aeronautics and Astronautics
The interest is in a general function f : D U mn
characterizing an engineering system to be
optimized, where D is here called the available design space and U the uncertain space. The function f represents the
model of the system budgets (e.g. power budget, mass budget, etc.), and depends on some uncertain parameters u
and design parameters d such that:
u U
n
; dD
m
(4)
A bpa is associated to the frame of discernment U of the uncertain parameters u. From the definition of Bel and
Pl and from Eq. (3) it is clear that the maximum and minimum of f over every focal element of U should be
computed and compared to . The threshold is the desired or expected value of the system budget. If the maximum
and minimum do not occur at one of the vertices of the focal element an optimization problem has to be solved for
every focal element and for each new design vector. Because the number of focal elements increases exponentially
with the number of uncertain parameters and associated intervals so does the number of optimization problems.
Furthermore, what designers are usually interested in are: a design solution d that optimizes performance (i.e. the
design budgets) and minimizes the impact of uncertainty, a quantification of the design margins on the system
budgets and a quantification of the reliability of the design solution. This information can be obtained for the worst
case scenario but that might lead to over conservative decisions. Therefore it is desirable to have also the variation
of the design margins and reliability with the threshold , i.e. with the expected value of the design budgets Indeed,
it may be relevant to take a little more risk (a slightly lower value of the belief) if the performance gain is
significant. Therefore, in practice, it would be desirable to have the trade-off curve, solution of the bi-objective
optimization problem:
max Bel f (d, u, )
dD uU
(5)
min
In previous works7,7, the bi-objective problem (5) was approached directly with a multi-objective evolutionary
optimizer working on the d and . The whole curve could be reconstructed with a population of agents converging
to the optimal pairs of values [Bel ]. However, the computational cost was driven by the identification of A and the
number of focal elements included in it. The assumption was that the maxima and minima of f were occurring only
at the corners of the focal element. The evaluation of the corners is in itself an operation that grows exponentially
with the number of dimensions and is, anyway, not applicable to a general case.
In this paper we propose a different way of approaching the problem. First of all, the computation of the Belief
function is performed by exploiting the following relationship:
(6)
Bel ( A) 1 Pl (A)
According to (2), the calculation of Pl (A) is computationally cheaper than the calculation of Bel(A). In fact,
any subset of U that contains at least one value (even a single sample) above the threshold contributes to Pl (A) .
The computation of Bel(A) instead requires that all the elements of A are below the threshold.
III COMPUTATIONAL APPROACH
Problem (5) would require the solution of a number of optimization problems that is exponentially increasing
with the number of focal elements. However, if one is interested only in the maximization of the Belief and in the f,
the exponential complexity can be avoided by solving the following two distinct problems over the Cartesian
product of the unit hypercube
U and D:
max min max f d, u
(7)
min min min f (d, u)
(8)
D
D
U
U
where U is the normalized collection of all the focal elements in U. In other words, all the focal elements in U are
normalized with respect to the maximum range of the uncertain parameters and collected into a compact unit
hypercube in which all the focal elements are adjacent and not overlapping. A point in the unit hypercube
mapped into the normal space U through the simple affine transformation:
xU ,i
b
b
u
U ,i
bUl ,i
u
U ,i
b
l
U ,i
xU ,i b
l
U ,i
b
b
u
U ,i
bUl ,i
u
U ,i
b
l
U ,i
bUl ,i
4
American Institute of Aeronautics and Astronautics
U is then
(9)
u
l
u
where bU ,i and bU ,i are the upper and lower boundaries of the i-th hypercube to which xU ,i belongs and bU ,i and
bUl ,i are the upper and lower boundaries of the i-th hypercube to which xU ,i belongs. The transformation is
relatively fast as it requires to scan only over the number of intervals per coordinate and not over the focal elements.
The computational complexity of the affine transformation is, therefore, linear with the number of dimensions. The
advantage is that each point within U belongs to at least one focal element, therefore by sampling U one is
guaranteed to sample only the focal elements and not other parts of U.
Problem (7) looks for the minimum possible threshold value max such that the entire unit hypercube is
admissible, hence the Belief is 1. The solution of problem (7) does not require the exploration or even the generation
of the focal elements and sets an upper limit on the value of the cost function. Problem (8) looks for the minimum
threshold value min above which the Plausibility is different from 0. As for problem (7), problem (8) does not
require the knowledge of the focal elements and sets a lower limit on the value of the cost function. Below that limit
the design is not feasible, given the current model and evidence on the design parameters.
The min/max problem is solved with a nested evolutionary process: an outer loop minimizes f over D and the
inner loop maximizes f over U . For each di vector an evolutionary process over U is run and the u vector with
maximum f is associated to di. The outer loop then proceeds till a maximum number of function evaluations is
reached. For the inner loop a Matlab implementation of Inflationary Differential Evolution (IDEA) is used in this
paper15. For the outer loop a modified version of IDEA, called IDEA, is used. The modification is mainly in the
way the objective function is computed. Due to the stochastic nature of the inner loop and the possible presence of
multiple maxima the outcome of each inner loop might not be the global maximum or not even near to it. In order to
increase the chances to produce optimal results, the local maximum ui ,max , with objective value f max di , ui ,max ,
computed for d i is compared against the local maximum u j ,max , associated to design vector d j . If
f max di , ui ,max f max di , u j ,max then u j ,max and its related maximum are associated to d i . The underlying
assumption here is that the cost function is Lipschitz continuous. Furthermore, if the location of the maxima in U
does not change with d a full optimization for each d is not required. If instead the location of the maxima is
changing with d, then for every di either a local search or a complete optimization is started. In both cases, running a
full optimization or a simple local search depends on the vector difference between di at step k and at step k+1 of the
evolution. The probability of running a full optimization is P pd dik 1 dik . The assumption here is that for small
variations of d there are small variations of the location of the local maxima. This assumption is generally verified in
the real-life applications the authors have encountered so far. A further level of verification of the global maximum
is introduced npop inner loop calls with probability pd by running a full optimization over U with two times the
number of functions evaluations. IDEA implements a memetic type of evolutionary process in which a local search
is started when the population collapses to a small region of the search space. The population is then restarted after
the local search is completed and all the local minima are collected in an archive (for more details on the general
algorithm implemented in IDEA the interested reader can refer to Ref. 15). In the inner loop, the local search is
performed with a Quasi-Newton method and with a convergent Nelder-Mead approach in the outer loop. When the
Nelder-Mead algorithm calls the inner loop no global search is run. At every restart of the population the archive is
examined and the local minima are compared. As for the population, even for the archive the local maximum ui ,max ,
with objective value f min/max di ,min , ui ,max , computed for di ,min is compared against the local maximum u j ,max ,
associated to local minimum design vector d j ,min . If f min/max di ,min , ui ,max f min/max di ,min , u j ,max then u j ,max and its
related maximum are associated to di ,min . When a given number of function evaluations for the inner and the outer
loop is exceeded, the search terminates. The best individual of the final population is added to the archive, the
archive is ranked and the best values is validated running a final global optimization over U Because, the
maximization and the minimization are based on stochastic processes a global convergence is not guaranteed unless
the optimizer is globally convergent. Nonetheless, by using an evolutionary process the computational complexity
remains polynomial. The consequence of an incorrect estimation of the global maximum and global minimum is an
overestimation of the point with minimum Plausibility and an underestimation of the point with maximum Belief.
However, on could argue that if a maximum is difficult to be found it correspond to a very unlikely event that
legitimately provide little support to the evidence of a given proposition. Evolutionary process and uncertainty
5
American Institute of Aeronautics and Astronautics
quantification are therefore closely entangled as the probability of sampling a particular uncertain value is directly
related to the probability of the realization of the event that corresponds to that uncertain value.
B. Evolutionary Binary Tree (EBT)
Given a design vector d, the value of Bel and Pl, for any value within [min max], can be computed by building
a binary tree in which a branch is pruned if the max of f associated to a leave is below (above) a given threshold .
The binary tree is built as follows. The transformed uncertain space is partitioned by cutting every box Bl ,i (with
B0,0 U ) in two halves along the longest edge. The cutting point is the boundary of the uncertain interval that is the
closest to the middle point of the edge. The tree is structured in levels with index l and at each level the tree has a
number of leaves each one identified by the index i. Now let Bll,i and Blr,i be respectively the left and right halves
deriving from the partition of the i-th box (or leave) Bl ,i at level l. If one box contains a maximum below the
threshold or a minimum above the threshold the box is removed from the tree, otherwise it is added to the list of the
boxes (leaves) that need to be partitioned at the following iteration (see Figure 1). If no maximum or minimum is yet
computed either for Bll,i or Blr,i then the box is added to the list of those that need to be explored. The boxes that
need to be explored and partitioned are said to be undecidable as a decision cannot be made on whether they
contribute to the Belief or not. The exploration of an undecidable box is performed by running a global
maximization and a global minimization of f(d,u) for a fixed d and for u Bll,i ( u Blr,i respectively). In this
implementation, IDEA is used for both the global maximization and minimization.
B00
B 0r
B0l
B11
B21
B12
B22
B23
B24
Figure 1. EBRO Binary Tree
Figure 1 shows a simple binary tree in which the right box deriving from splitting the initial uncertain space
contains a maximum below the threshold or a minimum above the threshold. The whole branch of the tree is then
discarded. The left box instead is undecidable and the branch that descends from that box needs to be explored. The
branching and exploration proceed until a decision can be made for all the boxes. The exploration of a branch
generates smaller and smaller boxes that eventually coincide with the focal elements. The exploration and branching
process generates a number of focal elements clustered in macro boxes and a number of boxes that correspond to
individual focal elements. All the discarded boxes with minimum above the threshold and all the boxes with
maximum above the threshold are used to compute Pl (A) . All the boxes with a minimum below the threshold are
used to compute Pl ( A) . The interesting aspect of this procedure is that Pl (A) and Pl ( A) can be computed even if
all the maxima and minima are not identified exactly. In fact for a box to be included in the calculation of Pl (A) it
is enough that a single value within the box, even not the actual maximum, is above the threshold. Likewise the
calculation of Pl ( A) requires that even s single value, not necessary the actual minimum, is below the threshold.
Note that the collection of all the boxes generated for a given threshold can be used to compute the Belief
(Plausibility) values in the interval [ max ]. In fact, the min and max values computed for each box represent
additional values within [ max ]. However, while the value of Bel and Pl can be exactly computed for with the
selected set of boxes, any other value within [ max ] might result to be an underestimation of Bel and
overestimation of Pl because of a lack of resolution (i.e. each box includes an excessive number of focal elements).
6
American Institute of Aeronautics and Astronautics
Therefore, if the whole curve is required, a refinement process is run by iteratively applying the evolutionary binary
tree to the boxes that contain a given [ max ] . This refinement process provides an exact value for each but
can lead to a number of optimizations that is equal to two times the number of focal elements, if the entire curve is
max min , the boxes generated
required. It is, however, interesting to note that if one takes an arbitrary
min
2
by the EBT can provide an approximation to the Bel and Pl curves that tend to be good in a neighborhood of and
close to max. This observation allows for the generation of a good estimation of the two curves at a fraction of the
computational cost. Further to this approximation, two specific mechanisms have been devised to reduce the
computational complexity and compute approximated Bel and Pl curves: the integrated focal element filtering and
the approximated min and max evaluation.
1. Integrated Focal Element filtering
The bpa associated to each focal element decreases in magnitude as the number of dimensions increases. This
observation suggests that an approximation to the value of Bel and Pl can be computed by using only a selected
subset of focal elements. The reduction in Bel or increase in Pl due to this approximation can be quantified by
looking at the cumulative value of the discarded focal elements. Therefore, during the generation of the binary tree
boxes with a bpa below a given filter threshold are discarded until the cumulative bpa associated to those boxes is
below the required filter accuracy
2. Approximated Min and Max Evaluation
Because the calculation of Pl (A) and Pl ( A) does not always require the exam maximum and minimum, one
can consider using an approximation of the min and max values for each subdomain Bl,i to make a decision. The
evolutionary search for a maximum and a minimum proceed through an optimal sampling of the uncertain space. A
selected subset of all the samples taken during the global exploration can be saved into an archive Au , when
maximizing, and into an archive into Al when minimizing, Then when two new subdomains Bll,i and Blr,i are
generated by bisection of Bl,i, one can take the maximum element xlsup in Au Bll,i (respectively Au Blr,i )and the
minimum element xlinf in Al Bll,i (respectively Al Blr,i ) instead of running a full optimization for every new
box.
B00
B0r
B0l
B11
B21
B12
B22
B23
B24
Figure 2. Binary Tree with Approximated Min-Max Evaluation.
Figure 2 shows a modified version of the binary tree in which decisions are made using the values in the archive
instead of the actual maxima and minima. This technique produces a substantial reduction in the number of
optimizations loops required to build the Belief and Plausibility curves but can introduce a significant error
r
r
depending on the archiving procedure. In order to preserve accuracy the values xlinf , xlsup , xinf
and xsup
are trusted
to be representative of the minimum and maximum values contained in the box with a probability ptrust inversely
proportional to the bpa associated to the box, i.e. the higher the bpa the lower the trust in the values xlinf and xlsup
r
r
(respectively xinf
and xsup
). In this way boxes with high bpa are properly explored with high probability. In
7
American Institute of Aeronautics and Astronautics
r
particular, xlsup and xsup
are trusted if ptrust>(c+bpa), where c is a trust factor. The use of a this trust factor ensure
good accuracy for a single value of . During the refinement process, however, the EBT makes use of existing boxes
but with different thresholds. Boxes that are correctly decided, on the basis of suboptimal archive values, for one
threshold might not be correctly decided for a different threshold. When a new threshold is considered, the boxes
generated for previous thresholds for which maximum and minimum was not exactly identified, are ranked from the
one with highest bpa value to the one with the lowest. A maximization is then progressively run on each of them,
starting from the one with highest bpa, until the cumulative bpa of the remaining ones is above c.
IV TEST CASES
The effectiveness of the techniques to reduce the computational cost has been put to the test on a set of simple
but representative, scalable problems (see Table 1). All the problems present a number of maxima and minima that
grows with the number of dimensions. Problem MV1 for example has a number of maxima in U that grows as 2n.
Problem MV2 has maxima that change location with d while MV8 is multimodal and has the maxima that change
with d. Table 2 represents the bpa structure for all the problems, with the intervals for each uncertain parameter and
the associated bpa. Tests are run considering two possible different bpa structures: with a uniform distribution of
bpa (labeled as EQ) and one with a non-uniform distribution. The tests in this paper are limited to three
disconnected uncertain intervals per parameter but the results can be generalized to a higher number of intervals and
also to overlapping intervals.
Table 1. Standard Benchmark.
ID
MV1
Function
Parameters
d [1,5]n ;
n
f di ui2
u [5,3]n
i 1
MV2
n
f ui di
d [1,5]n ;
2
u [5,3]n
i 1
MV8
n
f 2 ui cos ui di
i 1
d [0,3]n ;
u [0, 2 ]n
Table 2. BBA structure for the standard benchmark
MV1, MV2
MV8
MV1EQ,MV2EQ
MV8EQ
Interval
bpa
Interval
bpa
Interval
bpa
Interval
bpa
[-5 -4]
0.1
[0 1]
0.1
[-5 -4]
0.33
[0 1]
0.33
[-3 0]
0.25
[2 4]
0.25
[-3 0]
0.33
[2 4]
0.33
[1 3]
0.65
[5 2]
0.65
[1 3]
0.34
[5 2]
0.34
Figure 3 presents the number of optimizations required to approximate the Bel and Pl curves for problem MV1
as a function of the number of dimensions, while Figure 4 presents the same result but for function MV2. Figure 5
presents the number of optimizations for problem MV8. Different trust factors and for a combination of the
integrated focal element filtering, with a filter accuracy of 0.1, and the approximated min and max evaluation. The
use of the approximated min and max evaluation, even with a trust factor of 0.99, leads to a reduction of the number
of optimizations down to 25-30% of the number of optimizations required to explore all the focal elements.
Reducing the trust factor to 0.90 leads to a further reduction and the combination with the integrated focal element
filtering improves the reduction by 5-10%. The reduction is more limited in the case of a uniform distribution of
bpa’s. Figure 6 shows the Pl and Bel for problem MV1 with n=6. This problem has a Bel curve with a steep drop at
a thresh value of about 50. The figure shows the approximated curves using the approximated min and max
evaluation and the combination approximated min and max evaluation+ integrated focal element filtering. The
approximated curves well represent the true cumulative Belief curve and correctly identify the steep drop with an
error of maximum 0.1 in the Belief values. Plausibility is underestimated but the approximation technique was
geared towards an accurate estimation of the Belief. Therefore, an underestimation of the Plausibility was expected.
The reduction in computational cost however, is substantial reaching over 80% (see Bel90 with filter 0.1, that
corresponds to trust factor of 0.90 and a filter accuracy of 0.1). Figure 7 shows the equivalent result for problem
8
American Institute of Aeronautics and Astronautics
MV8. In this case the error in Belief for a trust factor of 0.90 is larger than 0.1 but for a very small difference in .
The overall error in the decision of the reliability the estimation of a design margin would be contained but with a
massive reduction in computational cost, up to 85%. It is also interesting to note that for all test cases the solution of
the min/max problem returned the global maximum and an optimal d. IDEA was set with a population of 10
individuals for the inner loop and 500n/2 number of function evaluations. The population was restarted when the
maximum distance among individuals was reaching 10% of the maximum distance experienced during the whole
search. IDEA was set with a population of 10 individuals and 5000n/2 function evaluations. The population was
restarted when the maximum distance among individuals was reaching 10% of the maximum distance experienced
during the whole search, and the probability of running the global search in the inner loop was pd=0.5.
Figure 3. MV1: complexity reduction
a)
b)
Figure 4 MV2: complexity reduction a) non uniform bpa structure b) uniform bpa structure
a)
b)
Figure 5 MV8: complexity reduction a) non uniform bpa structure b) uniform bpa structure
9
American Institute of Aeronautics and Astronautics
Figure 6 MV1: Bel and Pl curves at different approximation levels
Figure 7 MV8: Bel and Pl curves at different approximation levels
C. POWER-TELECOM INTEGRATED DESIGN PROBLEM
The techniques proposed in this paper were applied to the solution of a realistic case in which an integrated space
system made of a power generation unit and telecom subsystem need to be designed under uncertainty. This section
describes the power and telecom models used in this paper. The tests in this section aim to show how to use
Evidence-Based Robust Optimization can provide a more precise quantification of the design margins, compared to
a more traditional approach using rule of the thumb margins. The models in this section are derived from Ref.
16,17,18 and 19.
3. Power System Model
The power system (POW) model consists of a solar arrays and a battery. Starting from the required power in
daylight and eclipse, the total required power is computed as:
PT
Pd Td
e e
X + X
(10)
Psa = e d
Td
where Pe is the power consumption during eclipse, Te is the orbital eclipse time, Xe is the energy transfer efficiency
during eclipse, Pd is the power consumption during daylight, Td is the orbital daylight time, Xd is the energy transfer
efficiency during daylight. The generated power at Beginning of Life (BOL)is:
Po = cellGS
(11)
PBOL = Po I d cos SA
(12)
where Po is the ideal power output per unit area of the solar arrays, cell the solar cell efficiency, GS is the solar flux,
Id is the inherent degradation and SA is the worst case angle of incidence of the Sun light. In order for the model to
10
American Institute of Aeronautics and Astronautics
calculate the End of Life (EOL) power output per unit area, a solar array degradation over satellite lifetime factor Ld
is calculated as follows:
Ld = 1 cell
Life
(13)
where cell is the array degradation per year, Life is the expected satellite lifetime. Once the satellite lifetime factor Ld
is computed, the power output during EOL, PEOL, can be calculated, based on the power output per unit area PBOL, as
follows:
(14)
PEOL = PBOL Ld
Then the required solar array area Asa can be easily calculated as:
P
(15)
Asa = sa
PEOL
The solar array mass Msa is then derived from the solar array area Asa as follows:
(16)
M sa = Asa sa
where SA is the specific mass of the panel. The cell efficiency cell defines the type of solar cell that will be used
including its intrinsic characteristics. For every value of cell a database of cells, see Table 3, is used to obtain the
rest of the cell characteristics.
Table 3. Solar cell intrinsic characteristics.
CdTe p c-Si u c-Si 3j GaAs
Conc. Multijunc. cells
3j GaAs
ηcell 0.165 0.203 0.25
0.30
0.38
0.41
cell
0.05
0.05
0.05
1
0.037 0.037
The PCU power output Ppcu is calculated as follows:
Ppcu = Psa / η pcu
(17)
where ηpcu is the PCU efficiency. Finally the PCU mass Mpcu is calculated as a fraction of the PCU power output:
(18)
M pcu = a pcu Ppcu
where apcu is a PCU mass coefficient. The battery mass Mbat_pack, is computed starting from the energy density Ed,
which defines the particular battery chemistry to be used (see Table 4). The efficiency depends on the type of battery
and therefore on Ed. The efficiency ηbatt is computed by linearly interpolating the data in Table 4. Furthermore,
using a simple linear relationship in logarithmic scale, the depth of discharge DOD is calculated as a function of
parameter q in Table 4 and the number of cycles Ncycles. The number of cycles is derived from the orbital
characteristics and a fixed input in this analysis.
Table 4. Battery intrinsic characteristics
NiCd NiH2
Ed (Wh/kg)
60
75
ηbatt (%)
85
86
q
145.8 176.3
The minimum required battery capacity Cmin can then be calculated as follows:
PT
e e
Cmin =
DODb
and the mass of the battery cells Mb is calculated as:
Mb =
Cmin
Ed
11
American Institute of Aeronautics and Astronautics
(19)
(20)
4. Telecom System Model
The mass and power of the telecom system (TTC) are computed starting from the link budget. The required
communication link characteristics are the Bit Error Rate BER, the modulation, and ground station antenna gain Gr.
From the BER and modulation, one can compute the required energy per bit to noise ratio EbNo. The EbNo plus the
data rate are used to compute the Carrier to Noise Ratio CNratio . The total amount of data to be transmitted is
assumed to be Tdata = 103 B where B is the total amount of data coming from the C&DH (Command & Data
Handling) system to telecom. Given the access time ta the required data rate Rt is calculated as follows:
Tdata
Rt = 10 log10
t – t
aq
a
where taq is the target acquisition time. Given the data rate and the bit to noise ratio, CNratio is simply:
CN ratio Eb N 0 Rt
With the Carrier to Noise Ratio one can compute the Equivalent Isotropic Radiated Power (EIRP) as follows:
EIRP CN ratio G / T LTOTAL k
(21)
(22)
(23)
where k = 228.6 dB, LTOTAL is the total signal loss and G/T is receiving system performance. The total signal loss is
computed adding up all the factors that lead to a loss of signal energy and an increase of the noise. Here most of
these losses or sources of noise have been modeled with simple equations or look-up tables. The free space losses
FSL are calculated from the distance from the ground station rGS as well as the frequency of the transmitter fT:
(24)
FSL 32.4 20 log10 rGS 20 log10 fT
The polarization mismatch (Ionospheric) losses PL can be computed from the Faraday rotation f using the
following relationship:
(25)
PL = – 20 log10 cos f
The atmospheric losses AL are a function of the ground station altitude HG, are collected in a look-up table (as in
Table 5) and interpolated. The dependency of the atmospheric losses on the elevation angle is modeled by
introducing a simple sinusoidal function of the elevation angle e:
A
(26)
ALH = L
sin e
Table 5. Atmospheric losses' change with ground station altitude
HG (km)
-2 to 2
2.1 to 6
6.1 to 10
10.1 to 14
14.1 to 18
AL (dB)
0.04
0.025
0.008
0.004
0.001
The Rain absorption losses RaL are then calculated by using the data in Ref 16 and 18. The worst case losses for
the Feeder loss FL, the Antenna misalignment loss AML and the implementation loss IL are reported in Table 6.
Table 6. Worst case losses
FL [dB]
2
AML [dB]
0.5
IL [dB]
2
Summing up all the individual losses provides the total loss LTOTAL:
LTOTAL = FSL + FL + AM L + ALH + PL + RaL + I L
(27)
The system noise is computed from the antenna noise temperature ANtemp and from the cabling and receiver
losses. The total noise gives the noise figure RNfig:
RN fig AN temp TAMP
10
LA /10
– 1 ko 10LA /10 10F /10 – 1 ko
10GAMP /10
(28)
where TAMP is the amplifier noise, LA the cable loss, GAMP the low noise amplifier gain, F the receiver noise figure
and k0 = 290. The transmitter noise temperature Stemp is:
12
American Institute of Aeronautics and Astronautics
Stemp ATtempT
10LT /10 – 1 k 10LT /10 10FT /10 – 1 k
o
o
TeT
GT /10
10
(29)
Here ATtempT is the transmitter antenna noise temperature, TeT is the transmitter amplifier noise, LT is the transmitter
cable loss, GT is the transmitter low noise amplifier gain, FT is the transmitter noise figure. The rain noise Nrain is
then calculated as follows:
1
N rain = 1– RA/10 ko
10
(30)
where RA is the rain absorption. The total system noise TSnoise then writes:
TSnoise 10log10 RN fig Stemp N rain
(31)
The receiving system performance G/T is then calculated as follows:
G / T Gr – TSnoise
(32)
where Gr is the ground station receiver gain. The required transmission power Pld onboard the spacecraft is defined
as:
PLd EIRP – Gt
(33)
where Gt is the transmitter antenna gain. The spacecraft antenna type is chosen on the basis of the required antenna
gain Gt. It is well know that the best antenna for 5 dB ≤ gains ≤ 10 dB is the patch one, while the best for 10 dB <
gains ≤ 20 dB belongs to the horn type set, therefore the mass of the antenna is computed as follows. The antenna
characteristic length (it is the diameter of the normal conical section for conical horns, parabolas, and circular
patches, and an equivalent diameter for pyramidal horns and square/rectangular patches) is:
Gt
10
10
Dant
ANT
0.5
c
fT
(34)
where ANT is the antenna efficiency and c is the speed of light. If 5 ≤ Gt ≤ 10dB the mass of the patch is:
M ant , patch
2
Dant
4
0.0015
diel
0.0005 copper
(35)
where ρdiel = 2000 kg/m3 and ρcopper = 8940 kg/m3 are the averaged value of a dielectric material density and the
copper density, respectively, considering a 2 mm total thickness, with 1.5 mm of dielectric material and 0.5 mm
copper. If 10 dB < Gt ≤ 20 dB the lateral surface of the horn, SLAT, is computed as a conical surface:
SLAT
and the mass, Mant,horn, is:
Dant
2
2
Dant
L2horn
4
M ant ,horn SLAT A
(36)
(37)
where Lhorn is the length of the horn antenna can be assumed equal to 2Dant from available data, and ρA is the areal
density, which has a mean value of approximately 15 kg/m2 (from available data18). If the gain of the antenna is >
20 dB, the parabola antenna is selected, the diameter of the antenna is computed with Eq.(34), and the mass of the
antenna, Mant,par, is:
M ant , par
2
Dant
4
A
2
(38)
where the surface density has a typical value of 10 kg/m .
The mass of the amplifier Mamp is a function of PLd (see Ref. 17) as well as the mass of the case Mcase. An
identification parameter T [0, 1] is used to identify the type of amplifier such that for TWTA type, T = 0 and for
solid state type T = 1. Finally, the casing mass Mcase is computed as a fraction of the amplifier mass:
M case = M amp CMR
(39)
where CMR is the ratio between the mass of the case and the amplifier mass.
13
American Institute of Aeronautics and Astronautics
5. Test Results
The bpa structure and the design space for both the TTC and POW system are summarized in Table 7 and Table
8. The assumption for the integrated system is that the power demand for the TTC, PLd, is added to a fixed power
demand of 900W in daylight and 400W in eclipse. In these tests, it is assumed that the spacecraft spends half of the
time in eclipse and half in daylight with a maximum solar aspect angle of 15degrees.
Figure 8 and Figure 9 show the Bel and Pl curves for the TTC system and a comparison to the margin
quantification using a traditional margin approach. The cost function f is the system mass, i.e. the mass of TTC.
Note that some intervals are overlapping. This is an interesting feature of Evidence Theory that allows one to deal
with what can be considered as the degree of ignorance on the bpa assignment. The assumption is that the spacecraft
is operating at 1.5e6 km from the Earth and has an access time of 1000s to a ground station with a receiving antenna
with a gain of 60dB. The volume of data is 120000 bits. The lifetime of the mission is assumed to be 4 years. The
Faraday rotation is assumed to be 9 degrees, the gain of the ground station antenna 60dB and the BER is 1e-6. The
ground station is assumed to be at altitude 0m with the spacecraft at 30 degrees of elevation angle. The gain of the
amplifier is 60dB with cable losses of 8dB, a noise temperature of 400K, and a noise figure of 10. The transmitter
amplifier gain is assumed to be 20dB with noise temperature of 400K and noise figure of 10. Note that the
characteristics of the POW and TTC subsystems were not selected to reflect a real mission scenario but only to test
the proposed methodology. With these values, the difference between the optimal and robust solution is about 1kg
for the TTC.
ANT
CMR
Lt
Tant
Table 7. TTC bpa structure
Interval
bpa
Interval
bpa
Interval
bpa
Interval
bpa
[0.5 0.6]
0.2
[0.1 0.2]
0.5
[1 2]
0.2
[200 250]
0.1
[0.65 0.75]
0.5
[0.25 0.3]
0.35
[2 3]
0.3
[300 370]
0.6
[0.6 0.8]
0.2
[0.1 0.3]
0.15
[3 5]
0.5
[400 500]
0.3
[0.8 0.95]
0.1
Table 8. POW bpa structure
Xe
Xd
Id
PCU
Interval
Bba
Interval
Bba
Interval
Bba
Interval
Bba
[0.5 0.6]
0.1
[0.65 0.7]
0.2
[0.8 0.81]
0.7
[0.5 0.6]
0.1
[0.65 0.7]
0.6
[0.75 0.8]
0.6
[0.82 0.83]
0.2
[0.65 0.7]
0.6
[0.72 0.75]
0.3
[0.8 0.85]
0.2
[0.83 0.9]
0.1
[0.8 0.9]
0.3
Table 9. Design space for TTC and POW
Parameter
fT (MHz)
Mod
T
Gt (dB)
cell
sa (kg/m2)
aPCU (kg/W)
Ed (Wh/kg)
Low bound
7e3
0
0
5
0.1
1
0.01
60
Upper bound
11e3
1
1
20
0.3
2
0.02
100
The Bel margin curve in Figure 8 was generated assuming that a designer is taking the min/min solution (best
absolute performance) from problem (8) and adding a 25% margin to the required TTC power and to the mass of the
casing of the electronics. Then a system level margin is added to the total mass of the TTC. The system level margin
can range from 0% to 25% of the nominal mass of the TTC. For each mass plus system margin the value of Belief
and Plausibility was computed (see red and green thick solid lines). In Figure 9 a more conservative choice is made.
14
American Institute of Aeronautics and Astronautics
A 25% margin is added to the mass of antenna and amplifier and then a system level margin is added as before to
the total mass of the TTC. The two figures show that the margin approach either underestimates the Belief or
overestimates the margin. Figure 8, in fact, shows that the Belief of the mass corresponding to the maximum system
level margin is less than 60%. In the more conservative case, Figure 9, the Belief is 1 but the mass is overestimated
by about 0.5kg.
Figure 8. TTC system : margin approach vs. EBRO: best case margins.
Figure 9. TTC system: margin approach vs. EBRO: worst case margins.
Figure 10 shows the Bel and Pl curves computed with different c from 0.7 to 0.99. The total number of focal
elements is 108 corresponding to 216 optimizations to compute the exact Bel and Pl. With c= 0.99 one can obtain a
reduction of the computational cost down to 26% and with no relevant error in Belief. A c= 0.70 bring a reduction
down to 8% of the cost for an exact computation but with a limited error. In particular the error is very contained for
high values of Belief with a difference of about 0.1 kg in mass for the same Belief of less than 0.1 difference in
Belief for the same mass.
15
American Institute of Aeronautics and Astronautics
Figure 10. Approximated Bel and Pl curves for the Telecom system
Figure 11 shows a similar result for the integrated Power and Telecom system. The total number of focal elements in
this case is 8748 corresponding to 17496 optimizations for an exact calculation. The simple application of the
integrated filtering technique beings a moderate reduction of computational effort down to 80% of the cost of the
exact computation. The resulting Belief is underestimated by maximum 0.1. The application of the approximated
min and max evaluation with c= 0.90 brings to a more substantial reduction, down to 22% and with a moderate
overestimation that reduces almost to zero close to the left and right extremes of the Belief curve.
Figure 11. Approximated Bel and Pl curves for the integrated Power and Telecom system.
VI PROBLEM DECOMPOSITION
The interesting aspect of space engineering systems is that although the overall design requires the contribution
of all the subsystems, some subsystems are relatively decoupled and exchange information only through their
specific design budgets. For example, the telecom system and the power system exchange information only through
the output power from the telecom system that becomes an input parameter to the power system.
Let us consider a function g : D U with the following form:
Nf
g (d1 , u1 ,..., di , ui , u3i ,..., d N f , u N f , u3N f ) f1 (d1, u1,..., di , ui ,..., d N f , u N f ) fi (di , ui , u 3i ) (40)
i 2
Now assume that the functional dependency of function f1 on design and uncertain parameters di and ui is realized
through a function hi such that:
Nf
g (d1 , u1 ,..., di , ui , u3i ,..., d N f , u N f , u3 N f ) f1 (d1, u1,..., hi (di , ui ),..., hN f (d N f , u N f )) fi ( hi (di , ui ), u3i ) (41)
i 2
16
American Institute of Aeronautics and Astronautics
If hi could be handled as independent variable the two functions f1 and fi could be decoupled and Bel(g<) could
be expressed as:
Beld ( g )
Nf
1 A1
i Ai
m(1 )m(i )
i
A1 u1 U1 U , h | g
(42)
Ai hi , u3i U 3i U | g
If the value of the design parameters and of hi is fixed then the two functions are completely decoupled and the
computation of the Belief associated to their sum requires the independent computation of the maxima and minima
of f1 and fi over the subspaces U1 and U 3i . If the range of hi is well defined then one could compute the values of fi
for different values of hi by solving the following constrained problems:
max f i d, u3i , hi
(43)
u3 i
hi d, ui hi
If hi is fixed the computational complexity grows linearly Nf and the computation of the focal elements for each
function fi can be performed in parallel. If the vector ui contains a single uncertain parameter the result is an exact
representation of the Belief curve. If the vector ui contains multiple uncertain parameters then one can verify that the
belief function Beld is equal to the full belief curve Bel only for the values for which the value of hi is verified. A
good choice to fix the value of hi is to take the solution of the min-max problem (7). As an illustrative example
consider the following toy problem:
2
2
f 2 d12 d 22 u21
u22
2 u32
h u21 u22
(44)
f1 d 32 u1 h
g f1 f 2
Each uncertain parameter is defined over the three intervals [0 0.1], [0.2 0.4] and [0.5 1] with two different bpa
assignments: a) 0.3, 0.6 and 0.1 respectively and b) 0.3, 0.1 and 0.6 respectively. Figure 12 shows a comparison
between the exact Bel and Pl curves and the approximated ones. The approximated Bel curves are very close to the
exact ones and are almost identical for some values of The Pl approximation is instead very poor for low values
of and good for high values of The main reason is that only one value of h as used and it was the one
corresponding to the solution of the min/max problem. All minimum values within the focal elements of the
decomposed problem are therefore overestimated while the maxima are exact for a large number of focal elements at
some specific . The main advantage of this approximation becomes clear when one looks at the computational cost.
The exact computation requires 162 optimizations, while the approximated computation requires two parallel sets of
optimizations with 3 optimizations per set, thus 6 in total. The computational cost of the approximation is therefore
3.7% of the cost of the exact computation and would grow linearly with the number of dimensions.
a)
Figure 12. Decomposition approximation for bpa structure a) and bpa structure b)
17
American Institute of Aeronautics and Astronautics
b)
VII
FINAL REMARKS
The paper presented some strategies to obtain an estimation of Belief and Plausibility at a fraction of the
computational cost for their exact calculation. The approach presented in this paper provides the computation of the
optimal range of the design margin at a cost that is polynomial with the number of dimensions. An estimation of the
full Belief and Plausibility corves could be obtained with a cost reduction by over 80% but maintaining a contained
error. The effectiveness of the proposed strategies was proven on some benchmark problems, presenting a number
of minima and maxima exponentially increasing with the number of dimensions. Furthermore, two space system
design cases are used to show how evidence-based design optimization can improve the design of space systems
compared to a more traditional system margin approach. From these test it appeared that, even in the ideal case in
which an optimal deterministic design solution is available, a traditional margin approach tend to underestimate the
reliability of the design margin or to overestimate their value, given the available information. This justifies the use
of a rigorous margin quantification. Finally, a problem decomposition technique was proposed to reduce the
computational complexity of space system design problems in which all the components contributing to the overall
design budgets are only weakly coupled through a single function (the power in the case of space systems). For
these particular problems, it seems possible to obtain massive reductions of the computational cost but, more
importantly, a computational cost that increases linearly with the number of integrated systems. The results in this
paper, however, are only preliminary and a more in depth investigation is underway.
Acknowledgments
This work is partially supported through an ESA/ITI grant AO/1-5679/08/NL/CB
.
References
1
W. Oberkampf and J. C. Helton. Investigation of evidence theory for engineering applications. In 4th Non-Deterministic
Approaches Forum, volume 1569. AIAA, April 2002.
2
H. Agarwal, J. E. Renaud, and E. L. Preston. Trust region managed reliability based design optimization using evidence
theory. Collection of Technical Papers - AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials
Conference, 5:3449 – 3463, 2003.
3
H.-R. Bae, R. V. Grandhi, and R. A. Canfield. Uncertainty quantification of structural response using evidence theory.
Collection of Technical Papers -AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference,
pages 2135 – 2145, 2002.
4
T. Fetz, M. Oberguggenberger, and S. Pittschmann. Applications of plausibility and evidence thoery in civil engineering. In
1st International Symposium on Imprecise Probabilities and Their Appplications, Ghent, Belgium, 29 June- 2 July 1999.
5
L.-P. He and F.-Z. Qu. Possibility and evidence theory-based design optimization: an overview. Kybernetes.
6
Z. P. Mourela and J. Zho. A design 0ptimization method using evidence theory. Journal of Mechanical Design, 128:901 –
908, 2006.
7
T. Denoeux. Inner and outer approximation of belief structures using a hierarchical clustering approach. International Journal
of Uncertainty, Fuzziness and Knowlege-Based Systems, 9(4):437 – 460, 2001.
8
J. C. Helton, J. Johnson, W. L. Oberkampf, and C. Sallaberry. Sensitivity analysis in conjunction with evidence theory
representations of epistemic uncertainty.Reliability Engineering & System Safety, 91(10-11):1414 – 34, October 2006.
9
J. C. Helton, J. Johnson, W. L. Oberkampf, and C. Storlie. A sampling-based computational strategy for the representation of
epistemic uncertainty in model predictions with evidence theory. Computer Methods in Applied Mechanics and Engineering,
196(37-40 SPEC ISS):3980 – 3998, August 2007.
10
J. C. Helton, J. Johnson, C. Sallaberry, and C. Storlie. Survey of sampling based methods for uncertainty and sensitivity
analysis. Reliability Engineering & System Safety, 91(10-11):1175 – 209, October 2006.
11
M. Vasile. Robust mission design through evidence theory and multiagent collaborative search. Annals of the New York
Academy of Sciences, 1065:152–173, Dec. 2005.
12
N. Croisard, M. Vasile, S. Kemble, and G. Radice. Preliminary space mission design under uncertainty. Acta
Astronautica, 66:5 – 6, 2010.
13
G. Shafer. A Mathematical Theory of Evidence. Princeton University Press, 1976.
14
B. Tessem. Approximations for efficient computation in the theory of evidence. Artificial Intelligence.
15
M. Vasile, E. Minisci, and M. Locatelli. An inflationary differential evolution algorithm for space trajectory optimization.
IEEE Transactions on Evolutionary Computation, 2011.
16
Dennis Roddy Satellite Communications 3rd ed., McGraw-Hill, 2001.
17
J. Larson, James R. Wertz Space Mission Analysis and Design 3rd ed. , Microcosm Press & Kluwer Academic
Publishers.
18
Constantine A. Balanis Antenna Theory Analysis and Design 3rd ed. , Wiley-Interscience, 2005.
19
Mukund R. Patel. Spacecraft Power Systems, CRC Press, 2004.
18
American Institute of Aeronautics and Astronautics
| 5cs.CE
|
arXiv:1609.07182v2 [math.RT] 13 Jan 2017
An Enumeration of the Supercharacter Theories of
Cp × C2 × C2 for Prime p
Alexander Lang
April 5, 2018
Abstract
The supercharacter theories of Cp ×C2 ×C2 were classified in the language of Schur
rings by Evdokimov, Kovács, and Ponomarenko in [EKP16]. It was shown that every
nontrivial supercharacter theory of Cp ×C2 ×C2 can be constructed as a wedge product,
a direct product, or is generated by automorphisms. We use this classification to give
a precise count of the distinct supercharacter theories of Cp × C2 × C2 and describe
when a supercharacter theory can be constructed by more than one method. We also
present an alternative proof of the classification using the language of supercharacter
theories.
1
Introduction
A supercharacter theory for a finite group G is a pair of set partitions, one of G and one
of the set of irreducible characters Irr(G), satisfying a small set of conditions. They correspond to subalgebras of the group algebra Z(CG) satisfying some special properties. They
were first defined by Diaconis and Isaacs [DI08] and were used to assist in understanding
the groups UTn (q) of unimodular upper triangular matrices over the finite field Fq . Since
then, there has been research connecting supercharacter theories to Hopf algebras and
number theory by examining specific classes of supercharacter theories.
One natural question that arises is what are all the possible supercharacter theories for
a given group G? One important family of groups for which the answer to that question
is known is the cyclic groups as described in [LM98] and [LM96]. In this case there are
three general methods of constructing supercharacter theories which suffice to construct all
possible nontrivial supercharacter theories. The first method generates the supercharacter
theory as the orbits of the action of a subgroup of the automorphisms of G. The other
two methods build the supercharacter theory from the supercharacter theories of smaller
groups. First, if G can be expressed as a direct product of two subgroups, we can form the
direct product of any pair of supercharacter theories of the two subgroups of G. Second,
we can use what are known as wedge products. We will only describe a special case as
it is sufficient for our discussion, although the full generality is required for cyclic groups.
For any normal subgroup N of G, the wedge product combines any supercharacter theory
for N and any supercharacter theory for G/N to form a supercharacter theory for G. It
is possible that a supercharacter theory can be constructed by more than one of these
1
methods, or by the same method in different ways. Also for general Abelian groups, it
is known that these three methods are not sufficient. In particular there are nontrivial
supercharacter theories for p-groups which cannot be constructed using these methods.
It was shown in the proof of Theorem 1.5 in [EKP16] that these three methods are
also sufficient to construct every nontrivial supercharacter theory of Cp × C2 × C2 where
p is prime. Using this classification, it was proved that Cp × C2 × C2 is a Schur group
for any prime p. We use this classification to determine the total number of distinct
supercharacter theories of Cp × C2 × C2 for p odd, and further which supercharacter
theories can be constructed by each of the methods, and which ones can be constructed
in more than one way. Many of these supercharacter theories are isomorphic, however
we will consider them distinct if the partitions of Cp × C2 × C2 are distinct regardless of
whether they are isomorphic or not. In Sections 3 through 7, we present an alternative
proof of the classification using the language and techniques of supercharacter theories.
Our proof often uses more elementary and detailed methods than the one presented in
[EKP16], however it is also much longer. A key element of our argument is the utilization
of the properties of the sums of roots of unity which occur when the supercharacters are
evaluated on superclasses. These techniques may potentially allow the classification to
be extended in different directions, particularly to supercharacter theories of nonAbelian
groups. The proof presented in Theorem 1.5 in [EKP16] is recommended for readers
familiar with the language of Schur rings.
2
Preliminaries
We will use the following notation. We will denote the cyclic group of order n by Cn .
As we will discuss (C2 )3 separately, let p be an odd prime. All groups will be written
multiplicatively with identity e. If H1 and H2 are normal subgroups of G such that H1 ∩
H2 = {e} and hH1 , H2 i = G then G ∼
= H1 × H2 and we call H1 and H2 a complementary
pair. Given a group G we will let Irr(G) represent the set of irreducible characters of G
over C. We shall denote the character of the trivial representation by triv. For K ⊆ G
define
K (i) = {ki |k ∈ K},
X
b =
g ∈ CG.
K
(1)
(2)
g∈K
We begin with the definition of a supercharacter theory for a finite group based on the
original description given by Diaconis and Isaacs in [DI08]:
Definition 1. [DI08] Given a finite group G, a supercharacter theory for G is a pair (X , K)
where X is a partition of Irr(G) and K is a partition of the set of conjugacy classes of G
satisfying the following conditions:
1. {e} is an element of K, and {triv} is an element of X ,
2. |K| = |X |,
2
3. for all X ∈ X and all K ∈ K the class functions σX =
X
χ(e)χ satisfy σX (g) =
χ∈X
σX (h) for all g, h ∈ K.
The elements of K are called superclasses and the σX , for all X ∈ X are called
supercharacters. We shall denote the superclass containing the element g by [g]K or by
[g] when K is clear from context. We shall denote the X ∈ X containing the irreducible
character χ by [χ]X or [χ].
We will say that a supercharacter theory (X ′ , K′ ) is a refinement of the supercharacter
theory (X , K) if X ′ is a refinement of X as partitions or equivalently K′ is a refinement of
K.
Proposition 2.1. [DI08, Th. 2.2] If (X , K1 ) and (X , K2 ) are supercharacter theories for
G, then K1 = K2 . Similarly if (X1 , K) and (X2 , K) are supercharacter theories for G then
X1 = X2 .
Lemma 2.1. [Hen08, Lemma 6.1(a)] If (X , K) is a supercharacter theory for G and K ∈ K,
then the subgroup of G generated by K is a union of superclasses.
We will require the use of an alternative description of a supercharacter theory based
on the bijection between the set of supercharacter theories for a finite group G and the
set of Schur rings of G which are contained in Z(CG) [Hen10]. For more information, see
[Wie64].
Definition 2. The Hadamard product ◦ is defined on CG by
X
X
X
(ag bg )g.
bg g =
ag g ◦
g∈G
g∈G
(3)
g∈G
b When necesNote that (CG, ◦) is a commutative associative algebra, with identity G.
sary, we will denote the usual product on CG by ∗, to distinguish it from the Hadamard
product.
Proposition 2.2. [Hen10, Prop. 2.4] For a finite group G there is a bijective correspondence
between supercharacter theories (X , K) and C-linear subspaces A of Z(CG) containing e
b which are closed under the operations ∗ and ◦.
and G
Because of the above bijection we shall also refer to such algebras as supercharacter
theories. Given such an algebra A we shall denote the corresponding partitions of Irr(G)
and G by XA and KA respectively.
b ∈ A is equivalent to H being a union of superclasses.
Remark 2.1. Note that for H ⊆ G, H
Also for a supercharacter θ and an irreducible character χ the inner product hθ, χi =
6 0 is
equivalent to χ a summand of θ.
We recall the following, for more details see [DI08]. We note that if N is a normal
b ∈ A then CN ∩ A is a supercharsubgroup of G with a supercharacter theory A, and N
acter theory for N . We shall call such a theory the restriction of A to N , denoted by
3
A|N . We also observe that A|N is the supercharacter theory of N defined by the superclasses {K ∈ KA |K ⊂ N }. The supercharacter theory A = Z(CG) is called the minimal
supercharacter theory. The supercharacter theory with K = {{e}, G \ {e}} is called the
maximal supercharacter theory, and we will also refer to it as the trivial supercharacter
theory.
We recall the following methods of constructing supercharacter theories. For cyclic
groups, these three methods are sufficient to construct all nontrivial supercharacter theories. We note that it is sometimes possible to construct a given supercharacter theory
using more than one of the following constructions.
Proposition 2.3. [DI08] A subgroup H of Aut(G) acts on both the set of conjugacy classes
of G and Irr(G). Letting K be the set of orbits of the action on the conjugacy classes and
X be the orbits of the action on Irr(G) yields a supercharacter theory.
We will say that such a supercharacter theory (X , K) is generated by H, or generated
by automorphisms.
Proposition 2.4. [Hen10, Prop 8.1] Suppose that (XG , KG ) is a supercharacter theory for
G and (XH , KH ) is a supercharacter theory for H. Then there is a supercharacter theory
for G × H in which the superclasses are given by {K × L|K ∈ KG , L ∈ KH } and the
supercharacters are given by {φ × σ|φ ∈ XG , σ ∈ XH }.
We will often refer to a supercharacter theory constructed in this way as the direct
product of the supercharacter theories (XG , KG ) and (XH , KH ).
Proposition 2.5. [Hen10, Th. 4.2] Let N be a normal subgroup of G, and let π : G → G/N
be the natural quotient map. Suppose that (XN , KN ) is a supercharacter theory for N ,
and (XG/N , KG/N ) is a supercharacter theory for G/N . Then there is a supercharacter
theory for G in which the set of superclasses is
KN ∪ {π −1 (K)|K ∈ KG/N \ {{e}}}.
(4)
For ψ ∈ Irr(N ) let Irr(G|ψ) be the set of χ ∈ Irr(G) such that hχ|N , ψi > 0. The
corresponding partition of Irr(G) is given by
[
XG/N ∪ {
Irr(G|ψ)|X ∈ XN , X 6= {triv}}.
(5)
ψ∈X
We will call such a supercharacter theory the wedge product of the supercharacter
theory for N and the supercharacter theory of G/N , to agree with the conventions in
[LM98] and [LM96].
We note that the above construction is a special case of a more general method, see
[Hen10]. When classifying the supercharacter theories of cyclic groups the full generality
is necessary, but in our case this version will suffice.
3
Structure of the Main Argument
We are now ready to present the classification given in the proof of Theorem 1.5 in [EKP16]
in our current terminology:
4
Theorem 3.1. [EKP16] Let p be prime, G = Cp × C2 × C2 . Then every nontrivial supercharacter theory A of G is at least one of the following:
1. generated by automorphisms,
2. the direct product of supercharacter theories for a pair of complementary subgroups
H1 , H2 ≤ G,
3. the wedge product of supercharacter theories for H ≤ G and G/H.
We will now outline the structure of our version of a proof. The following theorem is
a key part of our argument.
Theorem 3.2. [Wie64, Th. 25.4] If G is an Abelian group of composite order and there
exists a prime p such that the p-Sylow subgroup of G is nontrivial and cyclic, then for
every nontrivial supercharacter theory A of G there exists a proper nontrivial subgroup
b ∈ A.
H such that H
It is clear that this theorem applies to Cp × C2 × C2 when p is an odd prime. Our
classification will be split into four cases based on what collection of proper nontrivial
subgroups are unions of superclasses, and by this theorem we know that this collection is
nonempty. Every proper nontrivial subgroup of Cp × C2 × C2 is isomorphic to C2 , C2 × C2 ,
Cp , or Cp × C2 . There are three isomorphic copies of C2 and Cp × C2 respectively. There
is one copy of C2 × C2 and one copy of Cp .
We see that there are a large number of different possibilities for which subgroups
are unions of superclasses. However, we can reduce the number of cases which must be
considered in the following way. Recall that if G is Abelian then Irr(G) ∼
= G as groups,
although the isomorphism is non-canonical. In [Hen09] the notion of a dual supercharacter
theory is introduced using the canonical isomorphism G ∼
= Irr(Irr(G)), and it is shown that
every supercharacter theory A of an Abelian group G yields a unique supercharacter theory
B of Irr(G) where KB = XA . B is not in general isomorphic to A. We will make frequent
use of the following lemma, for convenience we present a proof:
Lemma 3.1. [Hen09, Lemma 11.1] Let A be a supercharacter theory for G and let B be its
b ∈ A then there exists N ≤ Irr(G)
dual supercharacter theory for Irr(G). If H ≤ G and H
∼
c′ ∈ A
b
such that N = G/H and N ∈ B. Further this relation is inclusion reversing: if H
c′ ∈ B.
and H ≤ H ′ then there exists N ′ ≤ N such that N ′ ∼
= G/H ′ and N
Proof. Consider the dimension three supercharacter theory of G with superclasses {e}, H \
{e}, G \ H. The corresponding partition of Irr(G) is {triv}, N \ {triv}, and Irr(G) \ N
where N is the subgroup of Irr(G) satisfying
\
ker χ = H.
(6)
χ∈N
Then by considering the superclasses, we see that A is a refinement of this theory. Hence
N is a union of elements of XA as desired. It is clear that if H ≤ H ′ , then N ′ ≤ N , as
every χ which contains H ′ in its kernel also contains H in its kernel, so we are done.
5
We also recall the following part of Corollary 11.6 of [Hen09]:
Lemma 3.2. [Hen09, Cor. 11.6] A supercharacter theory A of G can be constructed as a
wedge product iff the dual of A can be constructed as a wedge product.
See [Hen09] for details on the dual supercharacter theory, and [EP14] for duality in
the Schur ring setting.
We note that G = Cp × C2 × C2 has the convenient property that H, H ′ ≤ G with
∼
H = H ′ implies G/H ∼
= G/H ′ . It is now clear that if we describe the collection of all
b ∈ A then we also
supercharacter theories A of G = Cp × C2 × C2 such that H ≤ G and H
b ∈ B where
have a description of the collection of all supercharacter theories B of G with N
N∼
= G/H by considering the dual supercharacter theory and using any fixed isomorphism
G → Irr(G). We first list all possible sets of proper nontrivial subgroups which can occur
as unions of superclasses, excluding those sets which contain a complementary pair of
subgroups as we will use separate arguments for them. It is a simple exercise to verify
that this list is complete:
1. C2
2. C2 × C2
3. one C2 and C2 × C2
4. all three C2 and C2 × C2
5. Cp
6. Cp × C2
7. Cp and Cp × C2
8. Cp and all three Cp × C2
9. C2 and the Cp × C2 containing it
10. C2 , Cp , and the Cp × C2 containing them
11. C2 , the Cp × C2 containing it, and C2 × C2 .
We now match these cases into pairs as in Lemma 3.1:
(i) Cp × C2 ↔ C2
(ii) Cp ↔ C2 × C2
(iii) Cp × C2 and Cp ↔ C2 × C2 and one C2
(iv) Cp and all three Cp × C2 ↔ C2 × C2 and all three C2
(v) C2 and the Cp × C2 which contains it ↔ C2 and the Cp × C2 which contains it
(vi) Cp , C2 , and the Cp ×C2 containing them ↔ C2 , the Cp ×C2 containing it, and C2 ×
C2 .
6
By the above argument, we only need to consider one of the cases in each corresponding
pair. We shall handle Cases (ii), (iii), (iv), and (vi) in one argument, by assuming that
cp ∈ A. The remaining two cases, (i) and (v), we shall argue by assuming that one copy
C
of Cp × C2 satisfies C\
p × C2 ∈ A, and that the C2 contained in it is the only other proper
b ∈ A.
nontrivial subgroup H which may satisfy H
We shall organize our argument as follows. (i) and (v) will be considered in Case 1.
cp ∈ A and C\
(ii), (iii), (iv), (vi) will be in Case 2 where we will assume C
/ A. The
2 × C2 ∈
c1 , H
c2 ∈ A will be
situation where there exists a complementary pair H1 and H2 with H
considered in Cases 3 and 4. We will conclude our proof in section 7, where we consider
the case when G = (C2 )3 .
We introduce some more notation. Let Cp = hωi and C2 × C2 = {e, a, b, c}. We fix
a primitive pth root of unity ρ. We define χ ∈ Irr(Cp × C2 × C2 ) to be the irreducible
character defined by χ(ω) = ρ, χ(a) = χ(b) = χ(c) = 1. We let ψa , ψb , and ψc be the
irreducible characters defined by ψx (ω) = 1, ψx (x) = 1, ψx (y) = −1 for all x, y ∈ {a, b, c}
with x, y distinct. We will let A denote a supercharacter theory of Cp × (C2 × C2 ). If
G = H1 × H2 then for K ⊂ G and h ∈ H2 we define M (K, h) = {g ∈ H1 |gh ∈ K}. Hence
[
M (K, h)h.
(7)
K=
h∈H2
See Example 9.1. Similarly if Irr(G) = H1 × H2 , and χ ∈ H2 , then for a supercharacter
θ = σX we will define M (θ, χ) = {ψ ∈ H1 |ψχ ∈ X}. Note that ψ ∈ M (θ, χ) is equivalent
to hθ, ψχi =
6 0. Finally, we note the following observation about ρ:
Remark 3.1. Since 1 + t + t2 + . . . + tp−1 is the minimal polynomial of ρ over Q we have
p−1
X
ρiℓ zi ∈ Z where zi ∈ Q, then for all i, j 6≡ 0 (mod p) zi = zj .
that if ℓ 6≡ 0 (mod p) and
i=1
4
Case 1
We begin with the case where C\
p × C2 ∈ A and the C2 subgroup contained in this copy
b ∈ A. This
of Cp × C2 is the only other nontrivial proper subgroup H which may have H
\
corresponds to (i) and (v) above. Without loss of generality we let hωi
× hai ∈ A and hai
b
the only other proper nontrivial subgroup H which may have H ∈ A. By Lemma 3.1 this
is equivalent to {ψa } ∈ XA and hχ, ψa i is the only other proper nontrivial subgroup of
Irr(G) which may be a union of elements of XA . {b} is not a superclass so there exists
\
x 6= b such that x ∈ [b]. Since hωi
× hai ∈ A x 6= ω ℓ , ω ℓ a for any ℓ. Therefore x = bω ℓ or
x = cω ℓ for some ℓ. Similarly bω k ∈ [c] or cω k ∈ [c] for some k. {b, c} a superclass implies
[
that hb,
ci ∈ A which contradicts our assumption. So we may choose ℓ, k 6≡ 0 (mod p). Let
θ 6= ψa , triv be a supercharacter. We will express Irr(G) as H1 × H2 where H1 = hψa , ψb i
and H2 = hχi. Hence
p−1
X
θ=
M\
(θ, χi )χi .
(8)
i=0
7
Since θ(b) = θ(x) we have:
p−1
X
M\
(θ, χi )(b) =
ρℓi M\
(θ, χi )(x)
(9)
i=0
i=0
p−1
X
p−1
X
M\
(θ, χi )(b)
i=0
!
− M\
(θ, χ0 )(x) =
p−1
X
ρℓi M\
(θ, χi )(x).
(10)
i=1
The LHS of the equation above is an integer, therefore by Remark 3.1 for all i, j ≡
6 0
(mod p):
M\
(θ, χi )(x) = M\
(θ, χj )(x)
(11)
!
p−1
X
M\
(θ, χi )(b) − M\
(θ, χ0 )(x) = −M\
(θ, χ)(x).
(12)
i=0
Case 1a:
We begin with the situation where bω ℓ ∈ [b] and bω k ∈ [c] for some ℓ, k 6≡ 0 (mod p),
and further cω u ∈
/ [b] for any u 6≡ 0 (mod p) and cω v ∈
/ [c] for any v 6≡ 0 (mod p). Note
that for all i, r
M\
(θ, χi )(b) = M\
(θ, χi )(bω r ).
(13)
Then we have:
p−1
X
M\
(θ, χi )(b)
i=0
p−1
X
!
− M\
(θ, triv)(b) = −M\
(θ, χ)(b)
M\
(θ, χi )(b) = −M\
(θ, χ)(b)
(14)
(15)
i=1
(p − 1)M\
(θ, χ)(b) = −M\
(θ, χ)(b)
(16)
M\
(θ, χ)(b) = 0.
(17)
M\
(θ, χi )(b) = 0.
(18)
ρri M\
(θ, χi )(b) = M \
(θ, triv)(b).
(19)
Hence for all i 6≡ 0 (mod p)
Then we conclude that for all r
r
θ(bω ) =
p−1
X
i=0
So θ is constant on {b, bω, . . . , bω p−1 }. For all r, ψa (bω r ) = −1, so ψa is also constant on
{b, bω, . . . , bω p−1 }. Therefore every supercharacter is constant on the set {b, bω, . . . , bω p−1 },
hence it is a subset of a superclass. Since bω ℓ ∈ [b] and bω k ∈ [c], {c, b, bω, . . . , bω p−1 } is a
subset of a superclass. By our Case 1a assumption, this must be a superclass. This would
imply
!2
p−1
p−1
p−1
X
X
X
i
i
ω i ∈ A.
(20)
aω + p
= e+2
bω
c+
i=0
i=0
8
i=0
c ∈ A, which contradicts our assumption for Case 1. Similarly,
Since p 6= 2, this implies hωi
we have a contradiction in the case of cω ℓ ∈ [b], cω k ∈ [c] for some ℓ, k 6≡ 0 (mod p),
bω u ∈
/ [b] for any u 6≡ 0 (mod p) and bω v ∈
/ [c] for any v 6≡ 0 (mod p).
Case 1b:
By the above, we can choose x, y ∈ {b, c} so that x 6= y, xω ℓ ∈ [b], ℓ 6≡ 0 (mod p) and
yω k ∈ [c], k 6≡ 0 (mod p).
Recall that for all i, j 6≡ 0 (mod p) M\
(θ, χi )(b) = M\
(θ, χj )(b) and M\
(θ, χi )(c) =
M\
(θ, χj )(c). Then Equation (12) becomes
(p − 1)M\
(θ, χ)(b) + M \
(θ, triv)(b) − M \
(θ, triv)(x) = −M\
(θ, χ)(x).
(21)
Similarly we have
(p − 1)M\
(θ, χ)(c) + M \
(θ, triv)(c) − M \
(θ, triv)(y) = −M\
(θ, χ)(y).
(22)
If x = b and y = c, then M\
(θ, χ)(b) = 0 and M\
(θ, χ)(c) = 0. If x = c and y = b then
(p − 1)M\
(θ, χ)(b) + M \
(θ, triv)(b) − M \
(θ, triv)(c) + M\
(θ, χ)(c) = 0,
(23)
(p − 1)M\
(θ, χ)(c) + M \
(θ, triv)(c) − M \
(θ, triv)(b) + M\
(θ, χ)(b) = 0.
(24)
Adding these equations gives
pM\
(θ, χ)(b) + pM\
(θ, χ)(c) = 0
(25)
M\
(θ, χ)(b) = −M\
(θ, χ)(c).
(26)
We want to show that we cannot have M \
(θ, triv) ∈ {ψb , ψc }. Without loss of generality
\
\
we assume that M (θ, triv) = ψb and M (ψa θ, triv) = ψc . Then Equations (23) and (24)
yield
(p − 1)M\
(θ, χ)(b) + M\
(θ, χ)(c) = −2,
(27)
(p − 1)M\
(θ, χ)(c) + M\
(θ, χ)(b) = 2.
(28)
Hence by Equation (26), (p−2)M\
(θ, χ)(c) = 2. Since M\
(θ, χ)(c) ∈ Z, we have p = 3. Then
\
\
\
M (θ, χ)(c) = 2 implies M (θ, χ) = triv +ψc . Then M (θ, χ)(b) = 0 which is a contradiction.
Hence we must have M \
(θ, triv) = ψb + ψc or M \
(θ, triv) = 0, so Equations (23) and
(24) yield
(p − 1)M\
(θ, χ)(b) + M\
(θ, χ)(c) = 0,
(29)
(p − 1)M\
(θ, χ)(c) + M\
(θ, χ)(b) = 0.
(30)
(p − 2)M\
(θ, χ)(b) − (p − 2)M\
(θ, χ)(c) = 0
(31)
Subtracting gives
9
M\
(θ, χ)(b) = M\
(θ, χ)(c).
(32)
Using Equation (26), we see that this implies M\
(θ, χ)(b) = M\
(θ, χ)(c) = 0. Further we see
\
\
\
that M (θ, triv)(b) = M (θ, triv)(c) = 0 as well, since M (θ, triv) = 0 or M \
(θ, triv) = ψb +ψc .
\
\
i
j
Recalling that for all i, j 6≡ 0 (mod p) M (θ, χ )(b) = M (θ, χ )(b) and M\
(θ, χi )(c) =
M\
(θ, χj )(c), we can now conclude that for any r
θ(ω r b) =
p−1
X
ρir M\
(θ, χi )(b) = 0,
(33)
ρir M\
(θ, χi )(c) = 0.
(34)
i=0
θ(ω r c) =
p−1
X
i=0
Since ψa (ω r b) = ψa (ω r c) = −1, we have that
{b, ωb, . . . , ω p−1 b, c, ωc, . . . , ω p−1 c}
(35)
\
is a subset of a superclass. Since hωi
× hai ∈ A, we see that the above must in fact be a
superclass. Therefore this supercharacter theory is a wedge product of a supercharacter
theory for Cp × C2 and a supercharacter theory for the quotient (Cp × C2 × C2 )/(Cp × C2 ).
Since the quotient group is isomorphic to C2 , there is only one choice of a supercharacter
theory for it. Hence this supercharacter theory is completely determined by the choice
of a supercharacter theory for Cp × C2 . This completes the classification in the case
\
hωi
× hai ∈ A and hai is the only other proper nontrivial subgroup H which may have
b
H ∈ A.
5
Case 2
cp ∈ A and C\
We now consider the case where C
/ A. By Lemma 3.1 this is equivalent
2 × C2 ∈
c
to ψa + ψb + ψc is a sum of supercharacters, and hχi is not a sum of supercharacters.
Case 2a:
c2 ∈ A, and without loss of generality we let it be
We begin by assuming there is a C
\
hai which implies by Lemma 3.1 that hψ
a , χi is a sum of supercharacters. Since hψa , χi ∩
{ψa , ψb , ψc } = {ψa }, ψa is a supercharacter. Since C\
/ A, we know that {b}, {c},
2 × C2 ∈
\
and {b, c} are not superclasses. Since hω, ai ∈ A, we have that xω ℓ ∈ [b] for some x ∈ {b, c}
and ℓ 6≡ 0 (mod p).
Let θ be a supercharacter which is a sum of irreducible characters in hχ, ψa i. Since ψa
is a supercharacter, M (θ, triv) = 0. Then
θ(b) = θ(xω ℓ ) =
p−1
X
ρiℓ M\
(θ, χi )(x).
(36)
i=1
Since θ(b) ∈ Z, we conclude by Remark 3.1 that for all i, j 6≡ 0 (mod p),
M\
(θ, χi )(x) = M\
(θ, χj )(x).
10
(37)
M (θ, χi ) = 0, triv, ψa , or triv +ψa , so M\
(θ, χi )(x) = 0, 1, or −1. If M\
(θ, χi )(x) = 1,
then M\
(θ, χi ) = triv. If M\
(θ, χi )(x) = −1, then M\
(θ, χi ) = ψa . If M\
(θ, χi )(x) = 0 then
M\
(θ, χi ) = 0 or M\
(θ, χi ) = triv +ψa . If M\
(θ, χ) = triv, then θ = χ + . . . + χp−1 , which
p−1
is a contradiction as χ + . . . + χ
a sum of supercharacters is equivalent to C2 × C2
\
a union of superclasses. If M (θ, χ) = ψa , then θ = ψa (χ + . . . + χp−1 ), so again we
have a contradiction. Therefore we conclude that for every i either M\
(θ, χi ) = 0 or
M\
(θ, χi ) = triv +ψa . Hence for all r = 0, . . . , p − 1 we have
θ(bω r ) = θ(cω r ) = 0.
(38)
Let θ̃ be a supercharacter which is a sum of irreducible characters each of which is
contained in {ψb χ, ψb χ2 , . . . , ψb χp−1 , ψc χ, ψc χ2 , . . . , ψc χp−1 }. M (θ̃, triv) = 0, and we have
similarly:
p−1
X
ρiℓ M\
(θ̃, χi )(x).
(39)
θ̃(b) = θ̃(xω ℓ ) =
i=1
θ̃(b) ∈ Z, so by Remark 3.1 for all i, j,
M\
(θ̃, χi )(x) = M\
(θ̃, χj )(x).
(40)
\
If θ̃ is a sum of irreducible characters contained in hχ, ψb i, then hχ,
ψb i is a sum of supercharacters which by Lemma 3.1 contradicts the assumption that {b} isn’t a superclass.
Similarly θ̃ is not a sum of characters contained in hχ, ψc i because {c} isn’t a superclass.
Therefore for all i either M\
(θ̃, χi ) = 0 or M\
(θ̃, χi ) = ψ + ψ . Hence for all r = 0, . . . , p − 1
b
c
we have
θ̃(bω r ) = θ̃(cω r ) = 0.
(41)
Recall that hωi a union of superclasses is equivalent to ψa + ψb + ψc is a sum of
supercharacters, and hω, ai a union of superclasses is equivalent to ψa a supercharacter.
Since neither hω, bi or hω, ci is a union of superclasses, neither ψb or ψc is a supercharacter.
Therefore ψa and ψb + ψc are both supercharacters.
We now are able to conclude that all supercharacters are constant on the set
{b, bω, . . . , bω p−1 , c, cω, . . . , cω p−1 }
(42)
and we conclude that it is a superclass. Since it is a hω, ai-coset, we conclude that A is a
wedge product as in Case 1.
Case 2b:
c2 ∈ A. We want to show that every
We now assume that there is no C2 with C
superclass disjoint from hωi is a union of hωi-cosets. Since C2 × C2 ∈
/ A, {a, b, c} is not a
superclass, and by assumption {a}, {b} and {c} are not superclasses either. If {a, b} is a
[
superclass, then ha,
bi ∈ A which contradicts {a, b, c} not a superclass. Similarly {a, c} and
{b, c} are not superclasses. Therefore we conclude that if x ∈ {a, b, c}, then there exists
11
y ∈ {a, b, c} and ℓ 6≡ 0 (mod p) such that yω ℓ ∈ [x]. Let θ 6= triv be a supercharacter
which is not a summand of ψa + ψb + ψc . Then
θ(x) = θ(yω ℓ )
p−1
X
M\
(θ, χi )(x) =
i=1
p−1
X
M\
(θ, χi )(x) =
i=0
p−1
X
ρℓi M\
(θ, χi )(y) =
i=0
(43)
p−1
X
ρℓi M\
(θ, χi )(y).
(44)
i=1
Since θ(x) ∈ Z, by Remark 3.1 we see that for all i 6≡ 0 (mod p)
θ(yω i ) = −M\
(θ, χ)(y).
(45)
In particular, note that this implies θ is constant on {yω, yω 2 , . . . , yω p−1 }. For a superclass
K let IK = {x ∈ C2 × C2 |∃ℓ such that xω ℓ ∈ K}. Given superclasses K, K ′ which are not
subsets of hωi, we claim that IK and IK ′ are either equal or disjoint. Suppose they are
not disjoint. Then we have
p−1
XX
c
b
xω i ∈ A
(46)
K hωi =
x∈K i=0
c we
c′ hωi
c′ must be a summand, this implies IK ′ ⊂ IK . However, by considering K
since K
see that IK ⊂ IK ′ which proves the claim.
There are three cases for the classes K disjoint from hωi:
1. |IK | = 1 for all K,
2. there exists z ∈ {a, b, c} such that for every K either IK = {z} or IK = {a, b, c}\{z},
3. IK = {a, b, c} for all K.
If |IK | = 1 for all K then there must exist i 6≡ 0 (mod p) such that aω i ∈ [a]. Then we have
Equation (45) for y = a, and hence {a, aω, . . . , aω p−1 } is a subset of a superclass. Similarly
we have that {b, bω, . . . , bω p−1 } and {c, cω, . . . , cω p−1 } are subsets of superclasses. Clearly
they are in fact superclasses.
For the case where IK = {z} or IK = {a, b, c} \ {z}, without loss of generality assume
z = a. Then as above there must exist i 6≡ 0 (mod p) such that aω i ∈ [a], and again we
have {a, aω, . . . , aω p−1 } is a superclass. There must exist β ∈ {b, c} and j 6≡ 0 (mod p)
such that βω j ∈ [b]. If β = b then by Equation (45) {b, bω, . . . , bω p−1 } ⊂ [b]. Therefore if
b ∈ IK , then K = [b] so we see that {b, bω, . . . , bω p−1 , c, cω, . . . , cω p−1 } is a superclass. If
β = c then cω j ∈ [b] so by Equation (45) {b, cω, . . . , cω p−1 } ⊂ [b]. There also exists t 6≡ 0
(mod p) such that bω t ∈ [c], so {c, bω, . . . , bω p−1 } ⊂ [c]. We are then able to combine
Equations (44) and (45) twice, once with x = b, y = c and once with x = c, y = b:
We then have
(p − 1)M\
(θ, χ)(b) = −M\
(θ, χ)(c),
(47)
(p − 1)M\
(θ, χ)(c) = −M\
(θ, χ)(b).
(48)
(θ, χ)(b) = −M\
(θ, χ)(b).
− (p − 1)2 M\
(49)
12
Therefore
M\
(θ, χ)(b) = M\
(θ, χ)(c) = 0.
(50)
We conclude that [b] = [c], or equivalently {b, bω, . . . , bω p−1 , c, cω, . . . , cω p−1 } is a superclass.
Finally, we consider when IK = {a, b, c} for all K. In this case we want to show that
G \ hωi = {a, aω, . . . , aω p−1 , b, bω, . . . , bω p−1 , c, cω, . . . , cω p−1 }
(51)
is a superclass. There exists x ∈ {a, b, c} and i 6≡ 0 (mod p) such that xω i ∈ [a]. If x = a
then by Equation (45) {a, aω, . . . , aω p−1 } ⊂ [a]. Since a ∈ IK implies K = [a] we must
have G \ hωi is a superclass. We now suppose that x 6= a and without loss of generality
suppose that x = b. Then by Equation (45) with y = b we have {a, bω, . . . , bω p−1 } ⊂ [a].
There exists j such that aω j ∈ [b]. If j ≡ 0 (mod p), then [a] = [b]. b ∈ IK implies K = [a]
so G\hωi is a superclass. If j 6≡ 0 (mod p), then by Equation (45) {b, aω, . . . , aω p−1 } ⊂ [b].
As above, we can combine 44 and 45 twice to conclude that [a] = [b]. Since a ∈ IK implies
K = [a] we again have G \ hωi is a superclass.
We now see that every superclass disjoint from hωi is a union of hωi-cosets. Therefore
we see that A is a wedge product of a supercharacter theory for Cp and a supercharacter
theory for (Cp × C2 × C2 )/Cp . This supercharacter theory is determined by a choice of
the supercharacter theory for Cp and a choice of supercharacter theory for Irr(C2 × C2 ).
6
Complementary Subgroups
We now consider the cases where there exist complementary subgroups which are unions of
superclasses. We shall first need a few lemmas about supercharacter theories of Cp = hωi.
Recall that ρ is a chosen primitive pth root of unity.
Lemma 6.1. [Hen08, Lemma 6.9] If p is prime, then every supercharacter theory of Cp is
generated by automorphisms.
From this we have the following lemma. We will also provide a direct proof.
Lemma 6.2. Let A be a supercharacter theory of Cp where p is prime. Then every element
of KA other than {e} has the same size, and this is equal to the size of every element of
XA other than {triv}.
Proof. Fix a supercharacter theory A of Cp . Let θ 6= triv be a supercharacter of A
with r irreducible characters as summands such that r is minimal. For k 6≡ 0 (mod p),
θ(ω k ) is a sum of r distinct roots of unity. Let χ be an irreducible character which
is a summand of θ. Because {ρ, ρ2 , . . . , ρp−1 } is linearly independent over Q we have
χ(ω k ) = χ(ω v ) iff ω k = ω v . Hence we see that r is the maximal number of elements a
superclass can contain. Let m be the number of superclasses different from {e}, which is
equal to the number of supercharacters different from triv. Then since r is the minimal
number of irreducible characters which are summands of a nontrivial supercharacter we
have rm ≤ p − 1. However, since r is the maximal number of elements in a superclass we
see that p − 1 ≤ rm. We conclude that rm = p − 1. Hence every superclass other than
{e} has r elements, and every supercharacter other than triv has r irreducible characters
as summands.
13
Remark 6.1. The size of every superclass other than {e} is the order of the group of
automorphisms which generates the supercharacter theory. Further since Aut(Cp ) ∼
= Cp−1
for p prime, and a cyclic group has at most one subgroup of a given order, we see that
the size of the superclasses determines the supercharacter theory. Also the dimension
determines the supercharacter theory.
Lemma 6.3. Let θ 6= triv be a supercharacter for a supercharacter theory A of Cp where
p is prime. If χ1 6= χ2 are summands of θ and x 6= e, then there exists y ∈ [x], y 6= x, such
that χ1 (x) = χ2 (y).
Proof. Suppose that θ is a sum of r irreducible characters. Since x 6= e, θ(x) is a sum of
r distinct roots of unity. By Lemma 6.2, | [x] | = r. Since χ2 (x) = χ2 (z) iff x = z, we see
that there are r distinct values that χ2 takes when evaluating elements of [x]. Hence there
must be a y ∈ [x] such that χ2 (y) = χ1 (x).
Lemma 6.4. Let ω ℓ , ω k ∈ Cp for p prime,
and
let θ 6= triv be a supercharacter for some
supercharacter theory A of Cp . If ω ℓ ∈
/ ω k then the set of roots of unity which appear
as summands of θ(ω ℓ ) are disjoint from those which appear as summands of θ(ω k ).
Proof. If either ω ℓ = e or ω k = e, then the result clearly holds. Suppose that ω ℓ 6= e
and ω k 6= e. Clearly the result also holds if θ is a single irreducible character, so suppose
it is not. Suppose that ρi is a summand of both θ(ω ℓ ) and θ(ω k ). Then there exist χ1 ,
χ2 ∈ Irr(Cp ), χ1 6= χ2 such that χ1 , χ2 are summands
of θ and χ1 (ω ℓ ) = χ2 (ω k ) = ρi .
′
′
However, by the previous lemma there exists a ω ℓ ∈ ω k such that χ1 (ω ℓ ) = ρi , which
is a contradiction.
c1 , H
c2 ∈ A for a complementary
Suppose that A is a supercharacter theory such that H
pair H1 , H2 of Cp × C2 × C2 . This may occur in two ways: H1 ∼
= Cp × C2 and H2 ∼
= C2
∼
∼
or H1 = Cp and H2 = C2 × C2 . We leave the proof of the following to the reader:
c1 , H
c2 ∈ A, then A is a refinement
Lemma 6.5. If H1 and H2 are a complementary pair and H
of the direct product theory of A|H1 and A|H2 .
If H1 ∼
= Cp × C2 and H2 ∼
= C2 , then we see that A must be the direct product
supercharacter theory as there is no supercharacter theory A′ which is a refinement of
A such that A′ |H2 = A|H2 . Since Cp × C2 is cyclic, all its supercharacter theories have
already been classified, see [LM98] and [LM96], and Lemma 8.2. Hence the case we
must consider is H1 ∼
= Cp and H2 ∼
= C2 × C2 . Similarly if A|H1 or A|H2 is the minimal
supercharacter theory, then A must be the direct product supercharacter theory. We shall
divide the remaining possibilities into two cases by considering the dimension of A|ha,bi .
By definition the dimension can’t be one, and if the dimension is four A|ha,bi is the minimal
supercharacter theory. Case 3 will be when the dimension is equal to three, and Case 4
will be when the dimension is equal to two.
6.1
Case 3
c1 , H
c2 ∈ A. By Lemma 3.1, this implies that
Suppose that H1 = hωi, H2 = {e, a, b, c} and H
c
ψa + ψb + ψc and hχi are sums of supercharacters. We consider the case where A|hωi is not
14
the minimal supercharacter theory for hωi, and the dimension of A|ha,bi is equal to three.
Without loss of generality we let {a} and {b, c} be superclasses, which implies by Lemma
\
3.1 that ψa and ψb + ψc are supercharacters. Note that this implies hχ,
ψa i is a sum of
supercharacters. Let TK be the set of superclasses which are disjoint from hω, ai and ha, bi
and let K ∈ TK . Similarly let TX be the set of supercharacters which are not a summand of
\
\
Irr(hχ,
ψa i) or Irr(hψ
a , ψb i), and let θ ∈ TX . We note that |TK | = |TX |. {M (θ, ψb )}θ∈TX is a
partition of Irr(hωi) \ {triv}. Since M (θ, ψb ) = M (ψa θ, ψc ) and θ ∈ TX iff ψa θ ∈ TX we see
that {M (θ, ψc )}θ∈TX is the same partition. Similarly {M (K, b)}K∈TK = {M (K, c)}K∈TK
are partitions of hωi \ {e}. We want to show that these partitions define a supercharacter
theory of Cp which is a refinement of A|hωi .
Since θ(b) = θ(c) we have
|M (θ, ψb )| = |M (θ, ψc )|.
(52)
If M (θ, ψb ) ∩ M (θ, ψc ) 6= ∅, then θ = ψa θ so M (θ, ψb ) = M (θ, ψc ). Hence M (θ, ψb ) and
M (θ, ψc ) are either disjoint or equal.
Suppose they are disjoint. By Lemma 6.5, θ is a summand of β(ψb + ψc ) for some
\
supercharacter β a summand of Irr(hωi).
θ(ψb + ψc ) = (M\
(θ, ψb ) + M\
(θ, ψc )) + (M\
(θ, ψb ) + M\
(θ, ψc ))ψa .
(53)
M\
(θ, ψb ) + M\
(θ, ψc ) is a summand of β, so M\
(θ, ψb ) + M\
(θ, ψc ) = β. If ω ℓ b ∈ ω k b for
ℓ 6≡ k (mod p) then
θ(ω ℓ b) = θ(ω k b)
(54)
M\
(θ, ψb )(ω ℓ ) − M\
(θ, ψc )(ω ℓ ) = M\
(θ, ψb )(ω k ) − M\
(θ, ψc )(ω k ).
(55)
Since M (θ, ψb ) ∩ M (θ, ψc ) = ∅ this implies
M\
(θ, ψb )(ω ℓ ) = M\
(θ, ψb )(ω k ),
M\
(θ, ψc )(ω ℓ ) = M\
(θ, ψc )(ω k ).
Similarly if ω ℓ c ∈ ω k c for ℓ 6≡ k (mod p) the above equations hold.
Since |TX | = |TK |, we see that
(56)
(57)
({M (θ, ψc )}θ∈TX , {M (K, c)}K∈TK ) = ({M (θ, ψb )}θ∈TX , {M (K, b)}K∈TK )
(58)
is a supercharacter theory of Cp . It is clearly a refinement of A|hωi since A is a refinement
of a direct product theory by Lemma 6.5. We further observe that for ω ℓ b ∈ ω k c
θ(ω ℓ b) = θ(ω k c)
(59)
M\
(θ, ψb )(ω ℓ ) − M\
(θ, ψc )(ω ℓ ) = −M\
(θ, ψb )(ω k ) + M\
(θ, ψc )(ω k )
(60)
M\
(θ, ψb )(ω ℓ ) = M\
(θ, ψc )(ω k ).
(61)
15
Similarly, for K ∈ TK we have that M (K, b) and M (K, c) are either equal or disjoint.
Suppose that there exists K such that M (K, b) = M (K, c). Let ω k b, ω k c ∈ K and θ ∈ TX .
Then
θ(ω k b) = θ(ω k c)
(62)
θ(ω k b) = M\
(θ, ψb )(ω k ) − M\
(θ, ψc )(ω k ) = −θ(ω k c)
(63)
θ(ω k b) = θ(ω k c) = 0.
(64)
So M\
(θ, ψb ) and M\
(θ, ψc ) have to take the same value on M (K, b) = M (K, c). For a
supercharacter theory of Cp where p is prime, it is impossible for distinct supercharacters
to take the same value on a superclass. Therefore either M (θ, ψb ) = M (θ, ψc ) and hence
A is a direct product theory, or M (K, b) 6= M (K, c) for every K ∈ TK . Since |TK | = |TX |,
the latter case implies M (θ, ψb ) 6= M (θ, ψc ) for every supercharacter θ ∈ TX .
This concludes our consideration of the conditions that every supercharacter theory of
the form of Case 3 must satisfy. We now consider the sufficient direction and show that
any partition that satisfies these conditions is indeed a supercharacter theory. Given a
set X we will denote the set of all subsets of X by P(X). We will consider the potential
supercharacter theory (X , K) of G = hω, a, bi where X ⊂ P(Irr(G)), K ⊂ P(G), |X | = |K|
and further the following are satisfied:
• (X ∩ P(Irr(hωi)), K ∩ P(hωi)) is a non-minimal supercharacter theory for Cp and we
will denote the corresponding algebra by B.
• {a}, {b, c} ∈ K and {ψa }, {ψb , ψc } ∈ X .
• (X ∩ P(Irr(hω, ai)), K ∩ P(hω, ai)) is the direct product supercharacter theory of B
and the unique supercharacter theory of hai.
If (X , K) is a direct product supercharacter theory we are done, so assume that it is not.
Suppose that
{M (K, b)|K ∈ K, K ⊂ hωi × {b, c}} = {M (K, c)|K ∈ K, K ⊂ hωi × {b, c}}
(65)
is a supercharacter theory B ′ of Cp which is a refinement of B satisfying the following
condition: for every superclass U 6= {e} of B ′ there is a superclass V of B ′ satisfying
(A1) |U | = |V |
(A2) U ∩ V = ∅
(A3) U ∪ V is a superclass of B
(A4) (U × {b}) ∪ (V × {c}) ∈ K
(A5) (U × {c}) ∪ (V × {b}) ∈ K.
16
Also, the analogous conditions hold for the supercharacters as well, with σ, σ ′ in the roles
of U, V . We also assume that for every such supercharacter of B ′ σ =
6 triv we have for all
x ∈ U, y ∈ V
σ(x) = σ ′ (y).
(66)
This is well defined since σ + σ ′ is a supercharacter for B, and hence is constant on U ∪ V .
Therefore Equation (66) is equivalent to σ(y) = σ ′ (x). We can view these assumptions as
stating that every superclass of B ′ is a superclass of B split into two halves which satisfy
Equation (66). Note that U and V correspond to M (K, b) and M (K, c) and similarly σ
and σ ′ correspond to M\
(θ, ψb ) and M\
(θ, ψc ). This concludes our listing of assumptions,
we shall now show that (X , K) is a supercharacter theory for G.
We observe that |X | = |K|, hence we only need to show that all potential supercharacters are constant on all potential superclasses. We begin with K ∈ TK and θ ∈ TX .
M\
(θ, ψb ) and M\
(θ, ψc ) are constant on M (K, b) and M (K, c) because they are supercharacters and superclasses for B ′ . By Equation (66) we have that θ is constant on all of K.
Since a supercharacter of B is constant on M (K, b) ∪ M (K, c) by assumption (A3), it is
constant on M (K, b) and M (K, c), so it is constant on K. Similarly for the supercharacters of Irr(hωi) × {ψa }. Clearly ψa and ψb + ψc are constant on K. Hence every potential
supercharacter from X is constant on K ∈ TK .
Now let K ∈ K be a superclass which is a subset of hωi. If ω ℓ ∈ K, then for θ ∈ TX we
have θ(ω ℓ ) = (M\
(θ, ψb ) + M\
(θ, ψc ))(ω ℓ ). Since M\
(θ, ψb ) + M\
(θ, ψc ) is a supercharacter for
B, we see that θ is constant on K. The supercharacters of Irr(hωi) and Irr(hωi) × {ψa } are
clearly constant on K. ψb + ψc takes the constant value 2 and ψa takes the constant value
1 on K. In a similar fashion, we see that all potential supercharacters in X are constant
on the superclasses which are contained in hωi × {a}.
Finally we consider the superclass {b, c}. For θ ∈ TX we have
θ(b) = |M (θ, ψb )| − |M (θ, ψc )| = 0,
(67)
θ(c) = −|M (θ, ψb )| + |M (θ, ψc )| = 0.
(68)
Clearly any supercharacter of Irr(hωi) or Irr(hωi)×{ψa } is constant on {b, c}. ψa takes the
value −1, and ψb + ψc takes the value 0 on {b, c}. Therefore every potential supercharacter
is constant on every potential superclass. We conclude that (X , K) defines a supercharacter
theory of G.
The above argument describes the supercharacter theories of Case 3. We will now
present an alternate proof of the sufficient direction above, as it is useful to see that every
supercharacter theory of Case 3 which is not a direct product can be generated by a group
of automorphisms. Let A be such a supercharacter theory, and let B be the supercharacter
theory A|hωi . Let B ′ be a refinement of B satisfying (A1) through (A5). By Lemma 6.1
every supercharacter theory of Cp can be generated by a subgroup of Aut(Cp ) ∼
= Cp−1 .
Since every subgroup of a cyclic group is cyclic, there exists φ0 ∈ Aut(hωi) such thathφ0 i
generates B. Similarly, there exists φ′ where hφ′ i generates the supercharacter theory B ′ .
By Lemma 6.2 we know that every superclass other than {e} in B ′ is the same size, and
by (A3) the size of every superclass other then {e} in B is double that size. Further, it is
clear that this size equals |hφ′ i|. Since a cyclic group has at most one subgroup of a given
17
order we have hφ′ i = hφ20 i. We now let φ ∈ Aut(G) be defined by φ(ω) = φ0 (ω), φ(a) = a,
φ(b) = b, and φ(c) = c.
Let ψ ∈ Aut(G) be such that ψ(ω) = ω, ψ(a) = a, ψ(b) = c, and ψ(c) = b. Let A′ be
the supercharacter theory generated by the group of automorphisms hφ ◦ ψi. We will show
that A = A′ . We see that {a} and {b, c} are superclasses for A′ , and A|hωi = A′ |hωi . Let
ℓ 6≡ 0 (mod p) and let K0 = ω ℓ b ∈ K(A′ ). Then M (K0 , b) = {ω ℓ , φ2 (ω ℓ ), φ4 (ω ℓ ), . . .}
and M (K0 , c) = {φ(ω ℓ ), φ3 (ω ℓ ), φ5 (ω ℓ ), . . .}. Hence we see that {M (K, b)}K∈K(A′ ) =
{M (K, b)}K∈K(A) , {M (K, c)}K∈K(A′ ) = {M (K, c)}K∈K(A) , and M (K0 , b) ∪ M (K0 , c) ∈
K(A|hωi ). Hence A and A′ have the same superclasses, so we conclude that A = A′ .
For a supercharacter theory generated by automorphisms as above, see Example 9.1.
6.2
Case 4
We now consider the final case for p odd. Again suppose that H1 = hωi and H2 =
c1 , H
c2 ∈ A. Suppose that A|hωi is not the minimal supercharacter theory,
{e, a, b, c}, and H
and that the dimension of A|ha,bi is equal to 2 which is equivalent to {a, b, c} is a superclass.
Let TK be the set of superclasses of A which are disjoint from H1 and H2 , and let TX be the
\1 ) or Irr(H
\2 ). Again |TK | = |TX |.
set of supercharacters which are not a summand of Irr(H
Let K ∈ TK and θ ∈ TX . Since {a, b, c} is a superclass, there is a constant r such that
b ◦ (a + b + c) = r(a + b + c) ∈ A.
((ω + ω 2 + . . . + ω p−1 )K)
(69)
Hence |M (K, a)| = |M (K, b)| = |M (K, c)| = r. Similarly |M (θ, ψa )| = |M (θ, ψb )| =
b equals:
|M (θ, ψc )|. Then (a + b + c)K
(M\
(K, a) + M\
(K, b) + M\
(K, c)) + (M\
(K, c)a + M\
(K, a)b + M\
(K, b)c)+
(M\
(K, b)a + M\
(K, c)b + M\
(K, a)c).
(70)
By Lemma 6.5 K is a subset of α × {a, b, c} for some superclass α of A|hωi , so we have for
some constant t, M\
(K, a) + M\
(K, b) + M\
(K, c) = tb
α. Then the remaining terms
W = (M\
(K, c) + M\
(K, b))a + (M\
(K, a) + M\
(K, c))b + (M\
(K, b) + M\
(K, a))c ∈ A. (71)
b is not a summand of W then M (K, a), M (K, b), and M (K, c) are pairwise disjoint.
If K
b is a summand. Without loss of generality, suppose ω ℓ ∈ M (K, a) ∩ M (K, b)
Suppose K
with ℓ 6≡ 0 (mod p). Then
θ(ω ℓ a) = θ(ω ℓ b)
(72)
and further by evaluating we see that
M\
(θ, ψa )(ω ℓ ) − M\
(θ, ψb )(ω ℓ ) − M\
(θ, ψc )(ω ℓ ) =
− M\
(θ, ψa )(ω ℓ ) + M\
(θ, ψb )(ω ℓ ) − M\
(θ, ψc )(ω ℓ )
(73)
M\
(θ, ψa )(ω ℓ ) − M\
(θ, ψb )(ω ℓ ) = −M\
(θ, ψa )(ω ℓ ) + M\
(θ, ψb )(ω ℓ )
(74)
which becomes
18
M\
(θ, ψa )(ω ℓ ) = M\
(θ, ψb )(ω ℓ ).
(75)
That implies M (θ, ψa ) = M (θ, ψb ). Since M\
(θ, ψa ) + M\
(θ, ψb ) + M\
(θ, ψc ) is a multiple
\
of a Cp supercharacter, we have M (θ, ψc ) = M (θ, ψb ) because M (θ, ψa ) + M\
(θ, ψb ) +
\
\
\
M (θ, ψc ) = 2M (θ, ψb ) + M (θ, ψc ) implies that some terms have multiplicity at least 2,
hence all terms must have multiplicity 3. Therefore every supercharacter θ ∈ TX satisfies
θ = M\
(θ, ψa )(ψa + ψb + ψc ) hence A is a direct product supercharacter theory. Therefore
for all K ∈ TK we have M (K, a) = M (K, b) = M (K, c).
Hence either M (K, a), M (K, b), and M (K, c) are all equal or all pairwise disjoint.
Further, if there is one superclass K ∈ TK with M (K, a) = M (K, b) = M (K, c), then
it must be true for all superclasses in TK . Also if M (K, a) = M (K, b) = M (K, c) then
b since they both have 6r terms. It is clear that if M (K, a), M (K, b), and M (K, c)
W = 2K
are pairwise disjoint for all K ∈ TK , then M (θ, ψa ), M (θ, ψb ), and M (θ, ψc ) are pairwise
disjoint for all θ ∈ TX .
Suppose that M (K, a), M (K, b), and M (K, c) are pairwise disjoint for all K ∈ TK .
Fix a K ∈ TK and let ω s , ω t ∈ M (K, a). Then
θ(ω s a) = θ(ω t a).
(76)
Evaluating gives
M\
(θ, ψa )(ω s ) − M\
(θ, ψb )(ω s ) − M\
(θ, ψc )(ω s ) =
M\
(θ, ψa )(ω t ) − M\
(θ, ψb )(ω t ) − M\
(θ, ψc )(ω t ).
(77)
Since M (θ, ψa ), M (θ, ψb ), and M (θ, ψc ) are pairwise disjoint we have
M\
(θ, ψa )(ω s ) = M\
(θ, ψa )(ω t ).
(78)
Therefore M\
(θ, ψa ) is constant on M (K, a). Similarly M\
(θ, ψb ) is constant on M (K, b)
\
and M (θ, ψc ) is constant on M (K, c). We conclude that the three pairs
({M (θ, ψa )}θ∈TX , {M (K, a)}K∈TK ) ,
(79)
({M (θ, ψb )}θ∈TX , {M (K, b)}K∈TK ) ,
(80)
({M (θ, ψc )}θ∈TX , {M (K, c)}K∈TK ) ,
(81)
are all supercharacter theories for Cp after adjoining {triv} and {e}. Using Lemma 6.1
and recalling Remark 6.1, we see that since they are all the same dimension and p is prime,
they are all the same supercharacter theory.
Suppose ω k ∈ M (K, a) and ω ℓ ∈ M (K, b), then
θ(ω k a) = θ(ω ℓ b).
(82)
Evaluating gives
M\
(θ, ψa )(ω k ) − M\
(θ, ψb )(ω k ) − M\
(θ, ψc )(ω k ) =
− M\
(θ, ψa )(ω ℓ ) + M\
(θ, ψb )(ω ℓ ) − M\
(θ, ψc )(ω ℓ ).
19
(83)
For any j and ψx and ψy where x, y ∈ {a, b, c} and x 6= y we observe that the set of
roots of unity which appear as summands of M\
(θ, ψx )(ω j ) and M\
(θ, ψy )(ω j ) are disjoint.
Therefore
M\
(θ, ψa )(ω k ) = M\
(θ, ψb )(ω ℓ ).
(84)
Hence for a fixed superclass K, ω k ∈ M (K, a), ω ℓ ∈ M (K, b), and ω r ∈ M (K, c), we have
by symmetry
M\
(θ, ψa )(ω k ) = M\
(θ, ψb )(ω ℓ ) = M\
(θ, ψc )(ω r ).
(85)
By Lemma 6.4, we have for ω k ∈ M (K, a) and ω ℓ ∈ M (K, b) that the set of roots of
unity which appear as summands of M\
(θ, ψc )(ω k ) is disjoint from those in M\
(θ, ψc )(ω ℓ ).
Returning to Equation (83), this implies
M\
(θ, ψb )(ω k ) = M\
(θ, ψc )(ω ℓ ),
(86)
M\
(θ, ψa )(ω ℓ ) = M\
(θ, ψc )(ω k ).
(87)
Then we proceed in a similar fashion to get
M\
(θ, ψa )(ω r ) = M\
(θ, ψb )(ω k ),
(88)
M\
(θ, ψc )(ω k ) = M\
(θ, ψb )(ω r ).
(89)
Note that the final set of equations we get by symmetry is redundant, as we may already
conclude
M\
(θ, ψb )(ω k ) = M\
(θ, ψc )(ω ℓ ) = M\
(θ, ψa )(ω r ),
(90)
M\
(θ, ψc )(ω k ) = M\
(θ, ψa )(ω ℓ ) = M\
(θ, ψb )(ω r ).
(91)
Suppose that the supercharacter theories A|hωi and ({M (K, a)}K∈TK , {M (θ, ψa )}θ∈TX )
are given. Then given a superclass α ∈ A|hωi , we have α = α1 ∪ α2 ∪ α3 where α1 , α2 ,
and α3 are superclasses in the supercharacter theory ({M (K, a)}K∈TK , {M (θ, ψa )}θ∈TX ).
Then there are two possibilities: either we have for A the three superclasses
{α1 a ∪ α2 b ∪ α3 c}, {α2 a ∪ α3 b ∪ α1 c}, {α3 a ∪ α1 b ∪ α2 c}
(92)
{α1 a ∪ α3 b ∪ α2 c}, {α2 a ∪ α1 b ∪ α3 c}, {α3 a ∪ α2 b ∪ α1 c}.
(93)
or
The situation is similar for θ ∈ TX . However, we do not have independent possibilities for
each K. The above equations tell us that given a single fixed superclass K ∈ TK , for any
θ ∈ TX the value of M\
(θ, ψa ) on M (K, a), M (K, b), and M (K, c) determines the values
\
\
of M (θ, ψb ) and M (θ, ψc ). Hence the value of a single K as either Equation (92) or (93)
determines all supercharacters in TX and hence the value as Equation (92) or (93) for
every other K ∈ TK is determined.
To understand the two possibilities, consider the algebra automorphism φ : CG → CG
defined by φ(ω) = ω, φ(a) = a, φ(b) = c, and φ(c) = b. We see that φ exchanges the
superclasses in Equations (92) or (93), so both yield supercharacter theories for G.
20
As in Case 3, we are now ready to consider the sufficient direction. Let B be a supercharacter theory for hωi and let B ′ be a refinement of B such that each superclass not
equal to {e} and each supercharacter not equal to triv of B is partitioned into 3 equal
sized superclasses and supercharacters respectively in B ′ . We partition G and Irr(G) so
that the partition of hωi and Irr(hωi) matches that of B, {a, b, c} is an element of the
partition of G, and {ψa , ψb , ψc } is an element of the partition of Irr(G). As before, let
TK be the subset of the partition of G disjoint from hωi and ha, bi, and let TX be the
\ or Irr(hψ
\
set of potential supercharacters which are not summands of Irr(hχi)
a , ψb i). We
assume that for all K ∈ TK , M (K, a), M (K, b), and M (K, c) are pairwise disjoint, and for
all θ ∈ TX , M (θ, ψa ), M (θ, ψb ), and M (θ, ψc ) are pairwise disjoint. Further, we assume
that each of ({M (θ, ψa )}θ∈TX , {M (K, a)}K∈TK ), ({M (θ, ψb )}θ∈TX , {M (K, b)}K∈TK ), and
({M (θ, ψc )}θ∈TX , {M (K, c)}K∈TK ), after adjoining {triv} and {e}, match the supercharacter theory B ′ . For all K ∈ TK and θ ∈ TX we assume that M (K, a)∪ M (K, b)∪ M (K, c)
is a superclass of B and M (θ, ψa ) ∪ M (θ, ψb ) ∪ M (θ, ψc ) ∈ XB . Further for every K ∈ TK ,
θ ∈ TX , we suppose that for ω k ∈ M (K, a), ω ℓ ∈ M (K, b), and ω r ∈ M (K, c) the following
holds:
M\
(θ, ψa )(ω r ) = M\
(θ, ψb )(ω k ) = M\
(θ, ψc )(ω ℓ ),
(94)
M\
(θ, ψa )(ω k ) = M\
(θ, ψb )(ω ℓ ) = M\
(θ, ψc )(ω r ),
(95)
M\
(θ, ψa )(ω ℓ ) = M\
(θ, ψb )(ω r ) = M\
(θ, ψc )(ω k ).
(96)
We want to show that the partition KB ∪ TK ∪ {{a, b, c}} is a set of superclasses for a
supercharacter theory of G and that the corresponding supercharacters are those from B,
TX , and ψa + ψb + ψc . Clearly the partitions of G and Irr(G) are the same size. Hence
we only need to show that every potential supercharacter is constant on every potential
superclass. Let θ1 be a supercharacter for B, θ1 6= triv, and let θ2 ∈ TX .
b = M\
Let K ∈ TK . Then K
(K, a)a + M\
(K, b)b + M\
(K, c)c. Clearly ψa + ψb + ψc is a
constant −1 on K. θ1 is a supercharacter for B, so it is constant on M (K, a) ∪ M (K, b) ∪
M (K, c) ∈ KB , hence θ1 is constant on K. Let ω k , ω ℓ , ω r be as above, so ω k a, ω ℓ b, and
ω r c are elements of K. Then using Equations (94), (95), and (96) we have
θ2 (ω k a) = M\
(θ2 , ψa )(ω k ) − M\
(θ2 , ψb )(ω k ) − M\
(θ2 , ψc )(ω k ) =
M\
(θ2 , ψb )(ω ℓ ) − M\
(θ2 , ψc )(ω ℓ ) − M\
(θ2 , ψa )(ω ℓ ) =
θ2 (ω ℓ b) =
M\
(θ2 , ψc )(ω r ) − M\
(θ2 , ψa )(ω r ) − M\
(θ2 , ψb )(ω r ) =
θ2 (ω r c).
(97)
′
Now suppose that for x ∈ {a, b, c} we have xω ℓ , xω ℓ ∈ K. Since M (K, x) is a superclass for
B ′ and M\
(θ2 , ψa ), M\
(θ2 , ψb ), and M\
(θ2 , ψc ) are supercharacters for B ′ we have θ2 (xω ℓ ) =
′
ℓ
θ2 (xω ). Hence θ2 is constant on K.
Now suppose K ∈ KB . ψa + ψb + ψc takes the constant value 3 on K. θ1 is constant
on K because B is a supercharacter theory for Cp . If ω i ∈ K, then θ2 (ω i ) equals
M\
(θ2 , ψa )(ω i ) + M\
(θ2 , ψb )(ω i ) + M\
(θ2 , ψc )(ω i ) =
21
(M \
(θ2 , ψa ) + M\
(θ2 , ψb ) + M\
(θ2 , ψc ))(ω i ).
(98)
Since M\
(θ2 , ψa )+M\
(θ2 , ψb )+M\
(θ2 , ψc ) is a supercharacter for B, we see that θ2 is constant
on K.
Finally, let K = {a, b, c}. ψa + ψb + ψc takes the constant value of −1 on K. θ1 is
constant on K, taking the value equal to the number of irreducible characters which are
summands of θ1 which may be expressed as | [χ] | if χ is a summand of θ1 . We have
θ2 (a) = |M (θ2 , ψa )| − |M (θ2 , ψb )| − |M (θ2 , ψc )|,
(99)
θ2 (b) = −|M (θ2 , ψa )| + |M (θ2 , ψb )| − |M (θ2 , ψc )|,
(100)
θ2 (c) = −|M (θ2 , ψa )| − |M (θ2 , ψb )| + |M (θ2 , ψc )|.
(101)
Since |M (θ2 , ψa )| = |M (θ2 , ψb )| = |M (θ2 , ψc )|, we see that θ2 is constant on K. Hence
we conclude that all potential supercharacters are constant on all potential superclasses,
so the partition KB ∪ TK ∪ {{a, b, c}} is indeed a set of superclasses for a supercharacter
theory of Cp × C2 × C2 .
As in Case 3 we have presented a direct proof for the sufficient direction. However,
such a supercharacter theory can be generated by a group of automorphisms in a similar
fashion to Case 3. Let A be a supercharacter theory satisfying the conditions of the
sufficient direction above, which is not a direct product supercharacter theory. There
exists a φ0 ∈ Aut(hωi) such that hφ0 i generates the supercharacter theory A|hωi . Let
φ ∈ Aut(G) be defined by φ(ω) = φ0 (ω) and φ|ha,bi = id. Let ψ ∈ Aut(G) be given by
ψ(ω) = ω, ψ(a) = b, ψ(b) = c, and ψ(c) = a. Then by an analogous argument as in Case
3, we see that hφ ◦ ψi generates A.
For supercharacter theories generated by automorphisms as above, see Examples 9.2
and 9.3.
7
Case (C2)3
To complete our proof, we will now discuss the supercharacter theories of C2 × C2 × C2 =
hxi × hyi × hzi. We first note that there are 5 supercharacter theories for C2 × C2 : the
minimal supercharacter theory, the maximal supercharacter theory, and 3 isomorphic to
the supercharacter theory with superclasses
{e}, {x}, {y, xy}.
(102)
Let A be a nontrivial supercharacter theory for (C2 )3 . For this group we cannot use
Theorem 3.2 to show that there must exist a proper nontrivial subgroup H such that
b ∈ A. So we begin by supposing there is no such H. Because A is nontrivial, no
H
superclass has size 7. Since every element other than the identity has order 2, the only
superclass which can have size 1 is {e}. By considering the complement, this implies that
no superclass has size 6. We have for {u, v} a superclass
(u + v)2 = 2e + 2uv ∈ A.
(103)
This implies that {uv} is a superclass, so we cannot have a superclass of size 2. Hence we
cannot have a superclass of size 5, again by considering complements. This leaves only a
22
superclass of size 3 and a superclass of size 4 as a valid option. If {u, v, w} is the superclass
of size 3, then by assumption {e, u, v, w} is not a subgroup. Therefore {u, v, w} is disjoint
from {uv, uw, vw}. We have
(u + v + w)2 = 3e + 2(uv + uw + vw) ∈ A.
(104)
Hence {uv, uw, vw} is a union of superclasses and is different from {u, v, w} which is
a contradiction. Therefore we see that for every nontrivial supercharacter theory of (C2 )3
b ∈ A.
there exists a proper nontrivial subgroup H such that H
There are only two isomorphism types for H, either H ∼
= C2 × C2 or H ∼
= C2 . Suppose
∼
b ∈ A, and without loss of generality let
that there is at least one H = C2 × C2 such that H
H = hx, yi. We consider the possible partitions of the complement of H, {z, xz, yz, xyz}.
If {z, xz, yz, xyz} is a superclass then the supercharacter theory is a wedge product of the
supercharacter theory of H and the unique supercharacter theory of (C2 )3 /H ∼
= C2 . Such
3
a supercharacter theory of (C2 ) exists for every supercharacter theory of H, and the set
of superclasses is
KA|H ∪ {{z, xz, yz, xyz}}.
(105)
If the partition of {z, xz, yz, xyz} contains a singleton, then there exists a subgroup H ′ ∼
=
c′ ∈ A and H ∩ H ′ = {e}. By Lemma 6.5 we have that A must be the direct
C2 with H
product of the supercharacter theories of H and H ′ . The only remaining possible partition
of {z, xz, yz, xyz} is two pairs, let them be {t, u} and {v, w}. We note that tu = vw ∈ H,
without loss of generality let tu = vw = x. Then A|H is either the minimal supercharacter
theory of H or {{e}, {x}, {y, z}}. If it is the minimal supercharacter theory of H, then
it is in fact a direct product supercharacter theory of ht, ui and hyi, so we have already
accounted for this situation. If the supercharacter theory of H is {{e}, {x}, {y, z}}, then
we see that every superclass disjoint from hxi is a hxi-coset, so A is the wedge product
of the unique supercharacter theory of hxi and the minimal supercharacter theory of
(C2 )3 /hxi ∼
= C2 × C2 .
b ∈ A.
We now consider the case where there does not exist an H ∼
= C2 × C2 with H
c′ ∈ A. If there is a singleton superclass {u} different
There must exist H ′ ∼
= C2 with H
′
′ × hui ∼
′ × hui ∈ A which contradicts
from {e} and H 6= hui, then H\
= C2 × C2 and H\
′
our assumption. Hence H is the only proper nontrivial subgroup which is a union of
superclasses. We determine how many possibilities there are for A by considering the
dual supercharacter theory of A, which we will denote by B. There is exactly one proper
b ∈ B, and H ∼
nontrivial subgroup H such that H
= C2 × C2 . Hence B|H is the maximal
supercharacter theory, and by the above argument, the complement of H is a superclass.
Hence there is only one isomorphism type for B and so there is only one isomorphism type
for A. A must be the wedge product of the supercharacter theory of H ′ and the maximal
supercharacter theory of (C2 )3 /H ′ ∼
= C2 × C2 .
This completes the proof of Theorem 3.1, as we have considered every case.
23
8
Enumeration of the Supercharacter Theories of Cp ×C2 ×C2
for p odd
We will now determine the number of supercharacter theories of Cp ×C2 ×C2 when p is odd,
and how many can be obtained by each of the three constructions. Working through the
argument provides a method for constructing the complete set of supercharacter theories
for a particular choice of p.
We will let D(n) equal the number of positive integers that divide n.
Lemma 8.1. Let G be a finite group with a complementary pair H, N , and let A be a
supercharacter theory for G such that A is the direct product supercharacter theory of
A|H and A|N . Then A is generated by automorphisms iff A|H and A|N are both generated
by automorphisms.
Proof. Since H and N are a complementary pair, G ∼
= H × N . Clearly if A is generated by
b ∈ A, then A|K is generated by
automorphisms and K is a normal subgroup such that K
automorphisms. Hence A|H and A|N are both generated by automorphisms by restriction.
Conversely, suppose that there exists H ′ ≤ Aut(H) and N ′ ≤ Aut(N ) such that H ′ and N ′
generate the supercharacter theories A|H and A|N respectively. H ′ includes into Aut(G)
by acting trivially on N ≤ G, and similarly N ′ includes into Aut(G) by acting trivially on
H ≤ G. Then the subgroup of Aut(G) generated by the images of H ′ and N ′ generates
A.
Lemma 8.2. If p is an odd prime, then there are 3D(p − 1) + 1 supercharacter theories of
Cp × C2 . Further a supercharacter theory of Cp × C2 is generated by automorphisms iff it
is a direct product supercharacter theory.
Proof. Let Cp × C2 = hωi × hai. Since Cp × C2 ∼
= C2p is cyclic, by [LM96, Th. 3.7]
every nonmaximal supercharacter theory can be constructed as a wedge product, a direct
product, or is generated by automorphisms. Since p and 2 are relatively prime, it is clear
c ∈ A and hai
c ∈ A. By Lemma 6.5 A must
that if A is generated by automorphisms, hωi
be a refinement of the direct product supercharacter theory of A|hωi and A|hai . Since
a ∈ A, we see that A must be the direct product supercharacter theory. Conversely, if
c ∈ A and hai
c ∈ A. By Lemma 6.1
A is a direct product supercharacter theory, then hωi
every supercharacter theory of hωi and hai is generated by automorphisms. Therefore by
Lemma 8.1 A is generated by automorphisms.
The direct product supercharacter theories are in bijection with the supercharacter
theories of hωi, and by Lemma 6.1 there are D(p − 1) of them. There are 2D(p − 1) wedge
products, D(p − 1) where hωi is the normal group and D(p − 1) where hωi is the quotient
group. The only other supercharacter theory is the maximal supercharacter theory, hence
Cp × C2 has 3D(p − 1) + 1 supercharacter theories.
Lemma 8.3. [Hen09, Lemma 8.1] Let N and H be proper nontrivial normal subgroups of
G. If a supercharacter theory of G can be constructed as a wedge product of supercharacter
theories for N and G/N and also as a wedge product of supercharacter theories of H and
G/H then N ≤ H or H ≤ N .
24
Theorem 8.1. Let p be an odd prime and let p − 1 = 2k 3ℓ n where n is not divisible by 2
or 3. Then the number of distinct supercharacter theories of Cp × C2 × C2 is
3kD(3ℓ n) + 2ℓD(2k n) + 30D(p − 1) + 13.
(106)
Further, we can enumerate the supercharacter theories by method of construction. There
are 3kD(3ℓ n) + 2ℓD(2k n) + 5D(p − 1) supercharacter theories which can be generated by
automorphisms, and 11D(p − 1) + 6 supercharacter theories which are direct products,
including 5D(p − 1) which can also be generated by automorphisms. If a supercharacter
theory is a wedge product, then it is not a direct product or generated by automorphisms.
There are 19D(p − 1) + 6 supercharacter theories which are wedge products, and the only
remaining supercharacter theory is the maximal supercharacter theory.
Proof. By Theorem 3.1, every nonmaximal supercharacter theory of Cp × C2 × C2 is
generated by automorphisms, a direct product, or a wedge product. We begin by considering all of the supercharacter theories generated by automorphisms. Every subgroup
of Aut(Cp × C2 × C2 ) generates a supercharacter theory of Cp × C2 × C2 , although they
are not all distinct. Also note that the minimal supercharacter theory is generated by the
trivial subgroup of Aut(Cp × C2 × C2 ). Since Aut(Cp ) ∼
= Cp−1 , Aut(C2 × C2 ) ∼
= S3 , and p
is odd we have that
Aut(Cp × C2 × C2 ) ∼
= Aut(Cp ) × Aut(C2 × C2 ) ∼
= Cp−1 × S3 .
(107)
We will now consider all possible subgroups of Aut(Cp )×Aut(C2 ×C2 ). Let Aut(Cp ) = hψi,
let A3 be the order three subgroup of Aut(C2 × C2 ) and for x ∈ {a, b, c} let Hx be the
order two subgroup of Aut(C2 × C2 ) that fixes x. It follows from Goursat’s Lemma that
every subgroup of G1 × G2 can be expressed as
R = {(x, y) ∈ H1 × H2 |φ(xN1 ) = yN2 }
(108)
for a unique choice of N1 E H1 ≤ G1 , N2 E H2 ≤ G2 and an isomorphism φ : H1 /N1 →
H2 /N2 , and every such choice gives a subgroup. Let N1 E H1 ≤ Aut(Cp ) and N2 E H2 ≤
Aut(C2 × C2 ). We have Aut(Cp ) ∼
= C2k × C3ℓ × Cn where n is not divisible by 2 or 3.
H2 /N2 is isomorphic to the trivial group, C2 , or C3 . Case I: If H2 /N2 ∼
= {e}, then
H1 = N1 and H2 = N2 . There is only one choice for φ, so it is easy to see that in this
case the subgroup R is the direct product H1 × H2 . Every choice of H1 gives a distinct
supercharacter theory of hωi, so there are D(p − 1) choices for H1 . However, Aut(C2 × C2 )
and A3 ≤ Aut(C2 × C2 ) both generate the maximal supercharacter theory of C2 × C2 . We
then observe that there are five choices for H2 . Hence there are 5D(p − 1) possibilities.
Case II: There are four ways for H2 /N2 ∼
= C2 : H2 = Aut(C2 × C2 ) and N2 = A3 ,
or H2 is Ha , Hb , or Hc and N2 = {e}. We have H1 = hψ m i and N1 = hψ 2m i where
m|p − 1 and (p − 1)/m is even. There is only one choice for φ. If H2 = Aut(C2 × C2 )
and N2 = A3 then (id, σ) and (id, σ 2 ) are elements of the subgroup R where σ(a) = b,
σ(b) = c, σ(c) = a. We see that the supercharacter theory generated by this subgroup R
is the same one generated by the subgroup for H1′ = N1′ = H1 , H2′ = N2′ = A3 , which is
hψ m i × hσi. Hence this supercharacter theory has already been constructed above.
25
So we will only consider H2 = Hx and N2 = {id}. There is only one choice for φ so
the resulting subgroup R is hψ m ◦ τ i ∼
= C(p−1)/m , where Hx = hτ i. There are three choices
ℓ
for H2 , kD(3 n) choices for H1 , and one choice for N1 , which gives 3kD(3ℓ n) new contributions. Note that supercharacter theories of this form correspond to the supercharacter
theories in Case 3 of Theorem 3.1 which are not direct products. See also Example 9.1.
Case III: When H2 /N2 ∼
= C3 , we must have H2 = A3 and N2 = {e}. Then H1 = hψ m i,
3m
N1 = hψ i where m|(p − 1) and 3 divides (p − 1)/m. There are ℓD(2k n) possibilities for
H1 , a unique N1 and two choices for φ which yields 2ℓD(2k n) subgroups R = hψ m ◦ σφ i ∼
=
C(p−1)/m where φ(ψ m N1 ) = σφ . Note that supercharacter theories of this form correspond
to the supercharacter theories in Case 4 of Theorem 3.1 which are not direct products.
See Examples 9.2, 9.3. We conclude that there are 3kD(3ℓ n) + 2ℓD(2k n) + 5D(p − 1)
supercharacter theories generated by automorphisms.
We now consider the supercharacter theories which are direct products. By Lemma
8.1, we see that some direct product supercharacter theories of Cp × C2 × C2 can be
generated by automorphisms. In particular, by Lemma 6.1 every supercharacter theory of
Cp is generated by automorphisms, and it is easy to see that every supercharacter theory
of C2 × C2 is also generated by automorphisms. Hence we have already constructed the
5D(p − 1) direct product supercharacter theories of Cp × C2 × C2 with the complementary
pair Cp , C2 × C2 . The other complementary pair to be considered is Cp × C2 , C2 . Let
x, y be distinct elements of {a, b, c} and let Cp × C2 = hω, xi, and C2 = hyi. Note that if a
supercharacter theory A can be expressed as the direct product for the complementry pair
hω, xi, hyi then it is also a direct product for the complementary pair hωi, hx, yi iff A|hω,xi
is a direct product supercharacter theory. Therefore we want A|hω,xi to not be a direct
product. Hence we need to count the supercharacter theories A which are direct products
of A|hω,xi and A|hyi such that A|hω,xi must be a wedge product with hωi normal, a wedge
product with hxi normal, or the maximal supercharacter theory. There are six choices for
the pair x, y. In the case of hωi normal there are D(p − 1) choices for the supercharacter
theory A|hωi , and given y either choice of x yields the same supercharacter theory so there
are 3D(p − 1) possibilities. Similarly, for hxi normal there are D(p − 1) supercharacter
theories for the quotient, and either choice of y yields the same supercharacter theory so
there are 3D(p − 1) possibilities. For the maximal case, all six choices of x, y give distinct
supercharacter theories. Hence there are 6D(p − 1) + 6 direct product supercharacter
theories of Cp × C2 × C2 which cannot be generated by automorphisms.
We now consider the wedge products. First note that if A is a wedge product, there
does not exist a complementary pair of subgroups such that both are unions of superclasses
of A. Hence A cannot be either a direct product supercharacter theory or generated
by automorphisms. If A is a wedge product of the supercharacter theories of a normal
subgroup N and G/N , then N is isomorphic to one of C2 , C2 × C2 , Cp , or C2 × Cp . We
will avoid constructing a supercharacter theory more than once by using Lemma 8.3. We
begin with all wedge products with N ∼
= Cp . There is only one subgroup isomorphic to
Cp , so N = hωi. There are D(p − 1) supercharacter theories of N , and five supercharacter
theories of the quotient, hence there are 5D(p − 1) such supercharacter theories.
Now we consider all wedge products with N ∼
= C2 . There are three possibilities for N :
hai, hbi, and hci. N has only 1 supercharacter theory, and by Lemma 8.2 the quotient has
26
3D(p − 1) + 1 supercharacter theories. Hence there are 3(3D(p − 1) + 1) wedge products
with N ∼
= C2 normal.
We now want to count the wedge products with N ∼
= C2 × C2 which are not also wedge
′
∼
products of supercharacter theories of N = C2 and G/N ′ . We have N = {e, a, b, c}. It
is easy to check that a supercharacter theory A of Cp × C2 × C2 is a wedge product for
both N and N ′ iff the dimension of A|N is three. Hence A|N must be either the minimal
supercharacter theory, or the maximal supercharacter theory. By Lemma 6.1 there are
D(p − 1) supercharacter theories for the quotient, hence we have 2D(p − 1) supercharacter
theories.
Finally, we count the wedge products where N ∼
= Cp ×C2 which are not wedge products
′′
′
∼
∼
for N = Cp or N = C2 . If A is a wedge product for Cp or C2 then A|N is a wedge product
for Cp or C2 respectively. Hence A|N must be either a direct product supercharacter theory
or the maximal supercharacter theory, so there are D(p − 1) + 1 choices for A|N . There
is only one choice for the supercharacter theory of the quotient, and three choices for N :
hω, ai, hω, bi, or hω, ci. Hence there are 3(D(p − 1) + 1) such supercharacter theories.
We conclude that there are 19D(p − 1) + 6 wedge products. Adding to this the supercharacter theories generated by automorphisms, the direct product supercharacter theories
which are not generated by automorphisms, and the maximal supercharacter theory we
see that there are 3kD(3ℓ n) + 2ℓD(2k n) + 30D(p − 1) + 13 supercharacter theories for
Cp × C2 × C2 .
9
Examples
Example 9.1. Let p = 5. Let φ, ψ be automorphisms of C5 ×C2 ×C2 defined by φ(ω) = ω 2 ,
φ(a) = a, φ(b) = b, φ(c) = c, and ψ(ω) = ω, ψ(a) = a, ψ(b) = c, ψ(c) = b. Then the
supercharacter theory generated by the group hφ ◦ ψi has the following superclasses:
{e}, {ω, ω 2 , ω 3 , ω 4 },
(109)
{a}, {b, c},
(110)
2
3
4
{ωa, ω a, ω a, ω a},
(111)
{ωb, ω 4 b, ω 2 c, ω 3 c}, {ω 2 b, ω 3 b, ωc, ω 4 c}.
(112)
For K = [ωb] we have M (K, b) = {ω, ω 4 } and M (K, c) = {ω 2 , ω 3 }. We see that
{e}, {ω, ω 4 }, {ω 2 , ω 3 } are the superclasses of the supercharacter theory of hωi generated
by the group hφ2 |hωi i and that φ2 (ω) = ω −1 .
Example 9.2. Let p = 7. {{e}, {ω, ω 2 , ω 4 }, {ω 3 , ω 5 , ω 6 }} is the set of superclasses for the
supercharacter theory of C7 generated by hφ0 i where φ0 (ω) = ω 2 . Let φ, ψ ∈ Aut(C7 ×
C2 × C2 ) be defined by φ(ω) = ω 2 , φ(a) = a, and φ(b) = b, φ(c) = c, and ψ(ω) = ω,
ψ(a) = b, ψ(b) = c, ψ(c) = a. Then hφ ◦ ψi generates the supercharacter theory with
superclasses:
{e}, {ω, ω 2 , ω 4 }, {ω 3 , ω 5 , ω 6 },
(113)
{a, b, c},
27
(114)
{ωa, ω 2 b, ω 4 c}, {ω 2 a, ω 4 b, ωc}, {ω 6 a, ω 5 b, ω 3 c},
(115)
{ω 5 a, ω 3 b, ω 6 c}, {ω 3 a, ω 6 b, ω 5 c}, {ω 4 a, ωb, ω 2 c}.
(116)
Note that the supercharacter theory generated by hφ ◦ ψ −1 i is
{e}, {ω, ω 2 , ω 4 }, {ω 3 , ω 5 , ω 6 },
(117)
{a, b, c},
(118)
{ωa, ω 2 c, ω 4 b}, {ω 2 a, ω 4 c, ωb}, {ω 6 a, ω 5 c, ω 3 b},
(119)
{ω 5 a, ω 3 c, ω 6 b}, {ω 3 a, ω 6 c, ω 5 b}, {ω 4 a, ωc, ω 2 b}.
(120)
These supercharacter theories differ according to the 2 possibilities described in Equations
(92) and (93), and also the different choices for the isomorphism in Equation (108).
Example 9.3. Let ψ be defined as above and let σ ∈ Aut(C7 × C2 × C2 ) be defined by
σ(ω) = ω 5 , σ(a) = a, σ(b) = b, and σ(c) = c. Then hσ ◦ ψi generates the following
supercharacter theory:
{e}, {ω, ω 2 , ω 3 , ω 4 , ω 5 , ω 6 },
(121)
{a, b, c},
(122)
{ωa, ω 6 a, ω 2 b, ω 5 b, ω 3 c, ω 4 c},
(123)
{ω 2 a, ω 5 a, ω 3 b, ω 4 b, ωc, ω 6 c},
(124)
{ω 3 a, ω 4 a, ωb, ω 6 b, ω 2 c, ω 5 c}.
(125)
References
[DI08]
Persi Diaconis and I. M. Isaacs. Supercharacters and superclasses for algebra
groups. Transactions of the American Mathematical Society, 360(5):2359–2392,
2008.
[EKP16] Sergei Evdokimov, István Kovács, and Ilya Ponomarenko. On Schurity of Finite
Abelian Groups. Communications in Algebra, 44(1):101–117, 2016.
[EP14]
Sergei Evdokimov and Ilia Ponomarenko. Schur rings over a product of Galois rings. Beiträge zur Algebra und Geometrie / Contributions to Algebra and
Geometry, 55(1):105–138, 2014.
[Hen08] Anders O. F. Hendrickson. Supercharacter theories of cyclic p-groups. PhD
thesis, University of Wisconsin-Madison, 2008.
[Hen09] Anders O. F. Hendrickson. Construction of supercharacter theories of finite
groups. arXiv:math.GR/0905.3538v1, 2009.
[Hen10] Anders O. F. Hendrickson.
Supercharacter theories and Schur rings.
arXiv:math.GR/1006.1363v1, 2010.
28
[LM96]
Ka Hin Leung and Shing Hing Man. On Schur Rings over Cyclic Groups, II.
Journal of Algebra, 183(2):273–285, 1996.
[LM98]
Ka Hin Leung and Shing Hing Man. On schur rings over cyclic groups. Israel
Journal of Mathematics, 106(1):251–267, 1998.
[Wie64] Helmut Wielandt. Finite Permutation Groups. Academic Press, 1964.
Department of Mathematics, University of California, Davis, One Shields
Avenue, Davis, CA 95616-8633
E-mail address: [email protected]
29
| 4math.GR
|
Binary Search in Graphs Revisited
Argyrios Deligkas∗
George B. Mertzios†
Paul G. Spirakis‡
arXiv:1702.08899v1 [cs.DS] 28 Feb 2017
Abstract
In the classical binary search in a path the aim is to detect an
unknown target by asking as few queries as possible, where each query
reveals the direction to the target. This binary search algorithm has
been recently extended by [Emamjomeh-Zadeh et al., STOC, 2016]
to the problem of detecting a target in an arbitrary graph. Similarly
to the classical case in the path, the algorithm of Emamjomeh-Zadeh
et al. maintains a candidates’ set for the target, while each query
asks an appropriately chosen vertex– the “median”–which minimises a
potential Φ among the vertices of the candidates’ set. In this paper
we address three open questions posed by Emamjomeh-Zadeh et al.,
namely (a) detecting a target when the query response is a direction
to an approximately shortest path to the target, (b) detecting a target
when querying a vertex that is an approximate median of the current
candidates’ set (instead of an exact one), and (c) detecting multiple
targets, for which to the best of our knowledge no progress has been
made so far. We resolve questions (a) and (b) by providing appropriate
upper and lower bounds, as well as a new potential Γ that guarantees
efficient target detection even by querying an approximate median each
time. With respect to (c), we initiate a systematic study for detecting
two targets in graphs and we identify sufficient conditions on the queries
that allow for strong (linear) lower bounds and strong (polylogarithmic)
upper bounds for the number of queries. All of our positive results can
be derived using our new potential Γ that allows querying approximate
medians.
Keywords: binary search, graph, approximate query, probabilistic
algorithm, lower bound.
1
Introduction
The classical binary search algorithm detects an unknown target (or “treasure”)
t on a path with n vertices by asking at most log n queries to an oracle which
always returns the direction from the queried vertex to t. To achieve this upper
bound on the number of queries, the algorithm maintains a set of candidates for
the place of t; this set is always a sub-path, and initially it is the whole path.
Then, at every iteration, the algorithm queries the middle vertex (“median”)
of this candidates’ set and, using the response of the query, it excludes either
the left or the right half of the set. This way of searching for a target in a path
∗ Faculty of Industrial Engineering and Management, Technion, Israel.
Email:
[email protected]
† School of Engineering and Computing Sciences, Durham University, UK. Email:
[email protected]
‡ Department
of Computer Science,
University of Liverpool,
UK. Email:
[email protected]
1
can be naturally extended to the case where t lies on an n-vertex tree, again by
asking at most log n queries that reveal the direction in the (unique) path to
t [21]. The principle of the binary search algorithm on trees is based on the same
idea as in the case of a path: for every tree there exists a separator vertex such
that each of its subtrees contains at most half of the vertices of the tree [13],
which can be also efficiently computed.
Due to its prevalent nature in numerous applications, the problem of detecting an unknown target in an arbitrary graph or, more generally in a search space,
has attracted many research attempts from different viewpoints. Only recently
the binary search algorithm with log n direction queries has been extended to
arbitrary graphs by Emamjomeh-Zadeh et al. [9]. In this case there may exist
multiple paths, or even multiple shortest paths form the queried vertex to t.
The direction query considered in [9] either returns that the queried vertex q
is the sought target t, or it returns an arbitrary direction from q to t, i.e. an
arbitrary edge incident to q which lies on a shortest path from q to t. The main
idea of this algorithm follows again the same principle as for paths and trees: it
always queries a vertex that is the “median” of the current candidates’ set and
any response to the query is enough to shrink the size of the candidates’ set by a
factor of at least 2. Defining what the “median” is in the case of general graphs
now becomes more tricky: Emamjomeh-Zadeh et al. [9] define the median of a
set S as the vertex q that minimizes a potential function Φ, namely the sum of
the distances from q to all vertices of S.
Apart from searching for upper bounds on the number of queries needed
to detect a target t in graphs, another point of interest is to derive algorithms
which, given a graph G, compute the optimal number of queries needed to detect an unknown target in G (in the worst case). This line of research was
initiated in [17] where the authors studied directed acyclic graphs (DAGs). Although computing a query-optimal algorithm is known to be NP-hard on general
graphs [4, 7, 15], there exist efficient algorithms for trees; after a sequence of papers [1, 12, 16, 18, 25], linear time algorithms were found in [18, 21]. Different
models with queries of non-uniform costs or with a probability distribution over
the target locations were studied in [5, 6, 14].
A different line of research is to search for upper bounds and informationtheoretic bounds on the number of queries needed to detect a target t, assuming
that the queries incorporate some degree of “noise”. In one of the variations of
this model [2, 9, 10], each query independently returns with probability p > 12 a
direction to a shortest path from the queried vertex q to the target, and with
probability 1 − p an arbitrary edge (possibly adversarially chosen) incident to
q. The study of this problem was initiated in [10], where Ω(log n) and O(log n)
bounds on the number of queries were established for a path with n vertices.
This information-theoretic lower bound of [10] was matched by an improved
upper bound in [2]. The same matching bound was extended to general graphs
in [9].
In a further “noisy” variation of binary search, every vertex v of the graph
is assigned a fixed edge incident to v (also called the “advice” at v). Then, for
a fraction p > 21 of the vertices, the advice directs to a shortest path towards t,
while for the rest of the vertices the advice is arbitrary, i.e. potentially misleading
or adversarially chosen [3]. This problem setting is motivated by the situation
of a tourist driving a car in an unknown country that was hit by a hurricane
which resulted in some fraction of road-signs being turned in an arbitrary and
2
unrecognizable way. The question now becomes whether it is still possible to
navigate through such a disturbed and misleading environment and to detect
the unknown target by asking only few queries (i.e. taking advice only from
a few road-signs). It turns out that, apart from its obvious relevance to data
structure search, this problem also appears in artificial intelligence as it can
model searching using unreliable heuristics [3, 19, 22]. Moreover this problem
also finds applications outside computer science, such as in navigation issues in
the context of collaborative transport by ants [11].
Another way of incorporating some “noise” in the query responses, while
trying to detect a target, is to have multiple targets hidden in the graph. Even
if there exist only two unknown targets t1 and t2 , the response of each query
is potentially confusing even if every query correctly directs to a shortest path
from the queried vertex to one of the targets. The reason of confusion is that
now a detecting algorithm does not know to which of the hidden targets each
query directs. In the context of the above example of a tourist driving a car in
an unknown country, imagine there are two main football teams, each having its
own stadium. A fraction 0 < p1 < 1 of the population supports the first team
and a fraction p2 = 1 − p1 the second one, while the supporters of each team are
evenly distributed across the country. The driver can now ask questions of the
type “where is the football stadium?” to random local people along the way,
in an attempt to visit both stadiums. Although every response will be honest,
the driver can never be sure which of the two stadiums the local person meant.
Can the tourist still detect both stadiums quickly enough? To the best of our
knowledge the problem of detecting multiple targets in graphs has not been
studied so far; this is one of the main topics of the present paper.
The problem of detecting a target within a graph can be seen as a special case
of a two-player game introduced by Renyi [24] and rediscovered by Ulam [26].
This game does not necessarily involve graphs: the first player seeks to detect
an element known to the second player in some search space with n elements.
To this end, the first player may ask arbitrary yes/no questions and the second
player replies to them honestly or not (according to the details of each specific
model). Pelc [23] gives a detailed taxonomy for this kind of games. Group
testing is a sub-category of these games, where the aim is to detect all unknown
objects in a search space (not necessarily a graph) [8]. Thus, group testing is
related to the problem of detecting multiple targets in graphs, which we study
in this paper.
1.1
Our contribution
In this paper we systematically investigate the problem of detecting one or
multiple hidden targets in a graph. Our work is driven by the open questions
posed by the recent paper of Emamjomeh-Zadeh et al. [9] which dealt with
the detection of a single target with and without “noise”. More specifically,
Emamjomeh-Zadeh et al. [9] asked for further fundamental generalizations of
the model which would be of interest, namely (a) detecting a single target when
the query response is a direction to an approximately shortest path, (b) detecting
a single target when querying a vertex that is an approximate median of the
current candidates’ set S (instead of an exact one), and (c) detecting multiple
targets, for which to the best of our knowledge no progress has been made so
far.
3
We resolve question (a) in Section 2.1 by proving that any algorithm requires
Ω(n) queries to detect a single target t, assuming that a query directs to a path
with an approximately shortest length to t. Our results hold essentially for
any approximation guarantee, i.e. for 1-additive and for (1 + ε)-multiplicative
approximations.
Regarding question (b), we first prove in Section
√ 2.2 that, for any constant
0 < ε < 1, the algorithm of [9] requires at least Ω( n) queries when we query
each time an (1 + ε)-approximate median (i.e. an (1 + ε)-approximate minimizer
of the potential Φ over the candidates’ set S). Second, to resolve this lower
bound, we introduce in Section 2.3 a new potential Γ. This new potential
can be efficiently computed and, in addition, guarantees that, for any constant
0 ≤ ε < 1, the target t can be detected in O(log n) queries even when an
(1 + ε)-approximate median (with respect to Γ) is queried each time.
Regarding question (c), we initiate in Section 3 the study for detecting multiple targets on graphs by focusing mainly to the case of two targets t1 and t2 .
We assume throughout that every query provides a correct answer, in the sense
that it always returns a direction to a shortest path from the queried vertex
either to t1 or to t2 . The “noise” in this case is that the algorithm does not
know whether a query is returning a direction to t1 or to t2 . Initially we observe
in Section 3 that any algorithm requires n2 − 1 (resp. n − 2) queries in the worst
case to detect one target (resp. both targets) if each query directs adversarially
to one of the two targets. Hence, in the remainder of Section 3, we consider
the case where each query independently directs to the first target t1 with a
constant probability p1 and to the second target t2 with probability p2 = 1 − p1 .
For the case of trees, we prove in Section 3 that both targets can be detected
with high probability within O(log n) queries.
For general graphs, we distinguish between biased queries (p1 > p2 ) in Section 3.1 and unbiased queries (p1 = p2 = 21 ) in Section 3.2. For biased queries,
we observe that we can utilize the algorithm of Emamjomeh-Zadeh et al. [9]
to detect the first target t1 with high probability in O(log n) queries; this can
be done by considering the queries that direct to t2 as “noise”. Thus our objective becomes to detect the target t2 in a polylogarithmic number of queries.
Notice here that we cannot apply the “noisy” framework of [9] to detect the
second target t2 , since now the “noise” is larger than 12 . We derive a probabilistic algorithm that overcomes this problem and detects the target t2 with
high probability in O(∆ log2 n) queries, where ∆ is the maximum degree of
a vertex in the graph. Thus, whenever ∆ = O(poly log n), a polylogarithmic
number of queries suffices to detect t2 . In contrast, we prove in Section 3.2
that, for unbiased queries, any deterministic (possibly adaptive) algorithm that
detects at least one of the targets requires at least n2 − 1 queries, even in an
unweighted cycle. Extending this lower bound for two targets, we prove that,
assuming 2c ≥ 2 different targets and unbiased queries, any deterministic (possibly adaptive) algorithm requires at least n2 − c queries to detect one of the
targets.
Departing from the fact that our best upper bound on the number of biased
queries in Section 3.1 is not polylogarithmic when the maximum degree ∆ is
not polylogarithmic, we investigate in Section 4 several variations of queries
that provide more informative responses. In Section 4.1 we turn our attention
to “direction-distance” biased queries which return with probability pi both the
direction to a shortest path to ti and the distance between the queried vertex
4
and ti . In Section 4.2 we consider another type of a biased query which combines
the classical “direction” query and an edge-variation of it. For both query types
of Sections 4.1 and 4.2 we prove that the second target t2 can be detected with
high probability in O(log3 n) queries. Furthermore, in Sections 4.3 and 4.4 we
investigate two further generalizations of the “direction” query which make the
target detection problem trivially hard and trivially easy to solve, respectively.
1.2
Our Model and Notation
We consider connected, simple, and undirected graphs. A graph G = (V, E),
where |V | = n, is given along with a weight function w : E → R+ on its edges;
if w(e) = 1 for every e ∈ E then G is unweighted. An edge between two vertices
v and u of G is denoted by vu, and in this case v and u are said to be adjacent.
The distance d(v, u) between vertices v and u is the length of a shortest path
between v and u with respect to the weight function w. Since the graphs we
consider are undirected, d(u, v) = d(v, u) for every pair of vertices v, u. Unless
specified otherwise, all logarithms are taken with base 2. Whenever an event
happens with probability at least 1 − n1α for some α > 0, we say that it happens
with high probability.
The neighborhood of a vertex v ∈ V is the set N (v) = {u ∈ V : vu ∈ E}
of its adjacent vertices. The cardinality of N (v) is the degree deg(v) of v. The
maximum degree among all vertices in G is denoted by ∆(G), i.e. ∆(G) =
max{deg(v) : v ∈ V }. For two vertices v and u ∈ N (v) we denote by N (v, u) =
{x ∈ V : d(v, x) = w(vu) + d(u, x)} the set of vertices x ∈ V for which there
exists a shortest path from v to x, starting with the edge vu. Note that, in
general, N (u, v) 6= N (v, u). Let T = {t1 , t2 , · · · , t|T | } ⊆ V be a set of (initially
unknown) target vertices. A direction query (or simply query) at vertex v ∈ V
returns with probability pi a neighbor u ∈ N (v) such that ti ∈ N (u, v), where
P|T |
i=1 pi = 1. If there exist more than one such vertices u ∈ N (v) leading to ti
via a shortest path, the direction query returns an arbitrary one among them,
i.e. possibly chosen adversarially, unless specified otherwise. Moreover, if the
queried vertex v is equal to one of the targets ti ∈ T , this is revealed by the
query with probability pi .
2
Detecting a Unique Target
In this section we consider the case where there is only one unknown target
t = t1 , i.e. T = {t}. In this case the direction query at vertex v always returns a neighbor u ∈ N (v) such that t ∈ N (u, v). For this problem setting,
Emamjomeh-Zadeh et al. [9] provided a polynomial-time algorithm which detects the target t in at most log n direction queries. During its execution, the
algorithm of [9] maintains a “candidates’ set” S ⊆ V such that always t ∈ S,
where initially S = V . At every iteration the algorithm computes in polynomial
time a vertex v (called the median of S) which minimizes a potential ΦS (v)
among all vertices of the current set S. Then it queries a median v of S and
it reduces the candidates’ set S to S ∩ N (v, u), where u is the vertex returned
by the direction query at v. The upper bound log n of the number of queries in
this algorithm follows by the fact that always |S ∩ N (v, u)| ≤ |S|
2 , whenever v
is the median of S.
5
2.1
Bounds for Approximately Shortest Paths
We provide lower bounds for both additive and multiplicative approximation
queries. A c-additive approximation query at vertex v ∈ V returns a neighbor u ∈ N (v) such that w(vu) + d(u, t) ≤ d(v, t) + c. Similarly, an (1 + ε)multiplicative approximation query at vertex v ∈ V returns a neighbor u ∈ N (v)
such that w(vu) + d(u, t) ≤ (1 + ε) · d(v, t).
It is not hard to see that in the unweighted clique with n vertices any algorithm requires in worst case n − 1 1-additive approximation queries to detect
the target t. Indeed, in this case d(v, t) = 1 for every vertex v 6= t, while every
vertex u ∈
/ {v, t} is a valid response of an 1-additive approximation query at v.
Since in the case of the unweighted clique an additive 1-approximation is the
same as a multiplicative 2-approximation of the shortest path, it remains unclear whether 1-additive approximation queries allow more efficient algorithms
for graphs with large diameter. In the next theorem we strengthen this result
to graphs with unbounded diameter.
Theorem 1 Assuming 1-additive approximation queries, any algorithm requires at least n − 1 queries to detect the target t, even in graphs with unbounded
diameter.
Proof. To prove the theorem we will construct a graph and a strategy for the
adversary such that any algorithm will need n − 1 queries to locate the target t.
Consider a horizontal 2 × n2 grid graph where we add the two diagonals in every
cell of the grid. Formally, the graph has n2 “top” vertices v1 , . . . , v n2 and n2
“bottom” vertices u1 , . . . , u n2 . For every i ∈ {1, 2, . . . , n2 − 1} we have the edges
vi vi+1 , ui ui+1 , vi ui , vi+1 ui+1 , vi ui+1 , vi+1 ui .
The strategy of the adversary is as follows. If the algorithm queries a top
vertex vi , then the query returns the bottom vertex ui . Similarly, if the algorithm queries a bottom vertex ui , then the query returns the top vertex vi .
Observe that, in every case, the query answer lies on a path of length at most
one more than a shortest path from the queried vertex and the target t. To
see this assume that the algorithm queries a top vertex vi ; the case where the
queried vertex is a bottom vertex ui is symmetric.
If t = ui , then the edge vi ui clearly lies on the shortest path between vi and
t. If t = uj , where j 6= i, then the shortest path uses one of the diagonal edges
incident to vi . In this case the edge vi ui leads to a path with length one more
than the shortest one. Finally, if t = vj , where j 6= i, then the shortest path
has length |j − i| and uses either the edge vi vi−1 or the edge vi vi+1 . In both
cases the edge vi ui lies on the path from vi to T with length |j − i| + 1 which
uses the edge vi ui and one of the diagonal edges ui+1 vi−1 and ui+1 vi+1 .
Hence, after each query at a vertex different than t, the algorithm can not
obtain any information about the position of t. Thus, in the worst case the
algorithm needs to make n − 1 queries to detect t.
In the next theorem we extend Theorem 1 by showing a lower bound of n · 4ε
queries when we assume (1 + ε)-multiplicative approximation queries.
Theorem 2 Let ε > 0. Assuming (1 + ε)-multiplicative approximation queries,
any algorithm requires at least at least n · 4ε queries to detect the target t.
6
Proof. For the proof we use the same construction from Theorem 1, however
the adversary we use here is slightly modified. Assume that the distance between
the queried vertex and the target t is d. If d + 1 ≤ (1 + ε) · d, or equivalently, if
d ≥ 1ε , the adversary can respond in the same way as in Theorem 1.
Overall, the adversary proceeds as follows. Initially all vertices are unmarked. Whenever the algorithm queries a vertex vi (resp. ui ), the adversary
marks the vertices {vj , uj : |j −i| < 1ε } in order to determine the query response.
If at least one unmarked vertex remains in the graph, then the query returns
(similarly to Theorem 1) vertex ui (resp. vi ). In this case the adversary can
place the target t at any currently unmarked vertex. By doing so, the adversary
ensures that the distance between t and any of the previously queried vertices
is at least 1ε . If all vertices of the graph have been marked, then the adversary
places the target t at one of the last marked vertices and in this case the query
returns a vertex on the shortest path between t and the queried vertex.
With the above strategy, any algorithm needs to continue querying vertices
until there is no unmarked vertex left. Thus, since at every query the adversary
marks at most 2/ε new vertices, any algorithm needs to perform at least n/2
2/ε =
n · 4ε queries.
2.2
Lower Bound for querying the Approximate Median
The potential ΦS : V → R+ of [9], where S ⊆ V , is defined as P
follows. For any
set S ⊆ V and any vertex v ∈ V , the potential of v is ΦS (v) = u∈S d(v, u). A
vertex x ∈ V is an (1 + ε)-approximate minimizer for the potential Φ over a set
S (i.e. an (1 + ε)-median of S) if ΦS (x) ≤ (1 + ε) minv∈V ΦS (v), where ε > 0.
We prove that an algorithm querying at each
√ iteration always an (1 + ε)-median
of the current candidates’ set S needs Ω( n) queries.
Theorem 3 Let ε > 0. If the algorithm of [9] queries at each
√ iteration an
(1 + ε)-median for the potential function Φ, then at least Ω( n) queries are
required to detect the target t in a graph G with n vertices, even if the graph G
is a tree.
Proof.
We will construct a graph G = (V, E) with n vertices such that
√
Ω( n) queries are needed to locate the target. The graph G will be a tree
with a unique vertex of degree greater than 2, i.e. G
√
√ is a tree that resembles the structure of a star. Formally, G consists of n paths of length n
each, where all these paths have a vertex v0 as a common endpoint. Let
√
Pi = (v0 , vi,1 , vi,2 , . . . , vi,√n−1 , vi,√n ) be the ith path of G. For every i ≤ n
denote by Qi = {vi,2 , vi,3 , . . . , vi,√n } be the
v0 and
S
√ set of vertices of Pi without
vi,1 . Furthermore, for every k ∈ {0, 1, . . . , n} define V−k = V \ ( 1≤i≤k Qi ) to
be the set of vertices left in the graph by keeping only the first edge from each
path Pi , where i ≤ k. Note by definition that V−0 = V .
Having constructed the graph G we are now ready to prove the theorem.
The target will be the vertex v0 . The main idea for the proof is as follows. At
every iteration the central vertex v0 and all its neighbors, who have not yet been
queried, are (1 + ε)-medians, while v0 is the exact median for the potential Φ
7
of [9]. For every k ∈ {0, 1, . . . ,
√
n} we have
√
n
X
√
ΦV−k (v0 ) = k + ( n − k)
j
j=1
√
1 √
= k + ( n − k)(n + n).
2
(1)
Next we compute ΦV−k (vp,1 ) for every p > k. Note that d(vp,1 , vi,1 ) = 2 for
Pk
every i ≤ k, and thus i=1 d(vp,1 , vi,1 ) = 2k. Furthermore, for the vertices on
the path Pp we have
√
n−1
√
n
X
d(vp,1 , vp,i ) =
i=2
X
i=1
i=
√
1
(n − n).
2
Finally, for the remaining of the vertices in V−k (denote them by R) we have
√
X
u∈R
n+1
X
√
d(vp,1 , u) = 1 + ( n − k − 1) ·
j
j=2
√
1 √
= 1 + ( n − k − 1)(n + 3 n).
2
Therefore, it follows that
√
√
1 √
1
ΦV−k (vp,1 ) = 2k + (n − n) + 1 + ( n − k − 1)(n + 3 n).
2
2
(2)
Now note that, due to symmetry, v0 is the exact median of the vertex set
V (with respect to the potential Φ of [9]), that is, ΦV (v0 ) = minx∈V {ΦV (x)}.
√
Furthermore note by (1) and (2) that ΦV−k (vp,1 ) ≥ ΦV−k (v0 ) for every k < n.
Moreover, due to symmetry this monotonicity of ΦV−k (·) is extended to all
vertices vp,2 , vp,3 , . . . , vp,√n , that is, ΦV−k (vp,j ) ≥ ΦV−k (v0 ) for every 1 ≤ j ≤
√
n. Therefore√v0 remains the exact median of each of the vertex sets V−k ,
where 0 ≤ k < n.
Let ε > 0. √Then (1) and (2) imply that ΦV−k (vp,1 ) ≤ (1 + ε)ΦV−k (v0 )
for every k < n and for large enough n. Now assume that the algorithm
of [9] queries always an (1 + ε)-median of the candidates’ set S, where initially
S = V . Then the algorithm may query always a different neighbor of v0 . Due to
symmetry, we may assume without loss of generality that the algorithm queries
the vertices v1,1 , v2,1 , . . . , v√n,1 in this order. Note that these vertices are (1+ε)medians of the candidates’ sets V−0 , V−1 , . . . , V−(√n−1) , respectively. Therefore
√
the algorithm makes
√ at least n queries, where the total number of vertices in
the graph is n − n + 1.
2.3
Upper Bound for querying the Approximate Median
In this section we introduce a new potential function ΓS : V → N for every
S ⊆ V , which overcomes the problem occured in Section 2.2. This new potential
guarantees efficient detection of t in at most O(log n) queries, even when we
always query an (1 + ε)-median of the current candidates’ set S (with respect
8
to the new potential Γ), for any constant 0 < ε < 1. Our algorithm is based
on the approach of [9], however we now query an approximate median of the
current set S with respect to Γ (instead of an exact median with respect to Φ
of [9]).
Definition 1 ( Potential Γ ) Let S ⊆ V and v ∈ V .
max{|N (v, u) ∩ S| : u ∈ N (v)}.
Then ΓS (v) =
Theorem 4 Let 0 ≤ ε < 1. There exists an efficient adaptive algorithm which
log n
queries, by querying at each iteration
detects the target t in at most 1−log(1+ε)
an (1 + ε)-median for the potential function Γ.
Proof. Our proof closely follows the proof of Theorem 3 of [9]. Let S ⊆ V
be an arbitrary set of vertices of G such that t ∈ S. We will show that there
exists a vertex v ∈ V such that ΓS (v) ≤ |S|
2 . First recall the potential ΦS (v) =
P
d(v,
x).
Let
now
v
∈
V
be
a
vertex
such that ΦS (v0 ) is minimized,
0
x∈S
i.e. ΦS (v0 ) ≤ ΦS (v) for every v ∈ V . Let u ∈ N (v0 ) be an arbitrary vertex
+
adjacent to v0 . We will prove that |N (v0 , u)∩S| ≤ |S|
2 . Denote S = N (v0 , u)∩S
−
+
+
and S = S \S . By definition, for every x ∈ S , the edge v0 u lies on a shortest
path from v0 to x, and thus d(u, x) = d(v0 , x) − w(v0 u). On the other hand,
trivially d(u, x) ≤ d(v0 , x) + w(v0 u) for every x ∈ S, and thus in particular for
every x ∈ S − . Therefore ΦS (v0 ) ≤ ΦS (u) ≤ ΦS (v0 ) + (|S − | − |S + |) · w(v0 u),
−
+
and thus |S + | ≤ |S − |. That is, |N (v0 , u) ∩ S| = |S + | ≤ |S|
2 , since S = S \ S .
|S|
Therefore which then implies that ΓS (v0 ) ≤ 2 as the choice of the vertex
u ∈ N (v0 ) is arbitrary.
Let vm ∈ V be an exact median of S with respect to Γ. That is, ΓS (vm ) ≤
ΓS (v) for every v ∈ V . Note that ΓS (vm ) ≤ ΓS (v0 ) ≤ |S|
2 . Now let 0 ≤ ε < 1
and let va ∈ V be an (1 + ε)-median of S with respect to Γ. Then ΓS (va ) ≤
(1 + ε)ΓS (vm ) ≤ 1+ε
2 |S|. Our adaptive algorithm proceeds as follows. Similarly
to the algorithm of [9] (see Theorem 3 of [9]), our adaptive algorithm maintains
a candidates’ set S, where initially S = V . At every iteration our algorithm
queries an arbitrary (1 + ε)-median vm ∈ V of the current set S with respect
to the potential Γ. Let u ∈ N (vm ) be the vertex returned by this query; the
algorithm updates S with the set N (v, u) ∩ S. Since ΓS (va ) ≤ 1+ε
2 |S| as we
proved above, it follows that the updated candidates’ set has cardinality at
most 1+ε
2 |S|. Thus, since initially |S| = n, our algorithm detects the target t
log n
queries.
after at most log( 2 ) n = 1−log(1+ε)
1+ε
Notice in the statement of Theorem 4 that for ε = 0 (i.e. when we always
query an exact median) we get an upper bound of log n queries, as in this case the
size of the candidates’ set decreases by a factor of at least 2. Furthermore notice
that the reason that the algorithm of [9] is not query-efficient when querying
an (1 + ε)-median is that the potential ΦS (v) of [9] can become quadratic in
|S|, while on the other hand the value of our potential ΓS (v) can be at most
|S| by Definition 1, for every S ⊆ V and every v ∈ V . Furthermore notice that,
knowing only the value ΦS (v) for some vertex v ∈ V is not sufficient to provide
a guarantee for the proportional reduction of the set S when querying v. In
contrast, just knowing the value ΓS (v) directly provides a guarantee that, if we
(v)
query vertex v the set S will be reduced by a proportion of ΓS|S|
, regardless
of the response of the query. Therefore, in practical applications, we may not
9
need to necessarily compute an (exact or approximate) median of S to make
significant progress.
3
Detecting Two Targets
In this section we consider the case where there are two unknown targets t1 and
t2 , i.e. T = {t1 , t2 }. In this case the direction query at vertex v returns with
probability p1 (resp. with probability p2 = 1−p1 ) a neighbor u ∈ N (v) such that
t1 ∈ N (v, u) (resp. t2 ∈ N (v, u)). Detecting more than one unknown targets
has been raised as an open question by Emamjomeh-Zadeh et al. [9], while to
the best of our knowledge no progress has been made so far in this direction.
Here we deal with both problems of detecting at least one of the targets and
detecting both targets. We study several different settings and derive both
positive and negative results for them. Each setting differs from the other ones
on the “freedom” the adversary has on responding to queries, or on the power
of the queries themselves. We will say that the response to a query directs to
ti , where i ∈ {1, 2}, if the vertex returned by the query lies on a shortest path
between the queried vertex and ti .
It is worth mentioning here that, if an adversary would be free to arbitrarily
choose which ti each query directs to (i.e. instead of directing to ti with probability pi ), then any algorithm would require at least ⌊ n2 ⌋ (resp. n − 2) queries
to detect at least one of the targets (resp. both targets), even when the graph
is a path. Indeed, consider a path v1 , . . . , vn where t1 ∈ {v1 , . . . , v⌊ n2 ⌋ } and
t2 ∈ {v⌊ n2 ⌋+1 , . . . , vn }. Then, for every i ∈ {1, . . . , ⌊ n2 ⌋}, the query at vi would
return vi+1 , i.e. it would direct to t2 . Similarly, for every i ∈ {⌊ n2 ⌋ + 1, . . . , n},
the query at vi would return vi−1 , i.e. it would direct to t1 . It is not hard to
verify that in this case the adversary could “hide” the target t1 at any of the
first ⌊ n2 ⌋ vertices which is not queried by the algorithm and the target t2 on any
of the last n − ⌊ n2 ⌋ vertices which is not queried. Hence, at least ⌊ n2 ⌋ queries
(resp. n − 2 queries) would be required to detect one of the targets (resp. both
targets) in the worst case.
As a warm-up, we provide in the next theorem an efficient algorithm that
detects with high probability both targets in a tree using O(log2 n) queries.
Theorem 5 For any constant 0 < p1 < 1, we can detect with probability at
2
both targets in a tree with n vertices using O(log 2 n) queries.
least 1 − logn n
Proof. Let G = (V, E) be a tree on n vertices and let T = {t1 , t2 } be the
two targets. The algorithm runs in two phases. In each phase it maintains
a candidates’ set S ⊆ V such that, with high probability, S contains at least
one of the yet undiscovered targets. At the beginning of each phase S = V .
Let without loss of generality p1 ≥ p2 . Furthermore let α = − log1p1 ; note that
α ≥ 1.
The first phase of the algorithm proceeds in log n iterations, as follows. At
the beginning of the ith iteration, where 1 ≤ i ≤ log n, the candidates’ set is
Si ; note that S1 = V at the beginning of the first iteration. Let vi be a median
of Si (with respect to the potential Γ of Section 2.3). In the first iteration we
query the median v1 of V once; let u1 be the response of this query. Then we
know that one of the two targets belongs to the set N (v1 , u1 ), thus we compute
10
the updated candidates’ set S2 = N (v1 , u1 ). Furthermore, since v1 was chosen
to be a median of S1 , it follows that |S2 | ≤ |S21 | = n2 .
For each i ≥ 2, the ith iteration proceeds as follows. We query the median
vi of the set Si for α log n times. First assume that at least one of these α log n
queries at vi directs to a subtree of vi (within Si ) that does not contain the first
median v1 of S1 = V , and let u′i be the response of that query. Then we know
that the subtree of vi (within Si ) which is rooted at u′i contains at least one of
the targets that belong to Si . Thus we compute the updated candidates’ set
Si+1 = Si ∩ N (vi , u′i ), where again |Si+1 | ≤ |S2i | .
Now assume that all of the α log n queries at vi direct to the subtree of
vi that contains the median v1 of the initial candidates’ set S1 = V . Let
u′′i be the (unique) neighbor of vi in that subtree, that is, all α log n queries
at vi return the vertex u′′i . Then we compute the updated candidates’ set
Si+1 = Si ∩ N (vi , u′′i ), where again |Si+1 | ≤ |S2i | . In this case, the probability
that at least one of the targets of Si does not belong to the subtree of vi (within
log n
Si ) which is rooted at u′′i is upper bounded by the probability pα
that each
1
of the α log n queries at vi directs to a target that does not belong to Si . That
log n
is, with probability at least 1 − pα
, at least one of the targets of Si (which
1
we are looking for) belongs to the subtree of vi (within Si ) rooted at u′′i . Since
at each iteration the size of the candidates’ set decreases by a factor of 2, it
follows that |Slog n | = 1. The probability that at each of the log n iterations we
maintained a target from the previous candidates’ set to the next one is at least
log n
log n
= 1 − n1
1 − p1α log n
≥ 1 − logn n by Bernoulli’s inequality. That is,
with probability at least 1 − logn n we detect during the first phase one of the two
targets in log n iterations, i.e. in α log2 n queries in total.
Let t0 be the target that we detected during the first phase. In the second
phase we are searching for the other target t′0 ∈ T \ {t0 }. The second phase of
the algorithm proceeds again in log n iterations, as follows. Similarly to the first
phase, we maintain at the beginning of the ith iteration, where 1 ≤ i ≤ log n, a
candidates’ set Si with median vi , where S1 = V at the beginning of the first
iteration.
For each i ≥ 1, in the ith iteration of the second phase we query α log n
times the median vi of the set Si . First assume that at least one of these α log n
queries at vi directs to a subtree of vi (within Si ) that does not contain the
target t0 that we detected in the first phase, and let u′i be the response of
that query. Then we can conclude that the other target t′0 belongs to the set
N (vi , u′i ), thus we compute the updated candidates’ set Si+1 = Si ∩ N (vi , u′i ),
where |Si+1 | ≤ |S2i | .
Now assume that all of the α log n queries at vi direct to the subtree of
vi that contains the target t0 . Let u′′i be the (unique) neighbor of vi in that
subtree, that is, all α log n queries at vi return the vertex u′′i . Then we compute
the updated candidates’ set Si+1 = Si ∩ N (vi , u′′i ), where again |Si+1 | ≤ |S2i | .
In this case, the probability that the undiscovered target t′0 does not belong to
the subtree of vi (within Si ) which is rooted at u′′i is upper bounded by the
log n
probability pα
that each of the α log n queries at vi directs to t0 . That is,
1
log n
with probability at least 1 − pα
, the target t′0 belongs to the subtree of vi
1
′′
(within Si ) rooted at ui . Since at each iteration the size of the candidates’ set
decreases by a factor of at least 2, it follows that |Slog n | = 1. The probability
11
that at each of the log n iterations we maintained the target t′0 in the candidates’
log n
set is at least 1 − p1α log n
≥ 1 − logn n . That is, with probability at least
1 − logn n we detect in α log2 n queries during the second phase the second target
t′0 , given that we detected the other target t0 in the first phase.
2
Summarizing, with probability at least 1 − logn n
we detect both targets
in 2α log2 n queries.
Since in a tree both targets t1 , t2 can be detected with high probability
in O(log2 n) queries by Theorem 5, we consider in the remainder of the section
arbitrary graphs instead of trees. First we consider in Section 3.1 biased queries,
i.e. queries with p1 > 21 . Second we consider in Section 3.2 unbiased queries,
i.e. queries with p1 = p2 = 12 .
3.1
Upper Bounds for Biased Queries
In this section we consider biased queries which direct to t1 with probability
p1 > 21 and to t2 with probability p2 = 1 − p1 < 12 . As we can detect in this
case the first target t1 with high probability in O(log n) queries by using the
“noisy” framework of [9], our aim becomes to detect the second target t2 with
the fewest possible queries, once we have already detected t1 .
For every vertex v and every i ∈ {1, 2}, denote by Eti (v) = {u ∈ N (v) :
ti ∈ N (v, u)} the set of neighbors of v such that the edge uv lies on a shortest
path from v to ti . Note that the sets Et1 (v) and Et2 (v) can be computed in
polynomial time, e.g. using Dijkstra’s algorithm. We assume that, once a query
at vertex v has chosen which target ti it directs to, it returns each vertex of
Eti (v) equiprobably and independently from all other queries. Therefore, each
of the vertices of Et1 (v) \ Et2 (v) is returned by the query at v with probability
p1
1−p1
|Et (v)| , each vertex of Et2 (v) \ Et1 (v) is returned with probability |Et (v)| , and
1
2
1
each vertex of Et1 (v) ∩ Et2 (v) is returned with probability |Etp1(v)| + |E1−p
. We
t2 (v)|
1
will show in Theorem 6 that, under these assumptions, we detect the second
target t2 with high probability in O(∆ log2 n) queries where ∆ is the maximum
degree of the graph.
The high level description of our algorithm (Algorithm 1) is as follows.
Throughout the algorithm we maintain a candidates’ set S of vertices in which
t2 belongs with high probability. Initially S = V . In each iteration we first
compute an (exact or approximate) median v of S with respect to the potential Γ (see Section 2.3). Then we compute the set Et1 (v) (this can be done
as t1 has already been detected) and we query c∆ log n times vertex v, where
2
1)
c = p7(1+p
2 is a constant. Denote by Q(v) the multiset of size c∆ log n that
1 (1−p1 )
contains the vertices returned by these queries at v. If at least one of these
O(∆ log n) queries at v returns a vertex u ∈
/ Et1 (v), then we can conclude that
u ∈ Et2 (v), and thus we update the set S by S ∩ N (v, u). Assume otherwise
that all O(∆ log n) queries at v return vertices of Et1 (v). Then we pick a vertex
u0 ∈ N (v) that has been returned most frequently among the O(∆ log n) queries
at v, and we update the set S by S ∩ N (v, u0 ). As it turns out, u0 ∈ Et2 (v)
with high probability. Since we always query an (exact or approximate) median
v of the current candidates’ set S with respect to the potential Γ, the size of
S decreases by a constant factor each time. Therefore, after O(log n) updates
12
we obtain |S| = 1. It turns out that, with high probability, each update of the
candidates’ set was correct, i.e. S = {t2 }. Since for each update of S we perform
O(∆ log n) queries, we detect t2 with high probability in O(∆ log2 n) queries in
total.
Algorithm 1 Given t1 , detect t2 with high probability with O(∆ log2 n) queries
2
1)
S ← V ; c ← p7(1+p
2
1 (1−p1 )
while |S| > 1 do
Compute an (approximate) median v of S with respect to potential Γ;
Compute Et1 (v)
Query c∆ log n times vertex v; Compute the multiset Q(v) of these query
responses
if Q(v) \ Et1 (v) 6= ∅ then
Pick a vertex u ∈ Q(v) \ Et1 (v) and set S ← S ∩ N (v, u)
else
Pick a most frequent vertex u ∈ Q(v) and set S ← S ∩ N (v, u)
1:
2:
3:
4:
5:
6:
7:
8:
return the unique vertex in S
9:
Recall that every query at v returns a vertex u ∈ Et1 (v) with probability p1
and a vertex u ∈ Et2 (v) with probability 1 − p1 . Therefore, for every v ∈ V the
multiset Q(v) contains at least one vertex u ∈ Et2 (v) with probability at least
|Q(v)|
|c∆ log n|
1 − p1
= 1 − p1
. In the next lemma we prove that, every time we
update S using Step 8, the updated set contains t2 with high probability.
Lemma 1 Let S ⊆ V such that t2 ∈ S and let S ′ = S ∩ N (v, u) be the updated
set at Step 8 of Algorithm 1. Then t2 ∈ S ′ with probability at least 1 − n2 .
2
7(1+p1 )
1
Proof. Let δ = 1−p
1+p1 and c = p1 (1−p1 )2 be two constants. Recall that each of
the vertices of Et1 (v) \ Et2 (v) is returned by the query at v with probability
1−p1
p1
|Et (v)| , each vertex of Et2 (v) \ Et1 (v) is returned with probability |Et (v)| , and
1
2
1
each vertex of Et1 (v) ∩ Et2 (v) is returned with probability |Etp1(v)| + |E1−p
.
t2 (v)|
1
Observe that these probabilities are the expected frequencies for these vertices
in Q(v). Recall that Step 8 is executed only in the case where Q(v) ⊆ Et1 (v).
To prove the lemma it suffices to show that, whenever Q(v) ⊆ Et1 (v), the most
frequent element of Q(v) belongs to Et1 (v) ∩ Et2 (v) with high probability.
First note that, for the chosen value of δ,
1 − p1
p1
p1
(3)
< (1 − δ)
+
(1 + δ)
|Et1 (v)|
|Et1 (v)| |Et2 (v)|
Let u ∈ Et1 (v) \ Et2 (v), i.e. the query at v directs to t1 but not to t2 . We
define the random variable Zi (u), such that Zi (u) = 1 if u is returned by the i-th
Pc∆ log n
Zi (u).
query at v and Zi (u) = 0 otherwise. Furthermore define Z(u) = i=1
p1
Since Pr(Zi (u) = 1) = |Et (v)| , it follows that E(Z(u)) = c∆ log n |Etp1(v)| by the
1
1
13
linearity of expectation. Then, using Chernoff’s bounds it follows that
2
p1
δ
c∆ log n
Pr(Z(u) ≥ (1 + δ)E(Z(u))) ≤ exp −
3 |Et1 (v)|
(1 + p1 )2
log
n
< exp −2δ 2
(1 − p1 )2
= exp (−2 log n)
1
= 2.
n
(4)
Thus (4) implies that the probability that there exists at least one u ∈
Et1 (v) \ Et2 (v) such that Z(u) ≥ (1 + δ)E(Z(u)) is
p1
1
1
Pr ∃u ∈ Et1 (v) \ Et2 (v) : Z(u) ≥ (1 + δ)
< (∆ − 1) 2 < . (5)
|Et1 (v)|
n
n
Now let u′ ∈ Et1 (v) ∩ Et2 (v). Similarly to the above we define the
random variable Zi′ (u′ ), such that Zi′ (u′ ) = 1 if u′ is returned by the ith query at v and Zi′ (u′ ) = 0 otherwise. Furthermore define Z ′ (u′ ) =
Pc∆ log n ′ ′
1
Zi (u ). Since Pr(Zi′ (u′ ) = 1) = |Etp1(v)| + |E1−p
, it follows that
i=1
t2 (v)|
1
1−p1
p1
E(Z(u)) = c∆ log n |Et (v)| + |Et (v)| by the linearity of expectation. Then
1
2
we obtain similarly to (4) that
2
p1
1 − p1
δ
c∆ log n
+
Pr(Z ′ (u′ ) ≤ (1 − δ)E(Z ′ (u′ ))) ≤ exp −
2 |Et1 (v)| |Et2 (v)|
2
2 (1 + p1 )
< exp −3δ
log n
p1 (1 − p1 )2
< exp (−3 log n)
1
< 2.
n
(6)
Thus, it follows by the union bound and by (3), (5), and (6) that
Pr(∃u ∈ Et1 (v) \ Et2 (v) : Z(u) ≥ Z ′ (u′ )) ≤
2
.
n
(7)
That is, the most frequent element of Q(v) belongs to Et1 (v) ∩ Et2 (v) with
probability at least 1 − n2 . This completes the proof of the lemma.
With Lemma 1 in hand we can now prove the main theorem of the section.
Theorem 6 Given t1 , Algorithm 1 detects t2 in O(∆ log2 n) queries with probability at least (1 − n2 )O(log n) .
Proof. Since we query at each iteration an (1 + ε)-median for the potential
log n
function Γ, recall by Theorem 4 that after at most 1−log(1+ε)
= O(log n) iterations we will obtain |S| = 1. Furthermore, in every iteration the algorithm
queries c∆ log n times the (1 + ε)-median of the current set, and thus the algorithm makes O(∆ log2 n) queries in total. Whenever the algorithm updates S
14
in Step 6 the target t2 belongs to the updated set with probability 1. Moreover,
whenever the algorithm updates S in Step 8, Lemma 1 implies that the target
t2 belongs to the updated set with probability at least (1 − n2 ). Thus, the probability all the O(log n) updates of S were correct, i.e. t2 belongs to S after each
of the O(log n) updates, is at least (1 − n2 )O(log n) .
Note by Theorem 6 that, whenever ∆ = O(poly log n) we can detect both
targets t1 and t2 in O(poly log n) queries. However, for graphs with larger maximum degree ∆, the value of the maximum degree dominates any polylogarithmic
factor in the number of queries. The intuitive reason behind this is that, for
an (exact or approximate) median v of the current set S, whenever deg(v) and
Et1 (v) are large and Et2 (v) ⊆ Et1 (v), we can not discriminate with a polylogarithmic number of queries between the vertices of Et2 (v) and the vertices of
Et1 (v) \ Et2 (v) with large enough probability. Although this argument does not
give any lower bound for the number of queries in the general case (i.e. when
∆ is unbounded), it seems that more informative queries are needed to detect
both targets with polylogarithmic queries in general graphs. We explore such
more informative queries in Section 4.
3.2
Lower Bounds for Unbiased Queries
In this section we consider unbiased queries, i.e. queries which direct to each of
the targets t1 , t2 with equal probability p1 = p2 = 12 . In this setting every query
is indifferent between the two targets, and thus the “noisy” framework of [9]
cannot be applied for detecting any of the two targets. In particular we prove
in the next theorem that any deterministic (possibly adaptive) algorithm needs
at least n2 − 1 queries to detect one of the two targets.
Theorem 7 Let p1 = p2 = 12 . Then any deterministic (possibly adaptive)
algorithm needs at least n2 − 1 queries to detect one of the two targets, even in
an unweighted cycle.
Proof. Let G be the unweighted cycle with n = 2k vertices v0 , v1 , . . . , v2k−1 .
For simplicity denote vi = vi mod 2k for every i ∈ Z. The targets t1 and t2 are
placed by the adversary on two anti-diametrical vertices of the cycle, i.e. t1 = vi
and t2 = vi+k , for some i ∈ {0, 1, . . . , 2k − 1}. Thus, for any vertex vx ∈
/
{t1 , t2 }, the unbiased query at vx returns vx−1 with probability 21 and vx+1
/ {t1 , t2 } the response of the
with probability 21 . That is, for each vertex vx ∈
query at vx is exactly the same. Let A be a deterministic algorithm that queries
at most k−2 different vertices. Then there exist at least two pairs {vi , vi+k } and
{vj , vj+k } of anti-diametrical vertices such that none of these vertices is queried
by the algorithm. Then the adversary can place the two targets either at the
vertices {vi , vi+k } or at the vertices {vj , vj+k }, without affecting the validity of
the previous answers. Thus the algorithm A needs to query at least k−1 = n2 −1
different vertices to detect a target.
In the next theorem we generalize the lower bound of Theorem 7 to the case
of 2c ≥ 2 different targets T = {t1 , t2 , . . . , t2c } and the query to any vertex
1
for every i ∈ {1, 2, . . . , 2c}.
v∈
/ T is unbiased, i.e. pi = 2c
1
for
Theorem 8 Suppose that there are 2c targets in the graph and let pi = 2c
every i ∈ {1, 2, . . . , 2c}. Then, any deterministic (possibly adaptive) algorithm
15
requires at least
cycle.
n
2
− c queries to locate at least one target, even in an unweighted
Proof. Our proof is similar to that of Theorem 7. Let T = {t1 , t2 , . . . , t2c } be
the set of targets. Again, let G be the unweighted cycle with n = 2k vertices
v0 , v1 , . . . , v2k−1 . For each i ∈ {1, 2, . . . , c} the targets {ti , ti+c } are placed by
the adversary on two anti-diametrical vertices of the cycle, i.e. ti = vj and
ti+c = vj+k , for some j ∈ {0, 1, . . . , 2k − 1}. Thus, for any vertex vx ∈
/ T , the
unbiased query at vx returns vx−1 with probability 12 and vx+1 with probability
1
/ T the response of the query at vx is exactly
2 . That is, for each vertex vx ∈
the same (similarly to Theorem 7). Let A be a deterministic algorithm that
queries at most k − c − 1 different vertices. Then there exist at least c + 1
pairs {vi1 , vi1 +k }, {vi2 , vi2 +k }, . . . , {vic , vic +k } of anti-diametrical vertices such
that none of these vertices is queried by the algorithm. Then the adversary
can place the 2c targets any c of these c + 1 pairs of anti-diametrical vertices,
without affecting the validity of the previous answers. Thus the algorithm A
needs to query at least k − c = n2 − c different vertices to detect a target.
4
More Informative Queries for Two Targets
A natural alternative to obtain query-efficient algorithms for multiple targets,
instead of restricting the maximum degree ∆ of the graph (see Section 3.1), is
to consider queries that provide more informative responses in general graphs.
As we have already observed in Section 3.1, it is not clear whether it is possible
to detect multiple targets with O(poly log n) direction queries in an arbitrary
graph. In this section we investigate natural variations and extensions of the
direction query for multiple targets which we studied in Section 3.
4.1
Direction-Distance Biased Queries
In this section we strengthen the direction query in a way that it also returns the
value of the distance between the queried vertex and one of the targets. More
formally, a direction-distance query at vertex v ∈ V returns with probability
pi a pair (u, ℓ), where u ∈ N (v) such that ti ∈ N (u, v) and d(v, ti ) = ℓ. Note
P|T |
that here we impose again that all pi ’s are constant and that
i=1 pi = 1,
where T = {t1 , t2 , . . . , t|T | } is the set of targets. We will say that the response
(u, ℓ) to a direction-distance query at vertex v directs to ti if ti ∈ N (v, u) and
ℓ = d(v, ti ). Similarly to our assumptions on the direction query, whenever
there exist more than one such vertices u ∈ N (v) leading to ti via a shortest
path, the direction-distance query returns an arbitrary vertex u among them
(possibly chosen adversarially). Moreover, if the queried vertex v is equal to one
of the targets ti ∈ T , this is revealed by the query with probability pi . These
direction-distance queries have also been used in [9] for detecting one single
target in directed graphs.
Here we consider the case of two targets and biased queries, i.e. T = {t1 , t2 }
where p1 > p2 . Similarly to Section 3.1, initially we can detect the first target t1
with high probability in O(log n) queries using the “noisy” model of [9]. Thus, in
what follows we assume that t1 has already been detected. We will show that the
second target t2 can be detected with high probability with O(log3 n) additional
16
direction-distance queries using Algorithm 2. The high level description of our
algorithm is the following. We maintain a candidates’ set S such that at every
iteration t2 ∈ S with high probability. Each time we update the set S, its size
decreases by a constant factor. Thus we need to shrink the set S at most log n
times. In order to shrink S one time, we first compute an (1 + ε)-median v of
the current set S and we query log n times this vertex v. Denote by Q(v) the
set of all different responses of these log n direction-distance queries at v. As it
turns out, the responses in Q(v) might not always be enough to shrink S such
that it still contains t2 with high probability. For this reason we also query log n
times each of the log n neighbors u ∈ N (v), such that (u, ℓ) ∈ Q(v) for some
ℓ ∈ N. After these log2 n queries at v and its neighbors, we can safely shrink S
by a constant factor, thus detecting the target t2 with high probability in log3 n
queries.
For the description of our algorithm (see Algorithm 2) recall that, for every
vertex v, the set Et1 (v) = {u ∈ N (v) : t1 ∈ N (v, u)} contains all neighbors of v
such that the edge uv lies on a shortest path from v to t1 .
Algorithm 2 Given t1 , detect t2 with high probability with O(log3 n) directiondistance queries
1: S ← V
2: while |S| > 1 do
3:
Compute an (approximate) median v of S with respect to potential Γ;
Compute Et1 (v)
4:
Query log n times vertex v; Compute the set Q(v) of different query
responses
5:
if there exists a pair (u, ℓ) ∈ Q(v) such that u ∈
/ Et1 (v) or ℓ 6= d(v, t1 )
then
6:
S ← S ∩ N (v, u)
7:
else
8:
for every (u, ℓ) ∈ Q(v) do
9:
Query log n times vertex u; Compute the set Q(u) of different query
responses
10:
if for every (z, ℓ′ ) ∈ Q(u) we have ℓ′ = ℓ − w(vu) then
11:
S ← S ∩ N (v, u); Goto line 2
12:
return the unique vertex of S
In the next theorem we prove the correctness and the running time of Algorithm 2.
Theorem 9 Given t1 , Algorithm
2 detects
t2 in at most O(log3 n) queries with
n
.
probability at least 1 − O log n · plog
1
Proof. Throughout its execution, Algorithm 2 maintains a vertex set S that
contains the second target t2 with high probability. Initially S = V . Let v be
an (1 + ε)-median of the set S (with respect to the potential Γ of Section 2.3) at
some iteration of the algorithm, and assume that t2 ∈ S. We query log n times
vertex v; let Q(v) be the set of all different query responses. Since each query
directs to t1 with probability p1 and to t2 with probability p2 , it follows that at
n
least one of the queries at v directs to t2 with probability at least 1 − plog
.
1
17
Consider a response-pair (u, ℓ) ∈ Q(v). If this query directs to t1 , then
u ∈ Et1 (v) and ℓ = d(v, t1 ). Hence, if we detect at least one response pair
(u, ℓ) ∈ Q(v) such that u ∈
/ Et1 (v) or ℓ 6= d(v, t1 ), we can safely conclude
that this query directs to t2 (lines 5-6 of Algorithm 2). Therefore, in this case,
u ∈ Et2 (v) = {u ∈ N (v) : t2 ∈ N (v, u)}, and thus we safely compute the
updated set S ∩ N (v, u) at line 6.
Assume now that u ∈ Et1 (v) and ℓ = d(v, t1 ) for every response-pair (u, ℓ) ∈
Q(v) (see lines 8-11 of the algorithm). Then every query at v directs to t1 .
However, as we proved above, at least one of these queries (u, ℓ) ∈ Q(v) also
n
. Therefore
directs to t2 (i.e. u ∈ Et2 (v)) with probability at least 1 − plog
1
n
ℓ = d(v, t1 ) = d(v, t2 ) with probability at least 1 − plog
.
Note
that,
in this
1
case, we can not use only the response-pairs of Q(v) to distinguish which query
directs to t2 .
In our attempt to detect at least one vertex u ∈ Et2 (v), we query log n times
each of vertices u such that (u, ℓ) ∈ Q(v). For each such vertex u denote by Q(u)
the set of all different response-pairs from these log n queries at u. Similarly to
the above, at least one of these log n queries at u directs to t2 with probability
n
at least 1 − plog
. Recall that d(v, t2 ) = ℓ and let (z, ℓ′ ) ∈ Q(u). If u ∈ Et2 (v)
1
then d(u, t2 ) = ℓ − w(vu), otherwise d(u, t2 ) > ℓ − w(vu). Furthermore note
that d(u, t1 ) = ℓ − w(vu), since u ∈ Et1 (v). Therefore, if we detect at least
one response-pair (z, ℓ′ ) ∈ Q(u) such that ℓ′ > ℓ − w(vu), then we can safely
conclude that u ∈
/ Et2 (v). Otherwise, if for every response-pair (z, ℓ′ ) ∈ Q(u)
′
we have that ℓ = ℓ − w(vu), then u ∈ Et2 (v) (i.e. t2 ∈ N (v, u)) with probability
n
at least 1 − plog
.
1
Recall that there exists at least one query at v that directs to t2 with probn
, as we proved above. That is, among all response-pairs
ability at least 1 − plog
1
(u, ℓ) ∈ Q(v) there exists at least one vertex u ∈ Et2 (v) with probability at
n
least 1 − plog
. Therefore, we will correctly detect a vertex u ∈ Et2 (v) at lines
1
2
n
, i.e. with at least
10-11 of the algorithm with probability at least 1 − plog
1
this probability the updated candidates’ set at line 11 still contains t2 . Thus,
log n
= O(log n) times, we eventually
since we shrink the candidates’ set 1−log(1+ε)
detect t2 as the unique vertex in the final candidates’ set with probability at
O(log n)
n
n
≥ 1 − O(log n · plog
) by Bernoulli’s inequality. Finally,
least 1 − plog
1
1
it is easy to verify from the above that the algorithm will terminate after at
n
).
most O(log 3 n) queries with probability at least 1 − O(log n · plog
1
4.2
Vertex-Direction and Edge-Direction Biased Queries
An alternative natural variation of the direction query is to query an edge instead of querying a vertex. More specifically, the direction query (as defined in
Section 1.2) queries a vertex v ∈ V and returns with probability pi a neighbor
u ∈ N (v) such that ti ∈ N (u, v). Thus, as this query always queries a vertex, it can be also referred to as a vertex-direction query. Now we define the
edge-direction query as follows: it queries an ordered pair of adjacent vertices
(v, u) and it returns with probability pi YES (resp. NO) if ti ∈ N (v, u) (resp. if
ti ∈
/ N (v, u)). Similarly to our notation in the case of vertex-direction queries,
we will say that the response YES (resp. NO) to an edge-direction query at the
vertex pair (v, u) refers to ti if ti ∈ N (v, u) (resp. if ti ∈
/ N (v, u)). Similar
18
but different edge queries for detecting one single target on trees have been
investigated in [9, 12, 20, 25].
Here we consider the case where both vertex-direction and edge-direction
queries are available to the algorithm, and we focus again to the case of two targets and biased queries, i.e. T = {t1 , t2 } where p1 > p2 . Similarly to Sections 3.1
and 4.1, we initially detect t1 with high probability in O(log n) vertex-direction
queries using the “noisy” model of [9]. Thus, in the following we assume that
t1 has already been detected. We will show that Algorithm 3 detects the second target t2 with high probability using O(log2 n) additional vertex-direction
queries and O(log3 n) edge–direction queries, i.e. in total O(log3 n) queries.
Algorithm 3 Given t1 , detect t2 with high probability with O(log3 n) vertexdirection and edge-direction queries
1: S ← V
2: while |S| > 1 do
3:
Compute an (approximate) median v of S with respect to potential Γ;
Compute Et1 (v)
4:
Apply log n vertex-direction queries at vertex v; Compute the set Q(v)
of different query responses
5:
if there exists a vertex u ∈ Q(v) such that u ∈
/ Et1 (v) then
6:
S ← S ∩ N (v, u)
7:
else
8:
for every u ∈ Q(v) do
9:
Apply log n edge-direction queries at (v, u); Compute the set Q(v, u)
of different query responses
10:
if Q(v, u) = {YES} then
11:
S ← S ∩ N (v, u); Goto line 2
12:
return the unique vertex of S
In the next theorem we prove the correctness and the running time of Algorithm 3.
Theorem 10 Given t1 , Algorithm 3 detects t2 in at most O(log2 n) vertexdirection queries and O(log 3 n) edge–direction queries with probability at least
n
1 − O(log n · plog
).
1
Proof. The proof follows a similar approach as the proof of Theorem 9.
Throughout its execution, Algorithm 3 maintains a vertex set S that contains
the second target t2 with high probability. Initially S = V . Let v be an (1 + ε)median of the set S (with respect to the potential Γ of Section 2.3) at some
iteration of the algorithm, and assume that t2 ∈ S. We query log n times vertex
v; let Q(v) be the set of all different query responses. Similarly to the analysis
of Algorithm 2 in the proof of Theorem 9, at least one of the queries at v directs
n
to t2 with probability at least 1 − plog
.
1
Consider a response-vertex u ∈ Q(v). If this query directs to t1 , then u ∈
/ Et1 (v), we
Et1 (v). Hence, if we detect at least one u ∈ Q(v) such that u ∈
can safely conclude that this query directs to t2 (lines 5-6 of Algorithm 3).
Therefore, in this case, u ∈ Et2 (v) = {u ∈ N (v) : t2 ∈ N (v, u)}, and thus we
safely compute the updated set S ∩ N (v, u) at line 6.
19
Assume now that u ∈ Et1 (v) for every response u ∈ Q(v) (see lines 8-11 of
the algorithm). Then every query at v directs to t1 , although at least one of
n
,
them also directs to t2 (i.e. Q(v)∩Et2 (v) 6= ∅) with probability at least 1−plog
1
as we proved above. Note that, in this case, we can not use only the vertices of
Q(v) to distinguish which query directs to t2 .
In our attempt to detect at least one vertex u ∈ Et2 (v), we apply log n
edge-direction queries at each of the ordered pairs (v, u), where u ∈ Q(v). For
each such pair (v, u) denote by Q(v, u) the set of all different YES/NO responses
from these log n queries at (v, u). Similarly to the above, at least one of these
n
log n queries at (v, u) refers to t2 with probability at least 1 − plog
. Therefore,
1
if NO∈ Q(v, u), then we can safely conclude that u ∈
/ Et2 (v). Otherwise, if
Q(v, u) = {YES}, then u ∈ Et2 (v) (i.e. t2 ∈ N (v, u)) with probability at least
n
1 − plog
.
1
Recall that there exists at least one query at v that directs to t2 with proban
bility at least 1−plog
. That is, among all responses in Q(v) there exists at least
1
n
. Therefore, we will corone vertex u ∈ Et2 (v) with probability at least 1 − plog
1
rectly detect a vertex u ∈ Et2 (v) at lines 10-11 of the algorithm with probability
2
n
, i.e. with at least this probability the updated candidates’
at least 1 − plog
1
set at line 11 still contains t2 . Thus, similarly to the proof of Theorem 9, we
eventually detect t2 as the unique vertex in the final candidates’ set with proban
bility at least 1 − O(log n·plog
). Finally, it is easy to verify from the above that
1
the algorithm will terminate after at most O(log2 n) vertex-direction queries and
n
log3 n edge–direction queries with probability at least 1 − O(log n · plog
).
1
4.3
Two-Direction Queries
In this section we consider another variation of the direction query that was
defined in Section 1.2 (or “vertex-direction query” in the terminology of Section 4.2), which we call two-direction query. Formally, a two-direction query at
vertex v returns an unordered pair of (not necessarily distinct) vertices {u, u′ }
such that t1 ∈ N (v, u) and t2 ∈ N (v, u′ ). Note here that, as {u, u′ } is an unordered pair, the response of the two-direction query does not clarify which of
the two targets belongs to N (v, u) and which to N (v, u′ ).
Although this type of query may seem at first to be more informative than
the standard direction query studied in Section 3, we show that this is not
the case. Intuitively, this type of query resembles the unbiased direction query
of Section 3.2. To see this, consider e.g. the unweighted cycle where the two
targets are placed at two anti-diametrical vertices; then, applying many times
the unbiased direction query of Section 3.2 at any specific vertex v reveals with
high probability the same information as applying a single two-direction query
at v. Based on this intuition the next theorem can be proved with exactly the
same arguments as Theorem 7 of Section 3.2.
Theorem 11 Any deterministic (possibly adaptive) algorithm needs at least n2 −
1 two-direction queries to detect one of the two targets, even in an unweighted
cycle.
20
4.4
Restricted Set Queries
The last type of queries we consider is when the query is applied not only to
a vertex v of the graph, but also to a subset S ⊆ V of the vertices, and the
response of the query is a vertex u ∈ N (v) such that t ∈ N (v, u) for at least one
of the targets t that belong to the set S. Formally, let T be the set of targets.
The restricted-set query at the pair (v, S), where v ∈ V and S ⊆ V such that
T ∩S 6= ∅, returns a vertex u ∈ N (v) such that t ∈ N (v, u) for at least one target
t ∈ T ∩ S. If there exist multiple such vertices u ∈ N (v), the query returns one
of them adversarially. Finally, if we query a pair (v, S) such that T ∩ S = ∅,
then the query returns adversarially an arbitrary vertex u ∈ N (v), regardless of
whether the edge vu leads to a shortest path from v to any target in T . That
is, the response of the query can be considered in this case as “noise”.
In the next theorem we prove that this query is very powerful, as |T | · log n
restricted-set queries suffice to detect all targets of the set T .
Theorem 12 Let T be the set of targets. There exists an adaptive deterministic algorithm that detects all targets of T with at most |T | · log n restricted-set
queries.
Proof. To detect the first target we simply apply binary search on graphs. At
every iteration we maintain a candidates’ set S (initially S = V ). We compute
a median v of S (with respect to the potential Γ of Section 2.3) and we query
the pair (v, S). If the response of the query at (v, S) is vertex u ∈ N (v) then we
update the candidates’ set as S ∩ N (v, u). We know that there is at least one
target in the updated set S and that the size of the candidates’ set decreased by
a factor of at least 2 (cf. Theorem 4). Thus, after at most log n restricted-set
queries we end up with a candidates’ set of size 1 that contains one target.
We repeat this procedure for another |T | − 1 times to detect all remaining targets of T ,as follows. Assume that we have already detected the targets t1 , t2 , . . . , ti ∈ T . To detect the next target of T we initially set S =
V \ {t1 , t2 , . . . , ti } and we apply the above procedure. Then, after at most
log n restricted-set queries we detect the next target ti+1 . Thus, after at most
|T | · log n restricted-set queries in total we detect all targets of T .
5
Conlusions
This paper resolves some of the open questions raised by Emamjomeh-Zadeh
et al. [9] and makes a first step towards understanding the query complexity
of detecting two targets on graphs. Our results provide evidence that different
types of queries can significantly change the difficulty of the problem and make
it from almost trivial impossible to solve.
There are several interesting avenues for future research both for detecting
one target and for detecting multiple targets. The potential Γ we introduced
in this paper has several interesting properties that have not yet been fully
explored. As we mentioned in the paper, just knowing the value ΓS (v) for
a vertex v directly provides enough information to quantify the “progress” a
direction query can make by querying vertex v, without the need to know the
values ΓS (u) for any other vertex u 6= v. This property of Γ may be exploited
to provide computationally more efficient algorithms for detecting one target;
21
an algorithm might only need to compute ΓS (v) for all vertices v lying within a
wisely chosen subset such that one of these vertices is an approximate median.
Of course, this approach cannot break the log n lower bound on the number of
queries needed to detect the target (e.g. in the path of n vertices), but it could
potentially improve the computational complexity of the detection algorithm.
Furthermore, the potential Γ might be a useful tool for deriving an optimal
number of queries for classes of graphs other than trees, since every exact median
of Γ separates the graph into roughly equal subgraphs.
For the setting where two, or more, targets need to be detected there is a
plethora of interesting questions. We believe that the most prominent one is to
derive lower bounds on the number of queries needed to detect both targets in
the biased setting. Preliminary results suggest a lower bound of log n log log n
bound for a special type of algorithms. A general lower bound seems to require
new techniques. Another interesting question is to decide whether there exists
an algorithm that detects both targets with a polylogarithmic number of direction queries. Finally, an intriguing question is to find the minimal requirements
a query has to satisfy in order to detect even one target in the unbiased setting.
References
[1] Y. Ben-Asher, E. Farchi, and I. Newman. Optimal search in trees. SIAM
J. Comput., 28(6):2090–2102, 1999.
[2] M. Ben-Or and A. Hassidim. The bayesian learner is optimal for noisy
binary search (and pretty good for quantum as well). In 49th Annual IEEE
Symposium on Foundations of Computer Science, FOCS 2008, October 2528, 2008, Philadelphia, PA, USA, pages 221–230, 2008.
[3] L. Boczkowski, A. Korman, and Y. Rodeh. Searching on trees with noisy
memory. CoRR, abs/1611.01403, 2016.
[4] R. Carmo, J. Donadelli, Y. Kohayakawa, and E. S. Laber. Searching in
random partially ordered sets. Theor. Comput. Sci., 321(1):41–57, 2004.
[5] F. Cicalese, T. Jacobs, E. S. Laber, and M. Molinaro. On the complexity
of searching in trees and partially ordered structures. Theor. Comput. Sci.,
412(50):6879–6896, 2011.
[6] F. Cicalese, T. Jacobs, E. S. Laber, and C. D. Valentim. The binary
identification problem for weighted trees. Theor. Comput. Sci., 459:100–
112, 2012.
[7] D. Dereniowski. Edge ranking and searching in partial orders. Discrete
Applied Mathematics, 156(13):2493–2500, 2008.
[8] D. Du and F. K. Hwang. Combinatorial Group Testing and its Applications.
World Scientific, Singapore, 1993.
[9] E. Emamjomeh-Zadeh, D. Kempe, and V. Singhal. Deterministic and probabilistic binary search in graphs. In Proceedings of the 48th Annual ACM
SIGACT Symposium on Theory of Computing, STOC 2016, Cambridge,
MA, USA, June 18-21, 2016, pages 519–532, 2016.
22
[10] U. Feige, P. Raghavan, D. Peleg, and E. Upfal. Computing with noisy
information. SIAM J. Comput., 23(5):1001–1018, 1994.
[11] E. Fonio, Y. Heyman, L. Boczkowski, A. Gelblum, A. Kosowski, A. Korman, and O. Feinerman. A locally-blazed ant trail achieves efficient collective navigation despite limited information. eLife, page 23 pages, 2016.
[12] A. V. Iyer, H. D. Ratliff, and G. Vijayan. Optimal node ranking of trees.
Inf. Process. Lett., 28(5):225–229, 1988.
[13] C. Jordan. Sur les assemblages de lignes. Journal für die reine und angewandte Mathematik, 70:195–190, 1869.
[14] E. S. Laber, R. L. Milidiú, and A. A. Pessoa. On binary searching with
non-uniform costs. In Proceedings of the Twelfth Annual Symposium on
Discrete Algorithms, January 7-9, 2001, Washington, DC, USA., pages
855–864, 2001.
[15] T. W. Lam and F. L. Yue. Edge ranking of graphs is hard. Discrete Applied
Mathematics, 85(1):71–86, 1998.
[16] T. W. Lam and F. L. Yue. Optimal edge ranking of trees in linear time.
Algorithmica, 30(1):12–33, 2001.
[17] N. Linial and M. E. Saks. Searching ordered structures. J. Algorithms,
6(1):86–103, 1985.
[18] S. Mozes, K. Onak, and O. Weimann. Finding an optimal tree searching
strategy in linear time. In Proceedings of the Nineteenth Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2008, San Francisco,
California, USA, January 20-22, 2008, pages 1096–1105, 2008.
[19] N. J. Nilsson. Problem-Solving Methods in Artificial Intelligence. McGrawHill Pub. Co., 1971.
[20] R. Nowak. Noisy generalized binary search. In Y. Bengio, D. Schuurmans,
J. D. Lafferty, C. K. I. Williams, and A. Culotta, editors, Advances in
Neural Information Processing Systems 22, pages 1366–1374. Curran Associates, Inc., 2009.
[21] K. Onak and P. Parys. Generalization of binary search: Searching in trees
and forest-like partial orders. In 47th Annual IEEE Symposium on Foundations of Computer Science (FOCS 2006), 21-24 October 2006, Berkeley,
California, USA, Proceedings, pages 379–388, 2006.
[22] J. Pearl. Heuristics - intelligent search strategies for computer problem
solving. Addison-Wesley series in artificial intelligence. Addison-Wesley,
1984.
[23] A. Pelc. Searching games with errors - fifty years of coping with liars.
Theor. Comput. Sci., 270(1-2):71–109, 2002.
[24] A. Renyi. On a problem in information theory. Magyar Tud. Akad. Mat.
Kutato Int. Kozl, 6(B):505–516, 1961.
23
[25] A. A. Schäffer. Optimal node ranking of trees in linear time. Information
Processing Letters, 33(2):91–96, 1989.
[26] S. Ulam. Adventures of a Mathematician. University of California Press,
1991.
24
| 8cs.DS
|
International Journal of Programming Languages and Applications ( IJPLA ) Vol.4, No.1, January 2014
ISSN: 1839 - 6291
TOWARDS HIGH PERFORMANCE COMPUTING
(HPC) THROUGH PARALLEL PROGRAMMING
PARADIGMS AND THEIR PRINCIPLES
Dr. Brijender Kahanwal
Department of Computer Science & Engineering, Galaxy Global Group of Institutions,
Dinarpur, Ambala, Haryana, India
ABSTRACT
Nowadays, we are to find out solutions to huge computing problems very rapidly. It brings the idea of
parallel computing in which several machines or processors work cooperatively for computational tasks.
In the past decades, there are a lot of variations in perceiving the importance of parallelism in computing
machines. And it is observed that the parallel computing is a superior solution to many of the computing
limitations like speed and density; non-recurring and high cost; and power consumption and heat
dissipation etc. The commercial multiprocessors have emerged with lower prices than the mainframe
machines and supercomputers machines. In this article the high performance computing (HPC) through
parallel programming paradigms (PPPs) are discussed with their constructs and design approaches.
KEYWORDS
Parallel programming languages, parallel programming constructs, distributed computing, high
performance computing
1. INTRODUCTION
The numerous computational concentrated tasks of the computer science like weather forecast,
climate research, the exploration of oil and gas, molecular modelling, quantum mechanics, and
physical simulations are performed by the supercomputers as well as mainframe computer. But
these days due to the advancements in the technology multiprocessors systems or multi-core
processor systems are going to be resembled to perform such type of computations which are
performed by the supercomputing machines.
Due to the recent advances in the hardware technologies, we are leaving the von Neumann
computation model and adopting the distributed computing models which have peer-to-peer
(P2P), cluster, cloud, grid, and jungle computing models in it [1]. All these models are used to
achieve the parallelism and are high performance computing (HPC) models.
Concurrency and Parallelism: The terms concurrency and parallelism must be clear in our
minds first. It can be well explained with the help of the threads (light weight processes). When
two or more threads are in the middle of execution process at the same time, actually, they may
or may not be executing at the same time, but they are in the middle of it.
DOI : 10.5121/ijpla.2014.4104
45
International Journal of Programming Languages and Applications ( IJPLA ) Vol.4, No.1, January 2014
It is called concurrency [25]. The threads may or may not execute at the single processor or a
multiprocessor machine in it. When two or more threads are actually running at the same time
on different CPUs, it is known as parallelism [25]. To achieve parallelism, we always require at
least two CPUs which may be on a single machine (multiprocessor machine) or more machines.
The parallel events may also be called as concurrent events, but the reverse is not true always. It
is well described with the help of set theory that parallelism ⊂ concurrency (Parallelism is
contained in Concurrency) as shown in Fig.1.
Figure 1: Concurrency is the superset of Parallelism
On the Von Neumann computing machines, the programming is a single execution sequence.
But there might be various subroutines that can be executed simultaneously within a single
program. These are called sequential due to the execution of subroutines proceeds in
predetermined sequence. In general cases, the programs are termed as concurrent or parallel in
which the subroutines can be executed concurrently and these subroutines are known as tasks
[2].
Now a day, it is a common practice now to execute various programs concurrently by the
computing machines. It may have the architecture with multiprocessors (various CPUs) which
share the common memory space as shown in the Fig.2 (a) or another architecture that may
have multiprocessors with their independent memories or distributed memories as shown in the
Fig.2 (b).
(a)
Figure 2 (a): Shared Architecture.
(b)
Figure 2 (b): Distributed Architecture.
It is the big challenge for the scientists to utilize these hardware technologies efficiently,
effectively, and these processors may work cooperatively. In the present scenario, software
technologies (STs) are not having well compatibility with the hardware technology’s growth
that the STs can utilize them efficiently and effectively. Hence the parallel computing
community is going to be aware that they can build software technologies which are efficient
and effective. But till now the programmers as well as the scientists are incompetent to find the
solution. So it is the need to get more awareness regarding the parallelism, so we can find the
better solutions for high performance computing (HPC). The article contains more sections
which are organized as follows: the related works will be available in the section 2.
46
International Journal of Programming Languages and Applications ( IJPLA ) Vol.4, No.1, January 2014
2. RELATED WORKS
The parallelism is not a novel concept of computing. The Law of Amdahl is the key principle to
estimate maximum improvements in the components of the system [3] which brings the idea of
parallel computing to find the optimum performance. The time 1960-70 was the boom time for
the parallel computing and during this time, we have solved many problems in achieving the
optimum performance, but they have encountered today. Multi-core chips are a new paradigm
in parallelism. Parallel computation is the never-ending desire for much faster and much
cheaper computation of level of supercomputers as well as mainframe computers [4]. Until now
we have not remarkable progress in building the efficient and optimal softwares for utilizing the
parallel computer architectures of today [5].
3. PARALLEL PROGRAMMING CONSTRUCTS OR PRINCIPLES
It is complicated to write the parallel programs as compared to write sequential programs. We
design algorithms and express them in some programming languages to execute on the
computing machines. In the case of parallel programming we have to develop the same
functioning, but it also adds more challenges to it. Such types of challenges are as follows:
structured constructs[6]: structured region; thread based constructs[7]: synchronization, critical
sections, and deadlock; and object-oriented constructs [8]: object replication, latency hiding,
termination detection, and user-level scheduling; concurrency; data distribution; inter-process
communication; computational load balancing; variable definitions [2]; parallel compositions
[2]; program structures [2]; and easy implementation and debugging;. All of these are explored
in the following sub sections.
3.1. Structured Construct – Structured Region
The structured parallel programming construct is introduced as a structured region. It has a
region name and a region body which is enclosed with two barriers namely entry barrier as well
as exit barrier. A par (or parfor) block has the starting instruction as the region name. A region
name has a region keyword then an arbitrary name given to the region by the programmer and a
list of the participants (processes) [6]. If the participants list is declared explicitly then the
specified names becomes the participants and if the list is not mentioned then all the processes
are the participants.
This structured region semantics is very easy and clear. The entry in the region is done only if
all the participants reach at their specific entry points. There are the unique effects of the
execution of the region body. The region is exited by all the participants after the complete
execution of all the operations of the region body. A structured region has a single entry and
exit point [6]. It wraps up inter-process communication as well as synchronization operations in
it and makes the parallel programming easier to understand and less error-prone. It is opposite
to the concept of mutual exclusion in which only a single process can enter in the critical region.
But in the case hare all the processes can enter in the structured region.
3.2. Thread-Based Constructs
The process and the thread are much related terms. A process is program in execution and there
may be some independent units within a process which are known as the threads. A thread is
dispatch-able work unit. It is also known as the light-weight process. So it is concluded that the
47
International Journal of Programming Languages and Applications ( IJPLA ) Vol.4, No.1, January 2014
threads makes a process or it is a subset of a process [10]. Both the process as well as the thread
is an active entity and a simple program before execution is a passive entity. The threads of a
single process share same address space, so the context-switching as well as communication
between threads is inexpensive [10]. The sharing between the threads creates some difficulties
which are explored in the following sub-sections.
1. Synchronization: It is the construct which enforces the mechanism for controlling the
execution order of the threads and resolves the conflicts among the threads [7]. It is a way
of coordinating the execution of the threads and managing the shared address space.
In synchronization, mutual exclusion and condition synchronization operations are used
widely. In mutual exclusion, one of the threads block the critical section (shared data area
by the threads) and other threads will wait for getting their turns one by one. The scheduler
controls for the turns. But in the case of conditional synchronization, threads are blocked
until some particular condition is satisfied. Here the thread has to wait until a particular
condition is achieved. So the synchronization is well managed by the programmer or by the
programming system, it is a critical construct for multi-threaded programming.
2. Critical sections: These sections have shared dependency variables and many threads are
dependent on them [7]. It is the great programming construct for thread-based
programming, so the threads can use these sections mutually exclusively and prevent to use
these sections simultaneously. These sections should be minimized in size.
3. Deadlock: It is the situation when a thread holds a lock and waiting for another lock which
is held by another thread and this thread is waiting for the lock first to be released. Such as
the code: T1: lock (1); lock (2); and T2: lock (2); lock (1); in this code, the deadlock may
or may not occur. The four basic conditions need to be hold which are mutual exclusion;
hold-and-wait; no pre-emption; and circular wait.
3.3. Object-Oriented Constructs
The object-oriented parallel programming has complex computational as well as
communicational structures to achieve the efficiency or optimization. For improving the
performance in the object-oriented programming languages some of the constructs are discussed
in the following sub-sections [8].
1. Object Replication: This construct highly improves the performance in the distributed
memory architectures. When a program is frequently accessing an object then it is better to
create a local replica of it for the processor and then there is a big fall in the number of
remote messages [8].
2. Latency Hiding: It is an optimization technique which reduces the waiting time for the
remote messages. Here local computations and remote communications are overlapped. In it
we break up a single thread into multiple threads manually by modifying the program [8].
3. Termination Detection: In few parallel applications like search problems, it is a typical task
to detect the termination point because of the invoking of the many threads and finding their
termination points in absence of the global control on them [8].
4. User-level Scheduling: A proper scheduling at application level also improves the
performance of the parallelism. User-level scheduling facility is not offered in most of the
programming language systems, so it becomes necessary to provide it by the programmers
explicitly to control the order of execution [8].
48
International Journal of Programming Languages and Applications ( IJPLA ) Vol.4, No.1, January 2014
3.4. Concurrency
These days, the processors are inexpensive as compared to the previous time, so we are
constructing the distributed systems. Due to such things in the development the concurrency has
a little importance. The programmers are working on various types of applications like DBMS,
L-S parallel technical computations, real time applications, and embedded systems etc [9].
When a concurrent program shares one or more processors during execution is known as
multiprogramming; when its sub-processes are to be executed on independent processor then it
is known as multiprocessing; when there is the addition of the communication network then it is
known as distributed processing; and any combination of these is known as hybrid approach [9].
Still, it is the fundamental construct to utilize optimally the parallel computing resources. We
can't achieve the parallelism without dividing the operations to execute concurrently. A problem
has many sub-problems in it for concurrent execution; there is a need to differentiate the
concurrent tasks within the main problem. It is the dexterity of the programmers.
3.5. Data Distribution
It is a big challenge to distribute the data which creates problem. In the parallelism there are so
many processors which are working cooperatively. Now a day, the principle of locality is
important for the better performance of the systems. But in the case of parallelism it becomes
the problem or a decision making event which data to be localized for the particular processor.
It is due to the concept of independent cache memories for each processor in the shared memory
systems. For the parallel programmers, it becomes issue to manage it carefully. The
performance of the system increases as we store more data in the caches because the processor
can access it quickly as compared to the shared memory area.
3.6. Inter-process Communication
When we are going to execute a process on two or more processors, it becomes necessary to
make communication among them for transferring data from one processor's cache memory to
another processor's cache memory. So here is the need of maintaining caches of the processors
with the mechanism called cache coherence that may be implemented via hardware or cache
coherence protocols. Another case may be that the processors may have distributed memories
and all the processors need to be communicated properly. There may be the need of explicit
calls to a library which require transferring values among processors. There may be the
communication overheads which must be minimized to get the advantage of the parallelism.
3.7. Computational Load Balancing
In the parallelism, there are two or more processors or separate machines which are connected
through the network, to take the advantage of the parallelism all the processors or machines
must be utilized properly and equally. The total computation must be equally distributed among
the processors or machines for getting the benefits of high performance computing.
3.8. Variable Definitions
Two types of variables may be used in the programming languages namely mutable and
definitional. The mutable variables are the normal variables which are used in the sequential
programming languages. The assignment may be done to the variables and that may change
during the program execution. The definitional variables are those variables in which we can
49
International Journal of Programming Languages and Applications ( IJPLA ) Vol.4, No.1, January 2014
assign values only once and they can be accessed by any number of tasks. In such variables
there is no need to maintain synchronization.
3.9. Parallel Compositions
In the execution process, the statements are executed one after another and they also have
additional sequential as well as conditional statements in the sequential programming language.
To achieve the parallelism, the parallel statements must be added which becomes the additional
threads of control to start the execution.
3.10. Program Structures
There may be two types of parallel program execution models. Firstly, transformational in
which the main task is to transform the input data into the correct output value. Secondly,
reactive or responsive in which the programs works regarding the events which are the external
one.
3.11. Ease of Programming and Debugging
This is the issue for every type of programming language. The parallel programs must be easily
implemented by the programmers. They do not require thinking more about the parallelism.
The parallelism should be tackled by the programming language platforms. It is common to be
bugs in the program implementation and there are so many side-effects of these bugs. So these
may be removed easily with the help of good debugging tools.
4. PARALLEL PROGRAMMING APPROACHES
There are basically three approaches to program high performance computers (parallel
computers). These are as follows:
4.1. Implicit Parallelism
It is also known as automatic parallelism. This approach is headache free for the programmer’s
point of view; here the complete working is done by the compilers to make parallel all of the
executions [11]. All the parallel language constructs are inherently implemented by the
language platform. Such type of job is always done in the pure functional programming
languages. With the help of this approach the existing code is utilized on parallel systems. No
changing is required in the existing code. It saves the development costs. And it is attractive for
the vendors of the high performance computing.
Such type of parallelism has its own advantages as well as drawbacks. The advantages are as
follows: Firstly, programmer's attention is completely on the algorithms. Secondly, we require
very less code for programming. Thirdly, the productivity of the programmers increases as
he/she does not care about the parallel programming constructs. Fourthly, the definitions of the
algorithms are separated from the parallel executions. Fifthly, the legacy systems are utilized
properly and which is the concept of re-usability.
The drawbacks are as follows: Firstly, the complete parallelism is not achieved because the
programmers have much more information of the parallel potential (not efficient). Secondly, the
50
International Journal of Programming Languages and Applications ( IJPLA ) Vol.4, No.1, January 2014
programmers have not the exact control over the parallelism. Thirdly, there is no optimum
parallel efficiency achieved. Fourthly, the algorithms which are already implemented may be
executed with a low configuration system (architecture + memory) a few or more decade ago.
But the recent configurations are very high with more storage capacity and processor speeds.
Fifthly, it is a tough task for the scientists and researchers to design the parallel compilers.
4.2. Explicit Parallelism
In it the existing programming languages are utilized. The proper extensions are made to them
to achieve all the parallel programming constructs [11]. Here the parallel programming
principles are defined explicitly by the programmers. Explicit threading is a sub-approach of
explicit parallelism in which the programmers creates parallel threads explicitly [22, 23].The
explicit parallelism also has its own advantages and disadvantages.
The advantages are as follows: Firstly, the programmers are already trained in the existing
language. Secondly, it is totally under the understanding and control of the programmers. The
disadvantages are as follows: Firstly, it is very hard to debug and difficult to program for the
programmers because everything is dependent on the creativity and thinking the programmers.
Secondly, there is no standardization because there are so many extensions have been made by
the developers with same functionality with different look.
4.3. Hybrid Parallelism
It is the mixed up approach which combines the features of implicit as well as explicit
parallelism. It will take the advantages of both the above mentioned technique.
It is summarized that the language designers may design completely the new programming
language paradigms which have all the parallel programming principles or constructs in it.
5. PARALLEL PROGRAMMING PARADIGMS
There are too many paradigms available for utilizing the current machines with parallel
architectures. Some of the parallel programming languages are as follows:
5.1. Message Passing Interface (MPI)
It is a specification for message passing. It is de facto standard for the development of high
performance computing applications for the distributed systems (heterogeneous networks) as
well as parallel computers and clusters [12]. It has bindings for C, C++ and FORTRAN
programming languages. It is highly portable environment. The workload partitioning as well as
the work mapping are done explicitly be the programmers like Pthread [25] and UPC. All the
communications between the processes take place with the help of message passing paradigm.
In it one process sends the data to another process through message passing.
5.2. Fortress
It is also a thread-based specification programming language to design the HPC applications
[12]. The work management, workload portioning, as well as work mapping may be done
implicitly by the compiler as well as explicitly by the programmers. All for loops are parallel by
51
International Journal of Programming Languages and Applications ( IJPLA ) Vol.4, No.1, January 2014
default as implicit approach. The synchronization principles like reductions as well as atomic
expression are specified by the programmers when there is a data competition in a program.
5.3. POSIX Threads (Pthreads) Programming
It is actually a set of C language types as well as procedure calls and all these are maintained or
defined in a library named as pthread.h [12]. It is the duty of the programmers to maintain the
shared data among the threads for avoiding the deadlocks and data races [25]. The pthread's
create function has four parameters the task run thread, attribute, tasks to run in routine call, and
routine argument. All has been closed with the help of pthread’s exit function call. The
workload partitioning and work mapping is done explicitly by the programmers.
5.4. OpenMP
It is also thread based open specification for shared memory architectures. It provides compiler
directives, callable runtime library, and environment variables which extends the existing
programming languages C, C++, and FORTRAN. It is portable platform [12]. The worker
management is done impliedly and a little programmer's effort is required for the workload
partitioning and task mappings, they are also performed implicitly. Programmers are required to
tell the parallel region with the help of the compiler directives. The synchronization is also
maintained implicitly by the OpenMP.
5.5. CILK (pronunciation as ‘silk’)
It is a multi-threaded programming language. It is appropriate for the recent multi-core CPU
architectures. It is based on the traditional programming language C. Cilk a true parallel
extension to C semantically with good performance [13]. In 1994, it was designed by the MIT
scientists. In it the work-stealing scheduler is efficiently utilized. A Cilk program is a collection
of Cilk procedures and every procedure has a sequence of threads. Every thread is non-blocking
C language function which can run independently without waiting or suspension.
5.6. OpenMPI
It is the programming tool which is specially designed for the poor scientific programmers for
achieving simple and routine parallelism. It is based on the existing programming tool OpenMP
[14]. It provides the sufficient directives for achieving the parallelism. All the directives are
followed by the notation directive pragma ompi. The few of the directives are distvar
(dim=dimension, sleeve=s size) for the distributed array on parallel processes; global for
declaring the variable as global variable; for (reduction (operator: variable)) to parallelize the
for loop; syn sleeve (var=variable list) for exchanging the sleeve data of the distributed array for
correctness ; sync var (var=variable list, master=node id) for synchronizing the global variable
by coping the master data to others; and single (master=node id) for executing the next block by
one process only as a delegate for other processes.
5.7. JAVA
It is the most popular programming language these days because we can create common
applications on it and it also supports the parallelism through its multi-threading concept. It uses
the Just in time (JIT) compiler and automatic garbage collection to perform the critical task [15].
52
International Journal of Programming Languages and Applications ( IJPLA ) Vol.4, No.1, January 2014
For transparent communication between Java Virtual Machines, it has Remote Method
Invocation (RMI). It is utilized to develop high performance computing applications.
5.8. High Performance FORTRAN (HPF)
Its name conveys that it is the extension of Fortran 90. It supports the parallel programming
principles [16]. It supports the data parallel programming pattern in which one program has the
complete control for the distribution of data among all the processors. It works in the distributed
memory environment. It is a portable programming language.
5.9. Z-level Programming Language (ZPL)
It is a language with parallelized compiler. It is particularly for the high performance
computations such as the scientific as well as engineering. It abstracts the Flynn's MIMD
(Multiple Instructions and Multiple Data) parallel architecture [17]. The applications developed
in this language are portable and the performance is independent of the compiler as well as the
machine. It is a good programming language, but the scientists and the engineers have not
shown much interest in it.
5.10. Erlang
It is a functional programming language. Firstly, it was introduced by the telecommunication
giant Ericsson to build the telecommunication switches. Lately in 1998, it becomes open source
software [18]. Concurrency is achieved through threads. The applications developed in this
language are highly available as well as reliable. In this programming paradigm the explicit
threading parallel mechanism is utilized in which the programmers create the explicit threads to
achieve the parallelism [23].
5.11. Unified Parallel C (UPC)
It supports both types of architectures shared memory as well as distributed memory. It is based
on the partitioned memory principle [12]. The complete memory is partitioned into many small
memory areas for every thread. Every thread has a private memory as well as global memory
which are shared among the same class of threads. A new principle is used to get high
performance that is thread affinity in which memory access performance among the threads of
same class is optimized [19]. Workload management in it is implied and the work partitioning
as well as workers mapping may be implied or programmer controlled. The thread
communication is maintained with the help of pointers. Three types of pointers are utilized here
which are as follows: (i) the private pointers who works on their own address spaces, (ii)
sharing pointer who works on the shared memory area, and (iii) sharing pointers to share, these
are the sharing pointers who works on the other shared memory. So many synchronization
mechanisms are utilized in this language like barrier, split phase barriers, fence, locks, and
memory consistency control. It resemble with the MPI platform in workload partitioning and
worker mapping.
5.12. Streams and Iteration in a Single Assignment Language (SISAL)
It is a functional programming language. It offers automatic parallelism through its functional
semantics. In it user-defined names are identifiers rather than variables. These identifiers are
known as values rather than the memory locations [20]. These values are dynamic entities. The
53
International Journal of Programming Languages and Applications ( IJPLA ) Vol.4, No.1, January 2014
identifiers are defined and bound to the values only for the time of execution. The Sisal
Compiler is optimizing one which converts the source program into the object code with the
execution time system parts needed to automatic managing of memory, tasks, and input/output.
The parallelism can be controlled by the users also. In conclusion the programming language
has an optimizing compiler with better runtime performance.
5.13. Laboratory Virtual Instrumentation Engineering Workbench (LabVIEW)
It is visual programming language from National Instruments. It is a platform as well as
development environment. It is also a data-flow programming language in which the execution
is decided with the help of structure of a graphical block diagram by drawing the wires for the
function nodes [21].These wires propagate the variables and the nodes starts to execute as soon
as the input data is available. This programming language is basically utilized for the acquiring
data and processing signals, instrument control, automating test and validation systems, and
monitoring and controlling embedded systems. It may be run on a number of platforms like MS
Windows, UNIX, Linux, and Mac OS X. Multiprocessing as well as multi-threaded hardware
are automatically utilized by its inbuilt schedulers. We can also create distributed applications
on this platform. Hence it is a good high performance computing technology. The nonprogrammers can also develop the good applications by dragging and dropping the virtual
representations of the laboratory equipments to whom they are well-known.
5.14. Manticore Programming language
It is a new functional parallel programming language. It is a heterogeneous programming
language that provides the parallelism at multiple levels. It provides coarse-grained, explicit
parallelism based on Concurrent ML platform. It supports the explicit concurrency with finegrain and implicit threads [22]. The synchronization is provided with first class synchronization
message passing which well fits to the nature of the functional programming paradigms [23].
The locally-concurrent/globally-sequential garbage collector is implemented.
6. CONCLUSIONS
In this construct, a little survey of parallel programming languages, their design approaches, and
their constructs are presented. The current scenario is totally towards the parallelism to achieve
the high performance computing (HPC) and the developers must be aware about the new
concepts of the technology. And this article is good food for the novice parallel programming
lovers who want to do much more in this field.
REFERENCES
[1]
[2]
[3]
[4]
B. Kahanwal & T. P. Singh, (2012) “The Distributed Computing Paradigms: P2P, Grid, Cluster,
Cloud, and Jungle”, International Journal of Latest Research in Science and Technology, Vol. 1,
No. 2, pp183-187.
T. W. Pratt & M. V. Zelkowitz, (2009) Programming Languages: Design and Implementation,
Prentice Hall.
G. M. Amdahl, (1967) “Validity of Single Processor Approach to Achieving Large Scale
Computing Capabilities”, Proc. of AFIPS Spring Joint Computer Conference, pp 483-485.
P. J. Denning & J. B. Dennis, (2010) “The Resurgence of Parallelism”, Communication of the
ACM, Vol. 53, pp 30-32.
54
International Journal of Programming Languages and Applications ( IJPLA ) Vol.4, No.1, January 2014
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
E. Kumm & R. Lea, (1994) “Parallel Computing Efficiency: Climbing the Learning Curve”, proc.
of IEEE Region 10’s Ninth Annual International Conference (TENCON’ 94), Vol. 2, pp 728-732.
Z. Xu & K. Hwang, (1992) “Language Constructs for Structured Parallel Programming”, proc. of
International Parallel Processing Symposium, (IPPS), Beverly Hills, CA, pp 454-461.
S. Akhter & J. Roberts, (2008) “Threading and Parallel Programming Constructs”, publisher, Intel
Press, Intel Corporation, pp 1-22.
H. Masuhara, S. Matsuoka & A. Yonezawa, (1996) “Implementing Parallel Language Constructs
Using a Reflective Object-Oriented Language”, proc. of Reflection’96, pp 1-13.
G. R. Andrews & F. B. Schneider, (1983) “Concepts and Notations for Concurrent Programming”,
ACM Computing Survey, Vol. 15, pp 3-43.
R. K. Buyya, S. T. Selvi & X. Chu, (2009) “Chapter - 14: Object--Oriented Programming with
JAVA”, Tata McGraw Hill Education, Noida, pp 364-387.
H. J. Sips, (1996) “Programming Languages for High Performance Computers”, Proc. of 19th
CERN School of Computing, Netherland, pp 229-238.
H. Kasim, V. March, R. Zhang & S. See, (2008) “Survey on Parallel Programming Model”, Proc.
IFIP Int. Conf. Network and Parallel Computing, 5245, pp 266-275.
R. D. Blumofe, C. F. Joerg, B. C. Kuszmaul, C. E. Leiserson, K. H. Randall & Y. Zhou, (1996)
“Cilk: An Efficient Multithreaded Runtime System”, Journal Parallel and Distributed Computing,
Vol. 37, pp 55-69.
T. Boku, M. Sato, M. Matsubara & D. Takashashi, (2004) “OpenMPI -- OpenMP Like Tool for
Easy Programming in MPI”, Proc. 6th European Workshop on OpenMP (EWOMP'04), pp 83-88.
A. E. Walsh, J. Couch & D.~H. Steinberg, (2000) Java -- 2 Bible, Wiley Publishing.
C. Koebel, D. Loveman, R. Schreiber, G. Steele (Jr.) & M. Zosel, (1994) The High Performance
Fortran Handbook, MIT Press.
L. Snyder, (2007) “The Design and Development of ZPL”, Proc 3rd ACM SIGPLAN History of
Programming Languages Conf., pp 8--1--8--37.
S. Vinoski, (2007) “Reliability with Erlang”, IEEE Internet Computing, Vol. 11, pp 79-81.
P. Husbands, C. Iancu & K. Yelick, (2003) “A Performance Analysis of hte Berkeley UPC
Compiler”, In: ACM ICS 2003: Proc. of the 17th Annual Int. Conf. on Supercomputing, New
York, pp 63-73.
J. L. Gaudiot, W. Bohm, W. Najjar, T. DeBoni, J. Feo & P. Miller, (1997) “The Sisal Model of
Functional Programming and its Implementation”, Proc. 2nd Aizu Int. Symp. on Parallel
Algorithms Architectures Synthesis (pAs '97), Aizu-Wakamatsu, Japan, pp 1-12.
T. Bress, (2013) “Effective LabVIEW Programming”, National Technology and Science Press,
Allendale, NJ.
M. Fluet, N. Ford, M. Rainey, J. Reppy, A. Shaw & Y. Xiao, (2007) “Status Report: The
Manticore Project”, Proc. of ACM SIGPLAN Workshop on ML, New York, pp 15-24.
M. Fluet, L. Bergstrom, N. Ford, M. Rainey, J. Reppy, A. Shaw & Y. Xiao, (2010) “Programming
in Manticore, a Heterogenous Parallel Functional Language”, in Proc. of the 3rd Summer School
Conf. on Central European Functional Programming School, CEFP'09, Berlin, Heidelberg, pp 94145.
R. L. Graham, G. M. Shipman, B. W. Barrett, R. H. Castain, G. Bosilca & A. Lumsdaine, (2006)
“OpenMPI: A high-performance, heterogeneous MPI”, In 5th Int. Workshop on Algorithms,
Models and Tools for Parallel Computing on Heterogenous Networks (HeteroPar '06), Barcelona,
Spain, pp 1-9.
B. Lewis & D. J. Berg, (1996) “PThreads Primer”, SunSoft Press--A Printice Hall Title.
55
| 6cs.PL
|
Spectrum Sharing for LTE-A Network in
TV White Space
Meghna Khaturia∗ , Sweety Suman∗ , Abhay Karandikar and Prasanna Chaporkar
arXiv:1707.07676v1 [cs.NI] 24 Jul 2017
Department of Electrical Engineering, Indian Institute of Technology Bombay, Mumbai-400076
Email: {meghnak,sweetysuman,karandi,chaporkar}@ee.iitb.ac.in
Abstract—Rural areas in the developing countries are predominantly devoid of Internet access as it is not viable for operators to
provide broadband service in these areas. To solve this problem,
we propose a middle mile Long Term Evolution Advanced (LTEA) network operating in TV white space to connect villages
to an optical Point of Presence (PoP) located in the vicinity
of a rural area. We study the problem of spectrum sharing
for the middle mile networks deployed by multiple operators.
A graph theory based Fairness Constrained Channel Allocation
(FCCA) algorithm is proposed, employing Carrier Aggregation
(CA) and Listen Before Talk (LBT) features of LTE-A. We
perform extensive system level simulations to demonstrate that
FCCA not only increases spectral efficiency but also improves
system fairness.
I. I NTRODUCTION
The world has seen a vast growth in communication technology and yet 52% of the global population is still unconnected [1], majority of which live in developing countries.
The broadband penetration in the rural areas of developing
countries is even worse due to high cost of infrastructure,
difficult terrain, sparse population density and low Average
Revenue per User (ARPU). A low cost broadband access to
the end users in these areas can be provided by deploying WiFi Access Points (APs). However, laying fiber to backhaul each
and every Wi-Fi AP becomes infeasible as it is time consuming
and expensive. If an optical Point of Presence (PoP) is located
in the vicinity of a rural area, then a wireless middle mile
network, as proposed in [2], can be established to connect the
optical PoP to the WiFi APs in the villages.
A possible solution for connecting the PoP to the APs is to
use TV UHF band in the middle mile network as it is highly
underutilized in many developing countries. In India, more
than 100 MHz of TV UHF band (470-585 MHz) is unused
(commonly referred to as TV white space) [3]. Owing to the
propagation characteristics of this band, it is possible to obtain
large coverage area even with low power transmission. This
will enable the use of renewable energy sources which is a
highly desirable feature to develop an affordable technology
for rural areas. Consequently, the operators can be encouraged
to deploy a middle mile network in these areas. However,
a middle mile network using the TV UHF band is not a
plug and play solution. As multiple operators have to coexist
in the same band, there will be huge interference among
them which will lead to low spectral efficiency. Hence, it
is important to design a spectrum sharing scheme which not
only increases the spectral efficiency but also guarantees a
fair share of spectrum to the operators. The major challenge
∗ These authors have contributed equally to this work.
in designing the above scheme is that it should be based on
trivial information which can be easily availed from operators
since they will be unwilling to share sensitive information.
The above-mentioned spectrum sharing scenario is very
similar to the limited spectrum pooling where a limited number
of operators share a common pool of spectrum, obeying
the rules of spectrum access set by multi-lateral agreements.
Limited spectrum pooling has been studied in the context of
heterogeneous networks in [4] and [5]. In [4], two operators
pool equal bandwidth for sharing among small cells. An
operator has preemptive priority over its own share of pooled
spectrum. In [5], spectrum sharing is studied for a dense small
cell network of two operators. In our system, the operators
have equal priority over the spectrum in contrast to the preemptive nature of the scheme in [4]. Moreover, in [4], [5],
scheduling information has to be shared among operators
which is infeasible in general. In literature, game theoretic
models are also employed to solve the problem of spectrum
sharing. Two operators dynamically share the spectrum by
playing a non-zero sum game in [6]. In [7], the authors
model spectrum sharing among operators as a non-cooperative
repeated game. The main concern in the above models is that
it may result in inefficient Nash Equilibrium depending on
the utility function selected by the operator. Moreover, the
schemes discussed in [6], [7] do not guarantee fairness and
are also difficult to implement in realistic scenario. All the
above literature discusses the spectrum sharing among only
two operators. Generalization to multiple operators has not
been studied in literature and is a challenging problem that
we tackle in this paper. In this paper, we address the above
issues while solving the spectrum sharing problem in rural
setting.
The main contributions of this paper are:
•
•
•
Analysis of low power middle mile LTE-A network in
TV UHF band with its coverage radius estimates.
A graph theory based algorithm for spectrum sharing
among multiple operators employing Listen Before Talk
(LBT) and Carrier Aggregation (CA) features provided
by LTE-A standard.
Performance evaluation of the proposed algorithm by
system level simulations using Network Simulator-3
(ns3) [8].
The organization of the rest of the paper is as follows.
In Section II, we discuss spectrum sharing among multiple
operators and also establish mathematical formulation for the
same. In Section III, we propose a channel allocation algorithm
to solve the spectrum sharing problem. In Section IV, we
analyse the performance of the proposed algorithm using ns-3
simulations. Finally, Section V concludes the paper.
II. S PECTRUM S HARING P ROBLEM
A. Network Architecture
We consider a middle mile LTE-A network operating in
the TV UHF band. We assume that a portion of this band is
available to multiple operators to deploy their networks in rural
areas. This portion is divided into multiple orthogonal channels
of equal bandwidth. The multi-operator middle mile network
architecture is illustrated in Fig. 1. The network comprises of
a centralized entity called the Spectrum Manager (SM) which
is responsible for channel allocation to the operators. Multiple
low power evolved NodeBs (eNBs) are deployed in a given
area, preferably, in the vicinity of an optical PoP. Even though
an eNB transmits at a very low power, the coverage area
is typically large owing to the propagation characteristics of
TV UHF band. Each operator has an entity called Gateway
Controller (GC) which acts as an interface to communicate
with the SM. Multiple LTE-A Customer Premise Equipments
(CPEs) are served by each eNB. A CPE connects to one or
many WiFi APs installed in a village. An end user accesses
broadband services through a WiFi AP. We assume that the
end users are uniformly distributed in a given area.
The operator registers itself with the SM to access the
TV UHF band. GC collects the topology details like antenna
height, location and transmit power of each eNB under an
operator and communicates it to the SM. No other details
such as user scheduling information are shared to the SM for
channel allocation. The SM then maintains a database of the
information shared by GCs of all operators. The SM treats
all operators equally. As we have considered that the end
users are uniformly distributed in a given area, the average
throughput requirement at each eNB is equal. Hence, an
operator gives equal priority to all its eNBs. Note that in
further discussion we consider each eNB as an independent
network entity. Henceforth, we study the spectrum sharing
problem with respect to an eNB, irrespective of the operator.
The channel allocated by the SM is communicated to an eNB
of an operator through its GC.
Fig. 1: Overview of a multi-operator middle mile network
For simplicity, we consider the Protocol Interference
Model [9] to model the interference between eNBs. In accordance with this model, two eNBs interfere with each other
if they are operating on the same channel and the euclidean
distance between them is less than a certain threshold distance.
The protocol model formulates interference state as a binary
symmetric matrix, where each element of the matrix indicates
whether or not the two eNBs interfere with each other.
B. System Model
Consider a set K = {1, 2, ..., K} representing total number
of eNBs belonging to all the operators in the network. Let
Lk = {1, 2, ..., Lk } be the set of Lk CPEs served by the
eNBk ∀k ∈ K. The set of channels available at the SM is
given by M = {1, 2, ..., M }. Also, let Mk ⊂ M be the set
of channels assigned to eNBk . The eNBk allocates resources
to its users from the assigned channels.
The SM allocates channel to the eNBs depending on the
interference state of the network. Let C = {ck,j |ck,j ∈
{0, 1}}KxK be a binary symmetric KxK matrix where
(k, j) ∈ KXK, represents the interference state such that
ck,j = 1 when eNBk and eNBj interfere with each other,
else ck,j = 0. In addition to allocating channels, SM also
defines the mode in which the channel has to be used. The
mode of access can be shared or dedicated. If the mode of
a channel assigned to an eNB is dedicated, then that channel
does not get allocated to its neighbours. If the mode of access
of the assigned channel is shared, then it is to be shared with
the neighbours using some sharing mechanism. The channel
allocation is given by the two matrices A and B which are
defined as follows:
•
•
Channel Allocation Matrix (A): We define channel allocation matrix as A = {ak,m |ak,m ∈ {0, 1}}KxM where
k ∈ K and m ∈ M such that ak,m = 1, if channel m is
assigned to eNBk , otherwise ak,m = 0.
Mode Allocation Matrix (B): B = {bk,m |bk,m ∈
{0, 1}}KxM is a K by M binary matrix where k ∈ K
and m ∈ M. B represents the mode of access on the
allocated channel such that bk,m = 1, if allocated channel
ak,m is to be shared, otherwise bk,m = 0.
SM assigns a single channel or multiple channels to an
eNB. Carrier Aggregation (CA) feature is required at eNB for
cross channel scheduling of resources [10]. When the mode of
the allocated channel is shared, Listen Before Talk (LBT) is
used for sharing the channel. LBT is a mechanism in which a
radio transmitter performs Clear Channel Assessment (CCA)
to opportunistically transmit over an idle channel. The LBT
mechanism has been discussed for the coexistence of LTE-A
and Wi-Fi system [11]. We have used LBT for the coexistence
among LTE-A systems in this work.
Once the channel assignment is done by the SM, an eNB
allocates resources from the assigned channels to its associated
CPEs in a proportional fair manner. The sum throughput at
eNBk is a function of A and B and is given by Tk (A, B). We
quantify the fairness F of the system using Jain’s Fairness
Index (JFI) as below:
2
K
P
Tk (A, B)
F = k=1K
.
(1)
P
2
Tk (A, B)
K×
k=1
C. Problem Formulation
The spectrum sharing problem can be modeled as a system
throughput maximization problem under the fairness constraint. Mathematically, the problem can be stated as follows:
!
K
X
?
?
(A , B ) = arg max
Tk (A, B) ,
(2)
A,B
k=1
subject to F > δ.
where δ is the constrained value of fairness. There are two
major challenges in obtaining an optimal solution of this
problem. Firstly, this is a combinatorial optimization problem
which is known to be NP-complete. Secondly, to determine
an optimal solution, a closed form expression for throughput
is required at the eNB. The mathematical expression for LBT
throughput can be obtained only for a network which forms a
complete graph. In our case the network graph is not complete.
Therefore, in the following section, we propose a heuristic
based graph theoretic algorithm to solve the above problem
sub-optimally.
III. FAIRNESS C ONSTRAINED C HANNEL A LLOCATION
(FCCA)
We first review the traditional graph coloring problem and
then describe FCCA algorithm in detail. The system can be
modeled as a conflict graph G(V, E), where V represents
set of all eNBs and E denotes the set of edges. An edge
between any two eNBs implies that the eNBs are interfering
with each other i.e. E := {(k, j)|ck,j = 1, ∀k, j ∈ K} where
ck,j is an element of the interference matrix, C defined in
Section II-B. In the traditional Graph Coloring problem, colors
are to be assigned to the vertices such that vertices with an
edge between them do not get the same color. The colors
represent the available set of channels denoted by M. Note
that the numbers of colors i.e. the channels are considered to
be fixed.
We now exploit the above graph coloring technique in our
algorithm. The FCCA algorithm takes graph G as an input
and outputs the allocation matrices, A and B. Here, G is the
graph representing the network as discussed above. In this
method, the channels are assigned to the eNBs according to
two sub-algorithms which are described next.
1) Multiple Dedicated Channel Allocation (MDCA): In this
sub-algorithm, multiple dedicated channels are assigned
to an eNB by using greedy graph-coloring method
iteratively. It is possible to assign multiple channels to
an eNB if the total number of neighbours of an eNB is
less than the total number of channels.
2) One Dedicated Rest Shared Channel Allocation (ODRSCA): In this sub-algorithm, the channel assignment is
done in two steps. In the first step, a single dedicated
channel is assigned to each eNB. Then, the set Nk ,
containing all the channels which are not assigned to
the neighbours of eNBk is obtained. In the second step,
all the channels contained in Nk are assigned to eNBk
in shared mode.
For a given network topology, the output of the above mentioned sub-algorithms are compared to decide the final channel
allocation as described in Algorithm 1. There is always a
guarantee that at least one channel will be allocated to each
eNB irrespective of which sub-algorithm is chosen. Hence, a
certain level of fairness is always ensured. Ideally, the value of
δ should be equal to 1 for complete fairness. However, if we
give more preference to the fairness, the system throughput
will be compromised. Hence, we choose the above δ equal to
0.75 to strike a balance between throughput and fairness.
Algorithm 1 Fairness Constrained Channel Allocation
Require: Graph G
δ = 0.75
Sub-Algorithm 1 : MDCA
while Nk is non empty for all K do
for each k from 1 to K do
find Ek , set of channels assigned to neighbours of k
obtain Nk = {M} \ {Ek }, set of feasible channels
for eNBk ,
q ← min Nk , ak,q ← 1, bk,q ← 0
end for
end while
T1 ← T (A, B), F1 ← F (A, B), A1 ← A, B1 ← B
Sub-Algorithm 2 : ODRS-CA
for each k from 1 to K do
find Ek
obtain Nk = {M} \ {Ek }
q ← min Nk , ak,q ← 1, bk,q ← 0
end for
for each k from 1 to K do
find Ek
obtain Nk = {M} \ {Ek }
ak,q ← 1, bk,q ← 1 ∀q ∈ Nk
end for
T2 ← T (A, B), F2 ← F (A, B), A2 ← A, B2 ← B
Result: Check F1 and F2 and choose (A? , B ? ) such that the
fairness is greater than δ. If both are greater than δ then
choose (A? , B ? ) corresponding to max(T1 , T2 ).
return A∗ , B ∗
IV. P ERFORMANCE A NALYSIS
In this section, we present the results of ns-3 simulations to
assess the performance of FCCA algorithm. We also compare
the proposed approach with few other coexistence approaches.
TABLE I: Simulation Parameters
Parameters
Central Frequency(fc )
Transmit Power (Pt )
Receiver Sensitivity (RS)
Cable Loss (CL)
Receiver Noise Figure (N F )
Transmitter Antenna Gain (Gt )
Receiver Antenna Gain (Gr )
Transmitter Antenna Height (ht )
Receiver Antenna Height (hr )
Slot Time
Transmit Opportunity (T xOp)
Detection Threshold
Simulation time
where RS, Pt , Gt , Gr , CL, N F , ht , hr and fc are as per
the Table I. P L(d, ht , hr , fc ) is the path loss which is also
a function of distance d between transmitter and receiver.
Hata model for Suburban Areas is used to calculate path
loss [12]. The transmit power, Pt , of an eNB is 18 dBm
which is significantly low. For the values of the parameters
given in Table I, the coverage radius of eNB comes out to be
approximately 3 km.
C. Results
Values
500-520 MHz
18 dBm
−101 dBm [13]
2 dB
7 dB
10 dB
0 dB
30 m
5m
9 µs
10 ms
−62 dBm
30 s
A. Scenario Description
We assume that 20 MHz of the TV UHF band is available
for middle mile network. This band is further divided into 4
orthogonal channels of 5 MHz each. All channels are assumed
to be identical. As the rural areas are sparsely populated, we
assume that an eNB will not get interference from three or
more eNBs. The eNBs are deployed uniformly at random in an
area of 100 km2 as shown in Fig. 2. The CPEs are distributed
uniformly within the coverage area of an eNB. Each eNB
is assumed to serve 5 stationary CPEs. For constructing the
conflict graph using protocol interference model, we consider
a distance of 4 km between eNBs. If the distance between
eNBs is less than 4 km, then they interfere with each other.
We perform ns-3 simulations over 100 random topologies. All
the performance metrics are averaged over such realizations.
The simulation parameters are given in Table I. With a static
rural environment along with stationary CPEs, it is reasonable
to rule out fast fading effects in our scenario. We consider
only saturated downlink transmission in this work i.e. at each
eNB, saturated traffic is generated for each of the associated
CPEs.
10000
eNB
CPE
2
8000
Y Coordinates
3
6000
4000
1
4
2000
0
0
2000
4000
6000
8000
10000
X Coordinates
Fig. 2: An example topology of the network.
B. Coverage Radius of eNB operating in TV UHF band
The coverage radius of a transmitter is defined as the
maximum allowed distance between the transmitter and the
receiver such that they can communicate. We calculate the
coverage radius of an eNB using the equation:
RS = Pt + Gt + Gr − P L(d, ht , hr , fc ) − CL − N F, (3)
We analyze three performance metrics to assess the performance of the proposed FCCA algorithm: a) Spectral Efficiency
b) Average System Throughput per eNB and c) Jain’s Fairness
Index. The performance of these metrics are observed with
respect to an increase in the network density i.e. we increase
the number of eNBs from 3 to 10 in a fixed area of 100 km2 .
In our results we present spectral efficiency per eNB which
is measured in bits/s/Hz. We have used Jain’s Fairness Index
to quantify how fairly the available band is shared among the
eNBs.
In Fig. 3, we compare the spectral efficiency and the
average system throughout of FCCA with two other schemes
i) LTE with no-coexistence mechanism ii) LTE with LBT
as the coexistence mechanism. Clearly, the FCCA algorithm
outperforms both the schemes in both the metrics. In the first
scheme, the entire 20 MHz band is used by all eNBs without
any coexistence mechanism. Here the spectral efficiency is
poor due to interference among the eNBs. In the second
scheme, the entire 20 MHz band is shared among all the
eNBs using LBT. In this case, the performance is poor as
the transmission time is wasted in contention. The FCCA
algorithm performs better than the above schemes as it takes
the topology into consideration for allocating the channels. As
shown in Fig. 4, the fairness of the FCCA algorithm is also
better than the other two schemes. The proposed algorithm
guarantees an excellent fairness index of 0.76 even in the case
10 eNBs per 100 km2 .
It is important to analyse the performance of the proposed
algorithm under the best case and the worst case scenario.
Consider a very sparse network (best case) where only three
eNBs are deployed in the given area. When 100 random
topologies are simulated under this setting, it is observed that
in 50% of the cases MDCA is preferred and in the remaining
cases ODRS-CA is preferred. This result highlights the fact
that orthogonal channel allocation may not always give the
best result. In a very dense network (worst case) where 10
eNBs are deployed in the given area, in 61% cases ODRSCA is preferred between the two sub-algorithm. This result
further emphasizes the use of LBT as sharing mechanism for
better system performance in terms of spectral efficiency and
fairness.
D. Average Throughput vs Demand
We now consider a typical rural setting in India. An optical
PoP is present at the village office called Gram Panchayat (GP)
which typically serves 2 villages. Approximately 5 GPs are
present in an area of 100 km2 serving 10 villages. The average
(a) Spectral Efficiency of eNB vs. number of eNBs
deployed in 100 km2 area
(b) Average throughput per eNB vs. number of eNBs
deployed in 100 km2 area
Fig. 3: Comparative analysis of average system throughput and spectral efficiency for the three schemes.
operators in a network. We have also compared the obtained
average throughput with the throughput demand generated by
a rural setting in India. We note that the proposed scheme
easily meets the throughput demand in a rural area.
R EFERENCES
Fig. 4: An example topology of the network.
population of a village in India is 1000. There is one subscriber
per household i.e. one person in a house of 5 will subscribe to
broadband service. Consider a minimum broadband throughput of 2 Mbps with the contention ratio of 1 : 50. Thus, the
average throughput requirement under the above scenario is
(1000 people×10 villages×2 Mbps)/(50 × 5) = 80 Mbps. If
5 eNBs are deployed, each at one GP, the average throughput
demand that can be served using FCCA algorithm is approximately 83 Mbps. Hence, the average throughput requirement
of the above setting can be easily met.
V. C ONCLUSION
We have discussed the problem of poor broadband penetration in rural areas of developing countries. To solve this
problem, we have proposed a middle mile LTE-A network
operating in TV UHF band. We have presented a centralized
graph theory based channel allocation algorithm with a novel
concept of allocating a combination of shared and dedicated
channel to an eNB. The performance of the algorithm has been
studied using ns-3 simulations. The results demonstrate that it
increases both the spectral efficiency and the fairness among
[1] http://www.broadbandcommission.org/documents/reports/
bb-annualreport2016.pdf.
[2] A. Kumar, A. Karandikar, G. Naik, M. Khaturia, S. Saha, M. Arora and
J. Singh, “Toward enabling broadband for a billion plus population with
TV white spaces,” IEEE Communications Magazine, vol. 54, no. 7, pp.
28-34, July 2016.
[3] G. Naik, S. Singhal, A. Kumar and A. Karandikar, “Quantitative assessment of TV White Space in India,” National Conference on Communications (NCC), pp. 16, 2014.
[4] A. Alsohaily and E. S. Sousa, “Spectrum sharing LTE-advanced small
cell systems,” International Symposium on Wireless Personal Multimedia
Communications, Atlantic City, NJ, 2013, pp. 1-5.
[5] Y. Teng, Y. Wang and K. Horneman, “Co-primary spectrum sharing for
denser networks in local area,” International Conference on Cognitive
Radio Oriented Wireless Networks and Communications, Oulu, 2014,
pp. 120-124.
[6] H. Kamal, M. Coupechoux and P. Godlewski, “Inter-operator spectrum
sharing for cellular networks using game theory,” International Symposium on Personal, Indoor and Mobile Radio Communications, Tokyo,
2009, pp. 425-429.
[7] B. Singh, K. Koufos, O. Tirkkonen and R. Berry, “Co-primary interoperator spectrum sharing over a limited spectrum pool using repeated
games,” International Conference on Communications, London, 2015, pp.
1494-1499.
[8] https://www.nsnam.org.
[9] K. Jain, J. Padhye, V. N. Padmanabhan and L. Qiu. “Impact of interference
on multi-hop wireless network performance,” in Proceedings of the 9th
annual international conference on Mobile computing and networking,
MOBICOM, 2003, New York, pp. 66-80.
[10] http://www.3gpp.org/technologies/keywords-acronyms/
101-carrier-aggregation-explained.
[11] 3GPP Technical specifications TS 36.889: https://portal.3gpp.org/
desktopmodules/Specifications/SpecificationDetails.aspx?specificationId=
2579.
[12] M. Hata, “Empirical formula for propagation loss in land mobile radio
services,” IEEE Transactions on Vehicular Technology, vol. 29, no. 3, pp.
317-325, Aug 1980.
[13] http://www.etsi.org/deliver/etsi ts/136100 136199/136104/11.02.00 60/
ts 136104v110200p.pdf.
| 7cs.IT
|
c 2018 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all
other uses, in any current or future media, including reprinting/republishing this material for advertising
or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or
reuse of any copyrighted component of this work in other works.
Pre-print of article that will appear in the 2018 IEEE Robotics and Automation Letters (RA-L).
arXiv:1709.09905v3 [cs.RO] 16 Feb 2018
Please cite this paper as:
A. Gawel, C. Del Don, R. Siegwart, J. Nieto, and C. Cadena. (2018).
"X-View: Graph-Based Semantic Multi-View Localization" in IEEE Robotics and Automation Letters
(RA-L), 2018.
bibtex:
@inproceedings{gawel2018x-view,
title={X-View: Graph-Based Semantic Multi-View Localization},
author={Gawel, Abel and Del Don, Carlo and Siegwart, Roland and Nieto, Juan and Cadena, Cesar},
booktitle=IEEE Robotics and Automation Letters (RA-L)},
year={2018}
}
1
X-View: Graph-Based Semantic Multi-View
Localization
Abel Gawel∗ , Carlo Del Don∗ , Roland Siegwart, Juan Nieto and Cesar Cadena
Abstract—Global registration of multi-view robot data is
a challenging task. Appearance-based global localization approaches often fail under drastic view-point changes, as representations have limited view-point invariance. This work is based on
the idea that human-made environments contain rich semantics
which can be used to disambiguate global localization. Here,
we present X-View, a Multi-View Semantic Global Localization
system. X-View leverages semantic graph descriptor matching for
global localization, enabling localization under drastically different view-points. While the approach is general in terms of the
semantic input data, we present and evaluate an implementation
on visual data. We demonstrate the system in experiments on
the publicly available SYNTHIA dataset, on a realistic urban
dataset recorded with a simulator, and on real-world StreetView
data. Our findings show that X-View is able to globally localize
aerial-to-ground, and ground-to-ground robot data of drastically
different view-points. Our approach achieves an accuracy of up
to 85 % on global localizations in the multi-view case, while the
benchmarked baseline appearance-based methods reach up to
75 %.
Index Terms—Localization, Semantic Scene Understanding,
Mapping
I. INTRODUCTION
LOBAL localization between heterogeneous robots is
a difficult problem for classic place-recognition approaches. Visual appearance-based approaches such as [1, 2]
are currently among the most effective methods for relocalization. However, they tend to significantly degrade with
appearance changes due to different time, weather, season,
and also view-point [3, 4]. In addition, when using different
sensor modalities, the key-point extraction becomes an issue
as they are generated from different physical and geometrical
properties, for instance intensity gradients in images vs. highcurvature regions in point clouds.
Relying on geometrical information, directly from the
measurements or from a reconstruction algorithm, on the
other hand shows stronger robustness on view-point changes,
seasonal changes, and different sensor modalities. However,
geometrical approaches typically do not scale well to very
large environments, and it remains questionable if very strong
view-point changes can be compensated while maintaining
G
This work was supported by European Union’s Seventh Framework Programme for research, technological development and demonstration under
the TRADR project No. FP7-ICT-609763 and by the National Center of
Competence in Research (NCCR) Robotics through the Swiss National
Science Foundation.
∗ The authors contributed equally to this work.
Authors are with the Autonomous Systems Lab, ETH Zurich.
[email protected], [email protected],
{rsiegwart, nietoj, cesarc}@ethz.ch
Figure 1: X-View globally localizes data of drastically different view-points
using graph representations of semantic information. Here, samples of the
experimental data is shown, i.e., semantically segmented images from the
publicly available SYNTHIA and the Airsim datasets. The localization target
graph is built from data of one view-point (right images), while the query
graph is built from sequences of another view-point (left images). X-View
efficiently localizes the query graph in the target graph.
only a limited overlap between the localization query and
database [5, 6].
Another avenue to address appearance and view-point
changes are Convolutional Neural Network (CNN) architectures for place recognition [4, 7]. While these methods show
strong performance under appearance changes, their performance is still to be investigated under extreme view-point
variations.
Recently, topological approaches to global localization regained interest as a way to efficiently encode relations between
multiple local visual features [8, 9]. On the other hand,
the computer vision community has made great progress in
semantic segmentation and classification, resulting in capable
tools for extracting semantics from visual and depth data [10–
12].
Based on the hypothesis that semantics can help to mitigate
the effects of appearance changes, we present X-View, a novel
approach for global localization based on building graphs of
semantics. X-View introduces graph descriptors that efficiently
represent unique topologies of semantic objects. These can
be matched in much lower computational effort, therefore not
suffering under the need for exhaustive sub-graph matching
[13].
2
By using semantics as an abstraction between robot viewpoints, we achieve invariances to strong view-point changes,
outperforming CNN-based techniques on RGB data. Furthermore, with semantics understanding of the scene, unwanted
elements, such as moving objects can naturally be excluded
from the localization. We evaluate our global localization
algorithm on publicly available datasets of real and simulated
urban outdoor environments, and report our findings on localizing under strong view-point changes. Specifically, this paper
presents the following contributions:
• A novel graph representation for semantic topologies.
• Introduction of a graph descriptor based on random walks
that can be efficiently matched with established matching
methods.
• A full pipeline to process semantically segmented images
into global localizations.
1
• Open source implementation of the X-View algorithm .
• Experimental evaluation on publicly available datasets.
The remainder of this paper is structured as follows: Sec. II
reviews the related work on global localization, followed by
the presentation of the X-View system in Sec. III. We present
our experimental evaluation in Sec. IV and conclude our
findings in Sec. V.
II. RELATED WORK
In this section we review the current state-of-the-art in
multi-robot global localization in relation to our proposed
system.
A common approach to global localization is visual feature
matching. A large amount of approaches have been proposed
in the last decade, giving reliable performance under perceptually similar conditions [1–3]. Several extensions have been
proposed to overcome perceptually difficult situations, such as
seasonal changes [14, 15], daytime changes [4, 16], or varying
view-points using CNN landmarks [7, 17]. However, drastic
view-point invariance, e.g., between views of aerial and ground
robots continues to be a challenging problem for appearancebased techniques.
In our previous work, we demonstrated effective 3D heterogeneous map merging approaches between different viewpoints from camera and LiDAR data, based on overlapping
3D structural descriptors [5, 6]. However, 3D reconstructions
are still strongly view-point dependent. While these techniques
do not rely on specific semantic information of the scenes, the
scaling to large environments has not yet been investigated,
and computational time is outside real-time performance with
large maps.
Other approaches to global localization are based on topological mapping [18, 19]. Here, maps are represented as graphs
G = (V , E) of unique vertices V and edges E encoding
relationships between vertices. While these works focus on
graph merging by exhaustive vertex matching on small graphs,
they do not consider graph extraction from sensory data or ambiguous vertices. Furthermore, the computationally expensive
matching does not scale to larger graph comparisons.
1 https://github.com/ethz-asl/x-view
With the recent advances in learning-based semantic extraction methods, using semantics for localization is a promising
avenue [20–22]. In [21, 22] the authors focus on the data association problem for semantic localization using Expectation
Maximization (EM) and the formulation of the pose estimation
problem for semantic constraints as an error minimization. The
semantic extraction is based on a standard object detector from
visual key-points.
Stumm et al. [8] propose to use graph kernels for place
recognition on visual key-point descriptors. Graph kernels
are used to project image-wise covisibility graphs into a
feature space. The authors show that graph descriptions can
help localization performance as to efficiently cluster multiple
descriptors meaningfully. However, the use of large densely
connected graphs sets limitations to the choice of graph
representation. Motivated, by these findings, we propose to
use graph descriptors on sparse semantic graphs for global
localization.
III. X-VIEW
In this section, we present our Graph-Based Multi-View Semantic Global Localization system, coined X-View. It leverages
graph extraction from semantic input data and graph matching
using graph descriptors. Fig. 2 illustrates the architecture of the
proposed global localization algorithm, focusing on the graph
representation and matching of query semantic input data to
a global graph. The localization target map is represented as
the global graph. X-View is designed to operate on any given
odometry estimation system and semantic input cue. However,
for the sake of clarity, we present our system as implemented
for semantically segmented images, but it is not limited to it.
A. System input
We use semantically segmented images containing pixelwise semantic classification as input to the localization algorithm. These segmentations can be achieved using a semantic
segmentation method, such as [11, 12]. Also instance-wise
segmentation, i.e., unique identifiers for separating overlapping
objects of same class in the image space can be considered
for improved segmentation, but is not strictly necessary for
the approach to work. Furthermore, we assume the estimate
of an external odometry system. Finally, we also consider a
database semantic graph Gdb , as it could have been built and
described on a previous run of our graph building algorithm
as presented in the next sub-sections.
B. Graph extraction and assembly
In this step, we convert a sequence of semantic images I q
into a query graph Gq . We extract blobs of connected regions,
i.e., regions of the same class label lj in each image. Since
semantically segmented images often show noisy partitioning
of the observed scene (holes, disconnected edges and invalid
labels on edges), we smooth them by dilating and eroding the
boundaries of each blob. We furthermore reject blobs smaller
than a minimum pixel count to be included in the graph, to
3
Semantic
segmentation
Odometry system
Frame-wise
graph
extraction
Sub-graph
assembly
Random
walk
extraction
Descriptor
matching
Global
localization
estimation
loop closure
graph fusion
visualization
etc.
Global graph
Figure 2: X-View global localization system overview. The inputs to the system are semantically segmented frames (e.g., from RGB images) and the global
graph Gdb . First, a local graph is extracted from the new segmentation. Then, the sub-graph Gq is assembled and random walk descriptors are computed
on each node of Gq . The system matches the sub-graph random walk descriptors to Gdb , e.g., recorded from a different view-point. Finally, the matches
are transferred to the localization back-end module to estimate the relative localization between Gq and Gdb . Consecutively, the relative localization can be
used for various purposes such as loop closure, fusing Gq into Gdb or for visualization.
mitigate the effect of minor segments. This process removes
unwanted noise in the semantically segmented images. The
magnitude of this operation is 4 pixels, and has a minor
effect on the segmentation result. However, it ensures clean
boundaries between semantic segments. Furthermore, the center location pj of the blobs are extracted and stored alongside
the blob labels as vertices vj = {lj , pj }. In the case that also
instance-wise segmentation is available, it can be considered
in the blob extraction step, otherwise the extraction operates
only on a class basis.
The undirected edges ej between vertices are formed when
fulfilling a proximity requirement, which can be either in
image- or 3D-space. In the case of image-space, we assume
images to be in a temporal sequence to grow graphs over
several frames of input data. However, this is not required in
the 3D case.
Using a depth channel or the depth estimation from, e.g.,
optical flow, the neighborhood can be formed in 3D-space,
using the 3D locations of the image blobs to compute a
Euclidean distance. The process is illustrated for image data in
Fig. 3 (top). Then, several image-wise graphs are merged into
Gq by connecting vertices of consecutive images using their
Euclidean distance, see Fig. 3. To prevent duplicate vertices
of the same semantic instance, close instances in Gq are
merged into a single vertex, at the location of the vertices’
first observation. The strategy of merging vertices into their
first observation location is further motivated by the structure
of continuous semantic entities, such as streets. This strategy
leads to evenly spaced creation of continuous entities’ vertices
in Gq .
C. Descriptors
X-View is based on the idea that semantic graphs hold high
descriptive power, and that localizing a sub-graph in a database
graph can yield good localization results. However, since subgraph matching is an NP-complete problem [13], a different
regime is required to perform the graph registration under
real-time constraints, i.e., in the order of seconds for typical
robotic applications. In this work, we extract random walk
descriptors for every node of the graph [23], and match them in
a subsequent step. This has the advantage that the descriptors
can be extracted and matched in constant or linear time, given
a static or growing database-graph, respectively.
Figure 3: Extraction of semantic graphs from one image (top) and a sequence
of images (bottom). Vertices are merged and connected from sequences of
input data. Note that we omitted some vertices and edges in the sample graphs
on the right side for visualization purposes and reduced the graph to a planar
visualization, whereas the semantic graphs in our system are connected in
3D-space. The ellipses around each vertex were added for visualization and
represent a scaled fitted ellipse on a semantic instance of the segmentation
image.
Each vertex descriptor is an n × m matrix consisting of
n random walks of depth m. Each of the random walks
originates at the base vertex vj and stores the class labels of
the visited vertices. Walk strategies, such as preventing from
immediate returns to the vertex that was visited in the last step,
and exclusion of duplicate random walks can be applied to
facilitate expressiveness of the random walk descriptors. The
process of random walk descriptor extraction is illustrated in
Fig. 4.
D. Descriptor Matching
After both Gq and Gdb are created, we find associations
between vertices in the query graph and the ones in the
database graph by computing a similarity score between the
corresponding graph descriptors. The similarity measure is
computed by matching each row of the semantic descriptor
of the query vertex to the descriptor of the database vertex.
The number of identical random walks on the two descriptors
4
Figure 4: Schematic representation of the random walk extraction. (Left) From
a seed vertex, cyan star, the random walker explores its neighborhood. This
results in the descriptor of n random walks of depth m (here, m = 4).
The highlighted path corresponds to the last line of the descriptor on the
right. (Right) Each line of the descriptor starts with the seed vertex label and
continues with the class labels of the visited vertices.
reflects the similarity score s, which is normalized between 0
and 1. In a second step, the k matches with highest similarity
score are selected for estimating the location of the query
graph inside the database map.
E. Localization Back-End
The matching between query graph and global graph, the
robot-to-vertex observations, and the robot odometry measurements result in constraints θ i ⊆ Θ(pi , ci ) on the vertex
positions pi and robot poses ci with θ i = eTi Ωi ei , the
measurement errors ei , and associated information matrix Ωi .
Specifically these three types of constraints are denoted as
ΘM (pi ), ΘV (pi , ci ), and ΘO (ci ) respectively. The matching constraints ΘM (pi ) stem from the semantic descriptor
matching of the previous step, while the robot odometry
constraints ΘO (ci ) are created using the robots estimated
odometry between consecutive robot poses associated to the
localization graph. The robot-to-vertex constraints encode the
transformation between each robot-to-vertex observation. Using these constraints, we compute a Maximum a Posteriori
(MAP) estimate of P
the robot pose ci by minimizing a negative
log-posterior E =
θ i , i.e.,
c∗i = argmin
ci
X
Θ(pi , ci )
(1)
with Θ(pi , ci ) = {ΘM (pi ), ΘV (pi , ci ), ΘV (pi )} This optimization is carried out by a non-linear Gauss-Newton optimizer. Optionally, the algorithm also allows to reject matching
constraints in a sample consensus manner, using RANSAC on
all constraints between Gq and Gdb , excluding the specific
constraints from the optimization objective. We initialize the
robot position at the mean location of all matching vertices’
locations from Gdb .
IV. EXPERIMENTS
We evaluate our approach on two different synthetic outdoor
datasets with forward to rear view, and forward to aerial view,
and one real world outdoor dataset with forward to rear view.
In this section, we present the experimental set-up, the results,
and a discussion.
SYNTHIA
AdapNet
Airsim
StreetView
Figure 5: Sample images from the datasets used in the experiments: (top)
RGB image, (middle) Depth image, (bottom) Semantic segmentation. (left)
SYNTHIA with perfect semantic segmentation, (middle left) SYNTHIA with
AdapNet semantic segmentation, (middle right) Airsim with perfect semantic
segmentation, (right) StreetView with SegNet semantic segmentation.
A. Datasets
The first of the used datasets is the public SYNTHIA
dataset [24]. It consists of several sequences of simulated
sensor data from a car travelling in different dynamic environments and under varying conditions, e.g., weather and
daytime. The sensor data provides RGB, depth and pixel-wise
semantic classification for 8 cameras, with always 2 cameras
facing forward, left, backwards and right respectively. The
segmentation provides 13 different semantic classes which
are labelled class-wise. Additionally, dynamic objects, such
as pedestrians and cars are also labelled instance-wise. We
use sequence 4, which features a town-like environment. The
total travelled distance is 970 m.
In the absence of suitable public aerial-ground semantic
localization datasets, we use the photo-realistic Airsim framework [25] to generate a simulated rural environment2 . This
environment is explored with a top-down viewing Unmanned
Aerial Vehicle (UAV) and a car traversing the streets with
forward-facing sensors. Both views provide RGB, depth and
pixel-wise semantic classification data in 13 different classes
with instance-wise labelling. Furthermore, both trajectories are
overlapping with only an offset in z-direction and have a
length of 500 m each. Please note that we used a pre-built
environment, i.e., the objects in the environment have not
specifically been placed for enhanced performance.
Finally, we evaluate the system on a dataset gathered from
Google StreetView imagery. The RGB and depth data of
a straight 750 m stretch of Weinbergstrasse in Zurich are
extracted via the Google Maps API 3 . Analogously to the
SYNTHIA dataset, we use forward and backward facing camera
views.
While the travelled distance between two image locations in
the Airsim dataset is always 1 m, it varies between 0 m to 1 m
in the SYNTHIA dataset, and is approximately 10 m between
two frames in the StreetView dataset. Sample images of all
datasets are depicted in Fig. 5.
Our approach relies on semantic representations of scenes.
While we do not propose contributions on semantic extraction
from raw sensor data, recent advances on semantic segmentation show ever increasing accuracies on visual and depth
2 http://robotics.ethz.ch/~asl-datasets/x-view/
3 https://goo.gl/iBniJ9
5
data [10–12, 26]. We therefore evaluate the performance
on SYNTHIA both using semantic segmentation with AdapNet [11], and the ground truth as provided by the dataset. On
the Airsim data, we only use the segmentation from the dataset,
and on the StreetView dataset, we use semantic segmentation
with SegNet [12].
B. Experimental Setup
We evaluate the core components of X-View in different
experimental settings. In all experiments, we evaluate XView on overlapping trajectories and the provided depth and
segmentation images of the data. First, we focus our evaluation
of the different graph settings on the SYNTHIA dataset. We
then perform a comparative analysis on SYNTHIA, Airsim, and
StreetView.
In SYNTHIA, we use the left forward camera for building
a database map and then use the left backward camera for
localization. Furthermore, we use 8 semantic classes of SYNTHIA: building, street, sidewalk, fence, vegetation, pole, car,
and sign, and reject the remaining four classes: sky, pedestrian,
cyclist, lanemarking. The AdapNet semantic segmentation
model is trained on other sequences of the SYNTHIA dataset
with different seasons and weather conditions.
Analogously, we use the forward-view of the car in the
Airsim dataset to build the database map and then localize
the UAV based on a downward-looking camera. Here we use
6 classes (street, building, car, fence, hedge, tree) and reject
the remaining from insertion into the graph (powerline, pool,
sign, wall, bench, rock), as these are usually only visible by
one of the robots, or their scale is too small to be reliably
detected from the aerial robot.
Finally, in the StreetView data, we use the forward view to
build the database and localize using the rear facing view. Out
of the 12 classes that we extract using the pre-trained SegNet
model4 , we use five, i.e., (road, sidewalk, vegetation, fence,
car), and reject the remaining as these are either dynamic
(pedestrian, cyclist), unreliably segmented (pole, road sign,
road marking), or omni-present in the dataset (building, sky).
We build the graphs from consecutive frames in all experiments, and use the 3D information to connect and merge vertices and edges, as described in III-B. The difference between
graph construction in image- and 3D-space is evaluated in a
separate experiment. No assumptions are made on the prior
alignment between the data. The ground-truth alignment is
solely used for performance evaluation.
C. Localization performance
We generate the PR of the localization based on two thresholds. The localization threshold tL is applied on the distance
between the estimated robot position c∗i and the ground truth
position cgt . It is set as true, if the distance between c∗i and
cgt is smaller than tL , i.e., kc∗i − cgt k ≤ tL , and to false for
kc∗i − cgt k > tL . The margin tL on the locations is required,
since Gq and Gdb do not create vertices in the exact same spot.
The same node can be off by up to twice the distance that we
4 goo.gl/EyReyn
use for merging vertices in a graph. Here, we use tL = 20 m
for SYNTHIA and StreetView, and tL = 30 m for Airsim. For
the PR curves, we vary the consistency threshold tc that is
applied on the RANSAC-based rejection, i.e., the acceptable
deviation from the consensus transformation between query
and database graph vertices. The localization estimation yields
a positive vote for an estimated consensus value s of s ≤ tc
and a negative vote otherwise.
Firstly, we evaluate the effect of different options on the
description and matching using the random walk descriptors
(i.e., random walk parameters, graph coarseness, number of
query frames, dynamics classes, graph edge construction technique, and seasonal changes) as described in Sec. III-B - III-D.
To illustrate the contrast to appearance-based methods, we
also present results on two visual place recognition techniques
based on BoW, as implemented by Gálvez-López and Tardos
[2], and NetVLAD [4] on the datasets’ RGB data. To generate
the PR of the reference techniques, we vary a threshold on
the inverse similarity score for BoW, and a threshold on the
matching residuals of NetVLAD.
Furthermore, we show the performance of the full global
localization algorithm on the operating point taken from the
PR curves. Our performance metric is defined as the percentage of correct localizations over the Euclidean distance
between c∗i and cgt . As for BoW and NetVLAD, we take
localization as the best matching image. The localization error
is then computed as the Euclidean distance between associated
positions of the matched image and the ground truth image.
To improve performance of the appearance-based methods, we
select the operating points with high performances, i.e., high
precisions in the PR curves.
D. Results
While we illustrate the effects of different attributes of XView in Fig. 6 as evaluated on SYNTHIA, we then also show
a comparison on all datasets in Fig. 7.
Fig. 6a depicts the effect of varying the random walk
descriptors on the graph. Here, a descriptor size with number
of random walks n = 200 and walk depth m between 3 − 5,
depending on the size of Gq perform best. Both decreasing
n or increasing m leads to a decrease in performance. These
findings are expected, considering query graph sizes ranging
between 20 − 40 vertices. Under these conditions, the graph
can be well explored with the above settings. Descriptors with
larger walk depth m significantly diverge between Gq and
Gdb , as the random walk reaches the size limits of Gq and
continues exploring already visited vertices, while it is possible
to continue exploring Gdb to greater depth.
Secondly, Fig. 6b presents PR-curves for different sizes of
Gq , i.e., different numbers of frames used for the construction
of Gq . An increase in the query graph size leads to a considerable increase of the localization performance. Also this effect
is expected as Gq contains more vertices, forming more unique
descriptors. However, it is also desirable to keep the size of Gq
limited, as a growing query graph size requires larger overlap
between Gq and Gdb . Furthermore, the computational time
for descriptor calculation and matching grows with increased
query graph size.
1.0
1.0
0.8
0.8
0.8
0.6
0.4
n=200, m=3
n=500, m=3
n=500, m=5
0.2
0.0
0.0
0.2
0.4
BoW
NetVLAD
Recall
0.6
Precision
1.0
Precision
0.6
0.4
30 frames
15 frames
0.2
0.8
0.0
0.0
1.0
0.2
Recall
0.6
0.6
0.4
0.8
0.0
0.0
1.0
0.8
0.8
0.8
0.6
0.4
0.0
0.0
0.2
0.4
3D to image
BoW
NetVLAD
Recall
0.6
0.6
0.4
Static objects
Dynamic objects
All objects
0.2
0.8
1.0
0.0
0.0
(d) Construction type.
0.2
0.4
Recall
BoW
NetVLAD
0.6
0.4
Recall
BoW
NetVLAD
0.6
0.8
1.0
(c) Graph coarseness.
1.0
3D space
Image space
Image to 3D
0.2
(b) Query length.
1.0
0.2
Medium coarse
Coarse
Dense
0.2
1.0
Precision
Precision
(a) Descriptor parameters.
0.4
BoW
NetVLAD
Precision
Precision
6
0.6
0.4
X-View Summer-Summer
X-View Summer-Fall
BoW Summer-Summer
0.2
0.8
1.0
(e) Number of Semantic classes.
0.0
0.0
0.2
0.4
Recall
BoW Summer-Fall
NetVLAD Summer-Summer
NetVLAD Summer-Fall
0.6
0.8
1.0
(f) Seasonal changes on same view.
1.0
1.0
0.8
0.8
Precision
Precision
Figure 6: PR curves for localization of the rear view semantic images against a database graph built from the forward view on the SYNTHIA dataset (except
(f)). For all plots we accept a localization if it falls within a distance of 20 m from the ground-truth robot position. This threshold corresponds to the value
up to which query graph vertices of the same semantic instance can be off from their corresponding location in the database graph, caused by the graph
construction technique. (a) illustrates the effect of different descriptor settings on the localization performance. (b) shows the effect of increasing the amount
of frames used for query graph construction, while (c) depicts the effect of using coarser graphs, i.e., a large distance in which we merge vertices of same
class label. In (d) we compare the extraction methods in image-, and 3D-space and in (e) the effect of including all semantic objects against including a
subset of semantic classes. Lastly, in (f), we evaluate the localization performance on a configuration with the right frontal camera as query and the left frontal
camera for the database, under the effect of seasonal changes. In contrast to the other plots where we use the ground truth, we use semantic segmentation
with AdapNet on the data. The appearance-based techniques used are visual BoW [2] and NetVLAD [4].
0.6
0.4
X-View SYNTHIA
X-View Airsim
NetVLAD SYNTHIA
0.2
0.0
0.0
0.2
0.4
NetVLAD Airsim
BoW SYNTHIA
BoW Airsim
0.6
0.6
0.4
X-View AdapNet SYNTHIA
X-View SegNet StreetView
NetVLAD SYNTHIA
0.2
0.8
0.0
0.0
1.0
0.2
0.4
Recall
0.8
1.0
(b) CNN-based Semantic Segmentation.
1.0
1.0
0.8
Success rate [%]
Success rate [%]
0.6
Recall
(a) Perfect Semantic Segmentation.
X-View SYNTHIA
X-View Airsim
NetVLAD SYNTHIA
NetVLAD Airsim
BoW SYNTHIA
BoW Airsim
0.6
0.4
0.2
0.0
NetVLAD StreetView
BoW SYNTHIA
BoW StreetView
0.8
X-View AdapNet SYNTHIA
X-View StreetView
NetVLAD SYNTHIA
NetVLAD StreetView
BoW SYNTHIA
BoW StreetView
0.6
0.4
0.2
0
20
40
60
80
100
Localization Accuracy [m]
(c) Perfect Semantic Segmentation.
0.0
0
20
40
60
80
100
Localization Accuracy [m]
(d) CNN-based Semantic Segmentation.
Figure 7: Localization performance of X-View on the SYNTHIA, Airsim, and the StreetView data compared to the appearance-based methods [2, 4]. The operation
points are chosen according to the respective PR curves in (a) and (b), indicated as dots. (c) illustrates the performance on perfectly semantically segmented
data on SYNTHIA, and Airsim. (d) shows the system’s performance on the SYNTHIA, and StreetView datasets using CNN-based semantic segmentation.
Thirdly, Fig. 6c shows the impact of increased graph
coarseness, i.e., larger distances of merging vertices. Here, the
coarseness cannot be arbitrarily scaled to low or high values,
as it leads to either over- or under-segmented graphs. Our
best performing results were obtained with a vertex merging
distance of 10 m for the SYNTHIA dataset, and 15 m for Airsim
and StreetView datasets, respectively.
Fourthly, Fig. 6d illustrates the effect of graph extraction in
either image- or 3D-space. The extraction in 3D-space, taking
advantage of the depth information as described in Sec. III-B
shows superior performance. However, X-View still performs
well when localizing a graph built in one space against a graph
built in the other.
Fifthly, Fig. 6e explores the inclusion of different object
classes. The configurations are: Only static object classes,
static object classes plus dynamic object classes, and all object
7
Module
SYNTHIA
Airsim
Blob extraction
2.73 ± 0.65
1.76 ± 0.26
Construction of Gq
337.39 ± 92.81
257.40 ± 28.30
Random Walks Generation
1.38 ± 0.82
1.07 ± 0.56
Matching Gq to Gdb
7.30 ± 4.51
4.33 ± 1.25
Localization Back-End
22.50 ± 9.71
5.15 ± 0.63
Total
371.3 ± 108.5
269.71 ± 31.0
Table I: Timing results in ms, reporting the means and standard deviations
per frame on the best performing configurations on SYNTHIA and Airsim. The
timings were computed on a single core of an Intel Xeon E3-1226 CPU @
3.30GHz.
classes. Here, the results are not conclusive on the SYNTHIA
dataset and more evaluations will be needed in the future.
Lastly, Fig. 6f shows X-View’s performance under seasonal
change. We compare the performance of localizing the query
graph built from the right forward facing camera of one season
in the database graph built from the left forward facing camera
of another season. Here, we consider the summer and fall
sequences of SYNTHIA. The BoW-based techniques perform
well in this scenario if the seasonal conditions are equal.
However, its performance drastically drops for inter-season
localization, while X-View, and NetVLAD suffer much less
under the seasonal change.
The evaluation using PR-curves, and success rates over the
localization error is depicted in Fig. 7. X-View has higher
success rate in multi-view experiments than the appearancebased techniques on both synthetic datasets at our achievable
accuracy of 20 m for SYNTHIA and 30 m on Airsim and
using perfect semantic segmentation inputs as depicted in
Fig. 7c. These accuracies are considered successful as node
locations between Gq and Gdb can differ by twice the merging
distance with our current graph merging strategy. On the
considered operation point of the PR curve, X-View achieves
a localization accuracy of 85 % within 30 m on Airsim, and
85 % on SYNTHIA within 20 m.
Furthermore, X-View expresses comparable or better performance for multi-view localization than the appearancebased techniques using CNN-based semantic segmentation on
the SYNTHIA, and StreetView datasets respectively. Here we
consider successful localizations within 20 m for both datasets.
The achieved accuracies on the chosen operation points are
70 % on SYNTHIA, and 65 % on StreetView.
Finally, we also report timings of the individual components
of our system in Table I. Here, the construction of Gq has by
far the largest contribution, due to iteratively matching and
merging frames into Gq . As the graphs in SYNTHIA consider
more classes and smaller merging distances, these generally
contain more vertices and therefore longer computational
times.
E. Discussion
Global registration of multi-view data is a difficult problem
where traditional appearance based techniques fail. Semantic
graph representations can provide significantly better localization performance under these difficult perceptual conditions. We furthermore give insights how different parameters,
choices, and inputs’ qualities affect the system’s performance.
Our results obtained with X-View show a better localization
performance than appearance-based methods, such as BoW
and NetVLAD.
During our experiments, we observed that some of the
parameters are dependent on each other. Intuitively, the coarseness of the graph has an effect on the random walk descriptors
as a coarser graph contains fewer vertices and therefore
deeper random walks show decreasing performance as Gq
can be explored with short random walks. On the other hand,
an increasing amount of frames used for localization has
the reverse effect on the descriptor depth as Gq potentially
contains more vertices, and deeper random walks do not show
a performance drop as they do for smaller query graphs.
Also the success rate curves indicate that X-View outperforms the appearance based methods particularly in the
presence of strong view-point changes. While the appearancebased methods fail to produce interesting results for the Airsim
dataset, they have a moderate to good amount of successful
localizations on SYNTHIA and StreetView. On the other hand,
X-View has generally higher localization performance and
does not show a strong drop in performance among datasets.
While computational efficiency has not been the main focus
of our research, the achieved timings are close to the typical
requirements for robotic applications.
Finally, we performed experiments both using ground truth
semantic segmentation inputs, and CNN-based semantic segmentation. The performance with semantic segmentation using
AdapNet [11] shows to be close to the achievable performance with ground truth segmentation on SYNTHIA. Using the
SegNet [12] semantic segmentation on real image data from
StreetView demonstrates the effectiveness of our algorithm’s
full pipeline on real data, resulting in better performance than
the best reference algorithm. Despite the high performance,
our system still receives a moderate amount of false localizations, which is due to similar sub-graphs at different locations,
and we hope to mitigate this effect by including it into a full
SLAM system in the future.
Furthermore, 3D locations of the vertices are presently positioned at the blob centers of their first observation. We expect
a more precise positioning technique to further disambiguate
the associations between graphs.
V. CONCLUSIONS
In this paper we presented X-View, a multi-view global
localization algorithm leveraging semantic graph descriptor
matching. The approach was evaluated on one real-world
and two simulated urban outdoor datasets with drastically
different view-points. Our results show the potential of using
graph representations of semantics for large-scale robotic
global localization tasks. Alongside further advantages, such as
compact representation and real-time-capability, the presented
method is a step towards view-point invariant localization.
Our current research includes the investigation of more
sophisticated graph construction methods, the integration of
X-View with a full SLAM system to generate loop closures,
and learning-based class selection for discriminative representations.
8
R EFERENCES
[1] M. Cummins and P. Newman, “Fab-map: Probabilistic
localization and mapping in the space of appearance,”
The International Journal of Robotics Research, vol. 27,
no. 6, pp. 647–665, 2008.
[2] D. Gálvez-López and J. D. Tardos, “Bags of binary words
for fast place recognition in image sequences,” IEEE
Transactions on Robotics, vol. 28, no. 5, pp. 1188–1197,
2012.
[3] S. Lowry, N. Sünderhauf, P. Newman, J. J. Leonard,
D. Cox, P. Corke, and M. J. Milford, “Visual place
recognition: A survey,” IEEE Transactions on Robotics,
vol. 32, no. 1, pp. 1–19, 2016.
[4] R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and
J. Sivic, “Netvlad: Cnn architecture for weakly supervised place recognition,” in IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 5297–
5307.
[5] A. Gawel, T. Cieslewski, R. Dubé, M. Bosse, R. Siegwart, and J. Nieto, “Structure-based vision-laser matching,” in IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), 2016, pp. 182–188.
[6] A. Gawel, R. Dubé, H. Surmann, J. Nieto, R. Siegwart,
and C. Cadena, “3d registration of aerial and ground
robots for disaster response: An evaluation of features,
descriptors, and transformation estimation,” in IEEE International Symposium on Safety, Security, 2017.
[7] Z. Chen, A. Jacobson, N. Sunderhauf, B. Upcroft, L. Liu,
C. Shen, I. Reid, and M. Milford, “Deep learning features
at scale for visual place recognition,” 2017.
[8] E. Stumm, C. Mei, S. Lacroix, J. Nieto, M. Hutter, and
R. Siegwart, “Robust visual place recognition with graph
kernels,” in IEEE Conference on Computer Vision and
Pattern Recognition, 2016, pp. 4535–4544.
[9] Y. Su, F. Han, R. E. Harang, and X. Yan, “A fast kernel
for attributed graphs,” in SIAM International Conference
on Data Mining, 2016, pp. 486–494.
[10] A. Garcia-Garcia, S. Orts-Escolano, S. Oprea, V. VillenaMartinez, and J. Garcia-Rodriguez, “A review on deep
learning techniques applied to semantic segmentation,”
arXiv preprint arXiv:1704.06857, 2017.
[11] A. Valada, J. Vertens, A. Dhall, and W. Burgard, “Adapnet: Adaptive semantic segmentation in adverse environmental conditions,” in IEEE International Conference on
Robotics and Automation (ICRA), 2017, pp. 4644–4651.
[12] V. Badrinarayanan, A. Kendall, and R. Cipolla, “Segnet: A deep convolutional encoder-decoder architecture
for image segmentation,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, 2017.
[13] S. A. Cook, “The complexity of theorem-proving procedures,” in ACM symposium on Theory of computing.
ACM, 1971, pp. 151–158.
[14] M. J. Milford and G. F. Wyeth, “Seqslam: Visual routebased navigation for sunny summer days and stormy
winter nights,” in IEEE International Conference on
Robotics and Automation (ICRA), 2012, pp. 1643–1649.
[15] T. Cieslewski, E. Stumm, A. Gawel, M. Bosse, S. Lynen,
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
and R. Siegwart, “Point cloud descriptors for place
recognition using sparse visual information,” in IEEE
International Conference on Robotics and Automation
(ICRA), 2016, pp. 4830–4836.
M. Bürki, I. Gilitschenski, E. Stumm, R. Siegwart,
and J. Nieto, “Appearance-based landmark selection for
efficient long-term visual localization,” in IEEE International Conference on Intelligent Robots and Systems
(IROS), 2016, pp. 4137–4143.
N. Sunderhauf, S. Shirazi, A. Jacobson, F. Dayoub,
E. Pepperell, B. Upcroft, and M. Milford, “Place
recognition with convnet landmarks: Viewpoint-robust,
condition-robust, training-free,” Robotics: Science and
Systems, 2015.
W. H. Huang and K. R. Beevers, “Topological map merging,” The International Journal of Robotics Research,
vol. 24, no. 8, pp. 601–613, 2005.
D. Marinakis and G. Dudek, “Pure topological mapping
in mobile robotics,” IEEE Transactions on Robotics,
vol. 26, no. 6, pp. 1051–1064, 2010.
I. Kostavelis and A. Gasteratos, “Semantic mapping
for mobile robotics tasks: A survey,” Robotics and Autonomous Systems, vol. 66, pp. 86–103, 2015.
S. L. Bowman, N. Atanasov, K. Daniilidis, and G. J. Pappas, “Probabilistic data association for semantic slam,” in
IEEE International Conference on Robotics and Automation (ICRA), 2017, pp. 1722–1729.
N. Atanasov, M. Zhu, K. Daniilidis, and G. J. Pappas,
“Semantic localization via the matrix permanent.” in
Robotics: Science and Systems, 2014.
B. Perozzi, R. Al-Rfou, and S. Skiena, “Deepwalk:
Online learning of social representations,” 2014, pp. 701–
710.
G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and
A. M. Lopez, “The synthia dataset: A large collection
of synthetic images for semantic segmentation of urban
scenes,” in IEEE Conference on Computer Vision and
Pattern Recognition, 2016, pp. 3234–3243.
S. Shah, D. Dey, C. Lovett, and A. Kapoor, “Airsim: High-fidelity visual and physical simulation for
autonomous vehicles,” in Field and Service Robotics,
2017.
H. Noh, S. Hong, and B. Han, “Learning deconvolution
network for semantic segmentation,” in IEEE International Conference on Computer Vision, 2015, pp. 1520–
1528.
| 1cs.CV
|
1
Training Feedforward Neural Networks with
Standard Logistic Activations is Feasible
Emanuele Sansone, and Francesco G.D. De Natale, Member, IEEE
arXiv:1710.01013v1 [cs.NE] 3 Oct 2017
Abstract—Training feedforward neural networks with standard logistic activations is considered difficult because of the intrinsic
properties of these sigmoidal functions. This work aims at showing that these networks can be trained to achieve generalization
performance comparable to those based on hyperbolic tangent activations. The solution consists on applying a set of conditions in
parameter initialization, which have been derived from the study of the properties of a single neuron from an information-theoretic
perspective. The proposed initialization is validated through an extensive experimental analysis.
Index Terms—Deep Neural Networks, Recurrent Neural Networks, Sigmoid Activations, Initialization
F
1
I NTRODUCTION
D
EEP LEARNING has received a lot of attention in the last
decade due to the impressive performance achieved
in numerous computer vision tasks, including object detection [1], human action recognition [2], image restoration [3]
and image classification [4], and in natural language tasks,
including language modelling [5], parsing [6], machine
translation [7] and speech-to-text translation [8]. The success
of deep learning is due to the capability of transforming input data into representations that are increasingly
more abstract with depth and in a way that resembles the
brain structure of primates [9]. Recent theoretical analysis
provides partial confirmation on the experimental findings
obtained by deep learning models and show that there
is an exponential advantage in terms of the complexity
of functions computed by deep architectures over shallow
ones [10], [11], [12]. 1
Deep learning models are impactful in many real-world
applications, and the transfer of this technology to society
has created new emerging issues, like the need of model
interpretability [14]. The General Data Protection Regulation
approved in 2016 by the European parliament, which will
be effective in 2018, is a concrete example of the need to
provide human understandable justifications for decisions
taken by automated data-processing systems [15].
Research could probably be inspired by old literature
in neural networks to find better explanations about the
dynamics of deep learning and provide more human interpretable solutions. An example of such process is found
in standard logistic activation functions, that have been
studied extensively in the past, but tend to be substituted
by other activation functions in modern neural networks.
To understand why this may be the case, it is important
to recall the unique properties of the logistic function and
therefore analyze the reasons why it has been introduced in
neural networks. Firstly, the standard logistic function is biologically plausible. In fact, it is one of the best differentiable
approximations to the leaky integrate-and-fire model, used
1. We refer to shallow models to indicate networks with a single
hidden layer, while we refer to deep models for networks with more
than one hidden layer [13].
in neuroscience to model the spiking behaviour of biological
neurons [16]. Biological plausibility can be essential for
driving deep learning research towards the uncovering of
human learning dynamics and also providing explanations
that are effectively interpreted by humans [17]. Secondly,
there is theoretical work showing that a family of neural
networks provided with standard logistic activations can be
equivalently converted into fuzzy rule-based systems [18],
thus raising the possibility to perform reasoning using fuzzy
logic and potentially extract human interpretable explanations of predictions made by deep learning models [19].
While the standard logistic function is used extensively
as activation in shallow neural networks, it receives less attention in deep learning. A common justification supporting
this fact is that training these deep neural networks is very
challenging due to the intrinsic properties of the standard
logistic function, like its non-zero mean output value [20],
[21] and its low derivative score in zero [22]. The bounded
alternative to the standard logistic activation is the hyperbolic tangent, which allows an easier training. Nevertheless,
this function is not biological plausible and does not have
any relation with fuzzy logic. This work aims at showing that training deep feedforward neural networks with
standard logistic activations can be feasible through careful
initialization. In particular, we derive some conditions using information theory that are used as principled criteria
for initialization. We show through extensive experimental
analysis that our conditions guarantee a better propagation
of information through the whole network and that during
training no vanishing gradients are observed, thus boosting the convergence speed of the optimizer. The proposed
initialization outperforms the other existing strategies also
in terms of generalization performance and contribute to
bridge the gap between networks with standard logistic
activations and networks with hyperbolic tangents.
The rest of the paper is organized as follows. Section 3
provides a preliminary discussion on the statistical properties of a single neuron with standard logistic activation
function. Section 4 studies the neuron from an informationtheoretic perspective and derives initialization conditions
for its parameters. Section 5 relates the proposed conditions
2
with the problem of vanishing gradients. Finally, Section 6
validates the theory over different well-known benchmarks
and different networks.
1
x1
x2
2
L ITERATURE R EVIEW
In 1986, Rumelhart et al. [23] propose the backpropagation algorithm to train a feedforward neural network.2 In
this seminal work, the authors use random weight initialization to break the symmetry of parameters and allow to
perform credit assignment during training, namely knowing how to compute each weight contribution to the final
error. Nevertheless, the initial choice of parameters plays
an important role on determining the generalization performance of the final trained network, as it is demonstrated
by subsequent works (see [25] for an empirical comparison
among the main works of the period until 2000 and [26] for
an updated summary of the related work up to 2004). In that
period, the research about initialization focused mainly on
shallow architectures, motivated by the fact that (i) shallow
neural networks are universal function approximators [27],
[28] and (ii) deep networks are more difficult to train than
shallow counterparts [29], due to the problem of vanishing and exploding gradients. Authors in [20] are probably
among the few to propose initialization strategies for deep
learning. In particular, they use random weight initialization
in combination with hand-crafted activations for hidden
neurons. We will see in the experimental section that their
proposed initialization strategy is not particularly suited for
standard logistic activations, as it is strongly affected by
vanishing/ exploding gradients.
The first effective strategy to learn deep models appears
in 2006 [30] and consists of splitting the learning process
into two stages, called unsupervised pre-training and finetuning, respectively. In the first stage, an unsupervised
algorithm is applied layerwise to learn increasingly more
complex representations of the input features.3 In the second
stage, network parameters are updated/fine-tuned using a
supervised criterion and gradient-based optimization. An
explanation for this success appears later in [33], where
the authors show experimentally that unsupervised pretraining can be regarded as an effective initialization strategy for the subsequent optimization stage. In other words,
”pre-training guides the learning towards basins of attraction of minima that support better generalization from the
training data set” [33].
Unsupervised pre-training is extremely expensive from
a computational perspective and alternatives are proposed
to overcome the problem of vanishing/exploding gradients.
In particular, there are solutions, which modify the structure
of neural networks with skip connections between hidden
layers [34], [35], [36] or with new normalizing layers [37],
[38], [39], in order to guarantee the continuous flow of
information through the network. Other approaches study
the properties of the loss surface and develop training
algorithms able to find better minima. In particular, we find
2. Even if, the author in [24] argues that backpropagation is dated
back to the early 1960s.
3. Authors in [30] use Restricted Boltzmann machines at each layer
to learn deep belief nets, while authors in [31], [32] use stacked
autoencoders to learn deterministic networks.
..
.
b
w1
w2
Σ
z
g(·)
y
wd
xd
Fig. 1. Graphical visualization of a single neuron.
(i) optimization algorithms based on accelerated gradients,
like momentum [40], [41] and Adam [42], which combine
information about past and current gradients in order to
dampen the oscillations on the loss surface observed during
training, thus converging faster to the final solution, and (ii)
second-order optimization strategies, like Hessian-free [43]
and natural gradient methods [44], [45], which looks for
efficient approximations of the Hessian using the GaussNewton [46] and the Fisher information matrices, respectively. There are a plethora of works studying optimization in neural networks, therefore we invite the interested
reader to see the recent work in [47], which provides a
theorical comparative analysis between different accelerated
gradient-based algorithms, and the survey in [48], which
presents a more general overview of optimization strategies
in machine learning.
One of the most influential studies about initialization in
the last decade is the work of [21]. In particular, the authors
observes that ”the logistic sigmoid activation is unsuited
for deep networks with random initialization because of
its mean value, which can drive especially the top hidden
layer into saturation”. Other more recent works, see for
example [22], confirm the fact that the standard logistic
activation is more difficult to train than other activations and
proper rescaling, making the logistic function similar to the
hyperbolic tangent, is required for successful training. Alternative activations are therefore proposed in the literature.
The rectifier linear function is one of the most appealing
solutions [49],4 because of its unbounded nature that allow
the gradient not to vanish. Principled criteria based on
orthogonality [51] and normalization of weights [52] are
used in combination with random weight initialization to
find better starting solutions for training these networks.
Recent theoretical analysis on the properties of random
inizialization, see [53] for the analysis of rectifier linear
functions and [54] for a more general theory validated also
on hyperbolic tangents, reveals that there exists a range of
values for the variance of weights which are more suited for
the propagation of gradients, thus improving the trainability
of networks. Our work proposes to study the more difficult
problem of training standard logistic activations from the
initialization perspective. Furthermore, we shed light on a
more general criterion to derive initializating conditions,
that explicitly maximizes the amount of information propa-
3
6
µ = -1.5
µ = -0.5
µ = 0.5
µ = 1.5
5
4
3
2
1
0
0.0
0.2
0.4
0.6
0.8
1.0
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0.0
(a)
density py (y) approaches a Bernoulli distribution.6 This is
an extreme case, where information can be propagated, but
the ouput becomes discrete (see Fig. 2(b)).
In the next section, the behaviour of the neuron is analyzed more in detail from the perspective of information
theory.
σ = 1.0
σ = 1.5
σ = 2.0
σ = 2.5
4
0.2
0.4
0.6
0.8
1.0
(b)
Fig. 2. Visualization of density py (y) for different settings of µ, σ . (a)
σ = 0.5 and variable µ, (b) µ = 0 and variable σ .
gation in neural networks.
3
S TATISTICAL BACKGROUND
Let us recall the statistical properties of a single neuron
characterized by an input vector x ∈ Rd , a weight vector
w ∈ Rd , a bias b ∈ R and a standard logistic activation
function g(z) = 1/(1 + e−z ). Fig. 1 provides a graphical interpretation of the computational unit used in many neural
networks.
Consider z as the logit of the given neuron, namely
P
z = di=1 wi xi + b = wT x + b. By modelling the inputs as
independent random variables, with densities/distributions
characterized by finite means and finite variances, it is
possible to exploit the Lyapunov theorem5 and model z as a
Gaussian random
σ 2 . In
Pd variable with mean µ2 and
Pdvariance
2
this case, µ = i=1 wi E{xi } + b and σ = i=1 wi V ar{xi },
where E{·} and V ar{·} are the expected value and the
variance operators, respectively. It is interesting to note that
the mean value of the Gaussian density associated with z
is mainly dominated by the parameter b (especially when
E{xi } = 0, ∀i = 1, . . . , d, since µ = b), while its variance
is influenced only by the weight vector w. Therefore, all
subsequent considerations valid for µ and σ will be valid
also for b and w, respectively.
Due to its nonlinearity, the activation function g(z) produces an output with different statistical properties from the
ones associated with logit z . In fact, the density associated
with output y ∈ (0, 1) can be expressed by the following
relation, namely:
y
(ln 1−y
− µ)2
1
√
exp −
(1)
py (y) =
2σ 2
y(1 − y) 2πσ 2
I NFORMATION -T HEORETIC A NALYSIS
In this section, we ask the following question: which
region of the parameter space (µ, σ) guarantees that the
maximum amount of information is propagated through
the activation function g(z)? We address this question by
formulating the problem as an optimization.
Logit z and output y are modelled as continuous random variables distributed according to N (µ, σ 2 ) and py (y),
respectively. We choose the entropy of y , viz. H(y), as
the objective of the maximization problem and discard for
example the mutual information, since it is not defined for
this particular case.7 Therefore, the objective can be written
in the following way:
Z 1
.
H(y)=−
py (y) ln py (y)dy
0
=
Z
∞
−∞
e−
√
(z−µ)2
2σ 2
(
2πσ 2
(z − µ)2
+
2σ 2
)
i
√
2
+ ln g(z) 1 − g(z) 2πσ
dz
h
=
Z
∞
−∞
e−
√
(z−µ)2
2σ 2
2πσ 2
(
)
√
(z − µ)2
0
2
+ ln 2πσ + ln g (z) dz
2σ 2
=H(z) + Ez {ln g 0 (z)}
(2)
0
where H(z), Ez {·}, g (z) are the entropy, the expected
value and the derivative of g(·) computed on variable z ,
respectively. Note that the second line in (2) is obtained
from the first one by a simple change of variable, namely
y = g(z). Thus, the information coming out from the neuron
is proportial to the information of the logit and the shape of
the activation function.
In general, the term Ez {ln g 0 (z)} in (2) is difficult to
compute analytically and can be approximated using a
lower and an upper bound (see Appendix A). This brings
us to the following inequalities:
HB (µ, σ) − 2 ln 2 ≤ H(y) < HB (µ, σ)
(3)
which is not any more a Gaussian density, as can be seen
from Fig. 2. It is important to mention that µ and σ controls
the mean and variance of py (y), and therefore also the
amount of information that can be propagated through the
neuron. In fact, for very large or very small values of µ,
we have no information propagation, since the activation
function is saturated and the output variance tends to zero
(see Fig. 2(a)). The same happens for small values of σ ,
because in this case py (y) behaves similarly to a Dirac delta
centered at g(µ). It can be shown that, when σ → ∞, the
where erf (·) is the error function [55] and the function
HB (µ, σ) can be visualized in Fig. 3. From 3, HB (µ, σ) defines both the lower and upper bounds of H(y). Therefore,
it can be used as a surrogate objective for our maximization
4. Many different extensions are also proposed. For example, [50]
parameterize the rectifier linear function to allow non-zero gradient in
the negative region.
5. This is an extension of the central limit theorem, which relaxes the
assumptions over the random variables and require that the random
variables are independent but not necessarily identically distributed.
6. py (y) → 0 on the interval (0, 1) and the mass fully concentrates on
the extrema of the interval.
7. It is not defined because of Dirac delta distributions, which come
from the fact that g(·) is deterministic and that the variables are
continuous.
and
.1 1
HB (µ, σ) = + ln(2πσ 2 )−
2 2
µ2
µ
2σ
√
−µ erf
− √ e− 2σ2
σ 2
2π
(4)
4
0.4
0.0
0.4
0.8
1.2
1.6
2.0
2.4
2.8
1
0
1
2
3
3
2
1
5
lies in the linear region of its sigmoid activation, thus constraining the whole network to approximate only a linear
function.
In the next sections, we provide theoretical evidence
about the utility of these conditions.
4
5
4
0
µ
3
1
2
2
1
3 0
σ
Fig. 3. 3D plot of HB (µ, σ). The dotted curve on the surface represents
∂HB (µ,σ)
the set of points for which
= 0.
∂σ
problem to find the optimal values of µ and σ . Based on this
principle, we can enunciate the following theorem (proof in
Appendix B).
Theorem 1. The optimality conditions for HB (µ, σ), defined
in (4), are given by the following statements:
•
•
∀σ ∈ R+ ,
∀µ ∈ R,
∂HB (µ,σ)
∂µ
∂HB (µ,σ)
∂σ
= 0 ⇒ µ = 0,
=0 ⇒ σ=
pπ
2e
W
2µ2
π
2
,
where W (·) is the principal branch of the Lambert
W function [56]. The optimal solution, maximizing
HB (µ,
p σ), is therefore obtained at the point µ = 0 and
σ = π2 ≈ 1.2533.
The optimality result given by Theorem 1 is interpreted as
a condition on µ and σ to have maximum amount of information propagated through the sigmoid activation. On one
Pd
hand, the condition µ = 0 implies that b = − i=1 wi E{xi }
(following directly from Section 3), meaning that the expected value associated with the logit z must lie in the
central part of the sigmoid far away from its saturating
p
π/2 for
regions. On the other hand, the condition σ =
Pd
µ = 0 implies that i=1 wi2 V ar{xi } = π/2 (following
directly from Section 3). Note that, if the variances of inputs
are equal to the same constant, p
namely V ar{xi } = k for
all i = 1, . . . , d,8 then kwk2 = π/(2k). This means that
the maximum amount of information propagated thorugh
g(z) is obtained
when vector w lies on the hypersphere
p
of radius
π/(2k). In Appendix C, we show how to
deal with the more general case where the input variances are different from each other. In practice, condition
Pd
2
T
i=1 wi V ar{xi }. = π/2 may be rewritten as w Dw =
π/2, where D = diag(V ar{x1 }, . . . , V ar{xd }). If D is
positive definite, then the quadratic equation wT Dw = π/2
characterizes a multidimensional ellipsoid. Vector w must
therefore lie on this geometric locus of points to guarantee
the maximal information propagation.
We argue that the obtained results are useful to define
the initial conditions for learning the parameters of neurons
with standard logistic activations and that cannot be enforced during training, because they limit the expressivity
of the neural network. In fact, they impose that each neuron
8. This is a common assumption used in the theory of neural networks [57].
T HE P ROBLEM OF VANISHING G RADIENTS
In this section, we study the implications of maximizing
the mutual information at each neuron on the problem of
vanishing gradients [29], [58]. In particular, we will show
that the conditions established by Theorem 1 ensure that
the selected starting point lies far away from the critical
point that implies the occurrence of the vanishing gradients.
Although the proposed theoretical analysis does not imply
that this condition cannot be reached at some later stage of
the learning, the effectiveness of the proposed initialization
will be also supported by the extensive testing, presented in
the experimental section.
We study the problem of vanishing gradients by adopting the same methodological analysis of [29], [58] and focus
on recurrent neural networks, which can be seen as the
deepest version of feedforward neural nets. The obtained
results are therefore valid for traditional feedforward neural
networks [29], [58].
Recall that a recurrent neural network is fully described
by the following equations:
E t = g(Wout yt + bout )
yt = g(Wyt−1 + Win xt + b)
(5)
where t is used to identify time, or equivalently to indicate
the layer in a deep feedforward neural net. xt ∈ Rd , yt ∈ Rh
and E t ∈ Ro are the input, the hidden state and the output
vectors of the network, respectively. Win ∈ Rh×d , W ∈
Rh×h and Wout ∈ Ro×h are the weight matrices associated
with the connections of neurons, b ∈ Rh and bout ∈ Ro are
the bias vectors and g(·) is an element-wise operator that
applies the sigmoid function to the incoming vector.
Recurrent neural networks are usually trained by minimizing an objective L that sums all loss contributions
incurred over a time horizon of duration T , namely L =
PT
t
t
o
t=1 L , where L : R → R. The training requires computing the gradient of the objective with respect to parameters
W [23], namely:
T
∂L X ∂Lt
=
∂W t=1 ∂W
T
∂Lt X
∂Lt ∂yk
=
diag
∂W k=1
∂yk ∂W
tT
T
∂Lt
∂L ∂E t ∂yt
=
∂yk
∂E t ∂yt ∂yk
k+1
k+1
Y ∂yl
Y
∂yt
=
=
diag g 0 (zl ) W
∂yk l=t ∂yl−1
l=t
(6)
where zl = Wyl−1 + Win xl + b is the vector of logits.
By exploiting the fact that for any real matrices A, B and
for any real vector v, kAvk2 ≤ kAk∗ kvk2 and kABk∗ ≤
kAk∗ kBk∗ , where k · k∗ is the operator-2 norm, it is possible
5
to derive an upper-bound on the gradient norm
fact,
∂Lt
k ∂y
k k2 .
In
given that real-world applications usually require hundreds/thousands of hidden neurons per layer. Therefore,
even if we cannot conclude that the whole training is exempt
of vanishing gradients problems, we inizialize the process
with a sufficient margin to prevent the problem in the initial
phase, which is typically the most critical. Experimental
results will provide evidence of this, as shown later in the
paper.
(7)
6
T
∂Lt ∂E t ∂yt
∂E t ∂yt ∂yk 2
2
∂Lt
∂E t
∂yt
≤
∂E t 2 ∂yt ∗ ∂yk ∗
k+1
∂Lt
∂E t Y ∂yl
≤
∂E t 2 ∂yt ∗ l=t ∂yl−1
∂Lt
∂yk
=
∗
Note that from (6) and (7), we can derive the following
relations:
∂yl
∂yl−1
≤ diag g 0 (zl ) ∗ W
∗
= max g 0 (zil ) W
i=1,...,h
≤
kW
4
F
F
F
In this section, we evaluate the performance of the proposed initialization theory on several benchmarks. We start
by analyzing shallow networks, then consider the case of
deep networks and finally extend the analysis to recurrent
neural networks (RNNs, being even deeper than previous
cases). We compare our initialization against several competitors. Hereunder, we provide a summary of all strategies:
•
Lecun et al.
[57] initialize
the weights according to
(8)
In particular, the first inequality in (8) can be obtained using
the property kAk∗ ≤ kAkF , where k · kF is the Frobenius
norm. The equality in second line follows directly from the
fact that the operator-2 norm of a diagonal matrix is equal
to the maximum of its diagonal entries, whereas the last
inequality is due to the fact that the derivative of a sigmoid
function cannot be larger than 1/4.
By using the result in (8), the gradient in (7) is bounded
by the follwoing inequality:
∂Lt
∂yk
2
≤
∂Lt
∂E t
2
t
=
∂L
∂E t
∂E t
∂yt
k+1
Y
∗ l=t
t
2
∂E
∂yt
E XPERIMENTAL R ESULTS
∗
kW
4
kW
4
F
•
•
F
•
t−k−1
(9)
The vanishing gradient problem refers to the decay of
∂Lt
k ∂y
k k2 as the number of time instants t − k , or equivalently
the number of layers, becomes larger. A sufficient condition
for the occurrence of this problem is given by the condition
kW F < 4, due to the fact that the bound in (9) tends to
zero as t − k → ∞.
q
Ph
2
Note that kWkF =
i=1 kwi k2 , where wi is the i-th
row of matrix W corresponding to the weights of the i-th
neuron. Furthermore, by the conditions derived in Section 4,
viz. kwi k22 = π/(2k) for all i = 1, . . . , h, we have that
r
πh
kWkF =
(10)
2k
If the number of hidden neurons is less than the quantity
.
hcritic = 32k/π , which implies that (10) is less than 4,
then the vanishing gradient problem is guaranteed to occur.
In this case, k refers to the output variance of a sigmoid
activation and is less than or equal to 1/4 (derived from
the fact that the output density of a neuron is limited by a
Bernoulli distribution and its maximum variance is 1/4, see
Section 3). Therefore, hcritic ≤ 8/π ≈ 2.55.
This means that the sufficient condition for the occurrence of the vanishing gradient problem is met only when
the number of hidden neurons is less than 3. In practical
cases, this limit is overcome with a very large margin,
•
√1 , √1
nl
nl
l
wij
∼ U
−
l
wij
∼ U
− √
l
, where wij
is the weight
connecting neuron i with neuron j in layer l and nl
is the number of input neurons to layer l. The biases
are set to zero.
Glorot and Bengio [21] (also known as ”Xavier
initialization”)
initialize the weights
according to
√
6
,
nl +nl+1
√
√
6
nl +nl+1
. The biases are
set to zero.
Saxe et al. [51] propose an initialization where each
weight matrix is randomly generated from the family
of orthogonal matrices, namely Wl is satisfying the
T
relation Wl Wl = I. The biases are set to zero.
Mishkin and Matas [52] extend the initialization of
Saxe et al. by normalizing the weight matrix by the
output variance in each layer. The biases are set to
zero.
Our approach exploits the result of Theorem 1 in Secl
tion 4, and initializes the weights according to wij
=
l
p π w̃ij
9
l
2k kw̃l k , where w̃ij can be obtained by using
j 2
l
either random generation, namely w̃ij
∼ U (−1, 1),
or the orthogonal initialization in [51]. The biases
Pnl
l
are set according to blj = − i=1 wij
E{xli }, where
l
E{xi } is the expected value of input neuron xli .
In the experiments, we use lecun, glorot, ortho and lsuv to
identify the results achieved by [57], [21], [51] and [52],
respectively. random+EP and ortho+EP are instead used to
identify the two versions of our initialization procedure. In
this case, the acronym EP stands for ellyptical projection (see
Section 4 for a detailed discussion).
All experiments presented in the next sections are run on
a Linux machine equipped by 4 cpu @ 3.2 GHz, 16 GB RAM
and a GPU card (NVIDIA TITAN X).
6.1
XOR Case: Shallow Network
In this experiment, we generate 200 samples from four 2dimensional Gaussians. Classes are assigned to create a XOR
.
9. k = V ar{x} ≈ 0.0589, where V ar{x} is computed numerically
using py (y) at optimality (x is therefore the output of the previous
neuron).
1.0
lecun
glorot
ortho
lsuv
random+EP
ortho+EP
1.0
lecun
glorot
ortho
lsuv
random+EP
ortho+EP
0.8
0.6
0.4
0.2
0
500
1000
Iterations
1500
2000
0.0
lecun
glorot
ortho
lsuv
random+EP
ortho+EP
0.8
Test error
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
Train error
Objective
6
0.6
0.4
0.2
0
500
1000
Iterations
(a)
(b)
1500
2000
0.0
0
500
1000
Iterations
1500
2000
(c)
Fig. 4. Learning curves for (a) training objective, (b) training error and (c) test error on the XOR case over different initialization methods.
TABLE 1
Quantitative results on MNIST with shallow network [59] and different
initialization strategies
Method
(*sigmoid)
lecun*
glorot*
ortho*
lsuv*
ortho+EP*
random+EP*
T.E. (%)
1.91±0.03
1.86±0.04
1.86±0.02
1.87±0.01
1.89±0.02
1.78±0.06
Train Time
(103 secs)
2.0
2.0
2.0
2.0
2.0
2.0
TABLE 2
Quantitative results on MNIST with shallow network [59] and dropout
(p = 0.5) for different initialization strategies
Method
(*sigmoid)
[59]
lecun*
glorot*
ortho*
lsuv*
ortho+EP*
random+EP*
T.E. (%)
1.60
1.65±0.03
1.63±0.02
1.62±0.04
1.61±0.03
1.61±0.03
1.59±0.03
Train Time
(103 secs)
6.0
6.0
6.0
6.0
6.0
6.0
configuration. In this case, we use a neural network with
one hidden layer containing two hidden units. This is the
smallest network that can learn correctly the problem (4
modes can be represented more compactly with 2 binary
digits, namely 2 hidden neurons). Gradient descent with
momentum equal to 0.9 and learning rate equal to 0.05 is
applied to the minimization of the cross-entropy objective
function. Fig. 4 shows results over ten repeated experiments (generating new data from the same distribution
and initializing the network differently). It is clear from
these experiments that our proposed initialization allows to
achieve the best solution in a very efficient way. Also, lsuv
is able to achieve comparable performance to our method.
Moreover, it’s interesting to note that ortho obtains the
worst performance, in fact the test error rate, viz. T.E., is
about 25%, meaning that this initialization is not particularly
suited for this non-linear separable scenario.
6.2
MNIST: Shallow Network
In this section, we compare different initialization methods
for a shallow network with 800 hidden units characterized
by sigmoid activation functions and with softmax output
layer on the MNIST bechmark dataset [60]. Mini-batch
gradient descent with momentum equal to 0.9 and learning
rate equal to 0.001 is applied to the minimization of the
cross-entropy objective function. The size of each mini-batch
consists of 50 training samples and the training algorithm is
run for 900k iterations. All results, including the learning
curves, are averaged over 4 different random initializations.
Data is normalized and mean-centered to lie in the range
[−1, 1].
We plot the training curves for the objective computed
on both the training and the validation sets, see Fig. 5.
We see that our proposed initialization scheme allows to
converge faster even in the case of non-deep models. Furthermore, our strategy is quite robust to the initial generation of weights. Previous attempts have shown that random initializations produces solutions with very different
performance, thus requiring very expensive pre-processing
strategies to reduce the variance of generalization estimates, like unsupervised pre-training [33]. To the best of
our knowledge, this is the very first time that random
initialization without unsupervised pre-training allows to
converge to solutions with reduced variance in performance
(see random+EP Fig. 5a and Fig. 5d).
In Table 1, we show the test errors with the related
training times.10 Model selection is performed based on the
minimization of the validation objective. The performance
in terms of generalization are pretty similar for all methods,
thus the main advantage of our strategy consists of faster
convergence.
We apply also dropout [61] (where the dropping probability is set to 0.5) and compare with the state of the
art results reported in [59], which are obtained with the
same network architecture, but with hand-crafted activation
functions. In particular, the authors use squashed hyperbolic
tangent activations, defined as f (z) = Atanh(Bz), where
A = 1.7159 and B = 2/3 (the parametrization derives by
the requirement that f (1) = 1 and f (−1) = −1), because
their function gain is close to 1 in the nominal region and
the computed gradients are therefore less attenuated [60]
compared to sigmoids. Table 2 summarizes the results. To
the best of our knowledge, this is the first experimental trial
demonstrating that sigmoids achieve similar performance to
hyperbolic tangents [59].
10. All training times are equal, since they represent the time to
perform all training epochs.
7
Training objective (Log)
10 1
10 0
10 -1
10 -2
10 -3
Training objective (Log)
10 1
lecun
glorot
ortho
lsuv
random+EP
ortho+EP
10 0
10 -1
10 -2
10 -4
100 200 300 400 500 600 700 800
Iterations (1K mini-batches)
0
10 0
Iterations (1K mini-batches)
0
Iterations (1K mini-batches)
10 0
10 -1
(c)
Validation objective (Log)
10 1
lecun
glorot
ortho
lsuv
random+EP
ortho+EP
random+EP
ortho+EP
500 1000 1500 2000 2500 3000 3500
(b)
Validation objective (Log)
ortho
lsuv
10 -1
10 -3
100 200 300 400 500 600 700 800
(a)
10 1
lecun
glorot
10 -2
10 -3
0
Training objective (Log)
10 1
lecun
glorot
ortho
lsuv
random+EP
ortho+EP
10 0
Validation objective (Log)
10 1
lecun
glorot
ortho
lsuv
random+EP
ortho+EP
lecun
glorot
ortho
lsuv
random+EP
ortho+EP
10 0
10 -1
10 -1
10 -2
0
100 200 300 400 500 600 700 800
0
Iterations (1K mini-batches)
100 200 300 400 500 600 700 800
Iterations (1K mini-batches)
(d)
10 -2
0
500 1000 1500 2000 2500 3000 3500
Iterations (1K mini-batches)
(e)
(f)
Fig. 5. Learning curves for training and validation objectives on MNIST (logarithmic scale): (a) and (d) results with shallow network, (b) and (e)
results with deep network, (c) and (f) results with deep network with data augmentation.
TABLE 3
Quantitative results on MNIST with deep network [62] and different
initialization strategies
Method
(*sigmoid)
lecun*
glorot*
ortho*
lsuv*
ortho+EP*
random+EP*
6.3
T.E. (%)
3.11±0.14
2.94±0.08
3.32±0.11
2.07±0.07
1.85±0.07
1.92±0.08
Train Time
(103 secs)
7.3
7.3
7.3
7.3
7.3
7.3
MNIST: Deep Network
We repeat the experiments of previous subsection with same
setup, but with a deeper network. In particular, we use the
same architecture of [62], consisting of 6 layers (5 hidden
and 1 output layer), where the number of neurons in each
layer is chosen according to a pyramidal structure, namely
2500, 2000, 1500, 1000, 500 and 10 neurons, respectively.
From Fig. 5b and Fig. 5e, we observe immediately that
our initialization provides a gain in terms of both convergence speed and the quality of the obtained solution. It is
also interesting to see that the competitors observe very
slow learning in the early stages of training, probably due
to the fact that the network parameters are initialized on
flat regions of the training objective and the gradients are
therefore vanishing.
We also analyze the evolution of the statistics of each
layer over the first 90k iterations (we consider the mean and
the standard deviation of each layer obtained by averaging
neuron activities over mini-batches). Fig. 6 shows the behaviour of the network for different initialization strategies.
For almost all competitors, the activities in the last hidden
layer (layer 5) tend to be biased towards zero and no sig-
TABLE 4
Quantitative results on MNIST with deep network [62] and data
augmentation for different initialization strategies
Method
T.E. (%)
(*sigmoid)
[62]
0.32
[62]+
0.62
lecun*
0.80
glorot*
0.82
ortho*
0.80
lsuv*
0.79
ortho+EP*
0.81
random+EP*
0.68
+ Our implementation
Train Time
(103 secs)
412.2
35.4
35.4
35.4
35.4
35.4
35.4
35.4
nificant variation is appreciated in the output layer. This is
a symptomatic behaviour, that have been already observed
in [21] for networks with standard logistic activations. It is
interesting to mention that this problem is not visible for
lsuv and our proposed initializations. This is probably due
to the fact that both impose some conditions on the variance
of the weights of each neuron. In particular, lsuv imposes a
unit variance, while our conditions require a larger value,
which is based on an information-theoretically criterion.
This has also some beneficial impact on the generalization
performance as shown in Table 3. Therefore, our weight
initialization is sufficient to guarantee a better propagation
of gradients at all layers for the whole training process,
allowing to achieve faster convergence and better solutions.
We apply also data augmentation to compare against the
state of the art results reported in [62] with squashed hyperbolic tangent activations. In particular, training data are augmented following the methodology suggested in [62]. We
use elastic distortion, to emulate uncontrolled oscillations of
8
Train activities for lecun
1.0
Layer 1
Layer 2
0.8
Layer 3
Layer 4
Train activities for glorot
1.0
Layer 5
Layer 6
Layer 1
Layer 2
0.8
Layer 3
Layer 4
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
10
20
30
40
50
60
70
80
90
0.0
0
10
20
30
40
(a)
Layer 1
Layer 2
0.8
60
70
80
90
0.0
Layer 3
Layer 4
Layer 5
Layer 6
1.0
Train activities for random+EP
0.8
Layer 1
Layer 2
Layer 3
Layer 4
0.4
0.4
0.4
0.2
0.2
0.2
20
30
40
30
50
60
70
80
90
0.0
0
10
20
30
40
(d)
40
50
60
70
80
90
50
60
70
Train activities for ortho+EP
80
Layer 1
Layer 2
0.8
0.6
10
20
1.0
Layer 5
Layer 6
0.6
0
10
Layer 5
Layer 6
(c)
0.6
0.0
0
Layer 3
Layer 4
(b)
Train activities for lsuv
1.0
50
Layer 1
Layer 2
0.8
0.6
0.0
Train activities for ortho
1.0
Layer 5
Layer 6
90
0.0
0
10
20
(e)
Layer 3
Layer 4
30
40
50
Layer 5
Layer 6
60
70
80
90
(f)
Fig. 6. Temporal evolution of statistics (mean and standard deviation) of activations in each layer for different initialization methods over the first 90k
iterations.
Train activities for lecun
1.0
Layer 1
Layer 2
0.8
Layer 3
Layer 4
Train activities for glorot
1.0
Layer 5
Layer 6
Layer 1
Layer 2
0.8
Layer 3
Layer 4
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
50
100 150 200 250 300 350 400
0.0
0
50
100 150 200 250 300 350 400
(a)
Layer 1
Layer 2
0.8
Layer 3
Layer 4
0.0
Layer 5
Layer 6
1.0
Train activities for random+EP
0.8
Layer 1
Layer 2
Layer 3
Layer 4
Layer 5
Layer 6
100 150 200 250 300 350 400
0.4
0.4
0.4
0.2
0.2
0.2
100 150 200 250 300 350 400
0.0
0
50
100 150 200 250 300 350 400
(d)
(e)
Layer 1
Layer 2
0.8
0.6
50
Layer 5
Layer 6
Train activities for ortho+EP
1.0
0.6
0
50
Layer 3
Layer 4
(c)
0.6
0.0
0
(b)
Train activities for lsuv
1.0
Layer 1
Layer 2
0.8
0.6
0.0
Train activities for ortho
1.0
Layer 5
Layer 6
0.0
0
50
Layer 3
Layer 4
Layer 5
Layer 6
100 150 200 250 300 350 400
(f)
Fig. 7. Temporal evolution of statistics (mean and standard deviation) of activations in each layer for different initialization methods over the first 90k
iterations. Data augmentation is applied to improve performance.
hand muscles (σ = 5.5 and α = 37, see [62]),11 rotation (with
an angle randonly sampled from [−β, β], where β = 7.5 for
digits 1 and 7 and β = 15 for all other digits) and horizontal
and vertical scaling (with scaling factor randomly sampled
11. Here, σ does not refer to σ introduced in Section 3.
from [1 − γ/100, 1 + γ/100], where γ = 17.5). We also rerun
the experiments of [62] and use these as baseline, since we
were not able to replicate the results.
The learning curves in Fig. 5c and Fig. 5f confirms the
findings that our method achieves faster convergence and
better solutions. It is important to mention that the use of
9
1.09
1.08
1.07
1.06
1.05
1.04
1.03
1.02
1.01
Training objective
lecun
glorot
ortho
0
lsuv
random+EP
20
40
ortho+EP
LSTM
60
Iterations (1K mini-batches)
80
(a)
1.09
1.08
1.07
1.06
1.05
1.04
1.03
1.02
1.01
Test objective
lecun
glorot
ortho
0
lsuv
random+EP
20
40
60
ortho+EP
LSTM
Iterations (1K mini-batches)
80
(b)
Fig. 8. Training and test learning curves on the copy memory problem
for different strategies.
vt
We
xt
Win
LSTM
or
RNN
Wout
Et
Fig. 9. General architecture used in the experiments with recurrent neural networks. vt , E t are the input and the output vectors, respectively,
representing a charater or a word using one-hot encoding, while xt
represents the distributed embedding of the input vector.
data augmentation plays an important role in achieving
better generalization performance and that it can partially
overcome to problems incurred by using a wrong initialization. Nevertheless, the advantages of our initialization are
still visible. Fig. 7 shows the temporal evolution of statistics
for each layer, emphasizing the fact that almost all competitors are subject to the problem of saturation of the last
hidden layer in the early stages of training [21]. Table 4 summarizes the quantitative results. Our initialization strategy
(random+EP) allows to achieve 0.68% of test error rate, which
is very close to the performance (namely 0.62%) obtained
by using the hand-crafted hyperbolic tangent of [62]. This
provides further evidence that deep networks with standard
logistic activations can perform similarly to networks with
hyperbolic tangents and that the training is made feasible
through some simple conditions at initialization.
6.4 Copying Memory Problem: Recurrent Neural Network
In this section, we go even deeper and use RNNs as case
study to compare the different initialization strategies. It is
well known that these models have difficulties to remember
information about inputs for long time intervals. This is
due to the fact that during training, RNNs are affected by
the problem of vanishing/exploding gradients [29], [58].
The literature contains a plethora of proposed solutions.
In particular, the authors in [63], [64] propose a strategy
called Echo-State Networks (ESN), which consist in carefully initializing the recurrent weights and training only the
output parameters. In practice, the recurrent weight matrix
is initialized to have a spectral radius close to one, such
that inputs can ”echo” for long time. This represents a very
drastic solution, which doesn’t exploit the full potential of
RNNs. Authors in [65] show that long-term dependencies
can be learnt using Hessian-free optimization, which is provided with information about the curvature of the loss surface and is therefore able to deal with vanishing gradients.
Authors in [58] propose a solution to train RNNs, where
a regularizer, specifically designed to cope with vanishing
gradients, is applied to the loss objective and a simple
momentum-based optimizer is used in combination with
gradient clipping, that ensures that gradients do no explode.
In our experiments, we use the same strategies of [58] and
we focus on analyzing the impact of initialization on this
problem. It is also important to mention that there is a recent
line of research in RNNs, studying unitary recurrent weight
matrices, see for example [66], [67], [68]. Nevertheless, in
this work we want to study the most general case, where
the feasible set consists of the whole parameter space.
In order to study the capability of models in learning
long-term dependencies, we consider a similar pathological
task introduced for the first time in [69], called the copy
memory problem. In this task, we are given a dictionary
of ten characters, viz. {ai }9i=0 . The input is a T + 6 length
sequence containing characters from the dictionary. Specifically, the first three characters in the sequence are drawn
uniformly and independently with replacement from the
subset {ai }7i=0 and represent the sequence to be memorized.
The next T − 1 characters are set to a8 and represent
a dummy sequence. The next character is set to a9 and
represent a trigger to inform the model that it should start
to predict the memorized sequence. The last three characters
are set to a8 and represent a dummy sequence. The ground
truth output is a T + 6 length sequence, where the first T + 3
entries are set to a8 and the last three characters are the copy
of the first characters in the input sequence. Therefore, the
task consists in memorizing the input sequence for T time
instants and then output the memorized sequence. In the
experiments, we set T = 100 and generate training and test
datasets consisting of 1000 samples each.
We compare RNNs using different initialization strategies and report also the results of two baselines. The first
baseline consists in outputting a8 for the first T + 3 entries
and randomly sampling from the subset {ai }7i=0 for the last
three characters. This is equivalent to having a memoryless
strategy. The second baseline consists in predicting the output using a LSTM model [69], which is the most widely used
10
TABLE 5
Quantitative results on the language modelling task for different
strategies.
Method
(*sigmoid)
lstm
lecun*
glorot*
ortho*
lsuv*
ortho+EP*
random+EP*
T.P.
128.3±0.5
143.1±0.8
155.1±0.9
156.1±0.5
135.9±0.1
135.5±0.3
164.0±0.5
Train Time
(103 secs)
77.0
28.7
28.7
28.7
28.7
28.7
28.7
alternative to RNNs and is able to learn long-term dependencies. Performance are measured in terms of perplexity
and this quantity is also used as training objective. Note
that the perplexity of the first baseline can be analytically
8
computed and is equal to exp 3Tln
+6 ≈ 1.06.
The network architecture used in the experiments is
shown in Fig. 9, where the input and the output vectors
represent the one-hot encoding of any character in the
dictionary. The size of hidden layers in the recurrent model
(which is also used to determine the size of the embedding
vector) depends whether we are using RNN or LSTM. In
particular, RNN and LSTM have hidden layers of size 100
and 52, respectively, which is equivalent to have roughly
32000 parameters per model. Adam optimizer with learning
rate equal to 0.001 is used in training and the algorithm
is run for 100k iterations. All results are averaged over 4
different random initializations.
Figure 8 shows the learning curves for the different initialization strategies as well as the learning curves of the two
baselines. It is possible to see that eventually all approaches
perform better than the memoryless strategy, meaning that
the networks effectively learn to memorize some information. Furthemore, our initialization (random+EP) outperforms all other approaches during the first 30k iterations,
including LSTM. There is a phase between 30k and 50k
iterations, in which ortho achieves the best performance. As
discussed in [51], an orthogonal initialization can guarantee
that the recurrent weight matrix remains orthogonal during
the whole training process even in the case of nonlinear activation functions. This is particularly useful in the copying
memory problem, since the family of orthogonal matrices
contain the solution, that allows the RNN to learn the
identity function, namely to copy the exact input sequence
to the output. Note that after 50k iterations the performance
of ortho degrades due to the problem of exploding gradients.
It is interesting to note that at convergence we achieve the
best performance among the other initializations and are
able to get closer to the results of LSTM.
6.5
TABLE 6
Quantitative results on the language modelling task for different
strategies with dropout.
Language modelling: Recurrent Neural Network
In this section, we conduct experiments on a real-world task,
namely the language modelling problem, using the Penn
Tree Bank (PTB) dataset [70]. The dataset consists of 929k
training words, 73k validation words, 82k test words and
the vocabulary has 10k words. The dataset is downloaded
from Tomas Mikolov’s webpage.12
12. http://www.fit.vutbr.cz/ imikolov/rnnlm/simple-examples.tgz
Method
(*sigmoid)
lstm
lecun*
glorot*
ortho*
lsuv*
ortho+EP*
random+EP*
T.P.
128.7±0.4
147.8±0.6
159.2±0.4
161.2±0.3
139.8±0.4
139.3±0.3
130.1±0.4
Train Time
(103 secs)
80.0
32.3
32.3
32.3
32.3
32.3
32.3
We compare RNNs using different initialization strategies and report also the results of the baseline using
LSTM [69]. The aim of these experiments is to show that our
conditions allow to improve the generalization performance
of RNNs. With our initialization, we are able to achieve
comparable performance to the one obtained by LSTM.
The network architecture is the same as the one used in
the copying memory problem with different size of the hidden layers. In particular, we use 200 neurons for RNNs and
190 neurons for LSTM. This corresponds to have roughly
4090k parameters for each model.
Adam optimizer with learning rate equal to 0.0001 and
gradient clipping [58] is used for training the models and
the algorithm is run for 500k iterations.
Results are averaged over 4 different random initializations and are shown in Table 5 (the acronym T.P. stands
for test perplexity). lsuv and ortho+EP achieve performance
closer to the ones of LSTM, while random+EP obtains apparently very bad performance. We argue that the generation
of the initial matrices in random+EP allows to sample from
a much larger space of matrices and consequently Adam
has more chances to overfit. Note that the problem of
overfitting of the Adam optimizer is known and is discussed
from a theoretical perspective in [47]. To validate this claim
and potentially reduce the problem of overfitting, we run
also the experiments with dropout following the procedure
in [71], where the dropping probability is set to 0.5. Results,
shown in Table 6, confirm our initial claim. random+EP
outperforms all other initialization and is able to achieve
performance comparable to LSTM. Note that RNNs offer a
clear advantage in terms of computational complexity over
LSTM, as demonstrated by the training times in Table 6.
7
C ONCLUSION
This work shows that a careful initialization of the
parameters is sufficient to successfully train feedforward
neural networks with standard logistic activations. The initialization is based on some conditions, that are derived
by studying the properties of a single neuron through
information theory. The study is corroborated by numerous
experiments over different known benchmarks and different
networks.
A PPENDIX A
D ERIVATION OF B OUNDS (3)
Recall that
H(y) = H(z) + Ez {ln g 0 (z)}
(11)
11
where (we drop the extrema of integration for the sake of
brevity in the notation)
Z
.
H(z)= − N (µ, σ 2 ) ln N (µ, σ 2 )dz
Z
√
1
2
2
(z
−
µ)
N
(µ,
σ
)dz
+
ln(
2πσ 2 )
=
2σ 2
1 1
= + ln(2πσ 2 )
(12)
2 2
and
.
Ez {ln g 0 (z)}=
Z
N (µ, σ 2 ) ln g 0 (z)dz
µ2
e 2σ2
2
(17)
=√
σ
2π
. µ
By the change of variable t = σ√
, (17) simplifies in the
2
following equation:
2
µ
tet = √
(18)
π
whose solution is given by:
Z
N (µ, σ 2 ) ln g(z) 1 − g(z) dz
Z
ez
2
dz
= N (µ, σ ) ln
(1 + ez )2
Z
Z
= zN (µ, σ 2 )dz − 2 N (µ, σ 2 ) ln(1 + ez )dz
Z
(13)
= µ − 2 N (µ, σ 2 ) ln(1 + ez )dz
=
where the first line is zero if and only if µ = 0, thus proving
the first condition in Theorem 1, while, by setting the second
line to zero, we obtain the following equation:
Note that the integrand in (13) is always positive and that
max(0, z) < ln(1 + ez ) ≤ max(0, z) + ln 2 is true for any
z . Therefore, we can easily derive the following relations,
namely:
0
A(µ, σ) − 2 ln 2 ≤ Ez {ln g (z)} < A(µ, σ)
(14)
where
.
A(µ, σ)= µ − 2
Z
∞
Z−∞
∞
2
N (µ, σ ) max(0, z)dz
zN (µ, σ 2 )dz
µ2
µ
2σ
√
= µ − µ − µ erf
− √ e− 2σ2
σ 2
2π
2σ − µ22
µ
√
− √ e 2σ
= −µ erf
σ 2
2π
=µ−2
0
(15)
By adding H(z) to (14) and using (11), we obtain that
H(z) + A(µ, σ) − 2 ln 2 ≤ H(y) < H(z) + A(µ, σ)
HB (µ, σ) − 2 ln 2 ≤ H(y) < HB (µ, σ)
(16)
.
where HB (µ, σ) = H(z) + A(µ, σ).
A PPENDIX B
P ROOF OF T HEOREM 1
Recall the definition of the lower bound in (4), namely:
√
µ2
µ
2σ
. 1
2
√ − √ e− 2σ2
HB (µ, σ) = + ln 2πσ −µ erf
2
σ 2
2π
derf (t)
2
By the fact that
= √2π e−t and using standard
dt
calculus of derivatives, we can get that:
∂HB (µ, σ)
µ
√
= −erf
∂µ
σ 2
∂HB (µ, σ) 1
2 − µ22
= − √ e 2σ
∂σ
σ
2π
2µ2
π
2
W
µ
t = √ e−
π
(19)
where W (·) is the principal branch of the Lambert W function [56]. We can check whether (19) is the solution of (18),
by substituting (19) in (18) and see that the equality (18)
holds, namely:
2µ2
2µ2
µ2 −W
µ −W π
µ
π
e
2
π
√ e
=√
e
π
π
2
2µ2
e−
e−
e−
W
W
W
π
2
e
2
2µ
π
2
e
2
2µ
π
2
e
µ2
π
e
2µ2 e
π
W
2µ
π
−W
2µ2
π
−W
2µ2
π
2
=1
2
=1
=1
where the fourth line is obtained by the identity t =
W (t)eW (t) , or equivalently te−W (t) = W (t). Therefore,
solution (19) can be explicited in terms of σ , thus giving
us the second condition in Theorem 1.
To prove the optimality of the the obtained solution, we
perform a second-order derivative test, namely:
r
2 1 − µ22
∂ 2 HB (µ, σ)
=
−
e 2σ
2
∂µ
πσ
r
∂ 2 HB (µ, σ)
1
2 µ2 − µ22
=− 2 −
e 2σ
2
∂σ
σ
π σ2 r
∂ 2 HB (µ, σ) ∂ 2 HB (µ, σ)
2 µ − µ22
(20)
=
=
e 2σ
∂µ∂σ
∂σ∂µ
π σ2
Note that the first two equations in (20) are always negative,
while the third equation is zero at µ = 0. Therefore the
Hessian matrix for the obtained stationary point is negativedefinite.
Q.E.D.
A PPENDIX C
G ENERAL C ASE OF D IFFERENT I NPUT VARIANCES
Suppose that we are given an initial weight vector
denoted by w̃ ∈ Rd . In general, w̃ does not satisy
the optimality condition given by Theorem 1, namely
Pd
2
i=1 w̃i V ar{xi } 6= π/2. We need to find the closest vector
to w̃ that satisfies the condition in order to have maximum
amount of information propagation. This can be formulated
as the following optimization problem:
minkw − w̃k22
w
s.t. wT Dw = π/2
(21)
12
.
where D = diag(V ar{x1 }, . . . , V ar{xd }) and the solution
can be obtained by using existing iterative non-convex optimization procedures, due to the fact that the feasible set
in (21) is not convex.
ACKNOWLEDGMENTS
We gratefully acknowledge NVIDIA Corporation with the
donation of a Titan X Pascal machine to support this research. We thank Massimo Zanetti for insightful discussion.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Region-Based
Convolutional Networks for Accurate Object Detection and Segmentation,” TPAMI, vol. 38, no. 1, pp. 142–158, 2016.
S. Ji, W. Xu, M. Yang, and K. Yu, “3D Convolutional Neural
Networks for Human Action Recognition,” TPAMI, vol. 35, no. 1,
pp. 221–231, 2013.
C. Dong, C. C. Loy, K. He, and X. Tang, “Image Super-Resolution
Using Deep Convolutional Networks,” TPAMI, vol. 38, no. 2, pp.
295–307, 2016.
A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet Classification with Deep Convolutional Neural Networks,” in NIPS, 2012,
pp. 1097–1105.
R. Jozefowicz, O. Vinyals, M. Schuster, N. Shazeer, and Y. Wu,
“Exploring the Limits of Language Modeling,” arXiv preprint
arXiv:1602.02410, 2016.
O. Vinyals, Ł. Kaiser, T. Koo, S. Petrov, I. Sutskever, and G. Hinton,
“Grammar as a Foreign Language,” in NIPS, 2015, pp. 2773–2781.
I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to Sequence
Learning with Neural Networks,” in NIPS, 2014, pp. 3104–3112.
A. Graves and N. Jaitly, “Towards End-to-End Speech Recognition
with Recurrent Neural Networks,” in ICML, 2014, pp. 1764–1772.
N. Kruger, P. Janssen, S. Kalkan, M. Lappe, A. Leonardis, J. Piater,
A. J. Rodriguez-Sanchez, and L. Wiskott, “Deep Hierarchies in the
Primate Visual Cortex: What Can We Learn for Computer Vision?”
TPAMI, vol. 35, no. 8, pp. 1847–1871, 2013.
G. F. Montufar, R. Pascanu, K. Cho, and Y. Bengio, “On the number
of linear regions of deep neural networks,” in NIPS, 2014, pp.
2924–2932.
M. Raghu, B. Poole, J. Kleinberg, S. Ganguli, and J. Sohl-Dickstein,
“On the Expressive Power of Deep Neural Networks,” ICML, 2017.
B. Poole, S. Lahiri, M. Raghu, J. Sohl-Dickstein, and S. Ganguli,
“Exponential Expressivity in Deep Neural Networks Through
Transient Chaos,” in NIPS, 2016, pp. 3360–3368.
Y. Bengio et al., “Learning Deep Architectures for AI,” Foundations
and trends R in Machine Learning, vol. 2, no. 1, pp. 1–127, 2009.
F. Doshi-Velez and B. Kim, “Towards a Rigorous Science of Interpretable Machine Learning,” arXiv preprint arXiv:1702.08608, 2017.
B. Goodman and S. Flaxman, “European Union Regulations on
Algorithmic Decision-Making and a ’Right to Explanation’,” arXiv
preprint arXiv:1606.08813, 2016.
P. Dayan and L. F. Abbott, Theoretical Neuroscience. Cambridge,
MA: MIT Press, 2001, vol. 806.
D. Hassabis, D. Kumaran, C. Summerfield, and M. Botvinick,
“Neuroscience-Inspired Artificial Intelligence,” Neuron, vol. 95,
no. 2, pp. 245–258, 2017.
J. M. Benı́tez, J. L. Castro, and I. Requena, “Are Artificial Neural
Networks Black Boxes?” IEEE Transactions on Neural Networks,
vol. 8, no. 5, pp. 1156–1164, 1997.
S. Mitra and Y. Hayashi, “Neuro-fuzzy Rule Generation: Survey
in Soft Computing Framework,” IEEE Transactions on Neural Networks, vol. 11, no. 3, pp. 748–768, 2000.
Y. LeCun, L. Bottou, G. B. Orr, and K.-R. Müller, “Efficient BackProp,” in NIPS Workshop. Springer-Verlag, 1998, pp. 9–50.
X. Glorot and Y. Bengio, “Understanding the Difficulty of Training
Deep Feedforward Neural Networks,” in AISTATS, 2010.
B. Xu, R. Huang, and M. Li, “Revise Saturated Activation Functions,” ICLR Workshop, 2016.
D. E. Rumelhart and G. E. Hintonf, “Learning Representations by
Back-Propagating Errors,” Nature, vol. 323, p. 9, 1986.
J. Schmidhuber, “Deep Learning in Neural Networks: An
Overview,” Neural Networks, vol. 61, pp. 85–117, 2015.
M. Fernández-Redondo and C. Hernandez-Espinosa, “A Comparison among Weight Initialization Methods for Multilayer Feedforward Networks,” in IJCNN, vol. 4. IEEE, 2000, pp. 543–548.
[26] X. M. Zhang, Y. Q. Chen, N. Ansari, and Y. Q. Shi, “MiniMax Initialization for Function Approximation,” Neurocomputing,
vol. 57, pp. 389–409, 2004.
[27] G. Cybenko, “Approximation by Superpositions of a Sigmoidal
Function,” Mathematics of Control, Signals, and Systems (MCSS),
vol. 2, no. 4, pp. 303–314, 1989.
[28] K. Hornik, M. Stinchcombe, and H. White, “Multilayer Feedforward Networks are Universal Approximators,” Neural networks,
vol. 2, no. 5, pp. 359–366, 1989.
[29] Y. Bengio, P. Simard, and P. Frasconi, “Learning Long-Term Dependencies with Gradient Descent is Difficult,” IEEE Transactions
on Neural Networks, vol. 5, no. 2, pp. 157–166, 1994.
[30] G. E. Hinton and R. R. Salakhutdinov, “Reducing the Dimensionality of Data with Neural Networks,” Science, vol. 313, no. 5786,
pp. 504–507, 2006.
[31] Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle, “Greedy
Layer-Wise Training of Deep Networks,” in NIPS, 2007, pp. 153–
160.
[32] C. Poultney, S. Chopra, Y. L. Cun et al., “Efficient Learning of
Sparse Representations with an Energy-Based Model,” in NIPS,
2007, pp. 1137–1144.
[33] D. Erhan, Y. Bengio, A. Courville, P.-A. Manzagol, P. Vincent,
and S. Bengio, “Why Does Unsupervised Pre-training Help Deep
Learning?” JMLR, vol. 11, no. Feb, pp. 625–660, 2010.
[34] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for
Image Recognition,” in CVPR, 2016, pp. 770–778.
[35] R. K. Srivastava, K. Greff, and J. Schmidhuber, “Highway Networks,” ICML Workshop, 2015.
[36] G. Huang, Z. Liu, K. Q. Weinberger, and L. van der Maaten,
“Densely Connected Convolutional Networks,” CVPR, 2017.
[37] S. Ioffe and C. Szegedy, “Batch Normalization: Accelerating Deep
Network Training by Reducing Internal Covariate Shift,” in ICML,
2015, pp. 448–456.
[38] J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer Normalization,” NIPS
Deep Learning Symposium, 2016.
[39] T. Salimans and D. P. Kingma, “Weight Normalization: A Simple
Reparameterization to Accelerate Training of Deep Neural Networks,” in NIPS, 2016, pp. 901–901.
[40] B. T. Polyak, “Some Methods of Speeding up the Convergence of
Iteration Methods,” USSR Computational Mathematics and Mathematical Physics, vol. 4, no. 5, pp. 1–17, 1964.
[41] I. Sutskever, J. Martens, G. Dahl, and G. Hinton, “On the Importance of Initialization and Momentum in Deep Learning,” in
ICML, 2013, pp. 1139–1147.
[42] D. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” ICLR, 2014.
[43] J. Martens, “Deep Learning via Hessian-Free Optimization,” in
ICML, 2010, pp. 735–742.
[44] S.-I. Amari, “Natural Gradient Works Efficiently in Learning,”
Neural Computation, vol. 10, no. 2, pp. 251–276, 1998.
[45] R. Grosse and R. Salakhudinov, “Scaling up Natural Gradient by
Sparsely Factorizing the Inverse Fisher Matrix,” in ICML, 2015, pp.
2304–2313.
[46] N. N. Schraudolph, “Fast Curvature Matrix-Vector Products for
Second-Order Gradient Descent,” Neural Computation, vol. 14,
no. 7, pp. 1723–1738, 2002.
[47] A. C. Wilson, R. Roelofs, M. Stern, N. Srebro, and B. Recht,
“The Marginal Value of Adaptive Gradient Methods in Machine
Learning,” arXiv preprint arXiv:1705.08292, 2017.
[48] L. Bottou, F. E. Curtis, and J. Nocedal, “Optimization Methods for
Large-Scale Machine Learning,” arXiv preprint arXiv:1606.04838,
2016.
[49] X. Glorot, A. Bordes, and Y. Bengio, “Deep Sparse Rectifier Neural
Networks,” in AISTATS, 2011, pp. 315–323.
[50] K. He, X. Zhang, S. Ren, and J. Sun, “Delving Deep into Rectifiers:
Surpassing Human-Level Performance on Imagenet Classification,” in ICCV, 2015, pp. 1026–1034.
[51] A. M. Saxe, J. L. McClelland, and S. Ganguli, “Exact Solutions
to the Nonlinear Dynamics of Learning in Deep Linear Neural
Networks,” ICLR, 2014.
[52] D. Mishkin and J. Matas, “All You Need is a Good Init,” ICLR,
2016.
[53] D. Sussillo and L. Abbott, “Random Walk Initialization for
Training Very Deep Feedforward Networks,” arXiv preprint
arXiv:1412.6558, 2014.
[54] S. S. Schoenholz, J. Gilmer, S. Ganguli, and J. Sohl-Dickstein,
“Deep Information Propagation,” ICLR, 2017.
13
[55] A. Papoulis and S. U. Pillai, Probability, Random Variables, and
Stochastic Processes. Tata McGraw-Hill Education, 2002.
[56] R. M. Corless, G. H. Gonnet, D. E. Hare, D. J. Jeffrey, and D. E.
Knuth, “On the LambertW Function,” Advances in Computational
Mathematics, vol. 5, no. 1, pp. 329–359, 1996.
[57] Y. A. LeCun, L. Bottou, G. B. Orr, and K.-R. Müller, “Efficient
Backprop,” in Neural Networks: Tricks of the Trade. Springer, 2012,
pp. 9–48.
[58] R. Pascanu, T. Mikolov, and Y. Bengio, “On the Difficulty of
Training Recurrent Neural Networks,” in ICML, 2013, pp. 1310–
1318.
[59] P. Y. Simard, D. Steinkraus, J. C. Platt et al., “Best Practices for
Convolutional Neural Networks Applied to Visual Document
Analysis.” in ICDAR, 2003.
[60] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-Based
Learning Applied to Document Recognition,” Proceedings of the
IEEE, vol. 86, no. 11, pp. 2278–2324, 1998.
[61] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov, “Dropout: A Simple Way to Prevent Neural
Networks from Overfitting,” JMLR, vol. 15, no. 1, pp. 1929–1958,
2014.
[62] D. C. Cireşan, U. Meier, L. M. Gambardella, and J. Schmidhuber,
“Deep, Big, Simple Neural Nets for Handwritten Digit Recognition,” Neural Computation, vol. 22, no. 12, pp. 3207–3220, 2010.
[63] H. Jaeger and H. Haas, “Harnessing Nonlinearity: Predicting
Chaotic Systems and Saving Energy in Wireless Communication,”
Science, vol. 304, no. 5667, pp. 78–80, 2004.
[64] M. Lukoševičius and H. Jaeger, “Reservoir Computing Approaches to Recurrent Neural Network Training,” Computer Science
Review, vol. 3, no. 3, pp. 127–149, 2009.
[65] J. Martens and I. Sutskever, “Learning Recurrent Neural Networks
with Hessian-Free Optimization,” in ICML, 2011, pp. 1033–1040.
[66] M. Arjovsky, A. Shah, and Y. Bengio, “Unitary Evolution Recurrent
Neural Networks,” in ICML, 2016, pp. 1120–1128.
[67] L. Jing, Y. Shen, T. Dubček, J. Peurifoy, S. Skirlo, M. Tegmark, and
M. Soljačić, “Tunable Efficient Unitary Neural Networks (EUNN)
and Their Application to RNNs,” in ICML, 2017, pp. 1733–1741.
[68] Z. Mhammedi, A. Hellicar, A. Rahman, and J. Bailey, “Efficient
Orthogonal Parametrisation of Recurrent Neural Networks Using
Householder Reflections,” pp. 2401–2409, 2017.
[69] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,”
Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997.
[70] M. P. Marcus, M. A. Marcinkiewicz, and B. Santorini, “Building a
Large Annotated Corpus of English: The Penn Treebank,” Computational linguistics, vol. 19, no. 2, pp. 313–330, 1993.
[71] W. Zaremba, I. Sutskever, and O. Vinyals, “Recurrent Neural
Network Regularization,” ICLR, 2015.
| 9cs.NE
|
ÉCOLE POLYTECHNIQUE DE MONTRÉAL
ELECTRICAL ENGINEERING DEPARTMENT
arXiv:1608.05786v1 [cs.SY] 20 Aug 2016
AUTOMATION SECTION
DESIGN OF A TRAJECTORY TRACKING
CONTROLLER FOR A NANOQUADCOPTER
Luis, C., & Le Ny, J. (August, 2016). Design of a Trajectory Tracking Controller for a
Nanoquadcopter. Technical report, Mobile Robotics and Autonomous Systems Laboratory,
Polytechnique Montreal.
Author:
Carlos Luis
Supervisor:
Jérôme Le Ny
Abstract
The primary purpose of this study is to investigate the system modeling of a nanoquadcopter as well as designing position and trajectory control algorithms, with the ultimate
goal of testing the system both in simulation and on a real platform.
The open source nanoquadcopter platform named Crazyflie 2.0 was chosen for the
project. The first phase consisted in the development of a mathematical model that
describes the dynamics of the quadcopter. Secondly, a simulation environment was created to design two different control architectures: cascaded PID position tracker and
LQT trajectory tracker. Finally, the implementation phase consisted in testing the controllers on the chosen platform and comparing their performance in trajectory tracking.
Our simulations agreed with the experimental results, and further refinement of the
model is proposed as future work through closed-loop model identification techniques.
The results show that the LQT controller performed better at tracking trajectories, with
RMS errors in position up to four times smaller than those obtained with the PID. LQT
control effort was greater, but eliminated the high control peaks that induced motor
saturation in the PID controller. The LQT controller was also tested using an ultrawide band two-way ranging system, and comparisons with the more precise VICON
system indicate that the controller could track a trajectory in both cases despise the
difference in noise levels between the two systems.
ii
Contents
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
iv
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
vii
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
1.1
Main Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
1.2
Secondary Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
2. Model of the Quadcopter . . . . . . . . . . . . . . . . . . . . . . . . . .
4
2.1
Coordinate Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
2.2
Dynamic Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
2.2.1
Force Equations
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.2.2
Momentum Equations . . . . . . . . . . . . . . . . . . . . . . . . .
9
2.3
Physical Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4
Linearization and State Space Representation . . . . . . . . . . . . . . . . 13
2.5
Movement Decoupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.6
Motor Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3. Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1
19
Cascaded PID Position Tracker . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1.1
3.1.2
On-Board Control Architecture . . . . . . . . . . . . . . . . . . . . 20
3.1.1.1
Inner Loop: Rate Controller . . . . . . . . . . . . . . . . 21
3.1.1.2
Outer Loop: Attitude Controller . . . . . . . . . . . . . . 22
3.1.1.3
Control Mixer . . . . . . . . . . . . . . . . . . . . . . . . 23
Off-Board Position Controller . . . . . . . . . . . . . . . . . . . . . 24
3.1.2.1
Altitude Controller . . . . . . . . . . . . . . . . . . . . . . 25
3.1.2.2
X-Y Position Controller . . . . . . . . . . . . . . . . . . . 25
3.1.2.3
Yaw Position Controller . . . . . . . . . . . . . . . . . . . 27
3.1.2.4
Controllers Gains
. . . . . . . . . . . . . . . . . . . . . . 27
iii
3.1.3
3.2
Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Linear-Quadratic Tracker (LQT) . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.1
The Optimization Problem Setup . . . . . . . . . . . . . . . . . . . 32
3.2.2
Kalman Filter for Linear Velocity Estimation . . . . . . . . . . . . 37
3.2.3
Weight Matrices and Integral Action . . . . . . . . . . . . . . . . . 43
3.2.4
Trajectory Generation . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2.5
Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4. Hardware Implementation and Experimental Results . . . . . . . . .
4.1
4.2
4.3
4.4
50
PID Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.1.1
ROS Controller Node
. . . . . . . . . . . . . . . . . . . . . . . . . 50
4.1.2
Experimental Results
. . . . . . . . . . . . . . . . . . . . . . . . . 53
LQT Controller Implementation . . . . . . . . . . . . . . . . . . . . . . . . 60
4.2.1
ROS Controller Node Modifications . . . . . . . . . . . . . . . . . . 62
4.2.2
MATLAB Interface Details . . . . . . . . . . . . . . . . . . . . . . 63
4.2.3
Experimental Results
. . . . . . . . . . . . . . . . . . . . . . . . . 64
Controller Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3.1
Simulation vs Experimental - PID . . . . . . . . . . . . . . . . . . 68
4.3.2
Simulation vs Experimental - LQT . . . . . . . . . . . . . . . . . . 71
4.3.3
Controller Performance: PID vs LQT
. . . . . . . . . . . . . . . . 73
LQT with UWB Position Estimation . . . . . . . . . . . . . . . . . . . . . 82
5. Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . .
89
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
92
Appendix A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
95
List of Figures
2.1.1 Body-fixed frame and Inertial frame. . . . . . . . . . . . . . . . . . . . .
4
2.1.2 “+” configuration at the left and “X” configuration at the right. . . . . .
5
2.2.1 Euler angles in the quadcopter’s body. . . . . . . . . . . . . . . . . . . .
6
2.2.2 Force diagram in the body-fixed frame. . . . . . . . . . . . . . . . . . . .
7
iv
LIST OF FIGURES
2.2.3 Rotation direction of each motor, courtesy of Bitcraze “Crazyflie 2.0 user guide”. 10
3.1.1 Block Diagram of Simulation environment. . . . . . . . . . . . . . . . . . 19
3.1.2 On-board control architecture, image courtesy of Bitcraze. . . . . . . . . 20
3.1.3 Rate Controller diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1.4 Rate Controller diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.5 Onboard control architecture with Control mixer. . . . . . . . . . . . . . 24
3.1.6 Altitude Controller. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.7 X-Y Position Controller. . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1.8 Yaw Position Controller. . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1.9 Simulation results for Test #1. . . . . . . . . . . . . . . . . . . . . . . . 28
3.1.10 3D Trajectory for Test #1. . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.1.11 Compound movement interference. . . . . . . . . . . . . . . . . . . . . . 29
3.1.12 Simulation results for a circular trajectory. . . . . . . . . . . . . . . . . . 30
3.1.13 3D Circular Trajectory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.14 Helical Trajectory Time response. . . . . . . . . . . . . . . . . . . . . . . 31
3.1.15 3D Helical Trajectory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.1 LQT Closed-Loop System. . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.2 Experimental validation of the Kalman Filter using the VICON system
raw data of X-Y positions. . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.3 Experimental validation of the Kalman Filter using the UWB system
raw data of X-Y positions. . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.2.4 Experimental validation of the Kalman Filter altitude estimation from
VICON raw data of Z position. . . . . . . . . . . . . . . . . . . . . . . . 42
3.2.5 Trajectory generation GUI. . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2.6 Simulation for steps in x, y and z positions. . . . . . . . . . . . . . . . . 46
3.2.7 Kalman Filter simulation, with VICON and UWB simulated noise. . . . 47
3.2.8 Tracking for complex trajectories. . . . . . . . . . . . . . . . . . . . . . . 48
3.2.9 3D Diagram for a complex trajectory. . . . . . . . . . . . . . . . . . . . . 48
3.2.10 Kalman Filter simulation for a complex trajectory. . . . . . . . . . . . . 49
4.1.1 ROS nodes and topics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.1.2 Crazyflie 2.0 with Vicon Sphere and UWB module. . . . . . . . . . . . . 51
4.1.3 Steps in vertical command zc . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.1.4 3D Vertical Trajectory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
v
LIST OF FIGURES
4.1.5 Steps in zc and xc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.1.6 3D Diagonal Trajectory. . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.1.7 Steps in yc and ψc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.1.8 3D Trajectory with Y-Yaw compound movement. . . . . . . . . . . . . . 57
4.1.9 Time Response for circular trajectory. . . . . . . . . . . . . . . . . . . . 58
4.1.10 3D Circular trajectory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.1.11 Time Response for helix trajectory. . . . . . . . . . . . . . . . . . . . . . 59
4.1.12
3D Helix Trajectory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.2.1 Pitch angle calibration rig. . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.2.2 Implementation diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2.3 MATLAB Interface diagram to implement the LQT controller. . . . . . . 63
4.2.4 Position plots for Trajectory#1. . . . . . . . . . . . . . . . . . . . . . . . 64
4.2.5 3D Trajectory#1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2.6 Position plots for Trajectory#2. . . . . . . . . . . . . . . . . . . . . . . . 65
4.2.7 3D Trajectory#2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2.8 Position plots for Trajectory#3. . . . . . . . . . . . . . . . . . . . . . . . 66
4.2.9 3D Trajectory#3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2.10 Position plots for Trajectory#4. . . . . . . . . . . . . . . . . . . . . . . . 67
4.2.11 3D Trajectory#4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3.1 Trajectory using the PID controller to follow a Step in the X position. . 69
4.3.2 Trajectory using the PID controller to follow a Step in the Y position. . 70
4.3.3 Trajectory using the PID controller to follow a Step in the Z position. . 70
4.3.4 Trajectory using the LQT controller to follow a Step in the X position. . 71
4.3.5 Trajectory using the LQT controller to follow a Step in the Y position. . 72
4.3.6 Trajectory using the LQT controller to follow a Step in the Z position. . 72
4.3.7 X-Y Position and error comparison when following a unit step in the X
position. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.3.8 Motor commands comparison when following a unit step in the X position. 75
4.3.9 X-Y Position and error comparison when following a unit step in the Y
position. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3.10 Motor commands comparison when following a unit step in the Y position. 77
4.3.11 Z Position and error comparison when following a unit step. . . . . . . . 78
4.3.12 X-Y Position and error comparison when following a unit step in the Z
position. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
vi
4.3.13 Motor commands comparison when following a unit step in the Z position. 79
4.3.14 X-Y Position and error comparison when following a circular trajectory.
80
4.3.15 Motor command comparison when following a circular trajectory. . . . . 81
4.4.1 Crazyflie 2.0 with UWB module exposed . . . . . . . . . . . . . . . . . . 82
4.4.2 X-Y Position and error comparison while hovering around a point. . . . 83
4.4.3 X-Y Position and error comparison when following Trajectory #1.
. . . 84
4.4.4 Comparison of 3D Trajectory#1. . . . . . . . . . . . . . . . . . . . . . . 84
4.4.5 X-Y Position and error comparison when following Trajectory #2.
. . . 85
4.4.6 Comparison of 3D Trajectory#2. . . . . . . . . . . . . . . . . . . . . . . 86
4.4.7 X-Y Position and error comparison when following Trajectory #3.
. . . 87
4.4.8 Comparison of 3D Trajectory#3. . . . . . . . . . . . . . . . . . . . . . . 87
List of Tables
2.2.1 Notation for vectors and states. . . . . . . . . . . . . . . . . . . . . . . .
6
2.3.1 Physical parameters for the Crazyflie 2.0. . . . . . . . . . . . . . . . . . . 12
3.1.1 Rate Controller’s gains. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.2 Attitude Controller’s gains. . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.3 Gains for Off-Board Controllers. . . . . . . . . . . . . . . . . . . . . . . . 27
4.3.1 Error comparison when following a unit step in X position. . . . . . . . . 74
4.3.2 Motor effort comparison when following a unit step in the X position. . . 75
4.3.3 Error comparison when following a unit step in Y position. . . . . . . . . 76
4.3.4 Motor effort comparison when following a unit step in the Y position. . . 77
4.3.5 Error comparison when following a unit step in Z position. . . . . . . . . 79
4.3.6 Motor effort comparison when following a unit step in the Z position. . . 79
4.3.7 Error comparison when following a circular position. . . . . . . . . . . . 80
4.3.8 Motor effort comparison when following a circular trajectory.
. . . . . . 81
4.4.1 Error comparison while hovering around a point. . . . . . . . . . . . . . 83
4.4.2 Error comparison while following Trajectory #1. . . . . . . . . . . . . . 85
4.4.3 Error comparison while following Trajectory #2. . . . . . . . . . . . . . 86
vii
LIST OF TABLES
4.4.4 Error comparison while following Trajectory #3. . . . . . . . . . . . . . 87
viii
Chapter 1
Introduction
In the past decade quadcopters have been studied due to their relative simple fabrication in comparison to other aerial vehicles, which turns them into ideal platforms for
modeling, simulation and implementation of control algorithms. The fact that they are
unmanned vehicles naturally invites developers to explore tasks that require a high degree of autonomy.
Past works such as [17, 30] have set the base for developing quadcopter platforms from
their construction to the automation techniques necessary to control the highly non-linear
dynamics that characterize these vehicles.
The scope of the quadcopter technology has changed over the years. The cost and sizes
have been reduced, it is now a platform affordable for a broad type of public, from researchers to hobbyists. But beyond the economic revenue these vehicles generate, the
manufacturers are searching for more autonomy, longer flight time, high data processing capabilities and adaptation to changing environments, hence the active research on
quadcopters.
A fairly new type of quadcopters are the so called “nanoquads” that are of considerably
low size and weight, making them an ideal platform for indoor usage. The project
proposed here considers the study of a commercial platform of a nanoquadcopter called
“Crazyflie 2.0” developed by Bitcraze company [33]. Weighting only 27 grams and having
9.2 cm of length and width, this nanoquad has rapidly become one of the preferred
platforms for quadcopter research.
1
CHAPTER 1. INTRODUCTION
For indoor control of quadcopters different localization techniques can be employed, for
example the VICON motion capture system [33] is one of the preferred systems for
precise localization and it has been used widely in recent quadcopter studies [11, 26]. A
recent low-cost technology based on ultra-wide band radio modules has proven effective
for indoor localization in robotics systems and specially in quacopters [35]. Its low costs
are inviting developers to create their own implementations and the system is getting
more precise and robust. In a few words, the system measures the distance between two
ultra-wide band modules, normally called anchor and tag, by measuring the time of flight
of an electromagnetic wave. Thus, by the simple relationship between time, distance and
velocity (in this case, the speed of light), then the distance can be easily determined. By
having at least three anchors constantly calculating the distance between them and a
certain tag, a triangulation allows to calculate the position in space of the tag, knowing
beforehand the fixed position of each anchor with respect to a frame.
The UWB system can be implemented using a two-way ranging protocol or a one-way
ranging protocol. In two-way ranging, the tag communicates with each anchor individually following a sequence to go through all the anchors and calculate each distance.
On the other hand, in one-way ranging the tag constantly broadcasts messages that are
received by every anchor and by precisely synchronising the clocks of the anchors then
the distance between each of them and the tag are calculated. One-way ranging is particular useful for multi-robot localization applications as there exists no bottle-neck in
the number of tags the system can support. In particular, for this project the two-way
ranging system developed in [36] was used to test the control loop behavior using different localization techniques. This system was developed using the decaWave DMW1000
ultra-wide band module [37] which offers an accuracy of 10-20 centimeters in distance
measurements.
1.1
Main Objectives
The main objectives of the research project were:
1. Develop the mathematical model that describes the dynamics of the Crazyflie 2.0
quadcopter.
2. Create a simulation environment for testing position and trajectory tracking control
algorithms.
2
CHAPTER 1. INTRODUCTION
3. Implement, test and compare different control architectures.
4. Evaluate the performance of a low-cost UWB-based localization system when integrated in the control loop.
1.2
Secondary Objectives
A set of small milestones were defined to help achieve the main objectives of the project:
1. Investigate past works to identify the physical and aerodynamical parameters of
the Crazyflie 2.0.
2. Linearize the quadcopter’s dynamics around hover state.
3. Study and identify the control architecture inside the Crazyflie’s firmware.
4. Design, simulate and implement an off-board position controller using data from
the VICON positioning system.
5. Conceive a second control system, from simulation to implementation, to track
more demanding trajectories.
6. Compare the performance of both controllers with in-flight data.
7. Compare the performance of the LQT controller using both the VICON and the
UWB systems.
3
Chapter 2
Model of the Quadcopter
In this section a mathematical model of the Crazyflie 2.0 is proposed. This study was
the basis on which the simulation environment was built and an important component in
the design of controllers. Thus, it was important to dedicate enough time to understand
how the system works and identify correctly some physical parameters that were relevant
for the simulation to be useful in the real case scenario.
2.1
Coordinate Frames
Before any dynamic study of the quadcopter begins, it is necessary to define the coordinate frames of the body of the quadcopter (non-inertial frame) as well as the inertial
frame, also called “world frame”, which in the case of this project refers to the coordinate
frame set by the external positioning system (VICON/UWB). Following the conventions
set by the “Bitcraze” company when designing their quadcopter, as seen in Figure 2.1.1
the body-fixed frame is defined.
Figure 2.1.1: Body-fixed frame and Inertial frame.
In the aeronautic systems, a popular axes convention is to define a positive altitude
downwards, the Y axis pointing towards the east and the X axis pointing towards the
4
CHAPTER 2. MODEL OF THE QUADCOPTER
true north. These types of frames are called NED frames (North, East, Down). It was
decided to follow the convention used in the Crazyflie 2.0 firmware, meaning a positive
altitude upwards, which defines an ENU frame (East, North, Up). Another remark is
that the origin of the body-fixed frame matches with the center of gravity of the quadcopter.
Another important remark is knowing the flight configuration of the quadcopter as there
are two of them: configuration “+” or configuration “X”. The difference between them is
the orientation of the X-Y frame in terms of the arms of the quadcopter, as shown in
Figure 2.1.2 taken from the manufacturer’s website [32] and modified accordingly.
Figure 2.1.2: “+” configuration at the left and “X” configuration at the right.
In the modern conceptions of quadcopters the “X” configuration is prefered over the
“+” configuration, mainly because in “X” it is easier to add a camera functionality as
the quadcopter’s arms will not be interfering with the images captured. By default the
Crazyflie 2.0 is in X mode, so for the rest of this project and during the mathematical
modeling it will be considered that the quadcopter is in this configuration.
2.2
Dynamic Equations
The dynamic equations of the quadcopter proposed here take into account certain physical properties that are not necessarily perfectly valid in the real platform that is being
used in this work, but they are good approximations that simplify greatly the study and
comprehension of this type of vehicles. Here are the hypothesis:
1. The quadcopter is a rigid body that cannot be deformed, thus it is possible to use
the well-known dynamic equations of a rigid body.
2. The quadcopter is symmetrical in its geometry, mass and propulsion system.
5
CHAPTER 2. MODEL OF THE QUADCOPTER
3. The mass is constant (i.e its derivative is 0).
The mechanical classic laws of motion are valid in inertial systems, so to be able to translate these equations into the body frame it is necessary to define a rigid transformation
matrix from the inertial frame to the body-fixed frame, in which only the rotational part
is meaningful to the discussion and is given by three successive rotations: first a rotation
of an angle ψ around the z axis, then a rotation of an angle θ around the intermediate
y axis and finally a rotation of an angle φ around the intermediate x axis. Once these
three rotations are calculated, the resulting transformation matrix is defined as:
cos θ cos ψ
cos θ sin ψ
− sin θ
Rib = sin φ sin θ cos ψ − cos φ sin ψ sin φ sin θ sin ψ + cos φ cos ψ sin φ cos θ
cos φ sin θ cos ψ + sin φ sin ψ cos φ sin θ sin ψ − sin φ cos ψ cos φ cos θ
(2.2.1)
where φ, θ and ψ represent the roll, pitch and yaw angles of the quadcopter’s body.
Figure 2.2.1 shows the direction of said angles in the Crazyflie 2.0 body-fixed frame
defined previously.
Figure 2.2.1: Euler angles in the quadcopter’s body.
The notation convention used during the mathematical analysis of the quadcopter’s
dynamics is exhibited in Table 2.2.1, where the state variables are defined.
Vector
pCG/o
Φ
VCG/o
ωb/o
State
Description
x
X position of CoG in the inertial frame
y
Y position of CoG in the inertial frame
z
Z position of CoG in the inertial frame
φ
Roll angle
θ
Pitch angle
ψ
Yaw angle
u
X linear velocity of CoG in the body-fixed frame w.r to the inertial frame
v
Y linear velocity of CoG in the body-fixed frame w.r to the inertial frame
w
Z linear velocity of CoG in the body-fixed frame w.r to the inertial frame
p
Roll angular velocity in the body-fixed frame w.r to the inertial frame
q
Pitch angular velocity in the body-fixed frame w.r to the inertial frame
r
Yaw angular velocity in the body-fixed frame w.r to the inertial frame
w.r = with respect.
Table 2.2.1: Notation for vectors and states.
6
CHAPTER 2. MODEL OF THE QUADCOPTER
Furthermore, a left superindex such as o V̇CG/o will indicate in what frame a derivative is
taken, while a right superindex indicate a vector coordinates into the specified frame. If
a right superindex is not specified as in the example, then the vector does not experience
any rotations after the derivative is taken.
2.2.1
Force Equations
According to Newton’s Second Law, the expression for the sum of forces is:
X
(2.2.2)
F = mo V̇CG/o
The expression of this derivative of velocity can be determined using the Coriolis equation, which gives the following dynamic expression in the body-fixed frame:
X
F = mo V̇CG/o = m
b
V̇CG/o + ωb/o × VCG/o
(2.2.3)
Each propeller of the quadcopter creates an aerodynamical force as shown in Figure 2.2.2
that acts upwards in the body-fixed frame.
Figure 2.2.2: Force diagram in the body-fixed frame.
In a situation where the quadcopter is parallel to the ground, meaning its roll and pitch
angles are zero, the aerodynamic forces created by the propellers will search to counteract
the effect of the weight and then make the quadcopter move upwards, downwards or stay
in a hover position. In Figure 2.2.2 the vector “mg” actually represents the projection of
the weight vector from the inertial frame to the body-fixed frame. That being said, this
qualitative analysis of how the forces work in the quadcopter’s body can be translated
7
CHAPTER 2. MODEL OF THE QUADCOPTER
into (2.2.3) as:
0
0
u̇
p
u
b
0 − Ro 0 = m v̇ + q × v
Fz
mg
ẇ
r
w
(2.2.4)
From (2.2.4) it is possible to isolate the vector b V̇CG/o :
u̇
0
0
p
u
b
v̇ = 0 − Ro 0 − q × v
ẇ
Fz /m
g
r
w
(2.2.5)
This equation dictates how the velocity of the center of gravity of the quadcopter evolves
in its body-fixed frame. To determine another set of state space variables it is necessary
to project this vector in the inertial frame to calculate the velocity in this coordinate
system. Note: the matrix Rob is a rotation matrix, so it has the following property:
−1
Rob
= Rob T = Rbo . Applying it, the projection is calculated:
o b
ṗCG/o
b
= Rob o ṗCG/o ⇐⇒ o ṗCG/o = Rbo VCG/o
ẋ
u
⇐⇒ ẏ = Rbo v
ż
w
(2.2.6)
By integrating (2.2.6) it is possible to know the position of the quadcopter in the inertial
frame.
Concerning the equations of forces and their state variables, it is necessary to specify
the form of the aerodynamical force generated by the propellers. Following the diagram
in Figure 2.2.2, the force generated by each propeller has the form:
0
Fib = 0
Ti
(2.2.7)
where Ti represents the upward thrust force in Newtons generated by each propeller. It
is widely known that the thrust generated by a propeller is a function of the square of
its angular speed:
Ti = CT ωi2
(2.2.8)
CT is a thrust coefficient that will be specified in Section 2.3 and ωi is the rotation speed
of the i-th motor, in revolutions per minute. As Figure 2.2.2 suggests, each propeller
generates a thrust force following (2.2.8) and all in the same direction, which leads to a
8
CHAPTER 2. MODEL OF THE QUADCOPTER
sum of all thrust forces:
X
Fib =
CT
2.2.2
0
0
ω12
+
ω22
+
ω32
2
(2.2.9)
+ ω4
Momentum Equations
These equations dictate the rotational dynamics of the quadcopter. Following the theorem of angular momentum:
X
M o = o ḣ
(2.2.10)
where h denotes the angular momentum around the center of gravity. Using Coriolis
equation:
X
M o = o ḣ = b ḣ + ωb/o × h
(2.2.11)
It is desirable to express (2.2.11) in the body-fixed frame as the momentum equations
are more easily calculated, as explained in [31]:
X
M b = J b ω̇b/o + ωb/o × J ωb/o
(2.2.12)
here J denotes the inertia matrix of the quadcopter, which in general can be expressed
as:
Ixx −Ixy −Ixz
J = −Ixy Iyy −Iyz
−Ixz −Iyz Izz
(2.2.13)
but from the hypothesis that the body of the quadcopter is symmetrical around all its
axes, the inertia matrix has all crossed terms equal to zero, i.e.,
Ixx 0
0
J = 0 Iyy 0
0
0 Izz
(2.2.14)
From equation (2.2.12) it is possible to isolate the vector b ω̇b/o :
p
p
ṗ
Mx
−1
=
(J
)
−
×
J
My q
q
q̇
r
r
ṙ
Mz
9
(2.2.15)
CHAPTER 2. MODEL OF THE QUADCOPTER
The last state equations come from the relation between ωb/o and the Euler angles
derivative Φ̇
p
1
0
− sin θ
φ̇
q = 0 cos φ sin φ cos θ θ̇
r
0 − sin φ cos φ cos θ
ψ̇
(2.2.16)
with the inverse relation the state vector Φ̇ is isolated:
φ̇
1 sin φ tan θ cos φ tan θ
p
π
cos φ
− sin φ q for θ 6=
θ̇ = 0
2
ψ̇
0 sin φ/ cos θ cos φ/ cos θ
r
(2.2.17)
To calculare the total momentum generated in the quadcopter system, it is imperative
to know the rotation direction of each motor. As seen in Figure 2.2.3, the manufacturer
of the Crazyflie 2.0 provides this information [32].
Figure 2.2.3: Rotation direction of each motor, courtesy of Bitcraze “Crazyflie 2.0 user guide”.
The expression for the momentum is given as:
M=
4
X
Pi × Fi +
i=1
4
X
τi
(2.2.18)
i=1
where Pi represents the position of each motor in the body-fixed frame and τi represents
the induced momentum in the quadcopter’s body generated by the i-th motor. When
a rotor turns in a given direction, conservation of angular momentum dictates that the
quadcopter’s body will have a tendency to counteract the generated angular momentum,
being consistent with Newton’s third law of action and reaction. This reaction momentum due to the spin of a rotor is the induced moment τi .
If d denotes the distance from the center of gravity to the center of each motor, the
10
CHAPTER 2. MODEL OF THE QUADCOPTER
position of each motor is:
√
√
√
√
d/ 2
−d/ 2
−d/ 2
d/ 2
√
√
√
√
P1 = −d/ 2 , P2 = −d/ 2 , P3 = d/ 2 , P4 = d/ 2
0
0
0
0
(2.2.19)
Then the mometum generated by the thrust force of each motor can be calculated:
√
√
−CT ω12 d/ 2
−CT ω22 d/ 2
√
√
P1b × F1b = −CT ω12 d/ 2 P2b × F2b = CT ω22 d/ 2
0
0
√
√
CT ω32 d/ 2
CT ω42 d/ 2
√
√
P3b × F3b = CT ω32 d/ 2 P4b × F4b = −CT ω42 d/ 2
0
0
The induced moments τi act only in the Z axis and have an opposite magnitude from
the moment generated by each propeller, due to the conservation of angular momentum.
In this particular case, given the axis convention that is being used (z axis pointing
upwards), applying the right-hand rule indicate that a clockwise spinning rotor yields
a negative momentum (one thumb points downward, in the opposite direction of the z
axis), thus the induced momentum will be positive. Then, the sum of induced moments
in the quadcopter’s body is calculated:
4
X
τib =
i=1
CD
0
0
−ω12 + ω22 − ω32 + ω42
(2.2.20)
where CD denotes the aerodynamic drag coefficient that will be specified in Section 2.3.
Finally, the total moment has the following form:
√
Mx
dCT / 2 −ω12 − ω22 + ω32 + ω42
√
M b = My = dCT / 2 −ω12 + ω22 + ω32 − ω42
CD −ω12 + ω22 − ω32 + ω42
Mz
(2.2.21)
In the total momentum equation there are certain terms that include angular accelerations that have been neglected as they tend to be small compared to the other terms
of the equation. Gyroscopic moments have also been neglected using the argument that
the inertia moment of each motor tends to be small thus their contribution in the total
momentum is also small [21, 30].
11
CHAPTER 2. MODEL OF THE QUADCOPTER
2.3
Physical Parameters
The precise measurement of certain physical parameters is the key to create a simulation
environment that correctly describes the behavior of the quadcopter. In [1] a study
of said physical parameters was undertaken for the Crazyflie 2.0. The aerodynamical
coefficients were studied in [2] for the Crazyflie 1.0, but they are the same or at least
close to those of the Crazyflie 2.0 given the fact that these coefficients only depend on
the geometry of the propellers [3], which remained unchanged between the two models.
Results of both works are summarized in Table 2.3.1.
Parameter
mquad
muwb
mvicon
m
d
r
Ixx
Iyy
Izz
kT
kD
Description
Mass of the quadcopter alone
Mass of the UWB module
Mass of one VICON marker
Total mass
Arm length
Rotor radius
Principal Moment of Inertia around x axis
Principal Moment of Inertia around y axis
Principal Moment of Inertia around z axis
Non-dimensional thrust coefficient
Non-dimensional torque coefficient
Value
0.27 [Kg]
0.04 [Kg]
0.02 [Kg]
0.33 [Kg]
39.73 × 10−3 [m]
23.1348 × 10−3 [m]
1.395 × 10−5 [Kg × m2 ]
1.436 × 10−5 [Kg × m2 ]
2.173 × 10−5 [Kg × m2 ]
0.2025
0.11
Table 2.3.1: Physical parameters for the Crazyflie 2.0.
In addition, as explained in [3], the thrust generated by the propeller is often expressed
as:
T = kT ρn2 D4
(2.3.1)
where kT is the non-dimensional thrust coefficient, ρ is the density of air, n is the
propeller speed in revolutions per second and D is the diameter of the rotor. As it will
be evident later, it is convenient to express the propeller speed in RPM’s. Knowing that
1 revolution per second is the same as 60 revolutions per minute, (2.3.1) becomes:
T = kT ρ (ω/60)2 D4
(2.3.2)
where ω is the angular speed of the propeller in RPM. Comparing the above with (2.2.8),
it is possible to determine the thrust coefficient CT as:
CT = kT ρ (2r)4 /3600
12
(2.3.3)
CHAPTER 2. MODEL OF THE QUADCOPTER
taking the value of air density constant ρ = 1.225 [Kg/m3 ] and all the other constants
defined previously, finally this coefficient is:
CT = 3.1582 × 10−10 [N/rpm2 ]
(2.3.4)
Now for the torque coefficient, as specified in [3], the torque created by the propellers is
described by this equation:
Q = kD ρn2 D5
(2.3.5)
Operating the same variable change as in (2.3.3), then:
CD = kD ρ (2r)5 /3600
(2.3.6)
CD = 7.9379 × 10−12 [Nm/rpm2 ]
(2.3.7)
With the parameters specified in Table 2.3.1 and the constants calculated in (2.3.4)
and (2.3.7), all the basic physical parameters were determined. Here we use the word
“basic” as these parameters are the minimum necessary to be able to simulate the behavior of a quadcopter and because in most applications, this one included, are a good
approximation of the real physical system.
2.4
Linearization and State Space Representation
The state space representation of a system gives an idea of how the system evolves in
time by the following equations:
(
ẋ (t) = A (t) x (t) + B (t) u (t)
y (t) = C (t) x (t) + D (t) u (t)
(2.4.1)
In general, this equation describes the evolution of a linear time-varying system, where
x (t) is the vector of states, y (t) is the output vector and u (t) is the input vector. For
the state space representation of a quadcopter it is conventional to consider the following
linear time-invariant realisation of the system, meaning that matrices A, B, C and D are
static and don’t change over time:
(
∆ẋ = A∆x + B∆u
∆y = C∆x + D∆u
where the prefix ∆ means that the vector is the result of a linearisation process.
13
(2.4.2)
CHAPTER 2. MODEL OF THE QUADCOPTER
When linearizing a system the important question becomes at around which so-called
"trim" trajectory we desire to linearize, or better yet, what is the trim trajectory most
well-suited given the needs of the system. This trim trajectory has to be associated with
an equilibrium point in which the states of the system do not change over time, meaning
ẋe = 0. In the case of the quadcopter, the common trim trajectory is the hover, in
which the drone stays stationary at a certain altitude. This fact can be translated as an
equilibrium state:
xe =
h
xe ye ze ψe θe φe ue ve we re qe pe
iT
(2.4.3)
At the equilibrium point, the quadcopter’s linear position and yaw angle are indifferent
in terms of the linearization calculus, so they can be considered arbitrary constants. For
the roll and pitch angles, they need to be zero in order for the quadcopter to keep the
stationary position, and so does any linear or angular velocity. Finally the equilibrium
vector is as simple as:
xe =
h
xe ye ze ψe 0 0 0 0 0 0 0 0
iT
(2.4.4)
In order to keep the quadcopter flying in hover mode, knowing that the body is levelled
with the floor and there are no gravity components other that in the z axis, the force
generated by the propellers need to compensate exactly for the weight of the quadcopter
to stay stationary in the air, this means:
2
2
2
2
CT ωe1
+ ωe2
+ ωe3
+ ωe4
= mg
(2.4.5)
From the hypothesis that the quadcopter’s body is perfectly symmetrical, a quick deduction is that all motors need to rotate with the same speed in order to maintain the
body levelled and don’t create any angular momentum, this means that in equilibrium:
ωe1 = ωe2 = ωe3 = ωe4 = ωe
(2.4.6)
Combining (2.4.5) and eqrefeq:32 gives as result:
r
ωe =
mg
= 16073 [rpm]
4CT
14
(2.4.7)
CHAPTER 2. MODEL OF THE QUADCOPTER
This constant specifies the required speed of each rotor in order to maintain the hover
position and so the input vector in equilibrium is:
ue =
h
ωe ωe ωe ωe
iT
(2.4.8)
After applying a Taylor’s first order expansion, and taking into account the equilibrium
state vector specified in (2.4.4), the linearized equations are:
∆Fx = m∆u̇ − mg∆θ
∆Fy = m∆v̇ + mg∆φ
∆F = m∆ẇ
z
∆Mx = Ixx ∆ṗ
∆My = Iyy ∆q̇
∆M = I ∆ṙ
z
zz
(2.4.9)
The linearized forces and moments are:
∆Fx
0
∆F b = ∆Fy =
0
∆Fz
2CT ωe (∆ω1 + ∆ω2 + ∆ω3 + ∆ω4 )
(2.4.10)
√
2dCT ωe (−∆ω1 − ∆ω2 + ∆ω3 + ∆ω4 )
∆Mx
√
∆M b = ∆My = 2dCT ωe (−∆ω1 + ∆ω2 + ∆ω3 − ∆ω4 )
2CD ωe (−∆ω1 + ∆ω2 − ∆ω3 + ∆ω4 )
∆Mz
(2.4.11)
In hover position the body-fixed frame coincides with the inertial frame, meaning:
∆ẋ = ∆u
∆ẏ = ∆v
∆ż = ∆w
∆φ̇ = ∆p
∆θ̇ = ∆q
∆ψ̇ = ∆r
(2.4.12)
Merging (2.4.9) to (2.4.12) allows to write the state space representation of the linearized
quadcopter:
∆ẋ
∆ẏ
∆ż
∆ψ̇
∆θ̇
∆φ̇
∆u̇
∆v̇
∆ẇ
∆ṙ
∆q̇
∆ṗ
=
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0 0 1 0 0 0
0 0 0 1 0 0
0 0 0 0 1 0
0 0 0 0 0 1
0 0 0 0 0 0
0 0 0 0 0 0
g 0 0 0 0 0
0 −g 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
∆x
∆y
∆z
∆ψ
∆θ
∆φ
∆u
∆v
∆w
∆r
∆q
∆p
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
+ ωe
0
0
0
0
0
0
0
0
2CT /m
2CT /m
2CT /m
2CT /m
−2C /I
2C /I
−2C /I
2C /I
√ D zz
√ D zz
√ D zz
√ D zz
2dC /I
2dCT /Iyy − 2dCT /Iyy
−√2dCT /Iyy
√ T yy √
√
− 2dCT /Ixx − 2dCT /Ixx
2dCT /Ixx
2dCT /Ixx
15
∆ω1
∆ω2
∆ω3
∆ω4
CHAPTER 2. MODEL OF THE QUADCOPTER
2.5
Movement Decoupling
In the state space realization of the quadcopter’s linear model, the four inputs of the
system act directly in just four states of the system. Rewriting together (2.4.10) and
(2.4.11) show how each input of the system contributes to each force and momentum.
Fz
Mx
My
Mz
= 2ωe
CT
CT
CT
CT
√
√
√
√
−dCT / 2 −dCT / 2 dCT / 2 dCT / 2
√
√
√
√
−dCT / 2 dCT / 2 dCT / 2 −dCT / 2
−CD
CD
−CD
CD
∆ω1
∆ω2
∆ω3
∆ω4
(2.5.1)
From (2.5.1) a transformation matrix can be defined between the forces acting on the
quadcopter’s body and the angular speed from the motors:
Γ = 2ωe
CT
CT
CT
CT
√
√
√
√
−dCT / 2 −dCT / 2 dCT / 2 dCT / 2
√
√
√
√
−dCT / 2 dCT / 2 dCT / 2 −dCT / 2
−CD
CD
−CD
CD
(2.5.2)
If matrix Γ is invertible (i.e: Γ−1 exists) that means that all four lines are independent
and thus the vertical and angular forces of the quadcopter act independently from each
other. Inverting the matrix:
Γ−1 =
1
2ωe
√
√
1/ (4CT ) − 2/ (4dCT ) − 2/ (4dCT ) −1/ (4CD )
√
√
1/ (4CT ) − 2/ (4dCT )
2/ (4dCT )
1/ (4CD )
√
√
2/ (4dCT )
2/ (4dCT ) −1/ (4CD )
1/ (4CT )
√
√
1/ (4CT )
2/ (4dCT ) − 2/ (4dCT ) 1/ (4CD )
(2.5.3)
Now taking the result in (2.5.3), it is possible to find the inverse relation of (2.5.1):
∆ω1
∆ω2
∆ω3
∆ω4
=
1
2ωe
√
√
1/ (4CT ) − 2/ (4dCT ) − 2/ (4dCT ) −1/ (4CD )
√
√
1/ (4CT ) − 2/ (4dCT )
2/ (4dCT )
1/ (4CD )
√
√
1/ (4CT )
2/ (4dCT )
2/ (4dCT ) −1/ (4CD )
√
√
1/ (4CT )
2/ (4dCT ) − 2/ (4dCT ) 1/ (4CD )
Fz
Mx
My
Mz
(2.5.4)
Equation (2.5.4) dictates how each motor contributes to each of the forces acting on
the quadcopter’s body. This study confirms that the vertical, lateral, longitudinal and
directional (yaw) forces act independently from each other in the mathematical model
and thus the quadcopter’s dynamics are decoupled and can be studied as sub-systems.
16
CHAPTER 2. MODEL OF THE QUADCOPTER
• Vertical Subsystem
This subsystem describes the dynamics of the upward movements of the quadcopter,
following this state space equation:
"
∆ẇ
∆ż
#
"
=
0 0
1 0
#"
∆w
∆z
#
"
+
1/m
0
#
∆Fz
(2.5.5)
• Directional Subsystem
The yaw angle and its velocity dictate the dynamics of the quadcopter direction in the
XY plane, as suggests this following state space equation:
"
#
∆ṙ
∆ψ̇
"
=
0 0
1 0
#"
∆r
∆ψ
#
"
+
1/Izz
0
#
∆Mz
(2.5.6)
• Lateral Subsystem
The lateral dynamic governs the pitch movement of the quadcopter, as well as its Y
position in the inertial frame:
∆ṗ
∆φ̇
∆v̇
∆ẏ
=
0 0 0
1 0 0
0 −g 0
0 0 1
0
0
0
0
∆p
∆φ
∆v
∆y
+
1/Ixx
0
0
0
∆Mx
(2.5.7)
• Longitudinal Subsystem
Similar to the lateral subsystem, it rules the movement around the X axis of the bodyfixed frame of the quadcopter, and its X position and velocity in the inertial frame.
2.6
∆q̇
∆θ̇
∆u̇
∆ẋ
=
0
1
0
0
0
0
g
0
0
0
0
1
0
0
0
0
∆q
∆θ
∆u
∆x
+
1/Iyy
0
0
0
∆Mx
(2.5.8)
Motor Characterization
The inputs of the above state space realization are the angular speed of each motor in
RPM’s, but after exploring the Crazyflie’s 2.0 firmware it became evident that this is
not the real input of the system. The voltage sent to each DC motor is controlled using
a PWM signal specified as a 16 bit number, ranging from 0 to 65535, meaning that the
actual input of the system can be considered directly as this PWM signal and not the
17
CHAPTER 2. MODEL OF THE QUADCOPTER
actual voltage sent to the motors. In [2] experimental data was retrieved from the motors
to identify the relationship between the PWM signal sent to the motors and the RPM’s
generated. The experiments proved that the angular speed of the motors have a linear
relationship with the PWM input of the system, following the equation:
RP M = 0.2685 × P W M + 4070.3
(2.6.1)
The characterization of a DC motor usually derives in a first order transfer function that
specifies certain response time from the motors as they do not react immediately to the
commands sent, but it is a good approximation to assume that this response time is fast
enough and that it will not cause much delays in the system, thus (2.6.1) serves as a
good approximation for the motor characterization.
18
Chapter 3
Simulation
In this section a simulation environment of the quadcopter’s dynamics is proposed with
the intention of testing and designing control schemes. Two different controllers are
proposed: in the first phase of the project a position PID tracker was considered and in
a second phase a trajectory tracker known as the Linear-Quadratic Tracker (LQT) was
studied.
3.1
Cascaded PID Position Tracker
Figure 3.1.1 presents the simulation model created for this phase of the project.
Trajectory
xc , yc , zc , ψc
Controller
PWM
State+Noise
Motors
RP M
Sensor’s
Noise
Quadcopter’s
Dynamics
State
Figure 3.1.1: Block Diagram of Simulation environment.
Now an explanation of each simulation block will be given:
1. Trajectory: this block serves as the input of the overall system, specifying a
trajectory in the x,y,z and yaw positions. As of now the commands sent by this
block are mere constants or signal inputs such as sinusoids, random signals, ramps,
etc. It serves as input commands to the controller.
19
CHAPTER 3. SIMULATION
2. Controller: takes the desired trajectory and the quadcopter’s states as inputs and
computes the necessary 16-bit PWM signal to send to the motors.
3. Motors: implements the linear relation between the 16-bit PWM signal sent to
the motors and the actual angular speed in revolutions per minutes generated by
them, as specified in Section 2.6.
4. Quadcopter’s dynamics: this block implements the dynamic equations of Section 2.2. The vectorial form of the equations, as they were developed, is the
simplest, most common and elegant way of constructing this block. The non-linear
model was linearized using MATLAB’s command “linmod” to verify it was consistent with the theoretical state space model found in the previous section.
5. Sensor’s noise: allows to add additive white Gaussian noise to specific states
of the quadcopters, to simulate the sensors that give the real states used by the
control system. This block could be modified by the user to define more complex
models, e.g., including sensor bias.
3.1.1
On-Board Control Architecture
The first steps to control the quadcopter was understanding the already implemented
controllers that came with the stock firmware of the Crazyflie 2.0 (Firmware release
2016.02). As seen in Figure 3.1.2 the manufacturers specify the control architecture
used:
Figure 3.1.2: On-board control architecture, image courtesy of Bitcraze.
A two cascaded PID control scheme was found in the Crazyflie’s firmware in order to
control the pitch and roll angles. A cascaded control structure can be analysed by
decomposing the architecture in an inner and outer control loops, in which the outer
loop regulates the inner loop, which in turn regulates the plant of the system. As a
20
CHAPTER 3. SIMULATION
common rule in cascaded structures, the inner loop needs to regulate at a faster rate
than the outer loop. It is ideal for the inner loop output to reach an steady state
value before the outer loop changes the setpoint sent to the inner loop. Synchronization
problems will occur between the two controllers if the inner loop response is not as fast
as it should, or if the outer loop is faster than it should. In implementation terms this
is easily remediable by forcing the inner loop to be, as in this case, twice as fast as the
outer loop (Attitude controller running at 250Hz and Rate controller running at 500Hz).
3.1.1.1
Inner Loop: Rate Controller
The inputs and outputs of this block are shown in Figure 3.1.3.
p c , qc , r c
p, q, r
Rate
Controller
∆φ , ∆θ , ∆ψ
Figure 3.1.3: Rate Controller diagram.
The goal of this controller is to calculate the input variation from the equilibrium point
of the motors in order to create the angular momentum required for the state variables
p, q and r to get the values pc , qc and rc respectively. For that, three independent
controllers are used:
• Roll Rate Proportional controller: calculates ∆φ following this equation, the
desired value pc is calculated by the outer loop attitude controller:
∆φ (t) = KP,p (pc (t) − p (t))
(3.1.1)
• Pitch Rate Proportional controller: very similar to the roll rate controller,
calculates ∆θ from the setpoint value qc :
∆θ (t) = KP,q (qc (t) − q (t))
(3.1.2)
• Yaw Rate Proportional-Integral controller: calculates the desired deviation
from the base thrust, ∆ψ , from an external setpoint rc that can be specified through
a teleoperation system or as it is going to be specified later, by an off-board con-
21
CHAPTER 3. SIMULATION
troller. The control law for this compensator is:
Z
t
(rc (τ ) − r (τ )) dτ
∆θ (t) = KP,r (rc (t) − r (t)) + KI,r
(3.1.3)
0
Table 3.1.1 contains the gains for each of these controllers, taken directly from the
Crazyflie’s firmware.
Controller
Roll Rate
Pitch Rate
Yaw Rate
KP
70
70
70
KI
16.7
Table 3.1.1: Rate Controller’s gains.
3.1.1.2
Outer Loop: Attitude Controller
The inputs and outputs of the attitude controller are as show in Figure 3.1.4.
φc , θc
φ, θ
pc , qc
Attitude
Controller
Figure 3.1.4: Rate Controller diagram.
This controller act as a regulator of the rate controller, calculating the appropriate
setpoints for the angular velocities around the X and Y axis, in order to stabilize the
quadcopter at a certain desired angular position. The attitude controller uses the pitch
and roll angles estimates from the sensor fusion algorithm, compares them to the external
commands φc and θc (coming from teleoperation, off-board controller, etc) and feeds
them to a controller that calculates the desired angular velocities pc and qc . These
controllers are as follows:
• Roll Attitude Proportional-Integral controller: computes the desired roll
rate in the body frame, pc ,using the control law:
Z
pc (t) = KP,φ (φc (t) − φ (t)) + KI,φ
t
(φc (τ ) − φ (τ )) dτ
(3.1.4)
0
• Pitch Attitude Proportional-Integral controller: works in the same fashion
22
CHAPTER 3. SIMULATION
as the roll controller, using the corresponding variables:
Z
t
qc (t) = KP,θ (θc (t) − θ (t)) + KI,θ
(θc (τ ) − θ (τ )) dτ
(3.1.5)
0
Once again, the gains for these controllers were already specified in the Crazyflie’s
firmware release 2016.02, as seen in Table 3.1.2. The same values were used during
this project as they turned out to be well tuned according to the simulations and the
tests made with the real platform.
Controller
Roll Attitude
Pitch Attitude
KP
3.5
3.5
KI
2
2
Table 3.1.2: Attitude Controller’s gains.
As it is evident from the gain values in Tables 3.1.1 and 3.1.2, the roll and pitch gains
for both controllers are the same, which is consistent with the initial hypothesis that the
quadcopter is a symmetrical body around all its axes.
3.1.1.3
Control Mixer
The output of the rate controller is the total input variation of the motors from the
equilibrium state required to generate a torque in the desired direction of movement.
Afterwards this input variation has to be distributed to the motors in the same fashion
as in (2.4.11), for it to move and rotate appropriately using the PWM input it received.
Given that the quadcopter is in X configuration, the motor effort has to be distributed
halfway in each motor for a desired torque around the X or Y axis. All this analysis can
be translated in the following equations that are implemented on board of the Crazyflie
2.0, thus specifying a “Control mixer” block in the simulation environment:
P W Mmotor1
PWM
motor2
P
W
M
motor3
PWM
motor4
=
=
=
=
Ω − ∆φ /2 − ∆θ /2 − ∆ψ
Ω + ∆φ /2 − ∆θ /2 + ∆ψ
Ω + ∆φ /2 + ∆θ /2 − ∆ψ
Ω − ∆φ /2 − ∆θ /2 + ∆ψ
(3.1.6)
where Ω is the PWM base signal for maintaining a certain altitude, a value that will be
regulated from an altitude controller; ∆φ , ∆θ and ∆ψ represent the outputs of the Rate
Controller and at the same time give an idea of a deviation needed from the base thrust
in order to obtain a certain torque in the X, Y or Z axis. From (3.1.6) it is more clearly
23
CHAPTER 3. SIMULATION
how, for example, a command ∆φ > 0 will be distributed 50/50 as a reduction of the
PWM input supplied to motors 1 and 4, and an increase of the power supplied to motors
2 and 3, thus creating the appropriate angular momentum to obtain certain angle in the
pitch direction.
Figure 3.1.5 shows a complete diagram of the on-board control scheme.
φ, θ
∆φ , ∆θ , ∆ψ
pc , qc
φc , θc
Attitude
Controller
rc
p, q, r
Rate
Controller
Ω
Control
Mixer
PWM
Figure 3.1.5: Onboard control architecture with Control mixer.
The red inputs in this diagram represent the actual inputs that can be controlled from
outside of the firmware, as it was suggested earlier, either by a teleoperation system
or by an automated position controller. The other signals come from the sensor fusion
algorithm or are intermediate variables in the control process.
3.1.2
Off-Board Position Controller
In hopes of controlling the quadcopter by sending waypoints or trajectories in a tridimensional space, it became necessary to add a position controller that in terms of the
implementation will be running off-board unlike the controller of the previous section.
The job of this controller can be divided into:
1. An altitude controller whose output is the thrust Ω required to maintain a certain
position in z.
2. An X-Y position controller whose outputs are the required roll and pitch angles
that will be regulated by the on-board controller.
3. A yaw position controller that sends the required angular velocity to the on-board
Yaw Rate Controller.
As seen in the dynamic analysis of the quadcopter, there exists a theoretical decoupling
between the vertical, lateral, longitudinal and yaw movement, meaning that each one
of these controllers can be tuned independently from each other, which simplifies the
controller design task.
24
CHAPTER 3. SIMULATION
3.1.2.1
Altitude Controller
It is common practice to add a feedforward term as in [4] that compensates the weight
of the quadcopter, in order to avoid the use of large controller gains that can lead to
saturation problems. The structure in Figure 3.1.6 was the one used for this controller.
zc
z
Altitude
Controller
∆Ω
+
Ω
+
Feedforward Ωe
Figure 3.1.6: Altitude Controller.
The altitude controller is a simple PID compensator whose inputs are the desired altitude
that comes from the trajectory block, and the altitude state that in terms of simulation
comes from the quadcopter’s dynamics block. The equation that describes this PID
controller is the following:
∆Ω (t) = KP,z (zc (t) − z (t)) + KI,z
Rt
0
d
(zc (t) − z (t))
(zc (τ ) − z (τ )) dτ + KD,z dt
(3.1.7)
The output of this PID controller is the 16-bit thrust deviation from the equilibrium
point set at the hover state. That means that the feedforward term Ωe in Figure 3.1.6
is the necessary PWM 16-bit signal needed for the quad to maintain its altitude. Using
the calculations at the equilibrium, as seen in (2.4.7) and (2.6.1), the feedforward term
can be calculated as:
Ωe =
3.1.2.2
ωe − 4070.3
= 44705
0.2685
(3.1.8)
X-Y Position Controller
The objective of this controller is to regulate the on-board Attitude Controller by calculating the necessary roll and pitch angles in order to move the quadcopter between
locations in the X-Y plane. The block diagram in Figure 3.1.7 shows the inputs and
outputs of this controller.
25
CHAPTER 3. SIMULATION
x c , yc
x, y
ψ
xbe , yeb
Error to
Body
Frame
u, v
X-Y
Position
Controller
φc , θc
Figure 3.1.7: X-Y Position Controller.
In [5] it was proven that a similar architecture as here gives a good performance in
position tracking. The first block calculates the error between the desired and actual
X-Y position and does the rotation operation needed to project this error vector in the
body frame. This operation is given as:
"
xe
ye
#b
"
=
cos (ψ) sin (ψ)
− sin (ψ) cos (ψ)
#"
xe
ye
#o
(3.1.9)
doing the calculations, the error in the body frame is defined as:
(
xbe = xoe cos (ψ) + yeo sin (ψ)
yeb = −xoe sin (ψ) + yeo cos (ψ)
(3.1.10)
Then the error in the body-fixed frame becomes the setpoint to the velocity in this same
frame, the logic behind this apparently odd choice of a setpoint is that the bigger the
error is, the more rapidly the quadcopter should move in order to arrive at the desired
point as quickly as possible. Otherwise, if the error is small, meaning the drone is near
the desired point, the setpoint for the velocity should be also small.
A more conventional method to do the position tracker is that the error in the bodyfixed frame dictates the setpoint of the position instead of the velocity, but in practice
it was found that in the first method proposed is easier to tune the PID gains as there
are only 2 out of 3 gains that need to be adjusted (derivative gain is not used), whereas
for the traditional tracker all three gains have to be used for a good performance. The
only disadvantage that might have the first method with respect to the second is that
in the real system the velocity in the body-fixed frame is not directly measured, thus it
is necessary to estimate this states through mathematical means.
The controllers that compute the desired attitude use the following control laws:
26
CHAPTER 3. SIMULATION
• X Position Proportional-Integral Controller:
Z t
φc (t) = KP,x xbe (t) − u (t) + KI,x
xbe (τ ) − u (τ ) dτ
(3.1.11)
0
• Y Position Proportional-Integral Controller:
θc (t) = KP,y
yeb (t)
− v (t) + KI,y
Z t
yeb (τ ) − v (τ ) dτ
(3.1.12)
0
3.1.2.3
Yaw Position Controller
The inputs and outputs of this controller are reflected in Figure 3.1.8:
ψc
ψ
Yaw
Position
Controller
rc
Figure 3.1.8: Yaw Position Controller.
The controller computes the error between the desired yaw position and the actual
position and feeds it to a proportional controller whose output is the desired yaw rate that
is regulated by the on-board rate controller. Thus the operation done by the controller
is given as:
rc (t) = KP,ψ (ψc (t) − ψ (t))
3.1.2.4
(3.1.13)
Controllers Gains
After validation both in simulation and in the real platform, the gains for each one of
the off-board controllers are summarized in Table 3.1.3.
Controller
Altitude
X
Y
Yaw
KP
11000
30
−30
3
KI
3500
2
−2
-
KD
9000
-
Output Limit
[−20000, 15000]
±30 [deg]
±30 [deg]
±200 [deg/s]
Table 3.1.3: Gains for Off-Board Controllers.
Note that the gains for the Y Position Controller are inverted, this is because a positive
roll angle makes the quadcopter move toward the negative Y axis, so the solution is
to invert the gains in order to send the correct attitude commands to the on-board
controller.
27
CHAPTER 3. SIMULATION
3.1.3
Simulation Results
The simulation environment was built in Simulink following the block diagram in Figure 3.1.1. The goal of said simulations was to test the control system for tracking a
desired position [xc , yc , zc , ψc ] and compare the behavior of the non-linear dynamics of
the quadcopter with the linear state model.
• Linear trajectories
This first simulation tests the response of the system when demanded to follow step
functions in all 4 trajectory inputs. The setpoints for Test #1 where: xc = 1m ; yc =
1m ; zc = 1m ; ψc = 60°. Figure 3.1.9 shows the simulation of a 15 seconds flight of the
quadcopter given the desired trajectory. For the X-Y position the time response is about
3 seconds, with almost no overshoot, whereas for the Z position the response is slower at
roughly 8 seconds for a 2% error margin and with a more pronounced overshoot. With
the experimental results, this values of PID gains gave the best response after some trial
and error in gain-tuning. As for the yaw response, it has a time response of about 2
seconds with no overshoot.
X Position
1
1
0.8
0.8
0.6
0.4
0.2
0.6
0.4
0.2
0
0
-0.2
-0.2
0
5
10
15
0
Time (s)
5
60
1
50
Angle (deg)
1.2
0.6
10
15
40
30
0.4
20
0.2
10
0
15
Yaw angle
70
0.8
10
Time (s)
Linear
Non-linear
Reference
Z Position
1.4
Position (m)
Y Position
1.2
Position (m)
Position (m)
1.2
0
0
5
10
15
0
5
Time (s)
Time (s)
Figure 3.1.9: Simulation results for Test #1.
Comparing the behavior of the non-linear model with the linear state space model, there
are some notable differences. As shown in Figures 3.1.10a and 3.1.10b, the trajectory
followed in the two cases is not exactly the same even though it is clear that in both
28
CHAPTER 3. SIMULATION
situations the desired final position was reached. The linear system follows a perfect line
from the starting point to the desired point, meaning that all movements are decoupled.
In the non-linear system there exist no such perfect decoupling and as suggested by
the trajectory followed, the motion dynamics are intertwined and influence each other,
meaning for example that the movement in the quadcopter’s X axis has some impact in
the Y axis and vice versa, even though they are small.
3D Trajectory
3D Trajectory
1.5
Z (m)
1
Z (m)
0.5
Linear
Non-linear
Reference
1.5
Linear
Non-linear
Reference
1
1
0.5
0.8
0
0
1
0.6
1
0.4
0.8
0.5
Y (m)
0
0
0.6
0.4
0.2
0.8
0.6
1.2
1
0.2
0.4
0.2
Y (m)
X (m)
(a) Standard View.
0
X (m)
0
(b) Top View.
Figure 3.1.10: 3D Trajectory for Test #1.
This coupled movement is better appreciated in Figures 3.1.11a and 3.1.11b, that show
simulation results of a trajectory purely in the X axis, with a rotation of 60 degrees in the
yaw angle, in order to study the influence in the Y axis. Simulation results confirm the
theory of the interference between movements as this trajectory generates a deviation of
3 centimeters in the Y axis that then returns to zero with the controller action.
3D Trajectory
X Position
1.2
Position (m)
1.5
1
1
0.5
Linear
Non-linear
Reference
Linear
Non-linear
Reference
0
0.8
0
5
10
15
0.6
Time (s)
Y Position
Position (m)
0.03
0.4
0.02
0.2
0.01
0
0
-0.01
0
5
10
15
0.4
Time (s)
0.3
0.2
0.1
Y (m)
(a) Time response.
(b) Top View.
Figure 3.1.11: Compound movement interference.
29
0
-0.1
X (m)
-0.5
CHAPTER 3. SIMULATION
• Circular trajectories
The simulated system was also tested to track positions that change over time, as a
circular trajectory for example. This trajectory is defined as:
xc (t) = 0.5 sin (2π0.05t − π/2)
y (t) = 0.5 sin (2π0.05t)
c
z
c (t) = 1
ψ (t) = 50t
(3.1.14)
This trajectory specifies a circle of frequency 0.05Hz and radius of 0.5 meters, at a
constant altitude of 1 meter and a constant angular velocity in the yaw angle of 50
degrees per second. Figure 3.1.12 displays the simulation results with these commands.
X Position
Y Position
0.5
Position (m)
Position (m)
0.5
0
-0.5
0
-0.5
0
5
10
15
20
25
30
35
40
0
Time (s)
Z Position
1.4
5
10
15
20
25
30
35
40
25
30
35
40
Time (s)
Linear
Non-linear
Reference
Yaw angle
2000
1.2
1500
Angle (deg)
Position (m)
1
0.8
0.6
0.4
1000
500
0.2
0
0
0
5
10
15
20
25
30
35
40
0
Time (s)
5
10
15
20
Time (s)
Figure 3.1.12: Simulation results for a circular trajectory.
As there is no trajectory tracker in the control system, the path taken to follow the
trajectory specified in (3.1.14) never accomplishes the task of minimizing the error between the quadcopter’s trajectory and the desired trajectory. The position tracker by
itself cannot follow a time-varying trajectory when the change rate of said trajectory is
too fast. For example if the frequency of the circular trajectory is augmented, the path
following will be less precise. Figures 3.1.13a and 3.1.13b show the 3D circular trajectory
described by the quadcopter in this simulation:
30
CHAPTER 3. SIMULATION
3D Trajectory
3D Trajectory
1
0.8
1.5
Linear
Non-linear
Reference
0.6
0.4
1
Y (m)
Z (m)
0.2
Linear
Non-linear
Reference
0.5
0
-0.2
-0.4
0
1
-0.6
0.5
1
0.5
0
-0.8
0
-0.5
-0.5
Y (m)
-1
-1
-1
-1
X (m)
-0.5
0
0.5
1
X (m)
(a) Standard view.
(b) Top View.
Figure 3.1.13: 3D Circular Trajectory.
A helical trajectory can be easily generated by making the altitude command a ramp
function, as specified in this next equation:
xc (t) = 0.5 sin (2π0.05t − π/2)
y (t) = 0.5 sin (2π0.05t)
c
zc (t) = 0.05t
ψ (t) = 50t
(3.1.15)
in which the altitude command refers to a linear velocity of 5 centimeters per second
in the vertical axis. Simulation results are presented in Figures 3.1.14 and 3.1.15. The
newly added time-varying command in altitude worked as expected, following the ramp.
X Position
0
-0.5
0
-0.5
0
5
10
15
20
25
30
35
40
0
Time (s)
5
10
15
25
30
35
40
25
30
35
40
Yaw angle
2000
2
20
Time (s)
Linear
Non-linear
Reference
Z Position
2.5
1500
Angle (deg)
Position (m)
Y Position
0.5
Position (m)
Position (m)
0.5
1.5
1
1000
500
0.5
0
0
0
5
10
15
20
25
30
35
40
0
Time (s)
5
10
15
20
Time (s)
Figure 3.1.14: Helical Trajectory Time response.
31
CHAPTER 3. SIMULATION
3D Trajectory
2
Z (m)
1.5
Linear
Non-linear
Reference
1
0.5
0
1
0.5
1
0.5
0
0
-0.5
Y (m)
-0.5
-1
-1
X (m)
Figure 3.1.15: 3D Helical Trajectory.
The deficiencies of this control architecture to follow more complicated trajectories justify
the need to conceive a higher performance controller for the task at hand.
3.2
Linear-Quadratic Tracker (LQT)
As the previous simulation results suggested, a more refined controller is required in
order to truly track trajectories that change over time, which can be seen as a problem
of being in the appropriate place, at the appropriate time. There exists a wide variety
of controllers suited to precisely track trajectories, e.g., Model-Predictive controllers
have proven to be quite robust in the case of quadcopters [26]. But this type of nonlinear approach requires heavy calculations and often require powerful processors in
order to be effective. On the other hand, the linear-quadratic tracker has proven to be
a versatile controller method for trajectory tracking with quadcopters in previous works
as [11] and [15], with the advantages of linear controllers and their rapid prototyping
and implementation.
3.2.1
The Optimization Problem Setup
The LQT algorithm is formulated as an optimization problem to reduce a cost function
in terms of the plant’s states, inputs and certain weight functions that must be specified
by the controller designer. It is indeed a problem very much alike the well-known LQR,
but in this case the states considered for the resolution of the Algebraic Ricatti Equation
(ARE) are time-varying, which leads to gains that also vary depending on the trajectory
to follow. These characteristics make the LQT controller more appropriate than the
32
CHAPTER 3. SIMULATION
LQR when trying to accomplish more precise trajectory following, which is exactly the
feature that the previous PID architecture lacked.
As the name suggests, the LQT algorithm is part of the family of linear controllers,
thus it uses linear state space models (or linearized models, as in this case). The design
of this controller was done directly in the discrete-time domain as it makes easier its
implementation on the real platform. The linear state space realization obtained in
Section 2.4 was first discretized using a time step Ts = 0.01s that corresponds to a
frequency of 100 Hz. This time step was chosen as it corresponds to the frequency the
controller was going to be working on when implemented in the real system. Considering
the state vector:
T
∆x =
∆x ∆y ∆z ∆ψ ∆θ ∆φ ∆u ∆v ∆w ∆r ∆q ∆p
(3.2.1)
the state space realization obtained in subsection 2.4 went through a discretization process in MATLAB, using the zero-order hold method and sample time Ts . The discretetime system obeys the following dynamic equation:
(
∆ẋ[k + 1] = Ad ∆x[k] + Bd ∆u[k]
∆y[k] = Cd ∆x[k] + Dd ∆u[k]
(3.2.2)
where the matrices Ad , Bd , Cd , Dd are the result of the discretization.
Following the procedure to set up the discrete-time linear quadratic tracking system
specified in [23], considering the state space system described by (3.2.2), the performance
index to be minimized Jd is defined as:
1
T
[Cd ∆x [kf ] − z [kf ]] F [Cd ∆x [kf ] − z [kf ]]
(3.2.3)
2
kf −1
o
1 Xn
T
+
[Cd ∆x [kf ] − z [kf ]] Q [Cd ∆x [kf ] − z [kf ]] + ∆uT [k] R∆u [k]
2
Jd =
k=k0
where F and Q are state weight matrices, R is the control weight matrix and z [k] is a
12x1 vector that specifies the time-varying trajectory for each state. The final time step,
kf , is fixed and the final state value ∆x [kf ] is not fixed nor specified, thus it is called in
the literature “free” state. Weight matrices have well-defined characteristics as to obtain
a stable close-loop system using the gains given by the optimization algorithm, mainly
that F and Q are both positive semidefinite symmetric n × n matrices and R is a p × p
33
CHAPTER 3. SIMULATION
positive definite symmetric matrix (in this case n = 12 and p = 4).
For simplicity in the next algorithm equations, the following matrices are defined:
E = Bd R−1 BdT ; V = CdT QCd ; W = CdT Q
(3.2.4)
Now, using results from optimal control theory in [23] it is possible to establish a matrix
Riccati Difference Equation (RDE) as follows:
P [k] = ATd [P [k + 1] + E]−1 Ad + V
(3.2.5)
this equation must be solved backwards in time using the final condition
(3.2.6)
P [kf ] = CdT F Cd
Also the algorithm requires to solve the following vector difference equation:
h
−1 i
g [k] = Ad I12×12 − P −1 [k + 1] + E
E g [k + 1] + W z [k]
(3.2.7)
The vector g [k] depends on the desired trajectory and thus contains all information
about it. This equation must also be solved backwards in time, with the final condition:
(3.2.8)
g [kf ] = CdT F z [kf ]
After resolving (3.2.5) and (3.2.7), the optimal control law can be computed by:
∆u [k] = −L [k] x + Lg [k] g [k + 1]
(3.2.9)
where the gain L corresponds to a state feedback gain given by the expression:
−1 T
L [k] = R + BdT P [k + 1] Bd
B P [k + 1] Ad
(3.2.10)
and the gain Lg is a trajectory feedforward gain that multiplies the vector g [k] which
contains the trajectory information. This gain can be calculated from the following
equation:
−1 T
Lg [k] = R + BdT P [k + 1] B
B
34
(3.2.11)
CHAPTER 3. SIMULATION
While developing the algorithm, it was noted that matrices P [k], L [k] and Lg [k] only
varied at the end to enforce the terminal condition. It was preferred to remove the final
"free state" enforcement of the algorithm as it lead to undesired behavior at the end
of the trajectory. Hence, the aforementioned matrices were considered time-invariant
by taking their constant values before the final state enforcement. Therefore, the only
factor in the optimal control law that changes over time is the feedforward vector g [k].
The versatility of this control method is that all gains can be computed offline, meaning
before the trajectory is executed.
Being a model-based algorithm, the LQT controller performance will be closely related
to the accuracy of the linear model of the quadcopter. The previous study with the PID
architecture showed that the coupling between the movements, as well as unmodeled
phenomena such as blade flapping, body and motor force asymmetry, all contribute as
model perturbations to the system. In the light of these real-life unfavorable conditions,
during the conception of the LQT controller the addition of integral action was necessary
to ensure disturbance rejection and thus a zero steady-state error in the 3D position of
the drone.
Normally the procedure dictate that an augmentation of the system should be executed
to include the state of the integral of the error, but in the LQT algorithm these states
can further complicate the task of designing the trajectory z [k] , as it will also require
to specify a trajectory for the position integral error. The first simulation trials with
this method were not satisfactory, specifying a trivial zero trajectory for the integral of
the error states did not yield good results. Instead of searching for a model that might
describe the evolution over time of the error, which is by itself a difficult task considering
that the position trajectory can take almost any form, a much more convenient solution
to resolve the disturbance rejection problem is to simply add the integral action directly
into the control vector ∆u [k], as proposed in a similar LQT formulation [27].
In addition to the integral action in the position, it was found in practice that using
integral action in the angular position improved the overall performance of the system,
by regulating the drone’s Euler angles to zero thus keeping it stable with an increased
35
CHAPTER 3. SIMULATION
level of robustness. Finally the command vector adopted the following form:
∆u0 [k] = −L [k] x + Lg [k] g [k + 1] + Kiang
kP
f −1
k=k0
(eang [k] ∆kang ) + Kipos
kP
f −1
(epos [k] ∆kpos )
k=k0
(3.2.12)
where eang [k] is the angular error vector in regulation mode (error with respect to zero)
defined as:
T
eang [k] =
(3.2.13)
−ψ [k] −θ [k] −φ [k]
and similarly epos [k] is the position error vector with respect to the desired trajectory
z [k]:
T
epos [k] =
z1 [k] − x [k] z2 [k] − y [k] z3 [k] − z [k]
(3.2.14)
The coefficients ∆kang and ∆kpos are the time steps corresponding to each integral gain.
The block diagram in Figure 3.2.1 represents the closed-loop control system proposed.
ue
+
u
+ ∆u0
P
Kipos
(epos ∆kpos )
+
+
+
quadcopter’s
Dynamics
−
∆u
L[k]
+
Kiang
P
x
Lg [k]
g[k + 1]
(eang ∆kang )
Integral Action
LQT Algorithm calculated offline
Figure 3.2.1: LQT Closed-Loop System.
With a better understanding of how to setup the LQT problem, the next logical step
was to test the algorithm in the quadcopter model and study the feasibility for practical
implementation in the Crazyflie 2.0 platform.
Before testing the actual control algorithm, it became necessary to address the problem
of the observation of the states, mainly to answer the question of how to reconstruct
all 12 states of the dynamic model given the data coming from different sensors used in
the implementation. The Inertial Measurement Unit inside the Crazyflie 2.0 gives good
estimates of the Euler angles and the body-fixed frame angular velocities, by a sensor
fusion algorithm that merges data coming from the accelerometer and the gyroscope. A
36
CHAPTER 3. SIMULATION
localization system such as the VICON estimates the position of an object with respect
to a certain fixed inertial frame, but a priori these type of systems do not directly measure
nor estimate the linear velocities. Thus the need of some algorithm that can reconstruct
the missing states from a model of how they evolve over time as well from the sensors’
data.
3.2.2
Kalman Filter for Linear Velocity Estimation
With the objective of simulating as close as possible real-life scenarios, white Gaussian
noise was added to the position values coming from the quadcopter non-linear dynamics,
as to reproduce the uncertainties given by any position-fixed system such as VICON or
an ultra-wide band system. In addition, the linear velocity is not directly given by any of
these systems, thus the need for a state observer capable of filtering the noise coming from
the position estimation while correctly computing estimations for the linear velocities in
the inertial frame defined by the positioning system.
The well-known Kalman Filter is ideal for these type of tasks. Using as reference the
example found in [24], the Kalman Filter problem was stated. First a vector containing
the estimated variables must be defined:
T
x̂ [k] =
ˆ ẏˆ żˆ
x̂ ŷ ẑ ẋ
(3.2.15)
As the estimation process only addresses the state variables for the position and linear
velocities, this reduced state vector is ideal to set-up the problem. The next step is to
describe the state space system based on the dynamics between the position and the
velocity of a given rigid body. The following equation defines the dynamics of the state
space in discrete-time:
(
x̂[k + 1] = Ahat x̂[k] + Gw [k]
yhat [k] = Chat x̂[k] + v [k]
(3.2.16)
where w [k] is the process noise vector, that multiplies a certain matrix G that models
how this noise interacts within the state evolution. Then v [k] is a random variable with
certain variance that simulates the measurements noise of the position-fixed system and
yhat [k] is the noise-infected output of the reduced position-velocity system. Matrix Ahat
37
CHAPTER 3. SIMULATION
is the state transition matrix, defined as:
Ahat
=
1
0
0
0
0
0
0
1
0
0
0
0
0 Ts 0 0
0 0 Ts 0
1 0 0 Ts
0 1 0 0
0 0 1 0
0 0 0 1
(3.2.17)
then matrix G was modeled:
G=
Ts /2
0
0
0
Ts /2
0
0
0
Ts /2
1
0
0
0
1
0
0
0
1
(3.2.18)
Thus, by examining the three lower rows of matrices Ahat and G, the evolution of the
linear velocities is defined as:
˙
˙
x̂ [k + 1] = x̂ [k] + w1 [k]
ŷ˙ [k + 1] = ŷ˙ [k] + w2 [k]
˙
ẑ [k + 1] = ẑ˙ [k] + w3 [k]
(3.2.19)
which describes a constant velocity model with and added random variable that accounts
for the process noise. As possible improvement, acceleration estimations in the world
frame could be added to the model, taking for example the accelerometer measurements
and the appropriate rotation estimate. Then, the position evolution was modeled starting
by the principle that the position in reality is a continuous variable and the velocity is
defined as the derivative of the position:
d
x̂ = x̂˙
dt
(3.2.20)
this relationship can be approximated in the discrete domain as:
x̂ [n + 1] − x̂ [n]
x̂˙ [n + 1] + x̂˙ [n]
=
Ts
2
(3.2.21)
Now, the left hand side of (3.2.21) is the approximation of the derivative in discrete time,
while the right hand side describes the approximation of a constant velocity between two
time steps. Basically the model proposed works with the idea of a constant velocity and
38
CHAPTER 3. SIMULATION
the accelerations that will introduce changes to the velocity are modeled as random
variables.
Matrix Chat selects which states are truly measured by the position-fixed system, in this
particular case that corresponds to the position states, then:
Chat
1 0 0 0 0 0
= 0 1 0 0 0 0
0 0 1 0 0 0
(3.2.22)
Now that the system is clearly defined, a noise characterization needs to be done in order
to get the best possible estimator. In practice that means to estimate the measurement
noise covariance from experimentally-retrieved data. This estimation gives a good approximation for the noise covariance matrix Rkal that is assumed to be diagonal from
the principle that there exists no correlation between the noises of the different positions.
Then the noise covariance matrix takes the following form:
Rkal = diag
h
σx2 σy2 σz2
i
(3.2.23)
As for the process noise covariance Qkal , it is analytically difficult to define or estimate its
weights. In practice, the values are often found by experimentation rather than modeling
process noise as a consequence of unmodeled phenomena. As the dynamical model for
each one of the axes is the same, and theoretically decoupled, the matrix Qkal takes also
the form of a diagonal,
Qkal = diag
h
2
σw
1
2
σw
2
2
σw
3
i
(3.2.24)
There exists an evident coupling between the position and the velocity process noises, but
this over complicates the task of a rather simple method of state observation, thus the
diagonal form was preferred over a full scale 3x3 matrix. Also note that one advantage
of using the matrix G is that it reduces the tuning parameters of the Kalman filter
algorithm, by considering a smaller dimension process covariance matrix. Normally,
without a matrix G in the model, the process covariance matrix is n × n, but if added,
a new matrix Q̄kal defined as:
Q̄kal = GQkal GT
(3.2.25)
is considered for the resolution of the Algebraic Riccati Equation associated with the
Kalman filter. Finally, the dimension of matrix Q̄kal is (n − r) × (n − r), where r is the
39
CHAPTER 3. SIMULATION
number of columns of matrix G and therefore, the dimension of the vector w [k].
In order to complete the Kalman filter design, the values for the noise matrices must be
given. Two scenarios were taken in account: one in which the data was taken from the
VICON system and another in which the position estimation came from an ultra-wide
band system.
1. VICON system: a simple test leaving the drone steady on the floor allowed the
retrieval of data during 1 minute to calculate through MATLAB the variance of
the position in each axis . The average results were:
σx2 = σy2 = σz2 = 5 × 10−9 [m2 ]
(3.2.26)
Having fixed the values of the noise covariance matrix Rkal , the process covariance
matrix was hand-tuned through simulation and experimentation. These values
were:
2
2
2
σw
= σw
= σw
= 8 × 10−8
1
2
3
(3.2.27)
The fact that the process covariance matrix has low values indicates that the state
space model is good at predicting future values for the estimated state vector.
2. Ultra-wide Band system: the system was used to estimate the X and Y position
of the quadcopter, while the altitude was tracked with the VICON. The same
procedure as before was applied, giving the following noise variance values:
σx2 = σy2 = 5 × 10−5 [m2 ] ; σz2 = 5 × 10−9 [m2 ]
(3.2.28)
these values suggested that the standard deviation of the UWB is around 100 times
greater that the VICON’s. Then for the process covariance matrix, the following
values were used:
2
2
σw
= σw
= 3 × 10−5 ; σz2 = 8 × 10−8
1
2
(3.2.29)
The validation of the filter was done using real data from a dummy test using the Crazyflie
2.0 and making a simulated flight by hand, just grabbing the drone and moving it around
the test area. Comparisons were made between the raw data of the position estimations
with the filter output, and for the velocity a discrete time derivative of the incoming
40
CHAPTER 3. SIMULATION
data estimations was taken to contrast it with the output of the filter.
First, while using the VICON positioning system and the values of variance in (3.2.26)
and (3.2.27), the experimental results obtained are shown in Figure 3.2.2. For the position estimation, the output of the filter superposes with the raw data as the VICON
system does already a lot of filtering and the raw data has low levels of noise, hence the
Kalman Filter algorithm output for the position is virtually the same as the raw data.
As for the velocity estimations, the Kalman filter reduces the noise levels while being
fast enough to follow the true dynamics.
X Position
0.5
Y Position
0.4
VICON
Kalman Filter
0.2
Position (m)
Position (m)
0
-0.5
0
-0.2
-0.4
-1
-0.6
-1.5
-0.8
0
5
10
15
20
25
30
35
0
5
10
15
Time (s)
X Velocity
4
20
25
30
35
25
30
35
Time (s)
Y Velocity
4
Discrete Derivative
Kalman Filter
2
Velocity (m/s)
Velocity (m/s)
2
0
-2
0
-2
-4
-4
0
5
10
15
20
25
30
35
0
5
10
Time (s)
15
20
Time (s)
Figure 3.2.2: Experimental validation of the Kalman Filter using the VICON system
raw data of X-Y positions.
Then, using the UWB system, the filter was adjusted with the appropriate variance
values for Rkal and Qkal , found in (3.2.28) and (3.2.29). The experimental results are
displayed in the time plots of Figure 3.2.3.
The position raw data is lightly filtered in order to keep low levels of lag in the estimations, while the velocity estimations of the Kalman Filter are more heavily filtered and
at the same time fast enough to keep a good convergence speed.
41
CHAPTER 3. SIMULATION
X Position
1
Y Position
1.5
UWB
Kalman Filter
1
Position (m)
Position (m)
0.5
0
-0.5
-1
0.5
0
-0.5
-1.5
-1
0
5
10
15
20
25
30
35
0
5
10
15
Time (s)
X Velocity
4
20
25
30
35
25
30
35
Time (s)
Y Velocity
4
Discrete Derivative
Kalman Filter
2
Velocity (m/s)
Velocity (m/s)
2
0
-2
0
-2
-4
-4
0
5
10
15
20
25
30
35
0
5
10
15
Time (s)
20
Time (s)
Figure 3.2.3: Experimental validation of the Kalman Filter using the UWB system raw
data of X-Y positions.
The last step of the filter validation was testing the altitude estimations, which in both
cases came from the VICON system. Figure 3.2.4 displays the experimental results for
this test.
Z Position
Position (m)
1.5
VICON
Kalman Filter
1
0.5
0
-0.5
0
5
10
20
25
30
35
Time (s)
Z Velocity
5
Velocity (m/s)
15
Discrete Derivative
Kalman Filter
0
-5
0
5
10
15
20
25
30
35
Time (s)
Figure 3.2.4: Experimental validation of the Kalman Filter altitude estimation from
VICON raw data of Z position.
Similar as the results seen in Figure 3.2.2, the altitude estimation results confirm the
quality of the filter developed and therefore validate the conception proposed.
42
CHAPTER 3. SIMULATION
3.2.3
Weight Matrices and Integral Action
A simulation environment in MATLAB was created to test the LQT algorithm with the
non-linear dynamics of the quadcopter, the added noise to the position states and the
Kalman Filter. The LQT algorithm is calculated before the simulation runs, using the
discrete time linear model of the quadcopter and the following weight matrices:
Q = diag (2000, 2000, 4000, 4000, 4000, 4000, 20, 20, 10, 10, 10, 10)
F =Q
R = 0.00003 × I4×4
(3.2.30)
It is common practice to choose state weight matrices such as Q and F to be diagonal,
that is a way of imposing individual weights to each one of the states considering that
all of them are decoupled. In the case of the quadcopter it is known the existence of
some coupling between the movements, although it is minor.
Another remark is that being a linear algorithm, the performance will hold as long as
the quadcopter stays in the vicinity of the hover point, meaning by default that the three
Euler angles will always be regulated at zero degrees. This implies that with the LQT
algorithm, as proposed in this project, specifying a trayectory for the yaw angle will not
be possible as the linearization will not longer be a valid approximation.
Continuing with the setup of the simulation environment for the LQT controller, the
angular integral action gain, the matrix Kiang , takes the following format:
Kiang =
−kiang −kiang −kiang
kiang
kiang −kiang
−kiang kiang
kiang
kiang −kiang kiang
(3.2.31)
where kiang was a tuned parameter to compensate modeling errors and other unmodeled
phenomena. The tuning procedure is further explained in Section 4.2. The value used
both in simulation and in the experimental phase was kiang = 8660.
Similarly, the position integral action gain Kipos takes on the form:
43
CHAPTER 3. SIMULATION
Kipos =
kipos −kipos
−kipos −kipos
−kipos kipos
kipos
kipos
−kipos
−kipos
−kipos
−kipos
(3.2.32)
The value chosen for kipos was 5000, after thorough testing both in simulation and practice. The approach to find the correct weight matrices in (3.2.30) was a simple trialand-error method, taking in account some general knowledge about the dynamics of the
quadcopter, but most importantly through experimentation with the actual platform.
Basically the simulation served as a first good approximation to obtain decent performance in practice and then as a tool to know which parameters to tune further to improve
the control system in the real-life scenario.
3.2.4
Trajectory Generation
A trajectory for the 12-state vector must be specified before running the LQT algorithm.
For the generation of the trajectory, a small angle approach was taken meaning that the
trajectory for angular velocities and angular positions was considered as zero throughout
the whole trajectory. As for the trajectory of the states [u, v, w], the position trajectory
was simply fed to a Kalman Filter similar as the one previously designed to obtain a
velocity profile for each axis (in this case the filter acts only as a velocity estimator).
Note that in the small angle approximation, the vector [u, v, w] coincides with the vector
[ẋ, ẏ, ż], hence a projection from the inertial frame to the body-fixed frame was not
necessary.
The generation of feasible trajectories for a quadcopter is a complex and open research
subject that has been treated before in works such as [25]. In the case of this project
these types of studies were not considered and are suggested as future work.
The trajectories were generated via MATLAB through a GUI interface (modified version
of function get_curve.m) that allowed the user to specify a number of waypoints in the
X-Y plane and then a cubic interpolation calculated a trajectory between each one of
the waypoints. After this first step was done, a similar window opened to specify the
trajectory in the altitude of the drone. The interpolation was done with respect to a
time vector with a fixed time step of Ts = 0.01s and a range from 0 to a fixed final time
44
CHAPTER 3. SIMULATION
in which the trajectory was to be executed. Then the interpolation function created a
position vector of the same length as the time vector, specifying a discrete time trajectory
z[k] between the waypoints chosen through the interface. The interpolation method gives
two variants to the trajectory generated:
1. “spline” does the classic piecewise cubic interpolation between the waypoints.
2. “pchip” does a shape-preserving cubic interpolation.
Figure 3.2.5 shows a series of points chosen by the user and the corresponding interpolation and trajectory generated by the interface. The user must then choose between the
“spline” or “pchip” trajectories generated.
This method establishes an easy and automated way of generating trajectories, but the
feasibility of said trajectories depends mostly on the time allowed to execute them.
Use mouse clicks to pick points INSIDE the gridded area
2
1.5
1.5
1
1
0.5
0.5
Y (m)
2
0
Waypoints
Spline
Pchip
0
-0.5
-0.5
-1
-1
-1.5
-1.5
-2
X-Y Trajectory Generated
-2
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
-2
-1.5
-1
-0.5
When you are done, click OUTSIDE the gridded area
0
0.5
1
1.5
X (m)
(a) MATLAB GUI waypoint selection
(b) Generated Trajectory
Figure 3.2.5: Trajectory generation GUI.
3.2.5
Simulation Results
The control system was thoroughly tested in simulation before the implementation phase,
even though the final tuning of the controller was done using and getting to know the
experimental platform. First the step response was computed for the x, y and z positions,
using simulated noise for both the VICON and UWB system. Simulation results are
appreciated in Figure 3.2.6.
45
2
CHAPTER 3. SIMULATION
X Position
Position (m)
1.5
UWB
VICON
Reference
1
0.5
0
-0.5
0
5
10
15
20
25
30
35
40
25
30
35
40
25
30
35
40
Time (s)
Y Position
Position (m)
1.5
1
0.5
0
-0.5
0
5
10
15
20
Time (s)
Z Position
Position (m)
1.5
1
0.5
0
-0.5
0
5
10
15
20
Time (s)
Figure 3.2.6: Simulation for steps in x, y and z positions.
This first simulation of the LQT algorithm shows an interesting feature for a step response. Instead of obtaining a classical response that starts when the step is commanded,
the quadcopter actually starts moving beforehand as to minimize the overall error of the
trajectory. This feature is only possible because the trajectory was known by the controller and the algorithm calculated the appropriate feedforward gain g [k] to ensure
the “anticipatory” feature seen in simulation. Another important remark after playing
around with the simulation is that the overshoot can be reduced easily by decreasing
the integral gains Kiang and/or Kipos , but due to later difficulties in the implementation
they remained with the values specified before (this subject is further discussed in Section 4.3.2). As to the UWB vs VICON performance, the time plots suggests that both
have almost the same exact response in position.
The Kalman Filter performance in simulation to estimate the linear velocities is presented in Figure 3.2.7. The filters main job in the control system is to calculate reliable
state estimations for the linear velocities in the inertial frame, then the appropriate rotation projects this estimations in the body frame thus obtaining estimates for the states
[u, v, w]. Simulation results suggests that the levels of noise in the estimations for the
UWB X-Y velocities is greater, but this was expected knowing already that the noise
has at least two order of magnitude greater standard deviation than the VICON system.
46
CHAPTER 3. SIMULATION
X Velocity in body frame
Velocity (m/s)
1
Kalman Filter UWB
True Dynamics
Kalman Filter VICON
True Dynamics
0.5
0
-0.5
-1
0
5
10
15
1
Velocity (m/s)
20
25
30
35
40
25
30
35
40
25
30
35
40
Time (s)
Y Velocity in body frame
0.5
0
-0.5
-1
0
5
10
15
0.5
Velocity (m/s)
20
Time (s)
Z Velocity in body frame
0
-0.5
0
5
10
15
20
Time (s)
Figure 3.2.7: Kalman Filter simulation, with VICON and UWB simulated noise.
Nonetheless, the Kalman Filter in both cases manages a good compromise between the
filtering and the convergence speed. As seen in the plots of the X and Y velocities, the
UWB Kalman Filter introduced some delay (∼200ms) in the estimations with respect to
the true dynamics of the quadcopter. This was the cost for a more aggressive filtering
of the noise. In practice the proposed gains provided a satisfactory performance despite
the lag introduced, ultimately it was found that if the filtering was less aggressive then
too much noise entered the system and the overall performance of the tracker degraded.
To truly justify the need of using a more refined trajectory tracker, the simulated system
was subject to follow a more complex trajectory to test the tracking capabilities of the
LQT algorithm and also to evaluate the Kalman filter performance under more complex
situations. Figure 3.2.8 shows simulation results for a trajectory created by the user
interface presented in Section 3.2.4.
The LQT controller was capable of tracking the desired trajectory, with a low error even
in closed curves as suggests the 3D diagrams in Figure 3.2.9.
47
CHAPTER 3. SIMULATION
X Position
Position (m)
2
UWB
VICON
Reference
1
0
-1
-2
0
5
10
15
20
25
30
35
40
25
30
35
40
25
30
35
40
Time (s)
Y Position
Position (m)
2
1
0
-1
-2
0
5
10
15
20
Time (s)
Z Position
Position (m)
3
2
1
0
0
5
10
15
20
Time (s)
Figure 3.2.8: Tracking for complex trajectories.
3D Trajectory
3D Trajectory
2
1.5
2.5
1
2
Y (m)
Z (m)
0.5
UWB
VICON
Reference
1.5
1
-0.5
0.5
0
2
0
-1
2
1
UWB
VICON
Reference
-1.5
0
0
-1
Y (m)
-2
-2
-2
-1.5
X (m)
-1
-0.5
0
0.5
1
1.5
2
X (m)
(a) Standard view
(b) XY Plane
Figure 3.2.9: 3D Diagram for a complex trajectory.
As for the Kalman filter estimations, Figure 3.2.10 shows that the filter performed as
expected when asked to track more complex velocity profiles as the one generated by
this trajectory.
48
CHAPTER 3. SIMULATION
X Velocity in body frame
Velocity (m/s)
1
Kalman Filter UWB
True Dynamics
Kalman Filter VICON
True Dynamics
0.5
0
-0.5
-1
0
5
10
15
1
Velocity (m/s)
20
25
30
35
40
25
30
35
40
25
30
35
40
Time (s)
Y Velocity in body frame
0
-1
-2
0
5
10
15
0.4
Velocity (m/s)
20
Time (s)
Z Velocity in body frame
0.2
0
-0.2
-0.4
0
5
10
15
20
Time (s)
Figure 3.2.10: Kalman Filter simulation for a complex trajectory.
The simulation environment for the LQT control system proved to be a useful tool for
developing and understanding the mechanics of the newly adopted control technique for
trajectory tracking. Going hand by hand with the implementation phase, the simulation
proposed in this section served as a helpful guidance while fine tuning the controller in
practice, and at the same time serving as reference in terms of the performance to aim
for in the control system.
49
Chapter 4
Hardware Implementation and
Experimental Results
The implementation process was divided in two major phases: first a familiarization with
the Crazyflie 2.0 platform and implementation of a PID position controller and secondly
the implementation of a linear quadratic algorithm.
4.1
PID Controller
After a thorough analysis of the original Crazyflie 2.0 firmware it became evident that
some minor changes were needed to be made because the body-fixed frame defined in the
embedded system did not match the one adopted during the simulation phase, therefore,
for the sake of consistency with the body-fixed frame defined in Figure 2.2.3, certain lines
of the original firmware code were changed (see Appendix A: Firmware Modifications).
The first step towards the implementation was to test the on-board sensors such as
the inertial measurement unit and all its components. Even though the sensors’ data
analysis was not part of the project, it was important to at least follow the idea of how
the firmware captured said data, ran it, for example, through the sensor fusion algorithm
and estimated the states of the quadcopter that were fed to the on-board controllers.
After this familiarization phase, the research moved towards the off-board controller
implementation using ROS.
4.1.1
ROS Controller Node
Starting from the Open source ROS nodes presented in [6], the controller node was
modified to implement the equations proposed in Section 3.1.2. Figure 4.1.1 shows the
50
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
ROS nodes and topics concerning the controller implemented.
Figure 4.1.1: ROS nodes and topics.
A detailed explanation of the internal organization of each one of this nodes and interactions is beyond the scope of this report, but a qualitative analysis is adequate for an
overall understanding of the system:
• Crazyflie Server: this node defines the core interaction between ROS and the
Crazyflie 2.0, through a radio communication. Its primary function in the control process is to serve as a data bridge between the Crazyflie and the off-board
controller. On the one hand, it publishes sensors’ readings coming from the quadcopter, such as the IMU, that will serve as state estimations for the control process.
On the other hand, the node receives commands coming from the controller node
and sends them back to the Crazyflie. More specifically, the data sent in the
“/cmd_vel” topic contains the outputs of the off-board controller, that means,
messages of the form [ φc θc rc Ω ]. This messages will then be the inputs of
the on-board control system, as shown previously in Figure 3.1.5.
• VICON Listener: manages the communication with the VICON positioning
system. It publishes data concerning the inertial frame coordinates [x, y, z] of a
reflective sphere as seen in Figure 4.1.2, sitting on top of the Crazyflie 2.0.
Figure 4.1.2: Crazyflie 2.0 with Vicon Sphere and UWB module.
51
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
Ideally, the VICON system should be used with three or more of these spheres, in
order for the position estimations to be more resilient against other reflections in
the laboratory or even for the Euler angles estimations. However, up to this point,
all the tests were done using the configuration shown in Figure 4.1.2. and it did
not present any major drawbacks during the tests.
• Joystick: this nodes serves basically as an external emergency stop button. It is
always a nice idea to keep the security measures “software free” in case something
goes wrong during a flight, thus the preference of using a hardware stop button
instead of one, probably more elegant, implemented in software such as MATLAB.
• Goal: this is a user interactive topic created in MATLAB that lets the user choose
between a number of predefined trajectories for the quadcopter to follow. Once
the trajectory is selected, through a certain ID number that can be changed in
real-time, the MATLAB node publishes the desired waypoint [xc , yc , zc , ψc ]. An
additional feature allows the user to choose whether to send position or velocity
commands to the yaw angle of the drone.
• Controller: this is naturally the core of the real-time control system. The node
takes data from the IMU of the Crazyflie and from the VICON system in order to
have an estimate of the states of the quadcopter in the control algorithm presented
in Section 3.1.2 . However this data only gives estimations of 9 states: the three
linear positions in the inertial frame, the three Euler angles and the three angular velocities coming from the gyroscope. The linear velocities in the body frame
[u, v, w] are not directly measured by any of the sensors and they need to be estimated somehow as they are used in the X-Y Position Controller. Usually, in this
type of scenarios, a state observer (Luenberger’s, Kalman Filter, etc) is required to
reconstruct the missing states. However, given the fact that the position estimations of the VICON system are truly precise (error<1mm), the following equations
to estimate the velocities in the body-fixed frame using a pseudo-derivative in discrete time of the X-Y positions proved to be good enough for the control system
to behave adequately:
x[k] − x[k − 1]
∆t
y[k] − y[k − 1]
∆t
Vx [k] =
Vy [k] =
(4.1.1)
(4.1.2)
where Vx [k] and Vy [k] are the linear velocities estimations in the inertial frame.
52
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
The term ∆t is the time step taken between iterations in the control algorithm,
given that the controller node runs at 100 Hz, then ∆t = 0.01s. Normally to
obtain the velocities in the body-fixed frame a multiplication by the Euler matrix
Rob should be executed, but taking a small angle approximation for the pitch and
roll angles, then the calculation is simplified to just one rotation of the yaw angle
around the Z axis, as shown in the following equation:
"
u[k]
v[k]
#
"
=
cos (ψ[k]) sin (ψ[k])
− sin (ψ[k]) cos (ψ[k])
#"
Vx [k]
Vy [k]
#
(4.1.3)
Finally the linear velocities in the body-fixed frame can be expressed in terms of
the measured states:
u[k] = cos (ψ[k]) x[k]−x[k−1] + sin (ψ[k]) y[k]−y[k−1]
0.01
0.01
v[k] = − sin (ψ[k]) x[k]−x[k−1] + cos (ψ[k]) y[k]−y[k−1]
0.01
0.01
(4.1.4)
This approximation proved to be good enough for the control architecture proposed
to work correctly.
4.1.2
Experimental Results
The flight data was retrieved using the MATLAB node mentioned earlier, allowing for
a more analytical interpretation of the results. The following section presents a number
of these flights with the appropriate analysis.
• Linear trajectories
For this type of trajectories the tests consisted basically in a take-off, a linear movement
in one or more directions, and a landing. For the first test, steps of two different amplitudes were sent in the vertical position of the drone, trying to maintain its X-Y position.
The time plots of Figure 4.1.3 present the experimental data retrieved.
The two commands were sent to maintain an altitude of either 0.68 meters or 1.18 meter,
with all other commands set to zero. The X position stayed within a 6cm margin of error
while the Y position had a more prominent error, roughly a 10 cm margin from the initial
take-off position. The fact that the drone used for these tests is not entirely symmetrical
(see Figure 4.1.2) means that the position holding in one of the axes is better than in
the other, as it was in fact the case. Figure 4.1.4 shows different 3D perspectives of the
trajectory followed by the Crazyflie during the test.
53
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
X Position
0.08
0.1
0.04
Position (m)
Position (m)
0.06
0.02
0
0.05
0
-0.05
-0.02
-0.04
-0.1
0
5
10
15
20
25
30
35
40
0
Time (s)
5
10
15
25
30
35
40
25
30
35
40
Yaw position
12
10
Angular Position (deg)
1.2
20
Time (s)
PID
Reference
Z Position
1.4
1
Position (m)
Y Position
0.15
0.8
0.6
0.4
0.2
0
8
6
4
2
0
-2
0
5
10
15
20
25
30
35
40
0
5
10
15
Time (s)
20
Time (s)
Figure 4.1.3: Steps in vertical command zc .
3D Trajectory
3D Trajectory
2
1.8
2
1.6
PID
Reference
1.5
PID
Reference
1.4
Z (m)
Z (m)
1.2
1
1
0.8
0.5
0.6
0.4
0
1
0
Y (m)
-1
-1
0
-0.5
0.5
0.2
1
0
-1
X (m)
-0.5
0
0.5
1
X (m)
(a) Standard view
(b) ZX Plane
Figure 4.1.4: 3D Vertical Trajectory.
The next test consisted in sending commands both in the X position and in the altitude,
in order to study the quadcopter’s behavior when sending mixed movements. Experimental data is plotted in Figure 4.1.5, showing the in-flight response of the quadcopter.
54
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
X Position
1.2
Y Position
0.1
1
0.05
Position (m)
Position (m)
0.8
0.6
0.4
0.2
0
-0.05
0
-0.2
-0.1
0
5
10
15
20
25
30
35
0
5
10
15
Time (s)
20
25
30
35
25
30
35
Time (s)
PID
Reference
Z Position
1.4
Yaw position
15
Angular Position (deg)
1.2
Position (m)
1
0.8
0.6
0.4
0.2
10
0
5
0
-5
0
5
10
15
20
25
30
35
0
5
10
15
Time (s)
20
Time (s)
Figure 4.1.5: Steps in zc and xc .
Similar to the simulations, the response time for a unity step in the X position was around
4 seconds, whereas for the altitude is around 7 seconds. The Y position remained in a
10 cm error margin and the yaw angle stayed well within a 3 degree margin after the
initial take off (which is represented in the peak at 3 seconds). Figure 4.1.6 shows two
different 3D perspectives of the experimental flight.
3D Trajectory
3D Trajectory
2
1
PID
Reference
PID
Reference
0.8
0.6
1.5
Y (m)
Z (m)
0.4
1
0.2
0
-0.2
0.5
-0.4
-0.6
0
-0.8
1
0
Y (m)
-1 -1
-0.5
0
0.5
1
-1
-1
X (m)
-0.5
0
0.5
1
X (m)
(a) Standard view.
(b) XY Plane.
Figure 4.1.6: 3D Diagonal Trajectory.
It remained only to test step commands in the Y axis and in the yaw position, so this
last test in the linear trajectories consisted in sending both steps simultaneously and
study the response of the system knowing the movement interference analysed during
55
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
the simulation phase of the project. Test results are showcased in Figure 4.1.7.
X Position
0.3
0.4
Position (m)
Position (m)
0.2
0.1
0
-0.1
0.2
0
-0.2
-0.2
-0.4
0
10
20
30
40
50
0
Time (s)
10
20
40
50
40
50
Yaw position
100
1.2
30
Time (s)
PID
Reference
Z Position
1.4
Angular Position (deg)
80
1
Position (m)
Y Position
0.6
0.8
0.6
0.4
0.2
0
60
40
20
0
-20
0
10
20
30
40
50
0
10
Time (s)
20
30
Time (s)
Figure 4.1.7: Steps in yc and ψc .
While maintaining an altitude of 1 meter and a position in X of zero, the commands sent
were yc = 0.5m and ψ = 90°. The response time for the Y position was, as in the case of
the X axis, around 4 seconds. On the other hand, the time response for a command in
yaw is much faster, about 1 second. Studying the impact these two movements had in
the other positions, for the altitude it is noted how it remained roughly within the same
level during the whole trial, meaning that those commands did not have a major impact
and the altitude controller compensated any small interference those movements could
have generated. As for the X position, aside from the initial deviation due to take-off,
the position stayed around the 10 cm error margin, but in this case the effect caused
by the movements was much more notable, as seen in the 3D perspectives of Figure 4.1.8.
The error in the X position around the initial position is more important than in the
other trajectories that were tested. Even as a qualitative analysis from the experimental observations, the simultaneous movement of the Y position and the yaw rotation
generated light deviations in the X axis while moving from point A to point B of the
trajectory.
56
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
3D Trajectory
0.6
2
0.5
1.5
0.4
PID
Reference
PID
Reference
0.3
1
Y (m)
Z (m)
3D Trajectory
0.5
0.2
0.1
0
-1
0
-0.5
-0.1
0
-0.2
0.5
X (m)
1
-1
-0.5
0
0.5
1
-0.3
-1
Y (m)
-0.5
0
0.5
1
X (m)
(a) Standard view.
(b) XY Plane.
Figure 4.1.8: 3D Trajectory with Y-Yaw compound movement.
• Circular trajectories
The next step for testing the position controller was to generate more complicated trajectories and see if the system was capable of following the desired path. To describe a
circle, the following function in discrete time was implemented:
xc [k] = x0 + sin (2π0.1k)
y [k] = y + sin (2π0.1k + π/2)
c
0
zc [k] = 0.9
ψ[k] = −30k
(4.1.5)
where the values x0 and y0 represent the initial position of the drone in the X-Y plane.
The experimental data retrieved while performing the circular trajectory is presented in
Figure 4.1.9.
In the X-Y position it is observed that the amplitude of the sine waves did not reach
the value of 1m, meaning that the position controller failed to minimize the error as
a consequence of not being fast enough for more complicated trajectories. A similar
problem occurred with the simulated model and actually the explanation can be reused.
Given the fact that the controller is a position tracker, it will try to regulate at each
time step of the process the actual position with the desired position. In this case, the
rate and amplitude of the trajectory were too high for the position tracker developed to
actually make the quadcopter follow the path required.
57
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
X Position
1
0.5
Position (m)
0.5
Position (m)
Y Position
1
0
-0.5
0
-0.5
-1
-1
0
10
20
30
40
50
60
0
Time (s)
Z Position
1.2
10
20
30
40
50
60
40
50
60
Time (s)
PID
Reference
Yaw position
200
Angular Position (deg)
Position (m)
1
0.8
0.6
0.4
0.2
0
100
0
-100
-200
0
10
20
30
40
50
60
0
10
20
Time (s)
30
Time (s)
Figure 4.1.9: Time Response for circular trajectory.
The 3D trajectories exhibited in Figure 4.1.10 show more clearly the deficiencies in the
path followed by the quadcopter.
3D Trajectory
3D Trajectory
PID
Reference
1
0.8
1.2
0.6
1
0.4
Y (m)
Z (m)
0.8
0.6
0.4
PID
Reference
0.2
0.2
0
-0.2
-0.4
0
1
-0.6
0.5
1
0.5
0
Y (m)
-0.8
0
-0.5
-0.5
-1
-1
-1
-1
X (m)
-0.5
0
0.5
1
X (m)
(a) Standard view.
(b) Top view.
Figure 4.1.10: 3D Circular trajectory.
The last trajectory tested was the helix as seen during the simulation phase, the discrete
time function implemented in the controller was defined as:
58
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
xc [k] = x0 + sin (2π0.1k)
y [k] = y + sin (2π0.1k + π/2)
c
0
zc [k] = 0.9 + 0.04k
ψ[k] = −90k
(4.1.6)
meaning that in the X-Y plane it described the same circle as before, but this time
augmenting the altitude at a ratio of 4 centimeters per second. Experimental data
in Figure 4.1.11 shows the performance of the quadcopter while following the helical
trajectory.
X Position
1
0.5
Position (m)
Position (m)
0.5
0
-0.5
0
-0.5
-1
-1
0
5
10
15
20
25
30
35
40
0
Time (s)
5
10
15
Angular Position (deg)
1.5
1
0.5
0
25
30
35
40
25
30
35
40
Yaw position
200
2
20
Time (s)
PID
Reference
Z Position
2.5
Position (m)
Y Position
1
100
0
-100
-200
0
5
10
15
20
25
30
35
40
0
Time (s)
5
10
15
20
Time (s)
Figure 4.1.11: Time Response for helix trajectory.
The increase of the yaw angle velocity did not have a major impact in the performance of
the system, the Crazyflie could turn at 90 degrees per second without any complications.
The results in the Z position confirms that the altitude controller is good enough to follow a time-varying function such as a straight line, even though the amplitude variation
of this trajectory was small. For the X-Y position the same phenomenon occurred as
in the case of the circular trajectory. The 3D perspectives in Figure 4.1.12 displays the
helicoidal trajectory followed.
After take-off, the quadcopter began the regulation of the desired trajectory which
presents the same deficiencies as the previous test. The landing was successful as the
data shows that the landing site was just a few centimeters away from the take-off point.
59
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
3D Trajectory
3D Trajectory
1.5
0.5
Y (m)
Z (m)
PID
Reference
1
2
1
0
0.5
PID
Reference
0
1
0.5
0
Y (m)
-0.5
-1
-1
-0.5
0
0.5
-0.5
1
-1
-1
X (m)
-0.5
4.2
0.5
1
X (m)
(a) Standard view.
Figure 4.1.12:
0
(b) Top view.
3D Helix Trajectory.
LQT Controller Implementation
Having completed the first phase of the implementation process, a broader knowledge
of the embedded system and its limitations was gained which was vital in the task of
implementing the more delicate LQT system. The word “delicate” in this context refers
to the fact that now the design included some low level control that was not done in
the previous phase. With the PID controller, the two cascaded architecture embedded
in the drone were already designed by the manufacturer, thus the design was reduced
just to the off-board position controller. Instead, this new implementation required the
careful thought of how to implement every portion of the controller.
The design process began with the naive idea of implementing the whole 12 state feedback off-board using the MATLAB/ROS interface to send directly the motor commands
through the radio link communicating the computer with the Crazyflie 2.0. This approach became rapidly discarded after some failed trials, mainly due to the latency of
the radio link (around 2ms according to the manufacturer’s specifications) and to the refresh rate used in the MATLAB interface (100 Hz). The low level stabilization, meaning
the angle and angular velocities, is done in almost every quadcopter inside the embedded
system and not through any external software, and there are two main reasons for it:
1. The control must be done at a high frequency rate to compensate the quick angular
dynamics of the quadcopter. These rates are seldom obtainable through wireless
communication protocols.
60
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
2. The low level control is not robust to the sort of delays introduced by wireless
links. If done within the embedded system, this latency is easily eliminated.
In order to resolve this issue, the elements of the time-invariant state feedback gain L
that corresponded to the angles and angular velocities states were implemented in the
embedded system while the elements corresponding to the position and linear velocities
of L and also the feedforward gains g [k] and Lg were implemented in MATLAB (see
Section 4.2.2).
After including the state feedback control in the Crazyflie’s firmware, the first tests of the
on-board attitude stabilizer were conducted using a calibration rig as seen in Figure 4.2.1.
Although it could only be used to test the feedback of the pitch angle, the roll angle
theoretically has similar dynamics and therefore the same tuned gains were used for both
pitch and roll.
Figure 4.2.1: Pitch angle calibration rig.
Even though there exists some non-negligible tension force induced by the cables that
attached the quadcopter’s body to the posts, this is a good first test trial to avoid
potential crashes. After tuning the gains of the roll and pitch angles the tests suggested
that integral action was needed to compensate for different sources of perturbation, such
as the asymmetry of the body and asymmetry of the motor’s power. While these effects
might also be compensated by the position integral action, the overall performance of the
control system was improved when adding the integral action to the angular positions.
The implemented architecture is exposed in the block diagram of Figure 4.2.2. On top
it shows the angular stabilization loop that runs in the embedded system at a faster rate
of 500Hz, while below is the rest of the LQT architecture running off-board at a slower
rate of 100 Hz.
61
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
On-Board @ 500Hz
Lang [k]
∆uang
Kiang
P
+
(eang ∆kang )
P
−
u
Crazyflie2.0
x̂ang [k]
upos +
Kipos
(epos ∆kpos )
+
∆u
+
+
Lpos [k]
x̂pos [k]
−
+
Lg [k]
g[k + 1]
ue
Off-Board @ 100Hz
Figure 4.2.2: Implementation diagram.
The sensor fusion algorithm calculation for the three Euler angles and the angular velocities coming from the gyroscope compose the state vector x̂ang [k] in the following
manner:
x̂ang =
h
ψ̂ θ̂ φ̂ r̂ q̂ p̂
iT
(4.2.1)
The gain Lang is a 4x6 matrix that multiplies accordingly the state vector x̂ang . The
system input ∆uang is a deviation from the equilibrium point, ue , to maintain an angle
equal to zero in all three Euler angles. The on-board controller runs at 500Hz, thus the
time step ∆kang in Figure 4.2.2 is equal to 0.002s.
The off-board section of the controller computes the state feedback for the position and
linear velocities states captured by the VICON or UWB systems and the Kalman filter
developed in Section 3.2.2, therefore composing the following state vector:
x̂pos =
h
x̂ ŷ ẑ û v̂ ŵ
iT
(4.2.2)
Matrix gain Lpos is also a 4x6 matrix that multiplies the state vector x̂pos . For the
position integral action, the time step ∆kpos is equal to 0.01s
4.2.1
ROS Controller Node Modifications
For the LQT implementation, the ROS controller node presented in Section 4.1.1 that
previously implemented the PID controller was modified to only receive the incoming
commands from the MATLAB interface and send them to the Crazyflie 2.0 through the
62
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
radio link. The Crazyflie server node was then slightly modified to send the motor PWM
signals instead of the desired Euler angles as it did with the previous control system.
4.2.2
MATLAB Interface Details
The MATLAB interface was the heart of the LQT algorithm implementation. The
ROS and MATLAB communication was done using the Robotics System ToolboxTM , in
particular the Simulink blocks that implement ROS publishers and subscribers. All of
the off-board section of the controller was executed within this interface, as well as the
state vector reconstruction and flight data retrieval for further analysis. Figure 4.2.3
exhibits the structure of the interface developed.
MATLAB/ROS Interface @ 100Hz
VICON/ [x̂, ŷ, ẑ]
UWB
ROS
Subscribers
x̂ang
[x̂, ŷ, ẑ]
Kalman
Filter
[ψ̂, θ̂, φ̂, r̂, q̂, p̂]
[m1 , m2 , m3 , m4 ]
x̂pos
LQT
Algorithm
(Off-board Section)
Crazyflie 2.0
upos
ROS
Publisher
[u1 , u2 , u3 , u4 ]
Figure 4.2.3: MATLAB Interface diagram to implement the LQT controller.
Now a breakdown of each block:
• VICON/UWB: block that gives the position estimations of the drone in a previously defined inertial frame.
• ROS Subscribers: these subscribers serve as bridge between MATLAB, the positioning system and the Crazyflie angle data. Four subscribers are included: one
for the VICON position estimations, one for the UWB system, another one for the
IMU data of the Crazyflie 2.0 and the last one to retrieve the PWM commands
sent to the motors.
• State observer: takes data from the position, Euler angles and angular velocities
of the quadcopter to calculate an estimation of the state vector x̂pos . This is done
through the Kalman Filter developed in Section 3.2.2.
• LQT algorithm: implements the off-board section of the LQT algorithm as suggested in Figure 4.2.2.
63
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
• ROS Publisher: takes the input vector upos and decomposes in its four components, in order to send it to the Crazyflie via radio link in the format [u1 , u2 , u3 , u4 ].
• Crazyflie 2.0: the drone takes the output message of the ROS publisher and adds
it to the on-board control section of the LQT algorithm, as suggested once again
in Figure 4.2.2. The platform continuously outputs the IMU readings as well as
the total 16-bit PWM signal sent to the motors.
4.2.3
Experimental Results
A series of trajectories created using the GUI developed in Section 3.2.4 were tested
with the LQT algorithm. The RMS error between the desired trajectory and the actual
trajectory followed by the drone was used as a measure of the controller’s performance
as in past works in UAV control such as [25, 28] . Another performance index was
defined, from the basis that a 10 cm error margin from the desired position is ideal,
then it was relevant to calculate for each one of the spatial coordinates of the drone the
percentage in which each coordinate stayed within this margin of the desired position.
These performance indices will be further known as ξx , ξy and ξz .
• Trajectory #1: a simple trajectory with fixed altitude of 1 meter was commanded
for the quadcopter to follow. The time plots in Figure 4.2.4 present the test results,
while Figure 4.2.5 show different 3D perspectives of the flight.
X Position
Position (m)
2
LQT
Reference
1
0
-1
-2
5
10
15
20
25
30
35
40
25
30
35
40
25
30
35
40
Time (s)
Y Position
Position (m)
2
1
0
-1
-2
5
10
15
20
Time (s)
Z Position
Position (m)
1.5
1
0.5
0
5
10
15
20
Time (s)
Figure 4.2.4: Position plots for Trajectory#1.
64
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
3D Trajectory
3D Trajectory
2
1.5
LQT
Reference
1.2
1
1
LQT
Reference
Y (m)
0.5
0
0.6
0.4
-0.5
2
0.2
0
-1
0
-0.2
2
1.5
1
0.5
0
-0.5
Y (m)
-1
-2
-1.5
-1.5
-1.5
X (m)
-1
-0.5
0
0.5
1
1.5
X (m)
(a) Standard view.
(b) XY Plane.
Figure 4.2.5: 3D Trajectory#1.
For this simple trajectory the RMS errors were low, 6.2cm for the X position, 6.3cm
for the Y and 3.9cm for the altitude Z. The corresponding performance indices were
ξx = 92.65%, ξy = 90.72% and ξz = 95.87%. The tracking of these type of slow
trajectories had an outstanding level of performance.
• Trajectory #2: for this second trajectory a varying altitude was also commanded
to asses the behavior of the quadcopter while moving simultaneously in all three
spatial coordinates. The flight data is displayed in Figures 4.2.6 and 4.2.7.
X Position
Position (m)
1
LQT
Reference
0
-1
-2
5
10
15
20
25
30
35
40
25
30
35
40
25
30
35
40
Time (s)
Y Position
Position (m)
1
0
-1
-2
5
10
15
20
Time (s)
Z Position
2
Position (m)
Z (m)
0.8
1.5
1
0.5
0
5
10
15
20
Time (s)
Figure 4.2.6: Position plots for Trajectory#2.
65
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
3D Trajectory
3D Trajectory
1
0.5
2
LQT
Reference
LQT
Reference
0
Y (m)
1
-0.5
-1
0.5
1
0
1
-1.5
0
-1
0
-1
-2
Y (m)
-2
-2
-1.5
X (m)
-1
-0.5
0
0.5
1
X (m)
(a) Standard view.
(b) XY Plane.
Figure 4.2.7: 3D Trajectory#2.
Despise adding more difficulty to the trajectory the performance was satisfactory, with
RMS errors of 4.72cm, 7.81cm and 4.75cm respectively for the x, y and z positions. The
performance indices remained in the high end, with ξx = 93.89%, ξy = 87.30% and
ξz = 94.63%.
• Trajectory #3: a more complex trajectory is showcased in Figures 4.2.8 and 4.2.9
X Position
Position (m)
2
LQT
Reference
1
0
-1
-2
5
10
15
20
25
30
35
40
25
30
35
40
25
30
35
40
Time (s)
Y Position
Position (m)
2
1
0
-1
-2
5
10
15
20
Time (s)
Z Position
2
Position (m)
Z (m)
1.5
1.5
1
0.5
0
5
10
15
20
Time (s)
Figure 4.2.8: Position plots for Trajectory#3.
66
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
3D Trajectory
3D Trajectory
1.5
2
1
0.5
Y (m)
LQT
Reference
1
LQT
Reference
0
-0.5
0.5
-1
0
-2
-1
0
1
Y (m)
2
-1
0
1
2
-2
-1.5
-1.5
X (m)
-1
-0.5
0
0.5
1
1.5
X (m)
(a) Standard view.
(b) XY Plane.
Figure 4.2.9: 3D Trajectory#3.
Even though the controller kept a good trajectory tracking, it is evident how the performance starts degrading when adding more complexity to the trajectories. In this case
the RMS erros were of 7.51cm, 5.97cm and 7.34cm for the X, Y and Z positions. The
performance indices were lower than in the two previous trajectories, with ξx = 83.47%,
ξy = 88.49% and ξz = 85.12%.
• Trajectory #4: a complex spiral trajectory in 3D was commanded. The experimental results can be appreciated in Figures 4.2.10 and 4.2.11.
X Position
Position (m)
2
LQT
Reference
1
0
-1
5
10
15
20
25
30
35
40
25
30
35
40
25
30
35
40
Time (s)
Y Position
Position (m)
2
1
0
-1
-2
5
10
15
20
Time (s)
Z Position
2
Position (m)
Z (m)
1.5
1.5
1
0.5
0
5
10
15
20
Time (s)
Figure 4.2.10: Position plots for Trajectory#4.
67
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
3D Trajectory
3D Trajectory
1.5
LQT
Reference
2
1
1.5
0.5
1
0
0.5
Y (m)
Z (m)
LQT
Reference
0
-0.5
2
-1
1
0
-1.5
-1
Y (m)
-2
-1
-0.5
1
0.5
0
1.5
2
-2
-1
-0.5
0
X (m)
0.5
1
1.5
2
X (m)
(a) Standard view.
(b) XY Plane.
Figure 4.2.11: 3D Trajectory#4.
The RMS error incurred when following this trajectory was of 7.93cm in X, 11.79cm
in Y and 5.44cm in Z. The performance indices were ξx = 79.55%, ξy = 58.19% and
ξz = 93.12%. Even more clear than with Trajectory#3, this spiral trajectory shows a
performance degradation when the trajectories demand faster movements, closer curves
or harder brakes. For example, at the 30 second mark in the X position a sudden brake
was required but the controller was not as fast which caused an overshoot. As seen in
the simulation phase these type of overshoots are caused mainly by the integral action,
but otherwise, if lowered, in practice the position regulation would worsen.
4.3
Controller Comparisons
In this section the step response of the control system is used as a measure to compare
the performance of the two controllers synthesized in this project, as well as comparing
the simulated model and the experimental data to determine how close the simulation
predicted the actual test results. Then, to compare trajectory tracking capabilities, the
two controllers were tested on identical trials to follow sinusoidal waves.
4.3.1
Simulation vs Experimental - PID
Comparing the step response of the simulation model developed in subsection 3.1 and
the one obtained during a real flight of the drone was the chosen method to verify the
accuracy of the mathematical model of the quadcopter. Three different flights were executed, sending individual steps in the direction of X, Y and Z.
68
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
• Step command in the X direction
A step command of 1 meter was commanded in the X coordinate, while maintaining a
1 meter altitude and a zero position in the Y coordinate. The experimental results are
presented in Figure 4.3.1.
X Position
1.5
Simulation
Experimental
Reference
Position (m)
1
0.5
0
-0.5
10
15
20
25
30
35
25
30
35
25
30
35
Time (s)
Y Position
Position (m)
0.1
0
-0.1
-0.2
10
15
20
Time (s)
Z Position
Position (m)
1.2
1
0.8
0.6
10
15
20
Time (s)
Figure 4.3.1: Trajectory using the PID controller to follow a Step in the X position.
The simulation and experimental step responses had almost identical response time of
around 3 seconds with almost to no overshoot. The simulation response shows less
damping in the response. The clear difference is in the Y response, the simulation
predicted a perfect zero which is clearly impossible in practice, nonetheless the drone
maintained a 10 centimeters error margin within the initial Y position. The movement
had little impact in the altitude.
• Step command in the Y direction
The test was conducted in a similar fashion as the previous one, but this time sending
the appropriate command to the Y coordinate. Simulation and experimental results were
plotted together as show in Figure 4.3.2.
69
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
X Position
Position (m)
0.1
0
-0.1
-0.2
10
15
20
25
30
35
Time (s)
Y Position
1.5
Simulation
Experimental
Reference
Position (m)
1
0.5
0
-0.5
10
15
20
25
30
35
25
30
35
Time (s)
Z Position
Position (m)
1.2
1
0.8
0.6
10
15
20
Time (s)
Figure 4.3.2: Trajectory using the PID controller to follow a Step in the Y position.
Similar to the X step response, the dynamics for the Y direction matches up to some
extent those predicted by the simulation.
• Step command in the Z direction
Finally a step command of 1 meter in altitude was tested, as seen in Figure 4.3.3.
X Position
Position (m)
0.05
0
-0.05
-0.1
-0.15
10
15
20
25
30
35
25
30
35
Time (s)
Y Position
Position (m)
0.2
0.1
0
-0.1
-0.2
10
15
20
Time (s)
Z Position
Position (m)
2.5
Simulation
Experimental
Reference
2
1.5
1
0.5
10
15
20
25
30
Time (s)
Figure 4.3.3: Trajectory using the PID controller to follow a Step in the Z position.
70
35
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
The altitude dynamics were slower in practice than in simulation and with a the experimental data exposes a more pronounced overshoot.
4.3.2
Simulation vs Experimental - LQT
Similarly, with the simulation model developed in Section 3.2, the following plots show
the step response comparisons between the simulation and the experimental data.
• Step command in the X direction
Figure 4.3.4 exhibits a comparative plot between the simulation and the experimental
data.
X Position
1.5
Simulation
Experimental
Reference
Position (m)
1
0.5
0
-0.5
5
10
15
20
25
30
35
40
25
30
35
40
25
30
35
40
Time (s)
Y Position
Position (m)
0.1
0
-0.1
-0.2
5
10
15
20
Time (s)
Z Position
Position (m)
1.5
1
0.5
0
5
10
15
20
Time (s)
Figure 4.3.4: Trajectory using the LQT controller to follow a Step in the X position.
The simulation scenario in the X and Z positions matches the results obtained during
the experiment. The main difference is the perturbation around the zero position of the
Y coordinate, the simulation predicted an almost perfect hold of this position while in
reality the quadcopter oscillated in a 10 cm error margin.
• Step command in the Y direction
Doing a similar test, but sending a command to the Y coordinate gave the results seen
in Figure 4.3.5.
71
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
X Position
Position (m)
0.2
Simulation
Experimental
Reference
0.1
0
-0.1
5
10
15
20
25
30
35
40
25
30
35
40
25
30
35
40
Time (s)
Y Position
Position (m)
2
1
0
-1
5
10
15
20
Time (s)
Z Position
Position (m)
1.5
1
0.5
0
5
10
15
20
Time (s)
Figure 4.3.5: Trajectory using the LQT controller to follow a Step in the Y position.
The dynamics were similar to those of the X step, the simulation proved once again to
be accurate in predicting the test results.
• Step command in the Z direction
The behavior in Figure 4.3.6 corresponds with a 1 meter command in the Z coordinate.
X Position
Position (m)
0.3
0.2
Simulation
Experimental
Reference
0.1
0
-0.1
5
10
15
20
25
30
35
40
25
30
35
40
25
30
35
40
Time (s)
Y Position
Position (m)
0.2
0.1
0
-0.1
-0.2
5
10
15
20
Time (s)
Z Position
Position (m)
3
2
1
0
5
10
15
20
Time (s)
Figure 4.3.6: Trajectory using the LQT controller to follow a Step in the Z position.
72
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
The altitude step simulation matched almost perfectly the one obtained in practice, as
the comparison demonstrate.
After reviewing this data a few conclusions can be drawn. On the first hand, the PID
simulation is precise up to some extent at predicting the response of the system when
applying step commands, although it has some considerable differences on the altitude
estimation. On the other hand the LQT simulation proved to be more precise, giving
accurate information of how the system will behave in the real scenario. Both of these
simulations serve the common purpose of validating the mathematical model of the
quadcopter.
In the LQT simulations there is an important remark to be made before jumping to
an early conclusion about the accuracy of the mathematical model. In the X-Y step
responses there is an overshoot introduced mainly by the high integral gains Kiang and
Kipos , if those gains were to be lowered at least in simulation the overshoot would be
reduced giving more pleasant results, however various experimental results suggest that
lowering these gains will worsen the overall performance of the controller in practice.
The conclusion is that the high integral gain is necessary to compensate all the model
deficiencies and future efforts should be made in refining the model in order to lower the
integral gains without loosing performance.
4.3.3
Controller Performance: PID vs LQT
One of the main goals of the project was to objectively compare the performance of the
two controllers synthesized, in terms of tracking accuracy and command effort. First
the step responses were compared and then a sinusoidal trajectory showed the tracking
capabilities of each controller. For each controller, the results presented correspond to
the flight with the best performance after performing a series of trials under the same
conditions (gains, weights, etc.). The same performance indices as in Section 4.2.3 were
used. To quantify the control effort, the two-norm is used as in [29], then the control
effort is defined as:
k=kf
Ui =
X
u2i [k]
(4.3.1)
k=0
where ui [k] is the PWM signal ranging from 0-65536 sent to the i-th motor, and Ui is the
U
−UP ID
associated control effort. Also, the ratio percentage % LQT
is used to quantify
UP ID
73
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
the increase or decrease percentage in control effort of the LQT with respect to the PID.
• Step command in the X direction
Figure 4.3.7 exhibits the experimental results for both controllers when commanded to
follow a 1 meter step in the X coordinate.
X Position
Y Position
0.1
1.2
1
0.05
Position (m)
Position (m)
0.8
0.6
0.4
0.2
0
-0.05
-0.1
0
-0.2
-0.15
10
15
20
25
30
35
10
Time (s)
15
X Position Error
30
35
30
35
Y Position Error
0.15
1
0.1
0.5
0.05
Error (m)
Error (m)
25
Time (s)
1.5
0
-0.5
-1
10
20
PID
LQT
Reference
0
-0.05
-0.1
15
20
25
30
35
10
15
Time (s)
20
25
Time (s)
Figure 4.3.7: X-Y Position and error comparison when following a unit step in the X
position.
The error terms for the LQT controller are lower, at the expenditure of a higher overshoot
than the PID. By moving before the step command, the LQT controller manages to
reduce an otherwise big error of 1 meter. Table 4.3.1 summarizes the performance for
both controllers.
RMS (ex ) [cm]
RMS (ey ) [cm]
%ξx
%ξy
LQT
12.58
3.37
76.53
100
PID
24.63
6.41
74.98
95.08
Table 4.3.1: Error comparison when following a unit step in X position.
The RMS error values for the X-Y position with the LQT controller were almost half
of those obtained with the PID. The performance indices were slightly better with the
LQT controller.
74
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
The motor commands data for both trials were registered and compared as shown in
Figure 4.3.8.
7
×10 4
PWM Commands Motor M1
6
×10 4
PWM Commands Motor M2
5.5
6
PWM
PWM
5
5
4
4.5
4
3.5
3
3
2
10
15
20
25
30
2.5
10
35
15
20
Time (s)
25
30
35
30
35
Time (s)
PID
LQT
6
×10 4
PWM Commands Motor M3
6
PWM
PWM
PWM Commands Motor M4
5
5
4
3
2
10
×10 4
4
3
2
15
20
25
30
1
10
35
15
Time (s)
20
25
Time (s)
Figure 4.3.8: Motor commands comparison when following a unit step in the X position.
The control effort is overall greater with the LQT controller, but the PID has control
effort discontinuities when the step command goes into action. Around the 32 second
mark the motor M1 reached saturation using the PID controller. Furthermore, Table 4.3.2 quantifies the motor’s command effort in both cases.
LQT
PID
%
ULQT −UP ID
UP ID
U1 [×1012 ]
U2 [×1012 ]
U3 [×1012 ]
U4 [×1012 ]
5.26
4.66
5.86
4.40
4.66
3.87
5.42
3.46
12.88%
20.41%
8.12%
27.17%
Table 4.3.2: Motor effort comparison when following a unit step in the X position.
The results show a clear tendency of control effort increase while using the LQT controller
with respect to the PID controller.
• Step command in the Y direction
When commanded to follow a 1 meter step in the Y coordinate, the quadcopter behaved
as suggested in Figure 4.3.9.
75
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
X Position
Y Position
0.1
1.5
1
Position (m)
Position (m)
0.05
0
-0.05
0.5
0
-0.1
-0.15
-0.5
10
15
20
25
30
35
10
Time (s)
X Position Error
0.1
1
0.05
0.5
-0.05
25
30
35
30
35
Y Position Error
1.5
0
20
Time (s)
0.15
Error (m)
Error (m)
15
LQT
PID
Reference
0
-0.5
-0.1
10
15
20
25
30
-1
10
35
15
Time (s)
20
25
Time (s)
Figure 4.3.9: X-Y Position and error comparison when following a unit step in the Y
position.
Similar as the last test, the anticipatory feature of the LQT controller allows to reduce
the big error the system incurs when sending step commands. The graphical evidence is
supported by the error comparison contained in Table 4.3.3.
RMS (ex ) [cm]
RMS (ey ) [cm]
%ξx
%ξy
LQT
2.53
12.25
100
78.34
PID
3.50
28.39
99.68
65.69
Table 4.3.3: Error comparison when following a unit step in Y position.
Once again, the RMS error in the direction the step was send was cut by more than half
with the LQT controller. Performance indices show that the LQT was superior in terms
of reducing the tracking error.
As for the motor commands, Figure 4.3.10 show the same trend as the last test, the LQT
algorithm control outputs is overall bigger than the control outputs of the PID system,
while this last one displays discontinuities in the commands sent as a consequence of the
step demanded.
76
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
5.5
×10 4
PWM Commands Motor M1
5
5
×10 4
PWM Commands Motor M2
4.5
PWM
PWM
4.5
4
4
3.5
3.5
3
3
2.5
10
15
20
25
30
2.5
10
35
Time (s)
6
×10 4
PWM Commands Motor M3
5
25
30
35
×10 4
30
35
PWM Commands Motor M4
4
PWM
PWM
20
Time (s)
5
4
3
2
10
15
PID
LQT
3
2
1
15
20
25
30
0
10
35
15
Time (s)
20
25
Time (s)
Figure 4.3.10: Motor commands comparison when following a unit step in the Y position.
The data presented in Table 4.3.4 indicate that the control effort increment of the LQT
algorithm with respect to the PID controller can get up to more than a 50%.
LQT
PID
%
ULQT −UP ID
UP ID
U1 [×1012 ]
U2 [×1012 ]
U3 [×1012 ]
U4 [×1012 ]
5.66
5.13
6.35
4.88
4.89
3.71
5.49
3.23
15.75%
38.28%
15.66%
51.08%
Table 4.3.4: Motor effort comparison when following a unit step in the Y position.
• Step command in the Z direction
The last comparison for the linear trajectories was done sending a 1 meter step command
in the Z coordinate. The test result flight data is exposed in Figures 4.3.11 and 4.3.12,
where all three coordinate comparisons between the two controllers are illustrated.
The LQT algorithm outperforms by a good margin the PID, by greatly reducing the
overshoot and the response time in the Z position. The LQT also kept a more precise
X-Y position, both at a constant altitude and when applying the command to ascend 1
meter.
77
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
Z Position
Position (m)
2.5
PID
LQT
Reference
2
1.5
1
0.5
10
15
20
25
30
35
30
35
Time (s)
Z Position Error
2
Error (m)
1
0
-1
-2
10
15
20
25
Time (s)
Figure 4.3.11: Z Position and error comparison when following a unit step.
X Position
Y Position
0.1
0.2
0.15
0.05
Position (m)
Position (m)
0.1
0
-0.05
0.05
0
-0.05
-0.1
-0.1
-0.15
-0.15
10
15
20
25
30
35
10
Time (s)
15
20
25
30
35
30
35
Time (s)
PID
LQT
Reference
X Position Error
Y Position Error
0.15
0.15
0.1
0.1
0.05
Error (m)
Error (m)
0.05
0
0
-0.05
-0.1
-0.05
-0.15
-0.1
-0.2
10
15
20
25
30
35
10
Time (s)
15
20
25
Time (s)
Figure 4.3.12: X-Y Position and error comparison when following a unit step in the Z
position.
The analysis of the results is summarized in Table 4.3.5. The error comparison confirms
the superiority of the LQT with respect to the PID, with reduced RMS values of error
and better performance indices.
78
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
RMS (ex ) [cm]
RMS (ey ) [cm]
RMS (ez ) [cm]
%ξx
%ξy
%ξz
LQT
2.82
4.35
13.38
100
99.48
81.48
PID
6.13
5.46
28.59
86.43
91.83
52.68
Table 4.3.5: Error comparison when following a unit step in Z position.
The control commands of the motors are visualized in Figure 4.3.13. As in all the other
tests, the LQT control effort remained greater than that of the PID. However the LQT
plots do not present evident command peaks as the PID when sending the altitude
command.
5.5
×10 4
PWM Commands Motor M1
5.5
5
×10 4
PWM Commands Motor M2
5
PWM
PWM
4.5
4
4.5
4
3.5
3.5
3
2.5
10
15
20
25
30
3
10
35
15
20
Time (s)
25
30
35
30
35
Time (s)
PID
LQT
5.5
×10 4
PWM Commands Motor M3
5
4
4
3.5
3.5
3
10
PWM Commands Motor M4
4.5
4.5
PWM
PWM
5
×10 4
15
20
25
30
3
10
35
15
Time (s)
20
25
Time (s)
Figure 4.3.13: Motor commands comparison when following a unit step in the Z position.
The increase in control effort is exposed by the results in Table 4.3.6. The numbers
reveal a maximum increase of around 31% in control effort in one of the motors with the
LQT controller with respect to the PID.
LQT
PID
%
ULQT −UP ID
UP ID
U1 [×1012 ]
U2 [×1012 ]
U3 [×1012 ]
U4 [×1012 ]
5.32
4.42
5.78
4.13
4.05
4.18
4.98
3.73
31.36%
5.74%
16.06%
10.72%
Table 4.3.6: Motor effort comparison when following a unit step in the Z position.
79
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
• Circular trajectory
A circular trajectory composing a circle in the X-Y plane was conducted with both
controllers to test the tracking of rapidly varying trajectories. The time plot comparisons
of the X-Y positions are displayed in Figure 4.3.14.
X Position
1
1
0.5
0.5
0
-0.5
0
-0.5
-1
-1.5
-1
5
10
15
20
25
30
Time (s)
35
40
45
50
-1.5
55
10
15
20
25
30
Time (s)
35
40
45
50
55
40
45
50
55
Y Position Error
1.5
1
Error (m)
0.5
Error (m)
5
PID
LQT
Reference
X Position Error
1
0
-0.5
-1
Y Position
1.5
Position (m)
Position (m)
1.5
0.5
0
-0.5
5
10
15
20
25
30
Time (s)
35
40
45
50
-1
55
5
10
15
20
25
30
Time (s)
35
Figure 4.3.14: X-Y Position and error comparison when following a circular trajectory.
The LQT controller manages to track and stay in phase with the sinusoidal waves, while
the PID controller was not capable of such feat nor of obtaining the desired 1 meter
amplitude. The error plots expose the improvement in trajectory tracking of the LQT
controller with respect to the PID, and are validated by the data shown in Table 4.3.7.
RMS (ex ) [cm]
RMS (ey ) [cm]
%ξx
%ξy
LQT
10.32
16.69
55.74
55.00
PID
46.05
47.28
14.72
17.68
Table 4.3.7: Error comparison when following a circular position.
The RMS error is about 4 times greater with the PID controller than with the LQT algorithm, clearly showing the superiority of the latter in tracking more complex trajectories.
The motors’ 16-bit PWM signals of both controllers are compared in Figure 4.3.15.
80
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
6
×10 4
PWM Commands Motor M1
6
5
4.5
4.5
4
3.5
4
3.5
3
3
2.5
2.5
10
15
20
25
30
35
40
45
50
55
10
Time (s)
×10 4
15
20
25
PWM Commands Motor M3
6
6
5
5
4
4
30
35
40
45
50
55
45
50
55
Time (s)
PID
LQT
PWM
PWM
7
PWM Commands Motor M2
5.5
5
PWM
PWM
5.5
×10 4
3
×10 4
PWM Commands Motor M4
3
2
2
1
10
15
20
25
30
35
40
45
50
55
10
15
20
25
Time (s)
30
35
40
Time (s)
Figure 4.3.15: Motor command comparison when following a circular trajectory.
The tendency of a greater command effort for the LQT algorithm compared to the
PID remained as before, also confirmed by the values in Table 4.3.8. Note that for
this trajectory there are important control spikes with the PID controller exactly in the
points where the error reached its maximum points in Figure 4.3.14, causing for instance
a motor saturation at the 28 second mark in motor M3.
LQT
PID
%
ULQT −UP ID
UP ID
U1 [×1012 ]
U2 [×1012 ]
U3 [×1012 ]
U4 [×1012 ]
11.68
10.20
12.50
9.65
8.82
7.69
10.28
6.88
32.43%
32.64%
21.60%
40.26%
Table 4.3.8: Motor effort comparison when following a circular trajectory.
The comparison between the PID and LQT control systems gave insightful data to draw
conclusions about their advantages and disadvantages. Starting with the overall performance in position and trajectory tracking, the LQT was superior at keeping low levels
of error with respect to the desired position. The goal of improving the tracking performance of the PID controller was achieved with the LQT algorithm.
As for the command effort of the motors, even though the LQT algorithm does an
optimization process to minimize it, the experimental data exposed that the PID used less
81
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
effort in every trial. The explanation for this phenomenon are the elevated integral gains
and weighting factors used with the LQT algorithm to obtain the desired performance.
It is intuitive to think that the correct tracking of more demanding trajectories leads to a
greater control effort, as for most control systems design this ends up being a compromise
issue, in this case between performance and motor power. This increase in control effort
translated in a shorter battery life-time while using the LQT controller with respect to
the PID.
4.4
LQT with UWB Position Estimation
Using the two-way ranging ultra-wide band system developed in [36], four base stations
were placed forming a 7x4m rectangle at 2.5m of height from the floor. Figure 4.4.1
displays how the tag was incorporated to the quadcopter’s body.
Figure 4.4.1: Crazyflie 2.0 with UWB module exposed
A series of trajectories were followed using the VICON and then the UWB system to
test the LQT controller. In both cases the flight data was compared using the VICON
measurements as the ground truth in order to study the tracking performance. The
UWB estimations were used only for the X and Y position, while the altitude came
from the VICON system. The analysis was focused to the tracking in the X-Y plane,
while keeping a 1 meter altitude. With each system a certain amount of flights were
executed and the one with best performance is the one showcased in this section, so the
comparisons were made in the best-case scenario with each localization system. Note:
while using the UWB system it was decided to lower the position integral gains as it
improved the performance.
82
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
• Hover
The first test was a simple hover to asses the controller’s capability of keeping a fixed
position. Figure 4.4.2 presents the experimental results for the hover test.
X Position
Position (m)
0.2
0
-0.2
-0.4
5
10
15
20
25
30
35
Time (s)
Y Position
0.2
Position (m)
40
UWB
VICON
Reference
0
-0.2
-0.4
5
10
15
20
25
30
35
40
Time (s)
Figure 4.4.2: X-Y Position and error comparison while hovering around a point.
Although in both cases the controller performed similarly, with the UWB system there
were more oscillations, specially in the take-off and landing stages at the beginning and
end of the time plot. Table 4.4.1 quantifies the error values of the hover flight.
RMS (ex ) [cm]
RMS (ey ) [cm]
%ξx
%ξy
VICON
4.67
5.03
93.74
94.78
UWB
5.90
6.42
92.15
90.27
Table 4.4.1: Error comparison while hovering around a point.
Both hover flights were solid, although the RMS errors and performance indices suggests
that with the VICON system the drone hold more precisely its position.
• Trajectory #1
The results for the first trajectory are displayed in Figure 4.4.3.
83
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
X Position
Y Position
1.5
1.5
1
1
Position (m)
Position (m)
0.5
0.5
0
0
-0.5
-1
-0.5
-1.5
-1
-2
5
10
15
20
25
30
35
40
5
Time (s)
X Position Error
15
20
25
30
35
40
30
35
40
Time (s)
Y Position Error
0.3
0.4
0.2
0.3
0.2
Error (m)
0.1
Error (m)
10
UWB
VICON
Reference
0
-0.1
0.1
0
-0.1
-0.2
-0.2
-0.3
-0.3
5
10
15
20
25
30
35
40
5
10
15
20
Time (s)
25
Time (s)
Figure 4.4.3: X-Y Position and error comparison when following Trajectory #1.
Once again, the overall performance in both cases in terms of the errors was similar, but
the greater levels of noise of the UWB system with respect to the VICON is translated
into more oscillations in the position. The 3D perspectives in Figure 4.4.4 show the
smoothness of the system with the VICON with respect to the more oscillating trajectory
with the UWB.
3D Trajectory
3D Trajectory
1.5
1
1.2
1
UWB
VICON
Reference
0.5
0
0.6
Y (m)
Z (m)
0.8
UWB
VICON
Reference
0.4
0.2
-0.5
-1
0
-0.2
2
2
-1.5
1
1
0
0
-1
Y (m)
-2
-1
-2
-1
X (m)
-0.5
0
0.5
1
1.5
X (m)
(a) Standard view.
(b) XY Plane.
Figure 4.4.4: Comparison of 3D Trajectory#1.
As for the performance, the summary in Table 4.4.2 indicates that the overall performance with the VICON system is better, but nonetheless with the UWB system the
84
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
controller manages to follow the trajectory with RMS errors around 10 centimeters.
RMS (ex ) [cm]
RMS (ey ) [cm]
%ξx
%ξy
VICON
9.16
7.55
73.69
86.39
UWB
10.22
7.82
63.44
79.85
Table 4.4.2: Error comparison while following Trajectory #1.
• Trajectory #2
While following a second trajectory, similar in complexity as the first one, the results
presented in Figure 4.4.5 illustrate the tracking capabilities of the LQT controller in
both cases, despite the higher levels of noise when using the UWB system. The error
plots compare the smooth lines in the VICON flight with the more oscillating ones of
the UWB.
X Position
Y Position
2
2
1.5
1
Position (m)
Position (m)
1
0.5
0
-0.5
0
-1
-1
-1.5
-2
5
10
15
20
25
30
35
40
5
10
20
25
30
35
40
30
35
40
Time (s)
X Position Error
Y Position Error
0.4
0.4
0.2
0.2
0
0
Error (m)
Error (m)
15
UWB
VICON
Reference
Time (s)
-0.2
-0.4
-0.2
-0.4
-0.6
-0.6
5
10
15
20
25
30
35
40
5
Time (s)
10
15
20
25
Time (s)
Figure 4.4.5: X-Y Position and error comparison when following Trajectory #2.
The 3D perspectives in Figure 4.4.6 illustrate the trajectory followed by the quadcopter
in both cases, and specially the X-Y plane view shows the main difference between the
two flights: the smoothness and stability with which the quadcopter managed to follow
the desired trajectory.
85
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
3D Trajectory
3D Trajectory
2
UWB
VICON
Reference
1.5
1.2
1
1
0.5
Y (m)
Z (m)
0.8
0.6
UWB
VICON
Reference
0.4
0.2
0
-0.5
-1
0
-0.2
2
2
-1.5
0
1
0
-1
-2
Y (m)
-2
-2
-1.5
X (m)
-1
-0.5
0
0.5
1
1.5
2
X (m)
(a) Standard view.
(b) XY Plane.
Figure 4.4.6: Comparison of 3D Trajectory#2.
Table 4.4.3 summarizes the performance in trajectory tracking. The UWB system incurred in a greater RMS error than the VICON, around 3.4cm more for the X position
and 0.6cm for the Y position. The performance indices ξx and ξy were inferior while
using the UWB system, but still close enough to be acceptable.
RMS (ex ) [cm]
RMS (ey ) [cm]
%ξx
%ξy
VICON
7.71
9.50
82.08
73.17
UWB
11.16
10.15
71.04
70.32
Table 4.4.3: Error comparison while following Trajectory #2.
• Trajectory #3
The third trajectory time plots and 3D path are exposed in figures Figures 4.4.7 and 4.4.8.
A similar behavior as before is appreciated, both flights had some level of success in following the commanded trajectory, but the error plots suggests a greater noise in the
position while using the UWB system. Nonetheless, the top view in Figure 4.4.8b illustrate the trajectory tracking capability in both flights.
The position errors reflected in Table 4.4.4 suggest a similar performance in the X position, with a more pronounced discrepancy in the Y position, in both cases being the
VICON was the more precise system.
86
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
X Position
Y Position
1.5
1.5
1
0.5
0.5
Position (m)
Position (m)
1
0
-0.5
-1
0
-0.5
-1
-1.5
-1.5
-2
5
10
15
20
25
30
35
40
5
Time (s)
10
15
X Position Error
25
30
35
40
30
35
40
Y Position Error
0.2
0.3
0.1
0.2
0
0.1
Error (m)
Error (m)
20
Time (s)
UWB
VICON
Reference
-0.1
-0.2
-0.3
0
-0.1
-0.2
-0.4
-0.3
5
10
15
20
25
30
35
40
5
10
15
Time (s)
20
25
Time (s)
Figure 4.4.7: X-Y Position and error comparison when following Trajectory #3.
3D Trajectory
3D Trajectory
1.5
1
1.2
UWB
VICON
Reference
1
0.5
0
0.6
Y (m)
Z (m)
0.8
0.4
UWB
VICON
Reference
0.2
-1
0
-0.2
2
-1.5
2
1
-0.5
1
0
0
-1
Y (m)
-1
-2
-2
-2
-1.5
X (m)
-1
-0.5
0
0.5
1
1.5
X (m)
(a) Standard view.
(b) XY Plane.
Figure 4.4.8: Comparison of 3D Trajectory#3.
RMS (ex ) [cm]
RMS (ey ) [cm]
%ξx
%ξy
VICON
7.56
7.29
84.69
83.00
UWB
9.43
8.11
74.68
78.66
Table 4.4.4: Error comparison while following Trajectory #3.
The experimental data confirms the robustness of the control system while using a positioning system with around 100 times greater standard deviation noise than the initial
87
CHAPTER 4. HARDWARE IMPLEMENTATION AND EXPERIMENTAL RESULTS
system. The tuning of the Kalman Filter to adapt to this new source of noise was vital
to ensure a good compromise between filtering and precision in the estimations. Even
though there is a clear advantage on using more precise technology, such as the VICON,
the system could still track the desired trajectories with an acceptable degree of precision
while using a more cheaper system such as the UWB.
The fact that the position integral gains for the X and Y positions had to be lowered
while using the UWB system, was a necessary compromise to ensure less oscillations
while following the desired trajectory. Such a compromise did not arise while using the
VICON system, with which the quadcopter remained stable and followed smoothly the
trajectories for a wide range of gain values.
If the control system could only use the UWB system, then a more detailed study of how
to compensate the different sources of noise and biases should be made. For instance,
the UWB system looses precision when the tag is close to one of the anchors, or if the
tag is facing away from one of the anchors. All this subtleties, if taken in account while
designing the control system, could lead to a better performance than the one obtained
and presented in this work.
Finally, the video found in [38] shows a summary of the project’s simulation and experimental results.
88
Chapter 5
Conclusions and Future Work
The study was set out to explore the dynamics of an open source nanoquadcopter named
Crazyflie 2.0, as well as creating a simulation environment for control design and then
testing it in the real platform. This type of unmanned aerial vehicles is becoming the
preferred platform for testing control algorithms of diverse natures, thus the inherent
importance of conceiving a mathematical model of the vehicle that can predict, up to
some extent, how the system will evolve over time. Hence, the project started by a
modeling of the nanoquadcopter and an identification of certain physical parameters,
based in previous work. Working in parallel with the literature and the quadcopter’s
embedded firmware was the main key in describing the system behavior just as it is in
the real platform, an important milestone for future work as the dynamics of a system
is the heart of every simulation environment.
The second phase of the project was building the simulation that served as the first testbench of the control architectures proposed. Using both the non-linear dynamics and the
linearised state space realisation of the system, the simulation created is a solid testing
environment to conceive all types of control systems. It was incredibly useful during the
first stages of the project to get a better understanding of how the system worked. In
addition, the simulation was used for designing both the PID position controller and the
LQT trajectory tracker.
An important conclusion is that the initial belief that all dynamics were decoupled as
suggested during the linear modeling was not entirely true in the non-linear system. As
observed in the simulations, there exists some interference between movements that, for
instance, does not allow the quadcopter to describe a perfectly straight line trajectory
89
CHAPTER 5. CONCLUSIONS AND FUTURE WORK
when there are more than one movement involved (a yaw rotation for example).
The position PID tracker was tested for time-varying trajectories, such as circles and
helices. Even though the system could described these trajectories, there were some
drawbacks and performance issues, for example not getting fast enough to the desired
points which lead to errors in the desired trajectory. The fact that the task at hand was
managed by a position tracker and not a trajectory tracker was the main reason of these
discrepancies. To address the deficiencies of the PID controller, a new control system
was conceived using the LQT algorithm, which proved to have interesting characteristics
while following step responses, mainly that it started moving before the command was
asked in order to reduce the tracking error. The feature was possible thanks to the
off-line calculation of the algorithm and the knowledge of the trajectory beforehand.
The comparisons between the PID and the LQT controller indicate a clear superiority
of the LQT in terms of reducing the trajectory tracking error, specially in the more
demanding trajectories, in which the LQT algorithm reduced up to 4 times the RMS
errors obtained with the PID controller. Directly related to the better tracking, the
LQT incurred in higher levels of control effort than the PID, but it also eliminated the
great command peaks seen in the motor time plots of the PID, thus getting rid of the
undesired motor saturations that could lead to unstable states.
There are two main drawbacks of the LQT algorithm with respect to the PID: the first
one is the inability to specify trajectories for the heading (yaw angle) and the second
one is the need to know the trajectory before its execution. Taking in account these
shortcomings, it is proposed as future work for this research to incorporate a method to
control the yaw angle while keeping the good performance in the LQT algorithm, the
author proposes a gain-scheduling method being the yaw angle the scheduling variable
as a possible solution for this problem. As for the second drawback of the LQT algorithm, more research should be directed towards an on-line implementation thus making
the controller useful in more complex tasks such as planning and execution missions in
real-time.
The GUI created for trajectory generation proved to be a valuable asset to quickly test
different types of trajectories, with varying difficulty. But the tool can be improved by
90
CHAPTER 5. CONCLUSIONS AND FUTURE WORK
adding physical constraints to the trajectory generation, as to assure the trajectory is
feasible for the quadcopter to follow. Future work in this area should explore feasible
trajectory generation as proposed in works such as [25].
The simulation versus experimental comparative time plots show that the simulation
environment developed in this project was accurate to some extent, serving its purpose
as a useful design tool for the controllers synthesized, but it had its limitations mainly
due to unmodeled phenomena, which lead to the need of introducing high integral gains
in the controllers to compensate the model errors and other perturbations of the system.
As future work, it is suggested a more thorough model identification for the quadcopter,
for example using numerical methods such as the closed-loop "black box" identification
proposed in [11].
The Kalman Filter approach for estimating the linear velocities from the position data
proved to be successful using both the VICON and the UWB, specially with the latter in
which the data had 100 times greater standard deviation noise. The VICON versus UWB
experiments suggest that in both cases the LQT tracked the desired position, but with
obvious different levels of smoothness and precision. Even though both performances
were satisfactory in terms of the scope of this work, future research into improving the
control system while using the UWB position system would be ideal. Starting from an
identification of different sources of added noise and biases of the UWB system, upto
different filtering techniques that are more appropriate than the classic Kalman filter
proposed in this project are the author’s recommendations to improve the control system performance.
This work represents a solid base for future research using this platform, with enough
explanation in the calculus for newcomers in the area to understand the basic functioning
of the system. The simulation environment was developed in a fashion that corresponds
exactly with the equations shown in the mathematical model, which helps in the quick
understanding of how everything works and saves time in comprehending an otherwise
complex system, plus it is easily customizable for future users to develop their own controllers. The project successfully fulfilled its ultimate goal of characterizing the provided
quadcopter platform and doing all the steps needed to develop an efficient control system
for trajectory tracking.
91
Bibliography
[1] Hanna, W. (2014). Modelling and control of an unmanned aerial vehicle (B.Eng
Thesis, Charles Darwin University).
[2] Subramanian, G. P. (2015). Nonlinear control strategies for quadrotors and CubeSats
(M.S. Thesis, University of Illinois at Urbana-Champaign).
[3] Greitzer, E. M., Spakovszky, Z. S., & Waitz, I. A. (2006). Thermodynamics and
propulsion. Mechanical Engineering, MIT.
[4] Corke, P. (2011). Robotics, vision and control: fundamental algorithms in MATLAB
(Vol. 73). Springer.
[5] Hartman, D., Landis, K., Mehrer, M., Moreno, S., & Kim, J.(2014) Quadcopter
Dynamic Modeling and Simulation (Quadsim) v1.00 (Senior Design project, Drexel
University)
[6] Hoenig, W., Milanes, C., Scaria, L., Phan, T., Bolas, M., & Ayanian, N.
(2015). Mixed reality for robotics. In Intelligent Robots and Systems (IROS), 2015
IEEE/RSJ International Conference on (pp. 5382-5387). IEEE
[7] Elruby, A. Y., El-Khatib, M. M., El-Amary, N. H., & Hashad, A. I. (2012). Dynamic
modeling and control of quadrotor vehicle. In Fifteenth International Conference on
Applied Mechanics and Mechanical Engineering, AMME (Vol. 15).
[8] Karwoski, K. (2011). Quadrocopter Control Design and Flight Operation. (Internship
Final Report, NASA USRP)
[9] Sonnevend, I. (2010). Analysis and model based control of a quadrotor helicopter.
(BSc Diploma work, Péter Pázmány Catholic University, Faculty of Information
Technology, Budapest, Hungary (supervisor: G. Szederkényi) )
[10] Habib, M. K., Abdelaal, W. G. A., & Saad, M. S. (2014). Dynamic modeling and
control of a Quadrotor using linear and nonlinear approaches. (M.S. Thesis, The
American University in Cairo).
[11] Landry, B. (2015). Planning and control for quadrotor flight through cluttered environments (Master’s Degree Thesis, Massachusetts Institute of Technology).
[12] Dunkley, O., Engel, J., Sturm, J., & Cremers, D. (2014). Visual-inertial navigation for a camera-equipped 25g nano-quadrotor. In IROS2014 Aerial Open Source
Robotics Workshop.
[13] Xu, D., Wang, L., Li, G., & Guo, L. (2012, August). Modeling and Trajectory
Tracking Control of a Quad-rotor UAV. In Proceedings of the 2012 International
Conference on Computer Application and System Modeling. Atlantis Press.
92
[14] Meyer, J., Sendobry, A., Kohlbrecher, S., Klingauf, U., & Von Stryk, O. (2012).
Comprehensive simulation of quadrotor uavs using ros and gazebo. In Simulation,
Modeling, and Programming for Autonomous Robots (pp. 400-411). Springer Berlin
Heidelberg.
[15] Suiçmez, E. C. (2014). Trajectory Tracking of a quadrotor unmanned aerial vehicle (UAV) via attitude and position control (Master’s Degree Thesis, Middle East
Technical University).
[16] Oh, S. M. (2012). Modeling and Control of a Quad-rotor Helicopter. (M.S. Thesis,
University of Florida)
[17] Pounds, P. E. I. (2007). Design, construction and control of a large quadrotor micro
air vehicle. (Doctoral dissertation, Australian National University.)
[18] Tamami, N., Pitowarno, E., & Astawa, I. G. P. (2014). Proportional Derivative
Active Force Control for “X” Configuration Quadcopter. Journal of Mechatronics,
Electrical Power, and Vehicular Technology, 5(2), 67-74.
[19] Roo, M. (2015). Optimal Event Handling by Multiple UAVs. (M.S. Report, University of Twente)
[20] Lehnert, C., & Corke, P. (2013). µAV-Design and implementation of an open source
micro quadrotor. AC on Robotics and Automation, Eds.
[21] Sabatino, F.(2015). Quadrotor control: modeling, nonlinear control design, and simulation. (Master’s Degree Project, KTH Royal Institute of Technology).
[22] Kader, S. A., El-henawy, A. E., & Oda, A. N. (2014). Quadcopter System Modeling and Autopilot Synthesis. In International Journal of Engineering Research and
Technology (Vol. 3, No. 11 (November-2014)). ESRSA Publications.
[23] Naidu, D. S. (2002).Optimal control systems. CRC press.
[24] Mathworks®(2015).State
Estimation
Using
Time-Varying
Kalman
Filter.
Retrieved
May
16,
2016,
from
http://www.mathworks.com/help/control/getstart/estimating-states-of-timevarying-systems-using-kalman-filters.html
[25] Hoffmann, G. M., Waslander, S. L., & Tomlin, C. J. (2008).Quadrotor helicopter
trajectory tracking control. In AIAA guidance, navigation and control conference
and exhibit (pp. 1-14).
[26] Mueller, M. W., & D’Andrea, R. (2013).A model predictive controller for quadrocopter state interception. In European Control Conference (pp. 1383-1389).
[27] Mu, S., Zeng, Y., & Wu, P. (2008).Multivariable control of anaerobic reactor by
using external recirculation and bypass ratio. Journal of chemical technology and
biotechnology, 83(6), 892-903.
[28] Huang, H., Hoffmann, G. M., Waslander, S. L., & Tomlin, C. J. (2009).Aerodynamics and control of autonomous quadrotor helicopters in aggressive maneuvering. In
Robotics and Automation, 2009. ICRA’09. IEEE International Conference on (pp.
3277-3282). IEEE.
[29] Sujit, P. B., Saripalli, S., & Sousa, J. B. (2014).Unmanned aerial vehicle path following: A survey and analysis of algorithms for fixed-wing unmanned aerial vehicles.
IEEE Control Systems, 34(1), 42-59.
93
[30] Bouabdallah, S., Noth, A., & Siegwart, R. (2004).PID vs LQ control techniques
applied to an indoor micro quadrotor. In Intelligent Robots and Systems, 2004.(IROS
2004). Proceedings. 2004 IEEE/RSJ International Conference on (Vol. 3, pp. 24512456). IEEE.
[31] Peraire, J., & Widnall, S. (2009) Lecture L28 - 3D Rigid Body Dynamics. MIT
OpenCourseWare, Dynamics Fall 2009. Available online: http://ocw.mit.edu.
[32] Bitcraze®(2015). Crazyflie 2.0 assembly instructions. Retrieved August 3, 2016,
from https://wiki.bitcraze.io/projects:crazyflie2:userguide:assembly
[33] Bitcraze®(2015). Crazyflie 2.0 Main Page. Retrieved August 5, 2016. from
https://www.bitcraze.io/crazyflie-2/
[34] ©Vicon Motion Systems (2015). VICON Motion Capture System Main Page. Retrieved August 5, from https://www.vicon.com/
[35] Mueller, M. W., Hamer, M., & D’Andrea, R. (2015). Fusing ultra-wideband range
measurements with accelerometers and rate gyroscopes for quadrocopter state estimation. In 2015 IEEE International Conference on Robotics and Automation (ICRA)
(pp. 1730-1736). IEEE.
[36] Rafrafi, W., & Le Ny, J. (2016). Intégration d’un système radio à bande ultra-large
pour la navigation de robots mobiles. (Master’s Degree Thesis, École Polytechnique
de Montréal).
[37] ©decaWave(2015). ScenSor DWM1000 Module Product Page. Retrieved August 11,
from http://www.decawave.com/products/dwm1000-module
[38] Luis, C. (2016). Trajectory Tracking of a Crazyflie 2.0 Nanoquadcopter [Video File].
Retrieved August 13, from https://youtu.be/c-SXovCyhJQ
94
Appendix A: Firmware
Modifications
The firmware used during this project was “Release 2016.02” found in https://github.
com/bitcraze/crazyflie-release/releases, with the following changes:
• power_distribution_stock.c lines 58-64
# ifdef QUAD_FORMATION_X
i n t 1 6 _ t r = c o n t r o l −> r o l l / 2 . 0 f ;
i n t 1 6 _ t p = c o n t r o l −>p i t c h / 2 . 0 f ;
motorPower . m1 = l i m i t T h r u s t ( c o n t r o l −>t h r u s t
motorPower . m2 = l i m i t T h r u s t ( c o n t r o l −>t h r u s t
motorPower . m3 = l i m i t T h r u s t ( c o n t r o l −>t h r u s t
motorPower . m4 = l i m i t T h r u s t ( c o n t r o l −>t h r u s t
−
−
+
+
r
r
r
r
−
+
+
−
p
p
p
p
−
+
−
+
c o n t r o l −>yaw ) ;
c o n t r o l −>yaw ) ;
c o n t r o l −>yaw ) ;
c o n t r o l −>yaw ) ;
• controller_pid.c lines 64-84
a t t i t u d e C o n t r o l l e r C o r r e c t A t t i t u d e P I D ( s t a t e −>a t t i t u d e . r o l l , −s t a t e −>a t t i t u d e . p i t c h , s t a t e −>a t t i t u d e . yaw ,
s e t p o i n t −>a t t i t u d e . r o l l , s e t p o i n t −>a t t i t u d e . p i t c h , a t t i t u d e D e s i r e d . yaw ,
&r a t e D e s i r e d . r o l l , &r a t e D e s i r e d . p i t c h , &r a t e D e s i r e d . yaw ) ;
// Bypass A t t i t u d e c o n t r o l l e r if Rate mode active
i f ( s e t p o i n t −>mode . r o l l == m o d e V e l o c i t y ) {
r a t e D e s i r e d . r o l l = s e t p o i n t −>a t t i t u d e R a t e . r o l l ;
}
i f ( s e t p o i n t −>mode . p i t c h == m o d e V e l o c i t y ) {
r a t e D e s i r e d . p i t c h = s e t p o i n t −>a t t i t u d e R a t e . p i t c h ;
}
i f ( s e t p o i n t −>mode . yaw == m o d e V e l o c i t y ) {
r a t e D e s i r e d . yaw = s e t p o i n t −>a t t i t u d e R a t e . yaw ;
}
a t t i t u d e C o n t r o l l e r C o r r e c t R a t e P I D ( s e n s o r s −>g y r o . x , s e n s o r s −>g y r o . y , s e n s o r s −>g y r o . z ,
r a t e D e s i r e d . r o l l , r a t e D e s i r e d . p i t c h , r a t e D e s i r e d . yaw ) ;
a t t i t u d e C o n t r o l l e r G e t A c t u a t o r O u t p u t (& c o n t r o l −>r o l l ,
&c o n t r o l −>p i t c h ,
&c o n t r o l −>yaw ) ;
95
| 3cs.SY
|
BIG POLYNOMIAL RINGS AND STILLMAN’S CONJECTURE
arXiv:1801.09852v3 [math.AC] 9 Mar 2018
DANIEL ERMAN, STEVEN V SAM, AND ANDREW SNOWDEN
1. Introduction
The purpose of this paper is to prove that certain limits of polynomial rings are themselves
polynomial rings, and show how this observation can be used to deduce some interesting results in commutative algebra. In particular, we give two new proofs of Stillman’s conjecture.
The first is similar to that of Ananyan–Hochster [AH16], though more streamlined; in particular, it establishes the existence of small subalgebras. The second proof is completely
different, and relies on a recent noetherianity result of Draisma [Dra17].
1.1. Polynomiality results. For a commutative ring A, let AJJx1 , x2 , . . .KK be the inverse
limit of the standard-graded polynomial rings A[x1 , . . . , xn ] in the category of graded rings.
A degree d element of this ring is a (possibly infinite) formal A-linear combination of degree
d monomials in the variables {xi }i≥1 . Fix a field k, and let R = kJJx1 , x2 , . . .KK. Our first
polynomiality theorem is:
Theorem 1.1. Assume k is perfect. Then R is (isomorphic to) a polynomial ring.
The set of variables in the polynomial ring is uncountable; hence the phrase “big polynomial rings” in the title of the paper. We deduce Theorem 1.1 from the following general
criterion. For a graded ring1 R, we write R+ for the ideal of positive degree elements.
Theorem 1.2. Let R be a graded ring with R0 = k a perfect field. Assume:
• Characteristic 0: R has enough derivations (Definition 2.1), that is, for every nonzero x ∈ R+ there is a derivation ∂ of negative degree such that ∂(x) 6= 0.
• Positive characteristic: R has enough Hasse derivations (see Definition 2.10).
Then R is a polynomial ring. Precisely, for any set of positive degree homogeneous elements
2
{fi }i∈I whose images in R+ /R+
form a k-basis, the k-algebra homomorphism k[Xi ]i∈I → R
taking Xi to fi is an isomorphism.
The proof of Theorem 1.2 is elementary: essentially, if one had an algebraic relation among
some of the fi , then one could apply an appropriate (Hasse) derivation to get a lower degree
relation, and eventually reach a contradiction. To prove Theorem 1.1, we simply observe that
(Hasse) derivatives with respect to the variables xi extend continuously to R and furnish it
with enough (Hasse) derivations.
The inverse limit R is one way to make sense of a limit of finite polynomial rings. A
different way is through the use of ultrapowers, or, more generally, ultraproducts (see §4.1
Date: March 9, 2018.
2010 Mathematics Subject Classification. 13A02, 13D02.
DE was partially supported by NSF DMS-1302057 and NSF DMS-1601619.
SS was partially supported by NSF DMS-1500069 and DMS-1651327 and a Sloan Fellowship.
AS was supported by NSF DMS-1303082 and DMS-1453893 and a Sloan Fellowship.
1In this paper, all graded rings are supported in non-negative degrees.
1
2
DANIEL ERMAN, STEVEN V SAM, AND ANDREW SNOWDEN
for background). Let S be the graded ultrapower of the standard-graded polynomial ring
k[x1 , x2 , . . .]. We also prove:
Theorem 1.3. Assume k is perfect. Then S is a polynomial ring.
This also follows quickly from Theorem 1.2. The perfectness hypotheses in this section
can be relaxed; for instance, Theorems 1.1 and 1.3 hold if [k : kp ] is finite, see Remarks 2.12
and 5.4. However, we do not know if they can be eliminated entirely.
1.2. Connection to the work of Ananyan–Hochster. We recall (and slightly extend)
the notion of strength from [AH16]:
Definition 1.4. Let R be a graded ring with R0 = k a field, and let f be a homogeneous
element of R.P
The strength of f is the minimal integer k ≥ −1 for which there is a decomposition f = k+1
i=1 gi hi with gi and hi homogeneous elements of R of positive degree, or ∞
if no such decomposition exists. The collective strength of a set of homogeneous elements
{fi }i∈I of R is the minimal strength of a non-trivial homogeneous k-linear combination.
Example 1.5. (a) In k[x1 , . . . , xn ], non-zero elements of degree one have infinite strength,
while non-zero elements of degree > 1 have strength < n.
P
(b) In R, there are a wealth of interesting elements of infinite strength, such as i≥1 xdi
(if d is invertible in k).
2
(c) In any graded ring R, the ideal R+
is exactly the ideal of finite strength elements.
Many of the results of Ananyan–Hochster are instances of the following general principle:
elements in a polynomial ring of sufficiently large collective strength behave approximately
like independent variables. Theorem 1.1 shows that this approximation becomes exact in the
limiting ring R. Indeed, suppose {fi }i∈I are elements of R+ that form a basis modulo R2+ .
Thus no linear combination of the fi belongs to R2+ , i.e., has finite strength (Example 1.5(c)),
and so {fi } has infinite collective strength. The Ananyan–Hochster principle thus suggests
that the {fi } should be independent variables, and this is exactly the content of Theorem 1.1.
1.3. Stillman’s conjecture via ultraproducts. While ultraproducts may be less familiar
to readers than inverse limits, Theorem 1.3 leads to our most efficient proof of Stillman’s
conjecture [PS09, Problem 3.14]. As in [AH16] (see §4.3), both the existence of small subalgebras and Stillman’s conjecture can be reduced to the following statement:
Theorem 1.6. Fix integers d1 , . . . , dr . Then there exists an integer N with the following
property. If k is an infinite perfect field, and f1 , . . . , fr ∈ k[x1 , . . . , xn ] are polynomials of
degrees d1 , . . . , dr with collective strength at least N, then f1 , . . . , fr is a regular sequence.
Ananyan–Hochster prove this theorem via a multi-tiered induction, where elements of
increasingly high strength obtain an array of increasingly nice properties. Our proof using
Theorem 1.3 is more direct. Here is the idea. Suppose that f1,i , . . . , fr,i ∈ k[x1 , x2 , . . . ], for i ∈
N, are polynomials of the given degrees with collective strength tending to infinity. It suffices
to show that f1,i , . . . , fr,i eventually forms a regular sequence. For each j, the sequence fj,•
defines an element fj in the ultraproduct ring S. It is easy to see that f1 , . . . , fr has infinite
collective strength (Proposition 4.5). Thus, by Theorem 1.3, f1 , . . . , fr are independent
variables in S, and hence form a regular sequence. We then apply a result (Corollary 4.9)
comparing codimension in S to codimension in k[x1 , x2 , . . .] to conclude that f1,i , . . . , fr,i is
eventually a regular sequence.
BIG POLYNOMIAL RINGS AND STILLMAN’S CONJECTURE
3
As in [AH16], we show that the bound in Theorem 1.6 (and Stillman’s conjecture as well)
is independent of the field k. To do so, we prove a generalization of Theorem 1.3 (see §4.2)
where S is replaced with an ultraproduct of polynomial rings with variable coefficient fields.
1.4. Stillman’s conjecture via inverse limits. Returning to the inverse limit, Theorem 1.1 enables a proof of Stillman’s conjecture that follows the very general rubric in
algebraic geometry of proving a result generically, spreading out to an open set, and then
inductively treating proper closed subsets. The basic idea in characteristic zero is as follows. Suppose that A is a characteristic 0 domain with fraction field K, and M is a finitely
presented AJJx1 , x2 , . . .KK-module. Then K ⊗A M is a finitely presented module over the
ring K ⊗A AJJx1 , x2 , . . .KK. While K ⊗A AJJx1 , x2 , . . .KK is not isomorphic to KJJx1 , x2 , . . .KK,
Theorem 1.2 shows it is also an abstract polynomial ring.
It then follows from simple homological properties of infinite polynomial rings that K ⊗A M
has a finite length resolution by finite free modules. A flatness argument produces an open
dense subset U of Spec(A) such that My has the same Betti table as K ⊗A M for all y ∈ U.
We can then restrict our attention to Spec(A)\U, and apply the same argument. This shows
that there is some (perhaps infinite) stratification of Spec(A) such that on each stratum the
fibers of M have the same Betti table.
We apply this as follows. Fix positive integers d1 , . . . , dr , and let A be the symmetric
algebra on the vector space Symd1 (k∞ ) ⊕ · · · ⊕ Symdr (k∞ ). Then Spec(A) is the space of
forms f1 , . . . , fr ∈ kJJx1 , x2 , . . .KK of degrees d1 , . . . , dr . We let M be the universal module
AJJx1 , x2 , . . .KK/(f1 , . . . , fr ). The stratification constructed in the previous paragraph can be
made compatible with the GL∞ action on Spec(A). A recent theorem of Draisma [Dra17]
asserts that Spec(A) is GL∞ -noetherian, and hence this stratification is finite. We conclude
that there are only finitely many resolution types for ideals generated by f1 , . . . , fr of the
given degrees. This, in particular, implies Stillman’s conjecture in characteristic zero.
The same idea works in positive characteristic, but when K fails to be perfect, we need to
bootstrap from the perfect case to produce the open subset with constant Betti numbers.
1.5. Connections to other work. The Milnor–Moore theorem [MM65], and generalizations [Sjö80], establish that certain commutative graded rings are polynomial rings via properties of a comultiplication. While this, and its extensions to non-commutative rings, can be
applied to examples in commutative algebra, it is of a fairly distinct nature from the criteria
in the present paper.
Theorem 1.1 is an example of the meta-principle that inverse limits of free objects tend to
be free themselves. See [Ser97, §I.4.2, Corollary 4] for an example of this principle with prop-groups. Alexandru Chirvasitu informed us that he can prove a non-commutative version
of Theorem 1.1 where polynomial rings are replaced by non-commutative polynomial rings.
The use of ultraproducts in commutative algebra was famously employed in [vdDS84] to
establish a variety of bounds (with the number of variables fixed). See [Sch10] for more
discussion and examples.
The Gröbner theory of the inverse limit ring kJJx1 , x2 , . . .KK was studied by Snellman
in [Sne98b, Sne98a]. Shortly after a draft of this article was posted, [DLL18] applied Theorem 1.1 to obtain finiteness results for grevlex Gröbner bases over R, and then used this
to answer some questions raised by Snellman and to give a generic initial ideal proof of
Stillman’s Conjecture.
4
DANIEL ERMAN, STEVEN V SAM, AND ANDREW SNOWDEN
The use of GL∞ -noetherianity of spaces to prove the existence of uniform bounds in
algebraic geometry has been used in several papers. See [Dra14] for a survey.
1.6. Outline. In §2, we establish our polynomiality criteria (summarized in Theorem 1.2).
In §3, we prove some easy results concerning dimension theory in polynomial rings with an
infinite number of variables. In §4, we prove that the ultraproduct ring is a polynomial ring
(Theorem 1.3), and use this to deduce our first proof of Stillman’s conjecture. Finally, in
§5, we prove that the inverse limit ring is a polynomial ring (Theorem 1.1), and use this to
deduce our second proof of Stillman’s conjecture.
Acknowledgements. We thank Craig Huneke and Gregory G. Smith for useful conversations. We also thank Alexandru Chirvasitu for informing us about his work on the noncommutative analogue of Theorem 1.1 and the reference in [Ser97].
2. Criteria for polynomiality
Let R be a graded ring with R0 = k a field. We say that R is a polynomial ring if there
are elements {xi }i∈I of R, each homogeneous of positive degree, such that the natural map
k[Xi ] → R sending Xi to xi is an isomorphism. The xi ’s need not have degree 1, and the
set I need not be finite. The purpose of this section is to characterize polynomial rings via
derivations.
2.1. Characteristic 0. We first treat the case where k has characteristic 0, for which the
following definition and theorem constitute our criterion for polynomiality. We say that a
derivation ∂ of a graded ring R is homogeneous of degree d if deg ∂(x) = deg(x) + d for
all homogeneous x ∈ R.
Definition 2.1. Let R be a graded ring with R0 = k a field. We say that R has enough
derivations if for every non-zero homogeneous element x of positive degree there is a homogeneous derivation ∂ of negative degree such that ∂(x) 6= 0.
Theorem 2.2. Let R be a graded k-algebra with R0 = k a field of characteristic 0. Then R
is a polynomial ring if and only if R has enough derivations.
Proof. In this proof, “derivation” will mean “homogeneous derivation of negative degree.”
It is clear that a polynomial ring has enough derivations. We prove the converse.
2
Let E be a set of homogeneous elements of R+ that gives a basis of R+ /R+
. By Nakayama’s
lemma, E generates R as a k-algebra, so it suffices to show that E is algebraically independent.
Let E≤d (resp. Ed ) be the set of elements in E of degree ≤ d (resp. d). We prove that E≤d is
algebraically independent for all d by induction on d. Suppose that we have shown E≤d−1 is
algebraically independent. To prove that E≤d is algebraically independent, it suffices to prove
the following statement: if E≤d−1 ⊂ E ⊂ E≤d is algebraically independent and x ∈ Ed \ E,
then E ′ = E ∪ {x} is algebraically independent. Indeed, this statement implies that all sets
of the form E≤d−1 ∪ E ′′ with E ′′ a finite subset of Ed are algebraically independent, which
implies that E≤d is algebraically independent.
Thus let E, E ′ , and x as above be given. Let A ⊂ R be the k-subalgebra generated
by E.
Pn
′
i
To prove that E is algebraically independent, it suffices to show that if 0 = i=0 ai x with
ai ∈ A then ai = 0 for all i. Before proceeding, we note that if ∂ is any derivation of R then
∂(E≤d ) ⊂ A since ∂ decreases degrees, and so ∂(A) ⊂ A and ∂(x) ∈ A.
BIG POLYNOMIAL RINGS AND STILLMAN’S CONJECTURE
5
P
Suppose that 0 = ni=0 ai xi with ai ∈ A and an 6= 0. Of all such relations, choose one
of minimal degree (i.e., with deg(an xn ) minimal). Suppose that an has positive degree. By
assumption, there exists a derivation ∂ such that ∂(an ) 6= 0. Applying ∂ to our given relation
Pn−1 i
yields 0 = ∂(an )xn + i=0
bi x where the bi are elements of A. This is a contradiction, since
∂(an ) has smaller degree than an . Thus deg(an ) = 0, and so we may assume an = 1.
2
Since E is linearly independent modulo R+
, we see that x ∈
/ A, and so n ≥ 2 and nx+ an−1
is non-zero. It follows that there exists a derivation ∂ such
that
+ an−1 ) 6= 0. Applying
Pn−2 ∂(nx
i
n−1
This is a
∂ to our original relation gives 0 = ∂(nx + an−1 )x
+ i=0 bi x for some bi ∈ A. P
smaller degree relation, which is a contradiction. We thus see that no relation 0 = ni=0 ai xi
exists with an non-zero, which completes the proof.
2.2. Positive characteristic. Theorem 2.2 obviously fails in characteristic p: since pth
powers are killed by every derivation, no reduced ring has enough derivations. The most
obvious adjustment would be to ask that if x is a homogeneous element of R that is not a
pth power then there is a derivation ∂ such that ∂(x) 6= 0. The following two examples show
that this condition is insufficient to conclude that R is a polynomial ring.
Example 2.3. Let R = k[x]/(xp ) where k is perfect of characteristic p and x has degree 1.
d
Then dx
is a well-defined derivation on R, and thus R has enough derivations.
Example 2.4. Let R = k[x, y, xyp ] where k is perfect of characteristic p, x has degree 1,
d
d
and y has degree p + 1. Then dx
and xp dy
are well-defined derivations on R, and every
homogeneous element of R that is not a pth power is not annihilated by one of them.
To extend our criterion to the positive characteristic case, we employ the following extension of the notion of a derivation (see [Gol03, pp. 27–29] for additional discussion).
Definition 2.5. Let R be a k-algebra. A Hasse derivation on R is a sequence ∂ • = (∂ n )n≥0
where each ∂ n is a k-linear endomorphism of R such that ∂ 0 is the identity and
X
∂ n (xy) =
∂ i (x)∂ j (y)
i+j=n
holds for all x, y ∈ R. If R is graded then we say ∂ • is homogeneous of degree d if ∂ n (x)
has degree deg(x) + nd for all homogeneous x ∈ R.
Remark 2.6. Giving a Hasse derivation on R is equivalent to giving a ring homomorphism
ϕ : R → RJtK such that the constant term of ϕ(x) is P
x. If ∂ • is a Hasse derivation, then the
associated ring homomorphism is defined by ϕ(x) = i≥0 ∂ i (x)ti .
Example 2.7. Suppose R = k[x], with k any field. Define ∂ n (xk ) = nk xk−n . (Note
dn
•
that ∂ n = n!1 dx
is a Hasse derivation, called the Hasse
n if n! is invertible in k.) Then ∂
derivative. If R is graded with x of degree d then ∂ • is homogeneous of degree −d. The
homomorphism ϕ : R → RJtK associated to the Hasse derivative is given by x 7→ x + t.
Remark 2.8. Curiously, Hasse derivatives also play a key role in Draisma’s [Dra17], where
they are closely related to his directional derivatives.
Lemma 2.9. Let R be a k-algebra, where k is a field of characteristic p, and let ∂ • be a
Hasse derivation on R. Let q be a power of p. Then for x ∈ R and n ∈ N we have
(
(∂ n/q x)q if q | n
∂ n (xq ) =
.
0
if q ∤ n
6
DANIEL ERMAN, STEVEN V SAM, AND ANDREW SNOWDEN
Proof. We have
∂ n (xq ) =
X
∂ i1 (x) · · · ∂ iq (x).
(i1 ,...,iq )
i1 +···+iq =n
If i1 , . . . , iq are not all equal then the orbit of (i1 , . . . , iq ) under the symmetric group Sq has
cardinality divisible by p. All elements of this orbit contribute equally to the sum, and thus
they all cancel. We thus see that the only surviving term occurs when n is a multiple of q
and i1 = · · · = iq = n/q; this term is (∂ n/q x)q .
The following definition and theorem constitute our criterion for polynomiality in positive
characteristic.
Definition 2.10. Let R be a graded ring with R0 = k a field of characteristic p > 0. We say
that R has enough Hasse derivations if the following condition holds: if x is a positive
degree homogeneous element of R such that x 6∈ kRp (the k-span of the set Rp ) then there
exists a homogeneous Hasse derivation ∂ • of R of negative degree such that ∂ 1 (x) 6= 0.
Theorem 2.11. Let R be a graded ring with R0 = k a perfect field of characteristic p > 0.
Then R is a polynomial ring if and only if it has enough Hasse derivations.
Proof. In this proof, “Hasse derivation” will mean “homogeneous Hasse derivation of negative
degree.” We note that since k is perfect, kRp = Rp . If R is a polynomial ring then it has
enough Hasse derivations; one can see this using Hasse derivatives (Example 2.7). We now
prove the converse.
We first show that R is reduced. Suppose not, and let x ∈ R be a non-zero homogeneous
nilpotent element of minimal degree. Note that x 6∈ Rp , for if x = y p then y would be a
r
lower degree nilpotent element. Let r be such that xp = 0 and let ∂ • be a Hasse derivation
r
r
r
such that ∂ 1 (x) 6= 0. Then 0 = ∂ p (xp ) = (∂ 1 x)p (Lemma 2.9), and so ∂ 1 (x) is nilpotent,
contradicting the minimality of x. Thus R is reduced.
2
Let E be a set of homogeneous elements of R+ that forms a basis for R+ /R+
. It suffices
to prove that E is algebraically independent. For E ⊂ E, consider the following statement:
AE : Given distinct elements x1 , . . . , xr ∈ E and a polynomial F ∈ k[X1 , . . . , Xr ] such
that F (x1 , . . . , xr ) ∈ Rp , we have F ∈ k[X1 , . . . , Xr ]p .
Observe that if AE holds then E is algebraically independent. Indeed, suppose that F (x1 , . . . , xr ) =
0 is a minimal degree algebraic relation among distinct elements of E. Since 0 ∈ Rp , we see
that F (X1 , . . . , Xr ) = G(X1 , . . . , Xr )p for some G by AE , and so G(x1 , . . . , xr )p = 0. Since
R is reduced, it follows that G(x1 , . . . , xr ) = 0, contradicting the minimality of F . Thus to
prove the theorem it suffices to prove AE .
We prove that AE holds for all E by induction on E in the following manner. Let E≤d
(resp. Ed ) be the set of elements of E of degree ≤ d (resp. d). Suppose that E≤d−1 ⊂ E ⊂ E≤d
and let E ′ = E ∪ {x} for some x ∈ Ed \ E. Assuming AE , we prove AE ′ . This will establish
AE for all E by the same logic used in the proof of Theorem 2.2.
Fix E, E ′ , and x as above. Let A be the k-subalgebra of R generated by E. We claim
that AE ′ can be reduced to the following statement, for all n and m:
P
Bn,m : If ni=0 ai xi ∈ Rp with ai ∈ A and deg(an ) ≤ m then ai ∈ Rp and iai = 0 for all i.
Indeed, suppose Bn,m holds for all n and m, and suppose F (x1 , . . . , xr ) ∈ Rp for distinct
elements x1 , . . . , xrP
∈ E ′ . We may as well suppose xr = x and x1 , . . . , xr−1 ∈ E. Write
n
i
By Bn,m , we see that
F (X1 , . . . , Xr ) =
i=0 Gi (X1 , . . . , Xr−1 )Xr for polynomials Gi .
BIG POLYNOMIAL RINGS AND STILLMAN’S CONJECTURE
7
Gi (x1 , . . . , xr−1 ) ∈ Rp for all i and Gi (x1 , . . . , xr−1 ) = 0 if p ∤ i. By AE , it follows that
Gi (X1 , . . . , Xr−1 ) = G′i (X1 , . . . , Xr−1 )p for some polynomial G′i and that Gi (X1 , . . . , Xr ) = 0
if p ∤ i. We thus find
X
p
′
i/p
F (X1 , . . . , Xr ) =
Gi/p (X1 , . . . , Xr−1 )Xr
,
0≤i≤n
p|i
which establishes AE ′ .
We now prove Bn,m P
by induction on n and m. Clearly, B0,m holds for all m. We note
that if Bn,m holds and ni=0 ai xi = 0 with deg(an ) ≤ m then ai = 0 for all i; the proof is the
same as the proof given above that AE implies algebraic independence of E. We also note
that if ∂ • is any Hasse derivation then ∂ n (E≤d ) ⊂ A for all n > 0, and so ∂ n (A) ⊂ A and
∂ n (x) ∈ A.
We now prove B1,m for all m by induction on m. First suppose m = 0, and suppose that
2
ax + b ∈ Rp with a ∈ k and b ∈ A. We thus see that ax + b = 0 in R+ /R+
. Since E is linearly
2
independent in R+ /R+ , it follows that a = 0, and so B1,0 holds. Now suppose B1,m−1 holds,
and let us prove B1,m . Thus suppose that ax + b = y p for some y ∈ Rp with a, b ∈ A
and deg(a) = m. If ∂ • is any Hasse derivation of R then ∂ 1 (a)x + (a∂ 1 (x) + ∂ 1 (b)) = 0
(Lemma 2.9). Since deg(∂ 1 (a)) < m, we see that ∂ 1 (a) = 0 by B1,m−1 . Since this holds for
all ∂ • , we find a ∈ Rp . Suppose a 6= 0, and let q be the maximal power of p such that a ∈ Rq
(this exists since deg(a) > 0). Write a = cq , and note c 6∈ Rp . Let ∂ • be a Hasse derivation
such that ∂ 1 (c) 6= 0; note then that ∂ q (a) = (∂ 1 c)q 6= 0 (Lemma 2.9). Again by Lemma 2.9,
we have
∂ q (a)x + (a∂ q (x) + ∂ q (b)) = ∂ q (y p ) = (∂ q/p y)p ∈ Rp
By B1,m−1 , we have ∂ q (a) = 0, a contradiction. Thus a = 0 and B1,m holds.
P
We now prove Bn,m for n ≥ 2, assuming Bn−1,• and Bn,m−1 . Thus suppose that ni=0 ai xi ∈
Rp with ai ∈ A and deg(an ) ≤ m. Let ∂ • be a Hasse derivation of R. Applying ∂ 1 , we find
0 = ∂ 1 (an )xn + (nan ∂ 1 (x) + ∂ 1 (an−1 ))xn−1 + · · · ,
where the remaining terms have degree ≤ n − 2 in x. By Bn,m−1 , all the above coefficients
vanish. Thus ∂ 1 (an ) = 0 for all ∂ • , and so an ∈ Rp . We now see that the coefficient of xn−1
p
is ∂ 1 (nan x + an−1 ). Since this vanishes for all ∂ • , we find nan x + aP
n−1 ∈ R , and so nan = 0
n−1
ai xi ∈ Rp . Thus by
by B1,• . In particular, p | n if an 6= 0, so an xn ∈ Rp , and hence i=0
Bn−1,• we have ai ∈ Rp and iai = 0 for all 0 ≤ i ≤ n − 1. This proves Bn,m .
Remark 2.12. The perfectness hypothesis in Theorem 2.11 can be omitted. Indeed, letting
K be the perfection of k, the theorem shows that K ⊗k R is a polynomial ring, which implies
that R is a polynomial ring.
3. Dimension theory in polynomial rings
Fix a field k. For a ring A and a (possibly infinite) set U, we let A[U] be the polynomial
algebra over A in variables U. We aim to prove a number of basic results on codimension
in rings of the form A[U] where A is a finitely generated k-algebra. All of these results are
standard when U is finite. We do not impose any gradings in this section.
For a prime ideal p in a commutative ring R, the codimension (or height) of p is the
maximum integer c for which there exists a chain of prime ideals p0 ( · · · ( pc = p, or ∞ if
such chains exist with c arbitrarily large. The codimension of an arbitrary non-unital ideal
8
DANIEL ERMAN, STEVEN V SAM, AND ANDREW SNOWDEN
I of R is the minimum of the codimensions of primes containing I, or ∞ if I is not contained
in any prime of finite codimension. This will be denoted codimR (I). We start with a basic
fact that we will cite often.
Proposition 3.1. Let A ⊂ B be a flat integral extension of rings. For any ideal I ⊂ B, we
have codimB (I) = codimA (A ∩ I).
Proof. We first prove the statement assuming that I = p is prime. Suppose that codimB (p) ≥
c. Let p0 ( p1 ( · · · ( pc = p be a chain of distinct prime ideals. Let qi = A ∩ pi . By
incomparability [AK13, Theorem 14.3(2)], the qi are distinct and thus codimA (qc ) ≥ c.
In particular, if codimB (p) = ∞, this shows that codimA (A ∩ p) = ∞. Now suppose
that codimB (p) is finite and equal to c. If there were some longer chain of primes leading
up to qc , then by going down for flat extensions [AK13, Theorem 14.11], we would have
codimB (pc ) > c, which is a contradiction. Thus codimA (qc ) = c which finishes the special
case when I is prime.
Now consider the general case. Given a prime p containing I, we have just shown that
codimA (A ∩ p) = codimB (p). On the other hand, given a prime q containing A ∩ I, using
[AK13, Theorem 14.3(4)], there is a prime p ⊃ I such that A ∩ p = q. In particular, we
deduce that codimB (I) = codimA (A ∩ I).
Proposition 3.2. Let A be a finitely generated k-algebra and let p be a prime ideal of A[U]
of finite codimension. Then p is finitely generated.
Proof. Let c be the codimension of p. We prove the statement by induction on c. First
suppose that c = 0. If p = 0, then we are done. Otherwise, choose a nonzero element g ∈ p.
Let V ⊂ U be a finite subset such that g belongs to A[V]. Let p′ = A[U](A[V] ∩ p). Then we
have p′ ⊆ p. Since p is prime, so is A[V] ∩ p, and hence so is p′ since A[U] is obtained from
A[V] by adjoining variables. In particular, we have p′ = p, and so p is finitely generated.
Now suppose c > 0. Choose a prime ideal q ⊂ p of codimension c − 1. By induction, we
know that q is finitely generated; let f1 , . . . , fr be generators. Let g ∈ p \ q and let V ⊂ U
be a finite subset such that the fi ’s and g belong to A[V]. Let p′ = A[U](A[V] ∩ p). Note
that q = A[U](A[V] ∩ q). We thus have q ⊂ p′ ⊂ p. Since p is prime, so is its contraction
to A[V], and so is the extension of this back to A[U], since A[U] is obtained from A[V] by
adjoining variables. Thus p′ is either q or p; however, it is not q since it contains g. Thus
p = p′ , which shows that p is finitely generated.
Proposition 3.3. Let A be a finitely generated k-algebra and let V ⊂ U be sets. If I
is a finitely generated ideal of A[V], and J is its extension to A[U], then codimA[V] (I) =
codimA[U] (J).
Proof. We note that the result is classical if U is finite (as can be seen, for example, using
Hilbert polynomials). We will use this twice in the proof of the general case.
First suppose that V is finite and I = p is prime. Note then that J = q is prime as well.
If p0 ( · · · ( pc = p is a chain of primes in A[V], then letting qi be the extension of pi ,
we get a chain of primes q0 ( · · · ( qc = q in A[U], and so codimA[U] (q) ≥ c, which shows
codimA[U] (q) ≥ codimA[V] (p).
Next suppose that q0 ( · · · ( qc = q is a chain of primes in A[U]. For each 0 < i ≤ c,
pick fi ∈ qi \ qi−1 . Let V′ be a finite subset of U containing V and such that each fi belongs
to A[V′ ]. Then q• ∩ A[V′ ] is a strict chain of primes ideals in A[V′ ], and so we see that
codimA[V′ ] (pA[V′ ]) ≥ c (note that the contraction of q to A[V′ ] is equal to the extension of p
BIG POLYNOMIAL RINGS AND STILLMAN’S CONJECTURE
9
to A[V′ ]). However, codimA[V′ ] (pA[V′ ]) = codimA[V] (p) by classical theory. We thus see that
codimA[V] (p) ≥ c, and so codimA[V] (p) ≥ codimA[U] (q). In particular, we have equality and
this case has been proven.
Next, suppose still that V is finite, but let I be an arbitrary ideal. If p is a codimension
c prime of A[V] containing I then pA[U] is a codimension c prime of A[U], by the previous
paragraph, containing J. We thus see that codimA[U] (J) ≤ codimA[V] (I). Next, suppose
that q is a codimension c prime of A[U] containing J. By Proposition 3.2, there is a finite
subset V′ of U (which we can assume contains V) such that q is the extension of an ideal
(necessarily prime) q′ of A[V′ ]. By the previous paragraph, q′ has codimension c in A[V′ ].
Since q′ clearly contains the extension of I to A[V′ ], we see that codimA[V′ ] (IA[V′ ]) ≤ c. But
codimA[V′ ] (IA[V′ ]) = codimA[V] (I) by classical theory, and so codimA[V] (I) ≤ c. We thus see
that codimA[V] (I) ≤ codimA[U] (J).
Finally, we treat the case where V is arbitrary. Since I is finitely generated, there is a
finite subset V0 of V such that I is the extension of an ideal I0 of A[V0 ]. Thus
codimA[V] (I) = codimA[V0 ] (I0 ) = codimA[U] (J),
by two applications of the case where V is finite.
Corollary 3.4. Let A be a finitely generated k-algebra. Every finitely generated ideal of
A[U] has finite codimension.
Proof. Let J be a finitely generated ideal of A[U]. Then J is the extension of an ideal I of
some A[V] with V ⊂ U finite. Since codimA[V] (I) ≤ dim A[V] < ∞, Proposition 3.3 implies
that codimA[U] (J) is finite as well.
Corollary 3.5. Let U be a set, and let f1 , . . . , fr ∈ k[U]. Then f1 , . . . , fr form a regular
sequence if and only if the ideal (f1 , . . . , fr ) has codimension r.
Proof. Let V be a finite subset of U such that f1 , . . . , fr ∈ k[V]. Let I (resp. J) be the ideal
of k[V] (resp. k[U]) generated by the fi . The fi form a regular sequence in k[V] (or in k[U])
if and only the Koszul complex on the fi is exact; however, since k[U] ⊆ k[V] is faithfully
flat, the Koszul complex on the fi is exact over k[U] if and only if it exact over k[V].
Corollary 3.6. Let A be a finitely generated k-algebra, let U be a set, and let J be a finitely
generated ideal of A[U] containing a nonzerodivisor f . Let J be the image of J in A[U]/(f ).
Then codimA[U]/(f ) (J) = codimA[U] (J) − 1.
Proof. Let V be a sufficiently large finite subset of U such that A[V] contains f and some
finite generating set of J; thus J is the extension of some ideal I of A[V]. Let B = A[V],
which is a finitely generated k-algebra, and note that A[U] = B[U′ ], where U′ = U \ V. Let
I be the image of I in B = B/(f ). Since J is the extension of I to B[U′ ] = A[U]/(f ), we
obtain codimB (I) = codimA[U]/(f ) (J) and codimB (I) = codimA[U] (J) by two applications of
Proposition 3.3. Finally since codimB (I) = codimB (I) − 1 by classical theory (for example,
the principal ideal theorem and [Eis95, Corollary 13.4]), the result follows.
Proposition 3.7. Let A be a finitely generated k-algebra and let U be a set. Let J be a
finitely generated ideal of A[U][y]. Then A[U] ∩ J is also a finitely generated ideal.
Proof. Let V be a finite subset of U such that J is the extension of an ideal I of A[V][y].
Then A[V] ∩ I is finitely generated since A[V] is noetherian. One easily sees that A[U] ∩ J
is the extension of A[V] ∩ I, which proves the result.
10
DANIEL ERMAN, STEVEN V SAM, AND ANDREW SNOWDEN
Corollary 3.8. Let A be a finitely generated k-algebra, let U be a set, let R = A[U], and let
S = R[y]. Let I be a finitely generated ideal of S. Suppose that I contains a positive degree
Pn−1
monic polynomial, that is, an element of the form y n + i=0
ai y i with ai ∈ R and n > 0.
Then codimR (R ∩ I) = codimS (I) − 1.
Proof. Let f ∈ I be a monic polynomial. Let I be the image of I in S/(f ). Then R → S/(f )
is a finite flat extension of rings and R ∩I is the contraction of I along this map. We thus see
that codimR (R ∩ I) = codimS/(f ) (I) by Proposition 3.1. But codimS/(f ) (I) = codimS (I) − 1
by Corollary 3.6.
4. Stillman’s conjecture via the ultraproduct ring
4.1. Background on ultraproducts. For more details and references on ultraproducts,
see [Sch10, §2.1]. Let I be an infinite set. We fix a non-principal ultrafilter F on I, which is
a collection of subsets of I satisfying the following properties:
(a) F contains no finite sets,
(b) if A ∈ F and B ∈ F, then A ∩ B ∈ F,
(c) if A ∈ F and A ⊆ B, then B ∈ F,
(d) for all A ⊆ I, either A ∈ F or I \ A ∈ F (but not both).
We think of the sets in F as neighborhoods of some hypothetical (and non-existent) point ∗
of I, and refer to them as such. We say that some condition holds near ∗ if it holds in some
neighborhood of ∗.
Q Given a family of sets {Xi }i∈I , their ultraproduct is the quotient of the usual product
i∈I Xi in which two sequences (xi ) and (yi ) are identified if the equality xi = yi holds near
∗. If x is an element of the ultraproduct, we will write xi for the ith coordinate of x, keeping
in mind that this is only well-defined in sufficiently small neighborhoods of ∗; in other words,
we can think of x as a germ of a function around ∗.
Suppose that each Xi is a graded abelian group. We define the graded ultraproduct
of the Xi ’s to be the subgroup of the usual ultraproduct consisting of elements x such that
deg(xi ) is bounded near ∗. The graded ultraproduct is a graded abelian group; in fact, it is
the ultraproduct of the Xi ’s in the category of graded abelian groups. The degree d piece
of the graded ultraproduct is the usual ultraproduct of the degree d pieces of the Xi ’s. We
apply this construction in particular to the case where the Xi ’s are graded rings; the graded
ultraproduct is then again a graded ring.
Example 4.1. If K is the ultraproduct of {ki }i∈I , then the graded ultraproduct of ki [x1 , . . . , xn ]
(with standard grading) is K[x1 , . . . , xn ] (also with standard grading).
In this subsection, we develop a few basic properties of graded ultraproduct rings. We
begin with a simple observation on adjoining variables to ultraproducts.
Proposition 4.2. Let {Ri }i∈I be a family of graded rings with graded ultraproduct S. Let y
e be the graded ultraproduct of the rings Ri [y]. Then the
be a variable of degree 1, and let S
e is an isomorphism.
natural map S[y] → S
P
e Then
Proof. Suppose that f = dk=0 ak y k is an element of S[y], and let g be its image in S.
Pd
k
gi = k=0 ak,i y . If g = 0 then, passing to some neighborhood of ∗, we can assume gi = 0
for all i, which implies that ak,i = 0 for all i and k, which implies that ak = 0 for all k, which
shows that f = 0. Thus the map is injective.
BIG POLYNOMIAL RINGS AND STILLMAN’S CONJECTURE
11
e of degree d. Then we can write gi = Pd ak,i y k
Next, suppose that g is an element of S
k=0
for each i, where the ak,i ’s are elements of Ri . Let ak be the element of S defined by the
P
sequence (ak,i ). Then g is the image of f = dk=0 ak y k , and so the map is surjective.
We now examine how ideals in an ultraproduct relate to ideals in the original rings.
Given a family of graded rings {Ri }i∈I and a family of ideals {Ii }i∈I , we say that the Ii are
uniformly finitely generated if there exists an integer n such that Ii is generated by at
most n elements for all i ∈ I.
Proposition 4.3. Let {Ri }i∈I be a family of graded rings with graded ultraproduct S.
(a) Suppose that {Ii } is a uniformly finitely generated family of homogeneous ideals. Then
their graded ultraproduct I is a finitely generated ideal of S.
(b) Suppose that {Ii } and {Ji } are two uniformly finitely generated families of homogeneous ideals whose graded ultraproducts are equal. Then Ii = Ji for all i in some
neighborhood.
(c) Suppose that I is a finitely generated homogeneous ideal of S. Then there exists a
uniformly finitely generated family of homogeneous ideals {Ii } with ultraproduct I.
Proof. (a) Suppose that each Ii is generated by ≤ n elements; pick generators f1,i , . . . , fn,i of
each Ii . Let f1 , . . . , fn be the elements of S defined by these sequences. We claim that I is
generated by f1 , . . . , fn . Indeed, suppose that g is a homogeneous element of I; thus, passing
to a small P
enough neighborhood, we see that each gi is an element of Ii , and can thus be
written as nk=1 ak,i fk,i for some homogeneous
elements ak,i ∈ Ri . Let ak be the element of S
P
defined by the sequence ak,i . Then g = nk=1 ak fk , proving the claim. (Note that for k fixed,
each ak,i is homogeneous of some degree, but that the degree may depend on i. However, the
degree is bounded by the degree of g, and so in any small enough neighborhood, the degree
of ak,i will be independent of i.)
(b) Suppose that Ii and Ji are each generated by at most n elements for all i, and pick
generators f1,i , . . . , fn,i and g1,i , . . . , gn,i . Let f1 , . . . , fn and g1 , . . . , gn be the elements of S
these sequences define.
Pn By (a), the fk ’s and gk ’s generate the same
Pn ideal of S. Thus we have
an expression gk = j=1 aj fj for some aj ∈ S, and so gk,i = j=1 aj,i fj,i holds for all i in
some neighboorhood of ∗, and so gk,i belongs to the ideal Ii for all such i. Since there are
only finitely many f ’s and g’s, we can pass to some common neighboorhood of ∗ so that
gk,i ∈ Ii and fk,i ∈ Ji for all i and k, and so Ii = Ji .
(c) Let I be generated by f1 , . . . , fn . Let fk be represented by some sequence (fk,i ), and
let Ii be the ideal of Ri generated by f1,i , . . . , fn,i . Then the argument in (a) shows that I is
the ultraproduct of the Ii ’s.
Due to this proposition, we can unambiguously speak of the germ of a finitely generated
homogeneous ideal I of S. We denote these ideals by Ii , keeping in mind that they are only
well-defined for i sufficiently close to ∗. We next show that this construction interacts well
with contraction.
Proposition 4.4. Let {Ri } be a family of graded rings with graded ultraproduct S, and let
{Ri′ } be a family of graded subrings of {Ri } with graded ultraproduct S′ . Let I be a finitely
generated homogeneous ideal of S, and suppose that S′ ∩ I is a finitely generated ideal of S′ .
Then (S′ ∩ I)i = S′i ∩ Ii in a neighborhood of ∗.
Proof. Let g1 , . . . , gm be generators for S′ ∩ I. Then gk,i belongs to S′i ∩ Ii (in some neighborhood of ∗), and so (S′ ∩ I)i is contained in S′i ∩ Ii (in some neighborhood of ∗), since the
12
DANIEL ERMAN, STEVEN V SAM, AND ANDREW SNOWDEN
former is generated by g1,i , . . . , gm,i . We now claim that the inclusion (S′ ∩ I)i ⊂ S′i ∩ Ii is
an equality in some neighborhood of ∗. Assume not. Then we can find a sequence (hi ) such
that hi ∈ S′i ∩ Ii for all i, but in any neighborhood of ∗ there exists i such that hi 6∈ (S′ ∩ I)i .
Let h ∈ S be the element defined by (hi ). Then h ∈PS′ , since hi ∈ Ri′ for all i, and h ∈ I,
′
since P
hi ∈ Ii for all i. Thus h ∈ S′ ∩ I, and so h = m
k=1 ak gk for some ak ∈ S . But then
m
′
hi = k=1 ak,i gk,i holds in some neighborhood of ∗, which shows that hi ∈ (S ∩ I)i in some
neighborhood of ∗, a contradiction.
We close this subsection with a result on strength in ultraproducts:
Proposition 4.5. Let {Ri } be a family of graded rings with graded ultraproduct S. Suppose
that the degree 0 piece of Ri is a field ki , so that the degree 0 piece of S is the ultraproduct K
of these fields. Let f1 , . . . , fr ∈ S. Suppose that the collective strength of f1,i , . . . , fr,i tends
to infinity as i → ∗. Then f1 , . . . , fr has infinite collective strength.
P
P
Proof. Suppose we have a relation rj=1 aj fj = sk=1 gk hk where ai ∈ K are not all zero and
gk and hk are elements of positive degree. Represent everything by sequences:
Pr aj = (aj,i ),
gj = (gj,i), and hj = (hj,i ). Then, by definition of the ultraproduct, we have j=1 aj,i fj,i =
P
s
k=1 gk,i hk,i for all i sufficiently close to ∗, and moreover, not all the aj,i vanish. But this
shows that f1,i , . . . , fr,i has collective strength < s in this neighborhood of ∗.
4.2. The main theorems on ultraproduct rings. Let {ki }i∈I be a family of perfect fields
with ultraproduct K. The field K is also perfect, as if K has characteristic p > 0, then ki is
perfect of characteristic of p for all i sufficiently close to ∗, and so one can take pth roots in
K. Let Ri = ki [x1 , x2 , . . .] with standard grading, and let S be the graded ultraproduct of
the family {Ri }i∈I .
Theorem 4.6. The ring S is a polynomial ring.
Proof. We use the criteria of §2. First suppose that K has characteristic 0, and let us prove
that S has enough derivations (Definition 2.1). Let f ∈ S be a non-zero homogeneous
element of degree d > 0. Passing to a neighborhood of ∗, we can assume that each ki has
characteristic 0 or characteristic p with p > d, and that fi 6= 0. For each i, let a(i) be an
index such that xa(i) appears in some monomial in fi , and let ∂i be the derivation dxda(i) of
Ri . The derivations (∂i ) define a derivation ∂ on S. Since ∂i (fi ) 6= 0 near ∗, we see that
∂(f ) 6= 0, and so S has enough derivations. Thus S is a polynomial ring (Theorem 2.2).
Now suppose that K has characteristic p, and let us prove that S has enough Hasse
derivations (Definition 2.10). Let f ∈ S be a homogeneous element of positive degree that
is not a pth power. Passing to a neighborhood of ∗, we can assume that each fi is not a pth
power. For each i, let a(i) be an index such that xa(i) appears in some monomial in fi with
exponent not divisible by p, and let ∂i be the Hasse derivative on Ri with respect to xa(i)
(Example 2.7). The Hasse derivations ∂i on the Ri induce a Hasse derivation ∂ on S. Since
∂i (fi ) 6= 0 near ∗, we see that ∂(f ) 6= 0, and so S has enough Hasse derivations. Thus S is a
polynomial ring (Theorem 2.11).
Theorem 4.7. Assume that for all i sufficiently close to ∗, the fields ki are infinite. If I ⊂ S
is a finitely generated ideal, then codimS (I) = codimRi (Ii ) for all i sufficiently close to ∗.
Proof. Let c = codimS (I), which is finite by Corollary 3.4. If c = 0 then I = 0, and so Ii = 0
for all i sufficiently close to ∗, and so the formula holds. We now proceed by induction on c.
BIG POLYNOMIAL RINGS AND STILLMAN’S CONJECTURE
13
Suppose the result holds for c − 1, and let I be an ideal of S of codimension c > 0. Let
f ∈ I be a non-zero homogeneous element. For each i, let γi be an automorphism of Ri such
that γi (fi ) is monic in x1 , at least for i sufficiently close to ∗ (see Lemma 4.8). The family
{γi } induces an automorphism γ of S. Since codimension is invariant under automorphisms,
we may replace I with γ(I), and so we can assume that fi is monic in x1 for all i sufficiently
close to ∗.
Let Ri′ = ki [x2 , . . .] and let S′ be the ultraproduct of {Ri′ }. We have Ri = Ri′ [x1 ] for each
i, and so S ∼
= S′ [x1 ] by Proposition 4.2. Under this identification, f corresponds to a monic
polynomial in S′ [x1 ]. Let I ′ be the contraction of I to S′ , which is finitely generated by
Proposition 3.7. We note that Ii′ is the contraction of Ii to Ri′ , for all i sufficiently close to ∗,
by Proposition 4.4. Corollary 3.8 implies that codimS′ (I ′ ) = codimS (I) − 1 = c − 1. Thus,
by the inductive hypothesis, we have codimR′i (Ii′ ) = c − 1 for all i sufficiently close to ∗. By
Corollary 3.8 again, codimRi (Ii ) = codimR′i (Ii′ ) + 1 = c. The result follows.
Lemma 4.8. Let k be an infinite field, let R = k[x1 , x2 , . . .], and let f ∈ R be a non-zero
homogeneous element. Then there exists an automorphism γ of R (as a graded k-algebra)
such that γ(f ) is monic in x1 .
Proof. See [Eis95, Lemma 13.2(c)].
Corollary 4.9. Assume that for all i in a neighborhood of ∗, the fields ki are infinite. Let
f1 , . . . , fr be homogeneous elements of S. Then f1 , . . . , fr form a regular sequence in S if and
only if f1,i , . . . , fr,i form a regular sequence in Ri for all i sufficiently close to ∗.
Proof. This follows from Theorem 4.7 and Corollary 3.5.
4.3. Stillman’s conjecture.
Theorem 4.10. Given positive integers d1 , . . . , dr there exists an integer N = N(d1 , . . . , dr )
with the following property. If f1 , . . . , fr are homogeneous elements of k[x1 , . . . , xn ], for any
infinite perfect field k and any n, of degrees d1 , . . . , dr and collective strength at least N then
f1 , . . . , fr form a regular sequence.
Proof. Suppose that the theorem is false. Then for each j ∈ N, we can find fj,1, . . . , fj,r in
kj [x1 , x2 , . . . ], with kj infinite and perfect, which fails to be a regular sequence and where the
collective strength goes to ∞ as j → ∞. Choose some function n : I → N where n(i) → ∞
as i → ∗. For each i ∈ I, we let fi,1 , . . . , fi,r be any of the collections in our sequence of
collective strength at least n(i). We let f1 = (fi,1 ), . . . , fr = (fi,r ) be the corresponding
collection in S. By Proposition 4.5, f1 , . . . , fr has infinite collective strength. However, by
Corollary 4.9, f1 , . . . , fr fail to be a regular sequence. This contradicts Theorem 1.1.
For completeness, we now illustrate how Theorem 4.10 implies the existence of small
subalgebras and Stillman’s conjecture. This implication is essentially the same as in [AH16].
Theorem 4.11. Given positive integers d1 , . . . , dr there exists an integer s = s(d1 , . . . , dr )
with the following property. If f1 , . . . , fr are homogeneous elements of k[x1 , . . . , xn ], for any
infinite perfect field k and any n, with deg(fi ) = di , then:
(a) There exists a regular sequence g1 , . . . , gs in k[x1 , . . . , xn ], where each gi is homogeneous of degree at most max(d1 , . . . , dr ), such that f1 , . . . , fr are contained in the
subalgebra k[g1 , . . . , gs ].
(b) The ideal (f1 , . . . , fr ) has projective dimension at most s.
14
DANIEL ERMAN, STEVEN V SAM, AND ANDREW SNOWDEN
Proof. (a) To each sequence d = (d1 , . . . , dr ) we attach a monomial y(d) = y1b1 y2b2 · · · where
bj is the number of times j appears in d. If there is an ideal (f1 , . . . , fr ) of type d that
fails to be a regular sequence, then by Theorem 4.10 there is some N, depending only on d,
such that some k-linear homogeneous combination of the fi has strength ≤ N. Without loss
of generality, weP
may replace one of our elements with this linear combination, and call it
′
fi . Taking fi = N
j=1 aj gj , and replacing fi by the gj , we get an ideal of type d and where
′
y(d) < y(d ) in the revlex order, and where the difference in total degree is at most N −1. In
particular, given y(d) there are only a finite number of possible monomials y(d′ ) that could
arise in this way. The descendants of y(d) thus form a tree with finitely many branches at
each node and with no infinite chains, and there are thus only finitely many descendants of
y(d). Letting s be the max total degree of a descendant of y(d), it follows that f1 , . . . , fr
can be embedded in a subalgebra k[g1 , . . . , gs ] where the gi form a regular sequence.
(b) Choose g1 , . . . , gs as in (a). Since g1 , . . . , gs form a regular sequence, the extension
k[g1 , . . . , gs ] ⊆ k[x1 , . . . , xn ] is flat. Thus, if G is the minimal free resolution of (f1 , . . . , fs )
over k[g1 , . . . , gs ], then the extension of G to k[x1 , . . . , xn ] is the minimal free resolution of
this ideal over k[x1 , . . . , xn ]. In particular, the projective dimension of (f1 , . . . , fs ) is ≤ s.
5. The inverse limit ring
5.1. Inverse limit polynomial ring. Recall that AJJx1 , x2 , . . .KK denotes the inverse limit
of the standard-graded polynomial rings A[x1 , . . . , xn ] in the category of graded rings. We
let K denote a ring containing A, and we write αn : K ⊗A AJJx1 , x2 , . . .KK → K[x1 , . . . , xn ] for
the natural surjection. We set R = K ⊗A AJJx1 , x2 , . . .KK.
The following hypothesis will be used repeatedly.
Hypothesis 5.1. A is an integral domain with fraction field K. If the characteristic p of K
is positive, we assume furthermore that the Frobenius map on A is surjective (so that K is
perfect) and that K is infinite.
Remark 5.2. If A is normal and its fraction field K is perfect, then because a1/p satisfies
the integral equation xp −a, it lies in A. Thus, we can often arrange to satisfy Hypothesis 5.1
by replacing A with its integral closure in an algebraic closure of K.
The following is an analogue of Theorem 4.7 and Corollary 4.9. It implies Theorem 1.1.
Theorem 5.3. Suppose Hypothesis 5.1 holds (except we do not require K to be infinite).
Then R is a polynomial ring.
Proof. We use the criteria of §2. If p = 0, then the partial derivatives dxd i show that R has
enough derivations.
Now suppose that p > 0. We claim that the Hasse derivatives corresponding to dxd i
(Example 2.7) provide R with enough Hasse derivations. Let f ∈ R be such that dxd i f = 0
for all i. This implies that f ∈ K ⊗A AJJxp1 , xp2 , . . .KK. In particular, we can write f = g/a
where a ∈ A and g ∈ AJJxp1 , xp2 , . . .KK. Since the Frobenius map is surjective on A and K, both
g and a are pth powers, which implies that f is also a pth power.
Remark 5.4. The perfectness hypothesis in Theorem 5.3 can be relaxed. For example,
suppose k is a field of characteristic p such that k is a finite extension of the subfield kp ,
and let R = kJJx1 , x2 , . . .KK. Then kRp consists exactly of all (possibly infinite) k-linear
combinations of pth powers of monomials; this uses the hypothesis on k. Thus if f 6∈ kRp
BIG POLYNOMIAL RINGS AND STILLMAN’S CONJECTURE
15
then some Hasse derivative will not kill f , and so R has enough derivations, and is thus a
polynomial ring by Remark 2.12.
Theorem 5.5. Suppose Hypothesis 5.1 holds. Let f1 , . . . , fs ∈ R and let I = (f1 , . . . , fs ).
(a) For any n ≫ 0, we have that codimR (I) = codimK[x1,...,xn ] (αn (I)).
(b) The sequence f1 , . . . , fs forms a regular sequence if and only if αn (f1 ), . . . , αn (fs )
forms a regular sequence for all n ≫ 0.
(c) If αn (f1 ), . . . , αn (fs ) forms a regular sequence for some n, then αm (f1 ), . . . , αm (fs )
forms a regular sequence for all m ≥ n.
Proof. (a) We prove this by induction on c = codimR (I), which is finite by Corollary 3.4.
If c = 0 then I = 0 and the statement is immediate. Now let c > 0 and pick f ∈ I
nonzero. Let n large enough so that αn (f ) 6= 0. Since K is infinite, there is a graded Kalgebra automorphism γ of K[x1 , . . . , xn ] such that γαn (f ) is monic over K[x2 , . . . , xn ] [Eis95,
Lemma 13.2(c)]. If γ ′ is the automorphism of R which acts by γ on x1 , . . . , xn and which acts
trivially on the other xi , then αn (γ ′ f ) = γαn f . We may thus assume that f is monic over
K ⊗A AJJx2 , x3 , . . .KK. The rest of the proof is essentially identical to the proof of Theorem 4.7.
(b) This is an immediate consequence of (a) and Corollary 3.5.
(c) By Corollary 3.5, we have codim αn (I) = s and it suffices to show that codim αn+1 (I) =
s. Since K[x1 , . . . , xn+1 ]/(αn+1 (I) + (xn+1 )) is isomorphic to K[x1 , . . . , xn ]/αn (I), the principal ideal theorem implies that codim αn+1 (I) is either s or s + 1. However, αn+1 (I) is
generated by s elements, so its codimension is at most s. Thus codim αn+1 (I) = s.
Definition 5.6. Fix a ring A, a field k, and a point y ∈ Spec(A)(k). For f ∈ AJJx1 , x2 , . . .KK,
we let fy denote the image of f in kJJx1 , x2 , . . .KK. Similarly, for an AJJx1 , x2 , . . .KK-module M,
we let My = kJJx1 , x2 , . . .KK ⊗AJJx1 ,x2 ,...KK M. We note that k ⊗A AJJx1 , x2 , . . .KK is not generally
isomorphic to kJJx1 , x2 , . . .KK.
If instead f ∈ A[x1 , . . . , xn ], then we let fy denote its image in k[x1 , . . . , xn ]. If M is an
A[x1 , . . . , xn ]-module, then we let My = k[x1 , . . . , xn ] ⊗A[x1 ,...,xn] M.
Corollary 5.7. Suppose Hypothesis 5.1 holds. Let f1 , . . . , fs ∈ AJJx1 , x2 , . . .KK be elements
whose images in R form a regular sequence. There exists a dense open set U ⊆ Spec(A)
such that for any algebraically closed field k and any y ∈ U(k), the elements f1,y , . . . , fs,y
form a regular sequence in kJJx1 , x2 , . . .KK.
Proof. By Theorem 5.5, there is some n so that αn (f1 ), . . . , αn (fs ) ∈ K[x1 , . . . , xn ] forms
a regular sequence. Let gi = αn (fi ), considered as an element of A[x1 , . . . , xn ]. Let Q =
A[x1 , . . . , xn ]/(g1 , . . . , gs ) and let π : Spec(Q) → Spec(A). Since the generic fiber of π has
dimension n − s, it follows that the locus U ⊆ Spec(A) of points whose fibers have dimension
n − s is dense and Zariski open by semicontinuity of fiber dimension [Stacks, 05F6].
Let k be an algebraically closed field and let y ∈ U(k). Since dim(Q ⊗A k) = n − s,
it follows that g1,y , . . . , gs,y forms a regular sequence. But gi,y equals αn (fi,y ), and thus by
Theorem 5.5(c) and (b), we have that f1,y , . . . , fs,y forms a regular sequence.
Lemma 5.8. If k is a perfect field and f1 , . . . , fs ∈ kJJx1 , x2 , . . .KK is a regular sequence, then
i′ : k[f1 , . . . , fs ] → kJJx1 , x2 , . . .KK is faithfully flat.
Proof. By Theorem 5.3, we can write kJJx1 , x2 , . . .KK ∼
= k[V]. There is a finite subset H ⊆ V
such that f1 , . . . , fs ∈ k[H]. Since the fi form a regular sequence, we can extend this to a
16
DANIEL ERMAN, STEVEN V SAM, AND ANDREW SNOWDEN
maximal regular sequence, f1 , . . . , fs , g1 , . . . , gr on k[H]. The map i′ factors as
i
i
i
1
2
3
k[f1 , . . . , fs ] −→
k[f1 , . . . , fs , g1 , . . . , gr ] −→
k[H] −→
kJJx1 , x2 , . . .KK.
For each extension, the larger ring is free over the smaller ring. Both i1 and i3 are extensions
of polynomial rings. For i2 , freeness follows from [BH93, Proposition 2.2.11] (the statement
there is for a local ring, but the proof also works for a graded ring).
5.2. Constant Betti tables over an open subset. For a graded ring R with R0 = k a
field, we set βi,j (M) = dimk TorR
i (M, k)j . The Betti table of M is the collection of all βi,j .
Definition 5.9. Let A be a commutative ring and let U ⊆ Spec(A) be a locally closed
subset. Let M be a finitely presented, graded module over either AJJx1 , x2 , . . .KK or over a
polynomial ring over A. We say that M has a constant Betti table over U if for every
algebraically closed field k and every y ∈ U(k), the Betti table of My is the same. (Recall
that My is defined in Definition 5.6).
The following lemma, which is likely known to experts, shows that a finitely presented
module over a finite polynomial ring has a constant Betti table over an open subset.
Lemma 5.10. Let A be an integral domain and let R = A[y1 , . . . , yr ] be a graded polynomial
ring over A, with deg(yi ) ≥ 1 for 1 ≤ i ≤ r. If M ′ is a finitely presented, graded R-module,
then M ′ has a constant Betti table over some dense, open subset U ⊆ Spec(A).
∂p
∂
Proof. Let K be the fraction field of A and let G′K = [0 → G′K,p → · · · →1 G′K,0] be the
minimal free resolution of K ⊗A M ′ over K[y1 , . . . , yr ]. Represent each ∂i by a matrix
ϕi . The entries of each ϕi have positive degree and, by multiplying by an element in A
if needed, we may assume that the entries of each ϕi also lie in R. These matrices can
then be used to define a bounded, graded complex G′ of free R-modules. By construction,
K ⊗A coker(G′1 → G′0 ) is isomorphic to K ⊗A M ′ . Since both M ′ and coker(G′1 → G′0 ) are
finitely presented, this isomorphism extends to an isomorphism over Ag for some g ∈ A.
Let N be the direct sum of the homology modules Hi (G′ ) for 1 ≤ i ≤ p. Since N is a finitely
generated module over the finitely presented extension A → A[y1 , . . . , yr ], [Stacks, 051S]
implies that there is some h ∈ A such that Nh is free over Ah . But taking homology
commutes with localization, so K ⊗Ah Nh is isomorphic to the homology of the acyclic
complex K ⊗A G′ . Since K ⊗Ah Nh = 0 and Nh is free, this implies that Nh = 0 and G′h is
acyclic. By a similar argument, there exists k ∈ A such that the cokernel of G′ is free over
Ak . In sum, if f = ghk, then G′f is a free resolution of Mf′ , and Mf′ is a flat Af -module.
Let U = Spec(Af ). Let k be a field and y ∈ U(k). Since Mf′ is flat over Af , k ⊗Af G′f is a
free resolution of My′ . The resolution is minimal since each entry of ϕi had positive degree,
and this remains true under localization at f and specialization to k. The Betti table of My′
is thus determined by the free modules in G′f , and so it does not depend on y.
Lemma 5.11. Suppose Hypothesis 5.1 holds. Let M be a finitely presented, graded AJJx1 , x2 , . . .KKmodule. There exist:
(a) elements f1 , . . . , fs ∈ AJJx1 , x2 , . . .KK whose images in R form a regular sequence, and
(b) an element g ∈ A and a finitely presented Ag [f1 , . . . , fs ]-module M ′ such that the
extension of M ′ to Ag ⊗A AJJx1 , x2 , . . .KK is isomorphic to Ag ⊗A M.
BIG POLYNOMIAL RINGS AND STILLMAN’S CONJECTURE
17
Proof. Let U be a set of homogeneous elements of R+ such that R = K[U]. Since any f ∈ R
can be written as a fraction with numerator in AJJx1 , x2 , . . .KK and denominator in A, we may
rescale each element of U so that it lies in AJJx1 , x2 , . . .KK. Since R = K[U] is a polynomial
ring, for any element f ∈ R, there is a finite subset U′ ⊆ U, and an element γ ∈ A such that
f lies in Aγ [U′ ]. The same holds for any finite collection of elements in R.
Let Φ be a finite presentation matrix of M. By the above discussion, we can find distinct
elements f1 , . . . , fs ∈ U and g ∈ A such that each entry of Φ lies in Ag [f1 , . . . , fs ]. Let Φ′ be
the same matrix as Φ, but considered as a map of graded, free Ag [f1 , . . . , fs ]-modules and
let M ′ be the cokernel of Φ′ . By construction, the extension of M ′ to Ag ⊗A AJJx1 , x2 , . . .KK
is isomorphic to Ag ⊗A M. The elements f1 , . . . , fs ∈ R form a regular sequence as they are
“variables” (elements of U).
Theorem 5.12. Suppose Hypothesis 5.1 holds. If M is a finitely presented, graded AJJx1 , x2 , . . .KKmodule, then M has a constant Betti table over some dense open subset U ⊆ Spec(A).
Proof. We apply Lemma 5.11, and let M ′ be the Ag [f1 , . . . , fs ]-module satisfying the conclusion of that lemma. Applying Lemma 5.10 to M ′ , we can assume that M ′ has constant
Betti table over a dense open subset U1 ⊆ Spec(Ag ). By Corollary 5.7, we can find a dense
open subset U2 ⊆ Spec(Ag ) where for all algebraically closed fields k and all y ∈ U2 (k), the
sequence f1,y , . . . , fs,y forms a regular sequence. We let U = U1 ∩ U2 .
Let k be an algebraically closed field and let y ∈ U(k). We have a commutative diagram
Ag [f1 , . . . , fs ]
/
k[f1,y , . . . , fs,y ]
i′
AJJx1 , x2 , . . .KK
/
Ag ⊗A AJJx1 , x2 , . . .KK
/
kJJx1 , x2 , . . .KK
where the extension of the k[f1,y , . . . , fs,y ]-module My′ by i′ is My . By Lemma 5.8, i′ is
faithfully flat, and thus the Betti table of My′ is the same as the Betti table of My . Since M ′
has a constant Betti table over U, the module M also has a constant Betti table over U.
Corollary 5.13. (We do not assume Hypothesis 5.1.) Let A be an integral domain. If M
is a finitely presented, graded AJJx1 , x2 , . . .KK-module, then M has a constant Betti table over
some dense open subset U ⊆ Spec(A).
Proof. Let A be the integral closure of A in an algebraic closure of K, and let K be the
fraction field of A. Let M be the extension of M to AJJx1 , x2 , . . .KK. Since A and K satisfy
Hypothesis 5.1 (see Remark 5.2), Theorem 5.12 implies that M has a constant Betti table
over some dense open subset U ⊆ Spec(A).
Since the integral morphism Spec(A) → Spec(A) is closed [Stacks, 01WM], the image of
U in Spec(A) contains a dense open set U. Let k be an algebraically closed field and let
y ∈ U(k). By integrality, there is k-point y ′ lying over y, and by definition of U, y ′ ∈ U (k).
The map y ′ → y induces an isomorphism of M y′ and My as kJJx1 , x2 , . . .KK-modules, and they
therefore have the same Betti table. Thus M has a constant Betti table over U.
5.3. Connection with GL-noetherianity and Stillman’s conjecture. We now combine
Corollary 5.13 with [Dra17] to prove Stillman’s conjecture.
Throughout this section we fix a ground field k. Fix degrees d1 , . . . , dr . Let S be the set of
pairs (α, i) where 1 ≤ i ≤ r and α ranges over all exponent vectors of degree di in the variables
18
DANIEL ERMAN, STEVEN V SAM, AND ANDREW SNOWDEN
P
x1 , x2 , . . . . Let A = k[cα,i | (α, i) ∈ S]. For 1 ≤ i ≤ r, let fei = cα,i xα ∈ AJJx1 , x2 , . . .KK be a
universal polynomial of degree di . We let Q = AJJx1 , x2 , . . .KK/(fe1 , . . . , fer ). If k′ is a field over
k, then there is a bijection between Spec(A)(k′ ) and tuples f1 , . . . , fr ∈ k′ JJx1 , x2 , . . .KK where
deg(fi ) = di ; with notation from Definition 5.6, this bijection is given by y ∈ Spec(A)(k′ ) ↔
fe1,y , . . . , fer,y .
There is a natural change of basis action by the group scheme GL∞ on Spec(A). The
AJJx1 , x2 , . . .KK-module Q is equivariant with respect to this action.
Theorem 5.14. The space Spec(A) decomposes into a finite disjoint union of locally closed
subsets {Uj } such that Q has a constant Betti table over Uj for each j. In particular, there are
only finitely many distinct Betti tables among all ideals (f1 , . . . , fr ) ⊆ k′ [x1 , . . . , xn ] generated
in degrees d1 , . . . , dr , for all n and all fields k′ over k.
Proof. Applying Corollary 5.13, we have that Q has a constant Betti table over a dense, open
subset U ′ ⊆ Spec(A). Let U be the union of all GL∞ translates of U ′ . Since Betti tables are
GL∞ -invariant, Q has a constant Betti table over U. By [Dra17, Theorem 1], Spec(A) \ U
consists of finitely many irreducible components, each of which is GL∞ -invariant. Passing
to a component, we can apply the same argument. Continuing in this way, we obtain the
desired stratification of Spec(A), and it is finite by [Dra17, Theorem 1].
For any field k′ , the canonical map k′ [x1 , . . . , xn ] ⊗k′ k′ JJxn+1 , . . .KK → k′ JJx1 , x2 , . . .KK is an
isomorphism. It follows that for f1 , . . . , fr ∈ k′ [x1 , . . . , xn ] the Betti table of k′ [x1 , . . . , xn ]/(f1 , . . . , fr )
is the same as that of k′ JJx1 , x2 , . . .KK/(f1 , . . . , fr ), and this implies the final statement of the
theorem for algebraically closed fields k′ . To get the statement for arbitrary k′ , we let k′
be an algebraic closure of k′ and note that the extension k′ [x1 , . . . , xn ] → k′ [x1 , . . . , xn ] is
faithfully flat, and hence Betti tables are unchanged under this extension.
Remark 5.15. It would be interesting to extend [Dra17, Theorem 1] to spaces over Z, as
this would yield characteristic free bounds in the above result.
Remark 5.16. Theorem 5.14 slightly generalizes Stillman’s conjecture, as it also applies to
ideals (f1 , . . . , fr ) in kJJx1 , x2 , . . .KK that use an infinite number of variables.
Remark 5.17. The proof of Theorem 5.14 is much less self-contained than our ultraproduct
proof, however it is distinctly different in character: it does not rely on the notion of strength,
but rather on a generalized noetherianity principle. This is pursued in more detail in [ESS]
to obtain generalizations of Stillman’s conjecture.
References
[AK13] Allen Altman and Steven Kleiman, A term of commutative algebra, 2013.
http://web.mit.edu/18.705/www/13Ed.pdf. ↑8
[AH16] Tigran Ananyan and Melvin Hochster, Small subalgebras of polynomial rings and Stillman’s conjecture (2016). arXiv:1610.09268v1. ↑1, 2, 3, 13
[BH93] Winfried Bruns and Jürgen Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics, vol. 39, Cambridge University Press, Cambridge, 1993. ↑16
[Dra14] Jan Draisma, Noetherianity up to symmetry, Combinatorial algebraic geometry, Lecture Notes in
Math., vol. 2108, Springer, Cham, 2014, pp. 33–61. ↑4
[Dra17] Jan Draisma, Topological noetherianity for polynomial functors (2017). arXiv:1705.01419v1. ↑1,
3, 5, 17, 18
[DLL18] Jan Draisma, Michal Lasoń, and Anton Leykin, Stillman’s Conjecture via generic initial ideals
(2018). arXiv:1802.10139v1. ↑3
BIG POLYNOMIAL RINGS AND STILLMAN’S CONJECTURE
19
[Eis95] David Eisenbud, Commutative algebra with a view toward algebraic geometry, Graduate Texts in
Mathematics, vol. 150, Springer-Verlag, New York, 1995. ↑9, 13, 15
[ESS] Daniel Erman, Steven V Sam, and Andrew Snowden, Stillman type bounds via twisted commutative
algebra. In preparation. ↑18
[Gol03] David M. Goldschmidt, Algebraic functions and projective curves, Graduate Texts in Mathematics,
vol. 215, Springer-Verlag, New York, 2003. ↑5
[MM65] John W. Milnor and John C. Moore, On the structure of Hopf algebras, Ann. of Math. (2) 81
(1965), 211–264. ↑3
[PS09] Irena Peeva and Mike Stillman, Open problems on syzygies and Hilbert functions, J. Commut.
Algebra 1 (2009), no. 1, 159–195. ↑2
[Sch10] Hans Schoutens, The use of ultraproducts in commutative algebra, Lecture Notes in Mathematics,
vol. 1999, Springer-Verlag, Berlin, 2010. ↑3, 10
[Ser97] Jean-Pierre Serre, Galois cohomology, Springer-Verlag, Berlin, 1997. Translated from the French
by Patrick Ion and revised by the author. ↑3, 4
[Sjö80] Gunnar Sjödin, Hopf algebras and derivations, J. Algebra 64 (1980), no. 1, 218–229. ↑3
[Sne98a] Jan Snellman, Gröbner bases and normal forms in a subring of the power series ring on countably
many variables, J. Symbolic Comput. 25 (1998), no. 3, 315–328. ↑3
[Sne98b] Jan Snellman, A graded subring of an inverse limit of polynomial rings, 1998.
http://www.diva-portal.org/smash/get/diva2:195258/FULLTEXT01.pdf. ↑3
[Stacks] The Stacks Project Authors, Stacks Project, 2017. http://stacks.math.columbia.edu. ↑15, 16,
17
[vdDS84] L. van den Dries and K. Schmidt, Bounds in the theory of polynomial rings over fields. A nonstandard approach, Invent. Math. 76 (1984), no. 1, 77–91. ↑3
Department of Mathematics, University of Wisconsin, Madison, WI
E-mail address: [email protected]
URL: http://math.wisc.edu/~derman/
Department of Mathematics, University of Wisconsin, Madison, WI
E-mail address: [email protected]
URL: http://math.wisc.edu/~svs/
Department of Mathematics, University of Michigan, Ann Arbor, MI
E-mail address: [email protected]
URL: http://www-personal.umich.edu/~asnowden/
| 0math.AC
|
Under review as a conference paper at ICLR 2018
L OSS F UNCTIONS FOR M ULTISET P REDICTION
arXiv:1711.05246v1 [cs.LG] 14 Nov 2017
Sean Welleck, Zixin Yao, Yu Gai, Jialin Mao, Zheng Zhang
New York University Shanghai
{wellecks,zixin.yao,yg1246,jialin.mao,zz}@nyu.edu
Kyunghyun Cho
New York University
CIFAR Azrieli Global Scholar
[email protected]
A BSTRACT
We study the problem of multiset prediction. The goal of multiset prediction is
to train a predictor that maps an input to a multiset consisting of multiple items.
Unlike existing problems in supervised learning, such as classification, ranking
and sequence generation, there is no known order among items in a target multiset, and each item in the multiset may appear more than once, making this problem extremely challenging. In this paper, we propose a novel multiset loss function by viewing this problem from the perspective of sequential decision making.
The proposed multiset loss function is empirically evaluated on two families of
datasets, one synthetic and the other real, with varying levels of difficulty, against
various baseline loss functions including reinforcement learning, sequence, and
aggregated distribution matching loss functions. The experiments reveal the effectiveness of the proposed loss function over the others.
1
I NTRODUCTION
A relatively less studied problem in machine learning, particularly supervised learning, is the problem of multiset prediction. The goal of this problem is to learn a mapping from an arbitrary input
to a multiset of items. This problem appears in a variety of contexts. For instance, in the context of high-energy physics, one of the important problems in a particle physics data analysis is to
count how many physics objects, such as electrons, muons, photons, taus, and jets, are in a collision
event (Ehrenfeld et al., 2011). In computer vision, automatic alt-text, such as the one available on
Facebook,1 is a representative example of multiset prediction (Welleck et al., 2017; Lempitsky &
Zisserman, 2010).2
In multiset prediction, a learner is presented with an arbitrary input and the associated multiset of
items. It is assumed that there is no predefined order among the items, and that there are no further
annotations containing information about the relationship between the input and each of the items
in the multiset. These properties make the problem of multiset prediction unique from other wellstudied problems. It is different from sequence prediction, because there is no known order among
the items. It is not a ranking problem, since each item may appear more than once. It cannot be
transformed into classification, because the number of possible multisets grows exponentially with
respect to the maximum multiset size.
In this paper, we view multiset prediction as a sequential decision making process. Under this view,
the problem reduces to finding a policy that sequentially predicts one item at a time, while the
outcome is still evaluated based on the aggregate multiset of the predicted items. We first propose an
oracle policy that assigns non-zero probabilities only to prediction sequences that result exactly in
the target, ground-truth multiset given an input. This oracle is optimal in the sense that its prediction
1
https://newsroom.fb.com/news/2016/04/using-artificial-intelligenceto-help-blind-people-see-facebook/
2
We however note that such a multiset prediction problem in computer vision can also be solved as segmentation, if fine-grained annotation is available. See, e.g., (He et al., 2017).
1
Under review as a conference paper at ICLR 2018
never decreases the precision and recall regardless of previous predictions. That is, its decision is
optimal in any state (i.e., prediction prefix). We then propose a novel multiset loss which minimizes
the KL divergence between the oracle policy and a parametrized policy at every point in a decision
trajectory of the parametrized policy.
We compare the proposed multiset loss against an extensive set of baselines. They include a sequential loss with an arbitrary rank function, sequential loss with an input-dependent rank function, and
an aggregated distribution matching loss and its one-step variant. We also test policy gradient, as
was done by Welleck et al. (2017) recently for multiset prediction. Our evaluation is conducted on
two sets of datasets with varying difficulties and properties. According to the experiments, we find
that the proposed multiset loss outperforms all the other loss functions.
2
M ULTISET P REDICTION
A multiset prediction problem is a generalization of classification, where a target is not a single
class but a multiset of classes. The goal is to find a mapping from an input x to a multiset Y =
y1 , . . . , y|Y| , where yk ∈ C. Some of the core properties of multiset prediction are
1. the input x is an arbitrary vector.
2. there is no predefined order among the items yi in the target multiset Y.
3. the size of Y may vary depending on the input x.
4. each item in the class set C may appear more than once in Y.
2.1
R ELATED P ROBLEMS IN S UPERVISED L EARNING
Variants of this multiset prediction problem have been extensively studied. However, they differ
from our definition of the problem. Here, we go over each variant and discuss how it differs from
our definition of multiset prediction.
Power Multiset Classification Perhaps the most naive approach to multiset prediction is to transform the class set C into a set M (C) of all possible multisets. This transformation, or the size of
M (C), is not well defined unless some constraints are put in place. If the maximum size of a target
multiset is set to K, the number of all possible multisets is
K
X
(|C| + k − 1)!
k=1
k!(|C| − 1)!
.
With some constant |C|, we notice that this grows exponentially in the maximum size of the target
multiset.
Once the class set C is transformed, we can train a multi-class classifier π that maps an input x
to one of the elements in M (C). However, this is infeasible in practice and generally intractable.
For instance, for the COCO Medium dataset used later in the experiments (see section 4.1), M (C)
has roughly 20 thousand elements while the dataset only contains roughly 40 thousand training
examples. For the full MS COCO dataset, |M (C)| is on the order of 1049 , making it infeasible to
learn a classifier using this method.
Ranking A ranking problem can be considered as learning a mapping from a pair of input x and
one of the items c ∈ C to its score s(x, c). All the items in the class set are then sorted according
to the score, and this sorted order determines the rank of each item. By taking the top-K items
from this sorted list, we can turn this problem of ranking into set prediction. Similarly to multiset
prediction, the input x is arbitrary, and the target is a set without any prespecific order. However,
ranking differs from multiset prediction in that it is unable to handle multiple occurrences of a single
item in the target set.
Aggregated Distribution Matching Instead of considering the target multiset as an actual multiset, one can convert it into a distribution by computing the frequency of each item from the class set
2
Under review as a conference paper at ICLR 2018
in the target multiset. That is,
P
p(y|x) =
yi ∈Y
Iyi =y
,
|Y|
where I· is an indicator function. Then, we can simply minimize a divergence between this distribution and the predictive distribution from a model. This loss function works only when the conditional
distribution p(y|x) substantially differs from the marginal distribution p(y), since the model would
resort to a trivial solution of predicting the marginal distribution regardless of the input x.
We describe this approach in more detail in Sec. 3.1, and test it against our proposal in the experiments.
Sequence prediction A sequence prediction problem ischaracterized as finding a mapping from
an input x to a sequence of classes Y = y1 , . . . , y|Y| . Representative examples of sequence
prediction include machine translation, automatic speech recognition and other tagging problems,
such as part-of-speech tagging, in natural language processing. Similarly to multiset prediction, the
input x is arbitrary, and an item in the class set C may appear more than once in the target sequence.
It is, however, different from multiset prediction in that there is a clear, predetermined order of items
in the target sequence. We detail this sequence prediction approach later in Sec. 3.2.
2.2
M ULTISET L OSS F UNCTION FOR M ULTISET P REDICTION
In this paper, we propose a novel loss function, called multiset loss, for the problem of multiset
prediction. This loss function is best motivated by treating the multiset prediction problem as a
sequential decision making process with a model being considered a policy π. This policy takes as
input the input x and all the previously predicted classes ŷ<t at time t, and outputs the distribution
over the next class to be predicted. That is, πθ (yt |ŷ<t , x). This policy is parametrized with a set θ
of parameters.
We first define a free label multiset at time t as
Definition 1 (Free Label Multiset).
Yt ← Yt−1 \ {ŷt−1 }
ŷt−1 is the prediction made by the policy at time t − 1.
This free label multiset Yt contains all the items that remain to be predicted after t − 1 predictions
by the policy.
We then construct an oracle policy π∗ . This oracle policy takes as input a sequence of predicted
labels ŷ<t and the input x. It outputs a distribution whose entire probability (1) is evenly distributed
over all the items in the free label multiset Yt . In other words,
Definition 2 (Oracle).
1
|Yt | , if yt ∈ Yt
π∗ (yt |ŷ<t , x) =
0,
otherwise
An interesting and important property of this oracle is that it is optimal given any prefix ŷ<t with
respect to both precision and recall. This is intuitively clear by noticing that the oracle policy allows
only a correct item to be selected. We call this property the optimality of the oracle.
Remark 1. Given an arbitrary prefix ŷ<t ,
Prec(ŷ<t , Y) ≤ Prec(ŷ<t ∪ ŷ, Y) and Rec(ŷ<t , Y) ≤ Rec(ŷ<t ∪ ŷ, Y),
for any ŷ ∼ π∗ (ŷ<t , x).
The proof is given in Appendix A.
From the remark above, it follows that the oracle policy is an optimal solution to the problem of
multiset prediction in terms of precision and recall.
Remark 2.
Prec(ŷ≤|Y| , Y) = 1 and Rec(ŷ≤|Y| , Y) = 1,
for all ŷ≤|Y| ∼ π∗ (y1 |x)π∗ (y2 |y1 , x) · · · π∗ (y|Y| |y<|Y| , x).
3
Under review as a conference paper at ICLR 2018
The proof can be found in Appendix B.
It is trivial to show that sampling from such an oracle policy would never result in an incorrect
prediction. That is, this oracle policy assigns zero probability to any sequence of predictions that is
not a permutation of the target multiset.
Remark 3.
|Y|
Y
π∗ (yt |y<t , x) = 0, if multiset(y1 , . . . , y|Y| ) 6= Y.
t=1
In short, this oracle policy tells us at each time step t which of all the items in the class set C must
be selected. This optimality allows us to consider a step-wise loss between a parametrized policy πθ
and the oracle policy π∗ , because the oracle policy provides us with an optimal decision regardless
of the quality of the prefix generated so far. We thus propose to minimize the KL divergence from
the oracle policy to the parametrized policy at each step separately. This divergence is defined as
X 1
KL(π∗t kπθt ) = H(π∗t ) −
log πθ (yj |ŷ<t , x),
(1)
| {z }
|Yt |
const. w.r.t. θ
yj ∈|Yt |
H(π∗t )
where
is the entropy of the oracle policy at time step t. This entropy term can be safely
ignored when learning πθ , since it is constant with respect to θ. We define
Lt (x, Y, ŷ<t , θ) = KL(π∗t kπθt ) − H(π∗t )
(2)
and call it a per-step loss function. We note that it is indeed possible to use another divergence in
the place of the KL divergence.
It is intractable to minimize the per-step loss from Eq. (2) for every possible state (ŷ<t , x), since
the size of the state space grows exponentially with respect to the size of a target multiset. We
thus propose here to minimize the per-step loss only for the state, defined as a pair of the input x
and the prefix ŷ<t , visited by the parametrized policy πθ . That is, we generate an entire trajectory
(ŷ1 , . . . , ŷT ) by executing the parametrized policy until either all the items in the target multiset
have been predicted or the predefined maximum number of steps have passed. Then, we compute
the loss function at each time t based on (x, ŷ<t ), for all t = 1, . . . , T . The final loss function is
then the sum of all these per-step loss functions.
Definition 3 (Multiset Loss Function).
T
X
1 X
log πθ (yj |ŷ<t , x),
L(x, Y, θ) = −
|Yt |
t=1
yj ∈Yt
where T is the smaller of the smallest t for which Yt = ∅ and the predefined maximum number of
steps allowed.
As was shown by Ross et al. (2011), the use of the parametrized policy πθ instead of the oracle policy
π∗ allows the upper bound on the learned policy’s error to be linear with respect to the size of the
target multiset. If the oracle policy had been used, the upper bound would have grown quadratically
with respect to the size of the target multiset. To confirm this empirically, we test the following three
alternative strategies for executing the parametrized policy πθ in the experiments:
1. Greedy search: ŷt = arg maxy log πθ (y|ŷ<t , x)
2. Stochastic sampling: ŷt ∼ πθ (y|ŷ<t , x)
3. Oracle sampling: ŷt ∼ π∗ (y|ŷ<t , x)
Once the proposed multiset loss is minimized, we evaluate the learned policy by greedily selecting
each item from the policy.
Variable-Sized Target Multiset We have defined the proposed loss function for multiset prediction while assuming that the size of the target multiset was known. However, this is a major limitation, and we relax this constraint by introducing an additional termination policy. The termination
policy πs outputs a stop distribution given the predicted sequence of items ŷ<t and the input x. Because the size of the target multiset is known during training, we simply train this termination policy
in a supervised way using a binary cross-entropy loss. At evaluation time, we simply threshold the
predicted stop probability at a predefined threshold (0.5).
4
Under review as a conference paper at ICLR 2018
3
OTHER L OSS F UNCTIONS
In addition to the proposed multiset loss function, we propose three more loss functions for multiset
prediction. They serve as baselines in our experiments later.
3.1
AGGREGATED D ISTRIBUTION M ATCHING
In the case of distribution matching, we consider the target multiset Y as a set of samples from a single, underlying distribution q ∗ over the class set C. This underlying distribution can be empirically
estimated by counting the number of occurrences of each item c ∈ C in Y. That is,
1 X
q∗ (c|x) =
Iy=c ,
|Y|
y∈Y
where I is the indicator function as before.
Similarly, we can construct an aggregated distribution computed by the parametrized policy πθ .
As with the proposed multiset loss in Def. 3, we first execute πθ to predict a multiset Ŷ. This is
converted into an aggregated distribution qθ in the same way as we turned the target multiset into
the oracle aggregate distribution.
Learning is equivalent to minimizing the divergence between these two distributions. In this paper,
we test two types of divergences. The first one is from a family of Lp distances defined as
Lpdm (x, Y, θ) = kq∗ − qθ kp ,
where q∗ and q are the vectors representing the corresponding categorical distributions. The other is
a usual KL divergence defined earlier in Eq. (1):
X
LKL
q∗ (c|x) log qθ (c|x).
dm (x, Y, θ) =
c∈C
One major issue with this approach is that minimizing the divergence between the aggregated distributions does not necessarily result in the optimal policy (see the oracle policy in Def. 2.) That is,
a policy that minimizes this loss function may assign non-zero probability to an incorrect sequence
of predictions, unlike the oracle policy. This is due to the invariance of the aggregated distribution
to the order of predictions. Later when analyzing this loss function, we empirically notice that a
learned policy often has a different behaviour from the oracle policy, for instance, reflected by the
increasing entropy of the action distribution over time.
One-Step Variant We can train an one-step predictor with this aggregate distribution matching
criterion, instead of learning a policy πθ . That is, a predictor outputs both a point qθ (·|x) in a |C|dimensional simplex and the size ˆlθ (x) of the target multiset. Then, for each unique item c ∈ C, the
number of its occurrences in the predicted multiset Ŷ is
#(c) = round(qθc (x) · ˆl(x)).
The corresponding loss function is then
X
L1-step (x, Y, θ) =
q∗ (c|x) log qθ (c|x) + λ(ˆlθ (x) − |Y|)2 ,
c∈C
where λ > 0 is a coefficient for balancing the contributions from the two terms.
A major weakness of this one-step variant, compared to the approaches based on sequential decision
making, is the lack of modelling dependencies among the items in the predicted multiset. We test
this approach in the experiments later and observe this lack of output dependency modelling results
in substantially worse prediction accuracy.
3.2
S EQUENCE PREDICTION WITH A PREDEFINED ORDER
All the loss functions defined so far have not relied on the availability of an existing order of items
in a target multiset. However, by turning the problem of multiset prediction into sequential decision
5
Under review as a conference paper at ICLR 2018
making, minimizing such a loss function is equivalent to capturing an order of items in the target
multiset implicitly. Here, we instead describe an approach based on explicitly defining an order in
advance. This will serve as a baseline later in the experiments.
We first define a rank function r that maps from one of the unique items in the class set c ∈ C to a
unique integer. That is, r : C → Z. This function assigns the rank of each item and is used to order
items yi in a target multiset Y. This results in a sequence S = (s1 , . . . , s|Y| ), where r(si ) ≥ r(sj )
for all j > i, and si ∈ Y. With this target sequence S created from Y using the rank function r, we
define a sequence loss function as
Lseq (x, S, θ) = −
|S|
X
log πθ (st |s<t , x).
t=1
Minimizing this loss function is equivalent to maximizing the conditional log-probability of the
sequence S given x.
This sequence loss function has two clear disadvantages. First, it does not take into account the actual behaviour of the policy πθ (see, e.g., Bengio et al., 2015; Daumé III & Marcu, 2005; Ross et al.,
2011). This makes a learned policy potentially vulnerable to cascading error at test time. Second
and more importantly, this loss function requires a pre-specified rank function r. Because multiset
prediction does not come with such a rank function by definition, we must design an arbitrary rank
function, and the final performance varies significantly based on the choice. We demonstrate this
variation in section 4.3.
Input-Dependent Rank Function When the input x has a well-known structure, and an object
within the input for each item in the target multiset is annotated, it is possible to devise a rank
function per input. A representative example is an image input with bounding box annotations.
Here, we present two input-dependent rank functions in such a case.
First, a spatial rank function rspatial assigns an integer rank to each item in a given target multiset Y
such that
rspatial (yi |x) < rspatial (yj |x), if posx (xi ) < posx (xj ) and posy (xi ) < posy (xj ),
where xi and xj are the objects corresponding to the items yi and yj .
Second, an area rank function rarea decides the rank of each label in a target multiset according to
the size of the corresponding object inside the input image:
rarea (yi |x) < rarea (yj |x), if area(xi ) < area(xj ).
The area may be determined based on the size of a bounding box or the number of pixels, depending
on the level of annotation.
We test these two image-specific input-dependent rank functions against a random rank function in
the experiments.
3.3
R EINFORCEMENT L EARNING
In (Welleck et al., 2017), an approach based on reinforcement learning was proposed for multiset
prediction. Instead of assuming the existence of an oracle policy, this approach solely relies on a
reward function r designed specifically for multiset prediction. The reward function is defined as
1,
if ŷt ∈ Yt
r(ŷt , Yt ) =
−1, otherwise
The goal is then to maximize the sum of rewards over a trajectory of predictions from a parametrized
policy πθ . The final loss function is
" T
#
X
LRL = −Eŷ∼πθ
r(ŷ<t , Yt ) − λH(πθ (ŷ<t , x)) ,
(3)
t=1
where the second term inside the expectation is the negative entropy multiplied with a regularization
coefficient λ. The second term encourages the exploration during training. As in (Welleck et al.,
6
Under review as a conference paper at ICLR 2018
2017), we use REINFORCE (Williams, 1992) to stochastically minimize the loss function above
with respect to πθ .
This loss function is optimal in that the return, i.e., the sum of the step-wise rewards, is maximized
when both the precision and recall are maximal (= 1). In other words, the oracle policy, defined in
Def. 2, maximizes the expected return.
However, this approach of reinforcement learning is known to be difficult, with a high variance (Peters & Schaal, 2008). This is especially true here, as the size of the state space grows exponentially
with respect to the size of the target multiset, and the action space of each step is as large as the
number of unique items in the class set.
4
E XPERIMENTS AND A NALYSIS
In this section, we extensively evaluate the proposed multiset loss function against various baseline
loss functions presented throughout this paper. More specifically, we focus on its applicability and
performance on image-based multiset prediction.
4.1
DATASETS
MNIST Multi MNIST Multi is a class of synthetic datasets. Each dataset consists of multiple
100x100 images, each of which contains a varying number of digits from the original MNIST (LeCun et al., 1998). We vary the size of each digit and also add clutters. In the experiments, we
consider the following variants of MNIST Multi:
• MNIST Multi (4): |Y| = 4, 20-50 pixel digits
• MNIST Multi (1-4): |Y| ∈ {1, . . . , 4}, 20-50 pixel digits
• MNIST Multi (10): |Y| = 10, 20 pixel digits
Each dataset has a training set with 70,000 examples and a test set with 10,000 examples. We
randomly sample 7,000 examples from the training set to use as a validation set, and train with the
remaining 63,000 examples.
MS COCO As a real-world dataset, we use Microsoft COCO (Lin et al., 2014) which includes
natural images with multiple objects. Compared to MNIST Multi, each image in MS COCO has
objects of more varying sizes and shapes, and there is a large variation in the number of object
instances per image which spans from 1 to 91. The problem is made even more challenging with
many overlapping and occluded objects.
To control the difficulty in order to better study the loss functions, we create the following two
variants:
• COCO Easy: |Y| = 2, 10,230 training examples, 24 classes
• COCO Medium: |Y| ∈ {1, . . . , 4}, 44,121 training examples, 23 classes
In both of the variants, we only include images whose |Y| objects are large and of common classes.
An object is defined to be large if the object’s area is above the 40-th percentile across the train set
of MS COCO. After reducing the dataset to have |Y| large objects per image, we remove images
1
containing only objects of rare classes. A class is considered rare if its frequency is less than |C|
,
where C is the class set. These two stages ensure that only images with a proper number of large
objects are kept. We do not use fine-grained annotation (pixel-level segmentation and bounding
boxes) except for creating input-dependent rank functions from Sec. 3.2.
For each variant, we hold out a randomly sampled 15% of the training examples as a validation set.
We form separate test sets by applying the same filters to the COCO validation set. The test set sizes
are 5,107 for COCO Easy and 21,944 for COCO Medium.
7
Under review as a conference paper at ICLR 2018
4.2
M ODELS
MNIST Multi We use three convolutional layers of channel sizes 10, 10 and 32, followed by
a convolutional long short-term memory (LSTM) layer (Xingjian et al., 2015). At each step, the
feature map from the convolutional LSTM layer is average-pooled spatially and fed to a softmax
classifier. In the case of the one-step variant of aggregate distribution matching, the LSTM layer is
skipped.
MS COCO We use a ResNet-34 (He et al., 2016) pretrained on ImageNet (Deng et al., 2009) as a
feature extractor. The final feature map from this ResNet-34 is fed to a convolutional LSTM layer,
as described for MNIST Multi above. We do not finetune the ResNet-34 based feature extractor.
Training and evaluation For each loss, a
model was trained for 200 epochs (350 for
MNIST Multi 10). After each epoch, exact
match was computed on the validation set. The
model state with the highest validation exact
match was used for evaluation on the test set.
Table 1: Influence of the choice of a rank function
on the sequence prediction loss function
MNIST Multi (4) COCO Easy
EM
F1
EM
F1
Random 0.920
When evaluating a trained policy, we use
Area
0.529
greedy decoding and auxiliary stop prediction
Spatial
0.917
for determining the size of a predicted multiset.
Each predicted multiset is compared against the
ground-truth target multiset, and we report both the accuracy based
F-1 score (F1).
0.977
0.830
0.976
0.721
0.700
0.675
0.779
0.763
0.738
on the exact match (EM) and
More details about the model architectures and training are in the Appendix C.
4.3
E XPERIMENT 1: I NFLUENCE OF A R ANK F UNCTION ON S EQUENCE P REDICTION
First, we investigate the sequence loss function
Lseq from Sec. 3.2, while varying a rank funcTable 2: Selection Strategies
tion. We test three alternatives: a random rank
MNIST Multi (10) COCO Easy
function3 r and two input-dependent rank funcEM
F1
EM
F1
tions rspatial and rarea . We compare these rank
functions on MNIST Multi (4) and COCO Easy Greedy
0.920 0.992
0.702 0.788
validation sets.
Stochastic 0.907 0.990
0.700 0.790
Oracle
0.875 0.986
0.696 0.780
We present the results in Table 1. It is clear
from the results that the performance of the sequence prediction loss function is dependent on the choice of a rank function. In the case of MNIST
Multi, the area-based rank function was far worse than the other choices. However, this was not
true on COCO Easy, where the spatial rank function was worst among the three. In both cases, we
have observed that the random rank function performed best, and from here on, we use the random
rank function in the remaining experiments. This set of experiments firmly suggests the need of an
order-invariant multiset loss function, such as the multiset loss function proposed in this paper.
4.4
E XPERIMENT 2: E XECUTION S TRATEGIES FOR THE M ULTISET L OSS F UNCTION
In this set of experiments, we compare the three execution strategies for the proposed multiset loss
function, illustrated in Sec. 3. They are greedy decoding, stochastic sampling and oracle sampling.
We test them on MNIST Multi (10) and COCO Easy.
As shown in Table 2, greedy decoding and stochastic sampling, both of which consider states that are
likely to be visited by the parametrized policy, outperform the oracle sampling. This is consistent
with the theory by Ross et al. (2011). Although the first two strategies perform comparably to
each other, across both of the datasets and the two evaluation metrics, greedy decoding tends to
outperform stochastic sampling. We conjecture this is due to better matching between training and
3
The random rank function is generated before training and held fixed. We verified that generating a new
random rank function for each batch significantly decreased performance.
8
Under review as a conference paper at ICLR 2018
Table 3: Loss Function Comparison on the variants of MNIST Multi
MNIST Multi (4) MNIST Multi (1-4) MNIST Multi (10)
EM
F1
EM
F1
EM
F1
Proposed L
LRL
L1dm
LKL
dm
Lseq
L1-step
0.950
0.912
0.921
0.908
0.906
0.210
0.987
0.977
0.978
0.974
0.973
0.676
0.953
0.945
0.918
0.908
0.891
0.055
0.981
0.980
0.969
0.962
0.952
0.598
0.920
0.665
0.239
0.256
0.592
0.032
0.992
0.970
0.714
0.874
0.946
0.854
testing in the case of greedy decoding. Thus, from here on, we use greedy decoding when training a
model with the proposed multiset loss function.
4.5
E XPERIMENT 3: L OSS F UNCTION C OMPARISON
We now compare the proposed multiset loss function against the five baseline loss functions: reinforcement learning LRL , aggregate distribution matching–L1dm and LKL
dm –, its one-step variant L1-step ,
and sequence prediction Lseq .
MNIST Multi We present the results on the MNIST Multi variants in Table 3. On all three variants
and according to both metrics, the proposed multiset loss function outperforms all the others. The
reinforcement learning based approach closely follows behind. Its performance, however, drops as
the number of items in a target multiset increases. This is understandable, as the variance of policy
gradient grows as the length of an episode grows. A similar behaviour was observed with sequence
prediction as well as aggregate distribution matching. We were not able to train any decent models
with the one-step variant of aggregate distribution matching. This was true especially in terms
of exact match (EM), which we attribute to the one-step variant not being capable of modelling
dependencies among the predicted items.
MS COCO Similarly to the results on the
variants of MNIST Multi, the proposed multiset loss function matches or outperforms all
the others on the two variants of MS COCO,
as presented in Table 4. On COCO Easy, with
only two objects to predict per example, both
aggregated distribution matching (with KL divergence) and the sequence loss functions are
as competitive as the proposed multiset loss.
The other loss functions significantly underperform these three loss functions, as they did on
MNIST Multi.
Table 4: Loss function comparison on the variants
of MS COCO
COCO Easy COCO Medium
EM
F1
EM
F1
Proposed L
LRL
L1dm
LKL
dm
Lseq
L1-step
0.702
0.672
0.533
0.714
0.709
0.552
0.788
0.746
0.614
0.763
0.774
0.664
0.481
0.425
0.221
0.444
0.457
0.000
0.639
0.564
0.085
0.591
0.592
0.446
The performance gap between the proposed
loss and the others, however, grows substantially on the more challenging COCO Medium, which
has more objects per example. The proposed multiset loss outperforms the aggregated distribution
matching with KL divergence by 3.7 percentage points on exact match and 4.8 on F1. This is analogous to the experiments on the MNIST Multi variants, where the performance gap increased when
moving from four to ten digits.
4.6
A NALYSIS : E NTROPY E VOLUTION
One property of the oracle policy defined in Sec. 2.2 is that the entropy of the predictive distribution
strictly decreases over time, i.e., Hπ∗ (y|ŷ<t , x) > Hπ∗ (y|ŷ≤t , x). This is a natural consequence
from the fact that there is no pre-specified rank function, because the oracle policy cannot prefer any
item from the others in a free label multiset. Hence, we examine here how the policy learned based
9
Under review as a conference paper at ICLR 2018
on each loss function compares to the oracle policy in terms of per-step entropy. We consider the
policies trained on MNIST Multi (10), where the differences among them were most clear.
As shown in Fig. 1, the policy trained on MNIST Multi
(10) using the proposed multiset loss closely follows the
oracle policy. The entropy decreases as the predictions
are made. The decreases can be interpreted as concentrating probability mass on progressively smaller free labels sets. The variance is quite small, indicating that this
strategy is uniformly applied for any input.
The policy trained with reinforcement learning retains
a relatively low entropy across steps, with a decreasing
trend in the second half. We carefully suspect the low entropy in the earlier steps is due to the greedy nature of policy gradient. The policy receives a high reward more easily by choosing one of many possible choices in an earlier step than in a later step. This effectively discourages
the policy from exploring all possible trajectories during
training.
Figure 1: Comparison of per-step entropies of predictive distributions compared over the validation set.
On the other hand, the policy found by aggregated distribution matching (LKL
dm ) has the opposite behaviour. The entropy in general grows as more predictions
are made. To see why this is sub-optimal, consider the final (10th) step. Assuming the first nine predictions {ŷ1 , ..., ŷ9 } were correct (i.e. they form a subset of Y), there is only one correct class left
for the final prediction ŷ10 . The high entropy, however, indicates that the model is placing a significant amount of probability on incorrect sequences. We believe such a policy is found by minimizing
the aggregated distribution matching loss function because it cannot properly distinguish between
policies with increasing and decreasing entropies.
The increasing entropy also indicates that the policy has learned a rank function implicitly and is
fully relying on it. Given some unknown free label multiset, inferred from the input, this policy uses
the implicitly learned rank function to choose one item from this set. We conjecture this reliance
on an inferred rank function, which is by definition sub-optimal,4 resulted in lower performance of
aggregate distribution matching.
5
C ONCLUSION
We have extensively investigated the problem of multiset prediction in this paper. We rigorously
defined the problem, and proposed to approach it from the perspective of sequential decision making. In doing so, an oracle policy was defined and shown to be optimal, and a new loss function,
called multiset loss, was introduced as a means to train a parametrized policy for multiset prediction.
The experiments on two families of datasets, MNIST Multi variants and MS COCO variants, have
revealed the effectiveness of the proposed loss function over other loss functions including reinforcement learning, sequence, and aggregated distribution matching loss functions. The success of
the proposed multiset loss brings in new opportunities of applying machine learning to various new
domains, including high-energy physics.
R EFERENCES
Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence
prediction with recurrent neural networks. In Advances in Neural Information Processing Systems, pp. 1171–1179, 2015.
Hal Daumé III and Daniel Marcu. Learning as search optimization: Approximate large margin methods for structured prediction. In Proceedings of the 22nd international conference on Machine
learning, pp. 169–176. ACM, 2005.
4
Note that this synthetic data was created without any specific rank function.
10
Under review as a conference paper at ICLR 2018
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale
hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009.
IEEE Conference on, pp. 248–255. IEEE, 2009.
W Ehrenfeld, R Buckingham, J Cranshaw, T Cuhadar Donszelmann, T Doherty, E Gallas, J Hrivnac,
D Malon, M Nowak, M Slater, F Viegas, E Vinek, Q Zhang, and the ATLAS Collaboration. Using
tags to speed up the atlas analysis process. Journal of Physics: Conference Series, 331(3):032007,
2011. URL http://stacks.iop.org/1742-6596/331/i=3/a=032007.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.
770–778, 2016.
Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. arXiv preprint
arXiv:1703.06870, 2017.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to
document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
V. Lempitsky and A. Zisserman. Learning to count objects in images. In Advances in Neural
Information Processing Systems, 2010.
Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr
Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European
conference on computer vision, pp. 740–755. Springer, 2014.
J. Peters and S. Schaal. Reinforcement learning of motor skills with policy gradients. Neural
Networks, 21(4):682–697, May 2008.
Stéphane Ross, Geoffrey J Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In International Conference on Artificial Intelligence
and Statistics, pp. 627–635, 2011.
Sean Welleck, Kyunghyun Cho, and Zheng Zhang. Saliency-based sequential image attention with
multiset prediction. In Advances in neural information processing systems, 2017. to appear.
Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement
learning. Machine learning, 8(3-4):229–256, 1992.
SHI Xingjian, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo.
Convolutional lstm network: A machine learning approach for precipitation nowcasting. In Advances in neural information processing systems, pp. 802–810, 2015.
A
P ROOF OF R EMARK 1
Proof. Note that the precision with ŷ<t is defined as
P
Prec(ŷ<t , Y) =
y∈ŷ<t
Iy∈Y
|ŷ<t |
.
Because ŷ ∼ π∗ (ŷ<t , x) ∈ Yt ,
Prec(ŷ≤t , Y) =
1+
P
y∈ŷ<t
Iy∈Y
1 + |ŷ<t |
.
Then,
1 − Prec(ŷ<t , Y)
≥ 0,
1 + |ŷ<t |
because 0 ≤ Prec(ŷ<t , Y) ≤ 1 and |ŷ<t | ≥ 0. The equality holds when Prec(ŷ<t , Y) = 1.
Prec(ŷ≤t , Y) − Prec(ŷ<t , Y) =
11
Under review as a conference paper at ICLR 2018
Similarly, we start with the definition of the recall:
P
Rec(ŷ<t , Y) =
y∈ŷ<t
Iy∈Y
|Y|
.
Because ŷ ∼ π∗ (ŷ<t , x) ∈ Yt ,
Rec(ŷ≤t , Y) =
1+
P
y∈ŷ<t
Iy∈Y
|Y|
.
Since the denominator is identical,
Rec(ŷ≤t , Y) − Rec(ŷ<t , Y) =
B
1
≥ 0.
|Y|
P ROOF OF R EMARK 2
Proof. When t = 1,
Prec(ŷ≤1 , Y) = 1,
because ŷ1 ∼ π∗ (∅, x) ∈ Y. From Remark 1, we know that
Prec(ŷ≤t , Y) = Prec(ŷ<t , Y),
when Prec(ŷ<t , Y) = 1. By induction, Prec(ŷ≤|Y| , Y) = 1.
From the proof of Remark 1, we know that the recall increases by
that
1
,
Rec(ŷ≤1 , Y) =
|Y|
1
Y
each time, and we also know
when t = 1. After |Y| − 1 steps of executing the oracle policy, the recall becomes
|Y|
X 1
1
Rec(ŷ≤|Y| , Y) =
+
= 1.
|Y|
|Y|
0
t =2
C
M ODEL D ESCRIPTIONS
Figure 2: Graphical illustration of a predictor used throughout the experiments.
12
Under review as a conference paper at ICLR 2018
Model An input x is first processed by a tower of convolutional layers, resulting in a feature
0
0
volume of size w0 × h0 with d feature maps, i.e., H = φ(x) ∈ Rw ×h ×d . At each time step t,
0
0
we resize the previous prediction’s embedding emb(ŷt−1 ) ∈ R(w )(h ) to be a w0 × h0 tensor and
0
0
concatenate it with H, resulting in H̃ ∈ Rw ×h ×(d+1) . This feature volume is then fed into a stack
0
0
of convolutional LSTM layers. The output from the final convolutional LSTM layer C ∈ Rw ×h ×q
0 P 0
P
w
h
is spatially average-pooled, i.e., c = w01h0 i=1 j=1 Ci,j,· ∈ Rq . This feature vector c is then
turned into a conditional distribution over the next item after affine transformation followed by a
softmax function. When the one-step variant of aggregated distribution matching is used, we skip
Pw0 Ph0
the convolutional LSTM layers, i.e., c = w01h0 i=1 j=1 Hi,j,· ∈ Rd .
See Fig. 2 for the graphical illustration of the entire network. See Table 5 for the details of the
network for each dataset.
Data
MNIST Multi
MS COCO
Table 5: Network Architectures
CNN
emb(ŷt−1 )
conv 5 × 5 max-pool 2 × 2 feat 10
conv 5 × 5 max-pool 2 × 2 feat 10
conv 5 × 5 max-pool 2 × 2 feat 32
81
ResNet-34
361
ConvLSTM
conv 3 × 3 feat 32
conv 3 × 3 feat 32
conv 3 × 3 feat 512
conv 3 × 3 feat 512
Preprocessing For MNIST Multi, we do not preprocess the input at all. In the case of MS COCO,
input images are of different sizes. Each image is first resized so that its larger dimension has
600 pixels, then along its other dimension is zero-padded to 600 pixels and centered, resulting in a
600x600 image.
Training The model is trained end-to-end, except ResNet-34 which remains fixed after being
pretrained on ImageNet. For all the experiments, we train a neural network using Adam (Kingma
& Ba, 2014) with a fixed learning rate of 0.001, β of (0.9, 0.999) and of 1e-8. The learning rate
was selected based on the validation performance during the preliminary experiments, and the other
parameters are the default values. For MNIST Multi, the batch size was 64, and for COCO was 32.
13
| 1cs.CV
|
1
TPC Together with Overlapped Time Domain
Multiplexing System Based on Turbo Structure
Hao Zheng, Mingjun Xing, Yutao Yue, Xue Li, Daoben Li and Chunlin Ji
Kuang-Chi Institute of Advanced Technology, Shenzhen, China
arXiv:1708.02910v1 [cs.IT] 9 Aug 2017
Email: {hao.zheng, chunlin.ji}@kuang-chi.org
Abstract
Overlapped time domain multiplexing (OvTDM) is a novel technique for utilizing inter-symbol interference (ISI) to benefit
a communication system. We implement the OvTDM technique based on turbo structure and associate a turbo product code
(TPC) to construct a novel coded turbo-structure OvTDM system. Two schemes of the iterative receiver and soft input and soft
output (SISO) decoding algorithms are presented. Simulation results show the advantage of structures in this paper. In addition,
an attractive transmission rate and symbol efficiency of the designed system can also be observed.
I. I NTRODUCTION
It is well known that most traditional communication systems are designed based on Nyquist criterion [1], in which intersymbol interference (ISI) should be avoided between consequent symbols. In fact, the communication system without ISI is
physically unrealizable. On the other hand, people tend to design a communication system with controlled ISI, such as Fasterthan-Nyquist (FTN) signaling [2] and partial response signaling (PRS) [3]. However, these methods also treat the overlap
between symbols as interference and do not really utilize it to collect the extra gain.
Based on ISI to benefit a communication system, overlapped time domain multiplexing (OvTDM) is proposed in [4]-[6].
The idea of OvTDM is to shift a data-weighted and band-limited multiplexing waveform in the time domain to achieve an
overlap between different transmitted symbols and a high transmission rate. It can help to form a convolution structure among
consequent symbols, so OvTDM can also be regarded as one kind of waveform convolution coding. In the OvTDM system,
these overlapped parts are never regarded as ISI but rather as a beneficial encoding constraint relationship. Therefore, OvTDM
can show great performance of the transmission rate and the symbol efficiency that is defined as bits per symbol [4][5].
Maximum likelihood sequence detection (MLSD) [7] and maximum a posteriori (MAP) detection can be utilized to detect
OvTDM signals. From the point of view of waveform convolution coding, the detection process can also be called OvTDM
decoding.
Most previous studies have focused on the single structure of the OvTDM system. However, another way to improve the
OvTDM system is to extend the coding structure [8][9]. So, we construct a turbo structure inspired by the turbo code [10] for
OvTDM. In addition, turbo product code (TPC) is employed as the forward-error-correction (FEC) module together with the
turbo structure of OvTDM. In comparison to another popular FEC code, the low-density parity-check code (LDPC) [11][12],
TPC is suitable to be constructed with a shorter code length and requires fewer iterations for decoding [13][14]. Finally,
a coding system with three layers is formed, which contains the FEC code, the turbo structure and OvTDM respectively.
Comparative simulation studies with the coded QAM system show the significant advantage of the coded turbo-strucutre of
OvTDM.
2
Fig. 1. The transmitter structure of the turbo-structure OvTDM together with TPC.
II. S YSTEM D ESCRIPTION
A. OvTDM Scheme
In the OvTDM system, we artificially introduce ISI to form an overlap among different symbols. The mapping relationship
between original bits and constellation symbols can follow the rule of ordinary modulation methods. Assuming the transmitted
signals followed BPSK as x = [x0 , x1 , · · · , xL−1 ] with length L and the multiplexing waveform as h(t), t ∈ [0, Ts ) with
symbol duration Ts , then the transmitted signal after overlapping can be expressed as
s(t) =
L−1
X
i=0
xi h(t − iTs /K) =
L−1
X
i=0
xi h(t − i∆T )
(1)
where ∆T = Ts /K is the time shift between symbols. In (1), K is the number of overlapped symbols during ∆T , which is
named the overlapping coefficient or the constraint length. Notice that, the larger the coefficient K is, the more serious the
ISI introduced.
B. Turbo-Structure OvTDM with FEC
Fig.1 shows the transmitter of the turbo-structure OvTDM with FEC. The coded sequence that has passed the FEC encoder
and one interleaver is sent to the first OvTDM encoder for the I channel. Meanwhile, the same sequence is sent to pass the
other interleaver to form another sequence with a different order, which is encoded by the second OvTDM encoder for the Q
channel. The output sequences from both two OvTDM encoders are combined to form a complex sequence.
The decoding process at the receiver is the key to the system design. It is based on the idea of iteration and the extrinsic
information transformation [10]. During each iteration, the extrinsic information is exchanged between different decoders. The
interleaver and the de-interleaver are employed to match the order of received sequences. Exchanging extrinsic information
with low correlation can help to improve performance with the increase of iterations. Together with FEC, two schemes are
addressed as follows:
Scheme A: After one round of decoding between two OvTDM decoders, the soft information is sent to the FEC decoder.
Then the FEC decoder sends the soft information back to the OvTDM decoder. The model is shown in Fig.2. In this scheme,
the FEC decoder needs to be involved in every iteration of the turbo structure.
Scheme B: As shown in Fig.3, OvTDM decoders work iteratively and do not exchange soft information with the FEC decoder
at first. After several iterations, the soft information is sent to the FEC decoder. Unlike in Scheme A, the soft information is
only exchanged once between OvTDM decoders and the FEC decoder.
The Viterbi algorithm [15] is a good choice for detecting the OvTDM signals and selecting the possible sequence that is
nearest to the received sequence [4][5]. However, in the turbo structure, we need to exchange the extrinsic information for
original bits, so soft output detecting algorithms [16][17] are more appropriate.
3
Fig. 2. The receiver structure of the turbo-structure OvTDM together with TPC (Scheme A).
Fig. 3. The receiver structure of the turbo-structure OvTDM together with TPC (Scheme B).
III. D ECODING A LGORITHMS
A. MAP Algorithm for OvTDM
As discussed in the above sections, OvTDM utilizes the ISI as the encoding constraint. Thus, it can also be represented as
a trellis graph [5]. The BCJR algorithm [16] is regarded as an optimal MAP method based on the trellis graph, so it can be
modified to calculate the maximum a posteriori probability (APP) for OvTDM encoded bits.
Denoting the input bit at time t as xt and the received sequence at the receiver as r with length N , the log-likelihood-ratio
(LLR) of APP of xt is
λt = log
= log
p(xt = +1|r)
p(xt = −1|r)
P
p(St−1 , St , r)
(St−1 ,St ),xt =+1
P
(2)
p(St−1 , St , r)
(St−1 ,St ),xt =−1
where St and St−1 are assumed as states of time t and t − 1 based on the trellis graph. According to properties of OvTDM,
p(St−1 , St , r) can be further expressed by
p(St−1 , St , r) = αt (St−1 )γt (St−1 , St )βt (St )
where αt (St−1 ) and βt (St ) can be calculated by forward and backward recursion:
X
αt (St−1 ) =
αt−1 (St−2 )γt (St−1 , St )
(3)
(4)
St−1
βt (St ) =
X
βt+1 (St+1 )γt+1 (St , St+1 )
(5)
St+1
Assuming the corresponding output bit at time t after OvTDM encoding as yt , then
γt (St−1 , St ) = p(xt )p(rt |yt )
(6)
4
It is worth noting that there is no input bit in the tail part of OvTDM. So, βL (SL ) can be initialized directly. In the AWGN
channel,
1
βL (SL ) = √
K
2πσ
N
P
2
(rt − yt )
i=L
exp −
2
2σ
(7)
where σ 2 is the variance of noise.
Let LLR of a prior probability and the extrinsic information be µt and et , in the iterative decoder, the extrinsic information
can be obtained by
et = λt − µt
(8)
which is the output of the OvTDM decoding module.
B. FBBA for TPC
One mainstream method of TPC decoding algorithm is augmented list decoding (ALD) [13][14]. The key idea of ALD is to
form a list including the most likely codewords. Based on ALD, the Fang-Battail-Buda-Algorithm (FBBA) [14] is an efficient
SISO algorithm for TPC decoding to achieve near-optimum performance. FBBA is concluded as follows:
Step 1: Sort the received symbols d in a decreasing order according to the LLR metric l.
Step 2: Permute the check matrix H according to the permutation pattern from the Step 1. Then, it has to be adjusted by
Gauss-Jordan elimination to obtain a systematic one Hπ that is used to re-code the component codeword to generate a new
one cπ(0) .
Step 3: A codebook list is obtained through the reversal of certain positions of cπ(0) and sorted in an increasing order
according to
π(i)
Z(l, c
)=−
n−1
X
j=0
log
π(i)
)
π(0)
)
p(lj |cj
p(lj |cj
(9)
where cπ(i) and n are denoted as the ith codeword in the codebook list and the component codeword length.
Step 4: The soft output can be calculated by the first codeword cπ(0) and the opposite to the first codeword in jth position
in the codebook list.
ρj =
1
(||l − cπ(opp) ||2 − ||l − cπ(0) ||2 )
4
(10)
Following the above process, soft outputs can be calculated. Generally, four iterations are sufficient for the BER performance
to converge.
IV. S IMULATION S TUDIES
In this section, we need to investigate the performance through some comparative simulation. We choose the Chebyshev
waveform with attenuation level 80dB as the multiplexing waveform for OvTDM. Extended BCH(64, 57) is employed as the
2
component codeword to construct a squared TPC. Thus, the code rate is RT P C = (57/64) = 0.7932. The AWGN channel is
considered as the transmission channel in the simulation.
As mentioned before, the turbo-structure OvTDM uses the same information sequence for both I and Q channels, so its
equivalent coding rate with TPC is ROvT = 1/2 · RT P C · L/N . When the length of the information sequence is large enough,
5
−1
10
The turbo−structure OvTDM
with TPC (Scheme A)
The turbo−structure OvTDM
with TPC (Scheme B)
64−QAM with TPC
−2
10
−3
BER
10
−4
10
−5
10
−6
10
4
5
6
7
Eb/No(dB)
8
9
10
Fig. 4. BER performance of the turbo-structure OvTDM (K = 6) and 64-QAM with TPC(64, 57)2 .
−1
10
The turbo−structure OvTDM
with TPC (Scheme A)
The turbo−structure OvTDM
with TPC (Scheme B)
256−QAM with TPC
−2
10
−3
BER
10
−4
10
−5
10
−6
10
5
6
7
8
9
10
11
Eb/No(dB)
12
13
14
15
Fig. 5. BER performance of the turbo-structure OvTDM (K = 8) and 256-QAM with TPC(64, 57)2 .
L/N ≈ 1, so we ignore it in the simulation. BPSK is used as the original modulation for the OvTDM system. Thus, the symbol
efficiency of the turbo-structure OvTDM with TPC is ηOvT = ROvT · 2K = RT P C · K (bits/symbol). On the other hand, if
we select M -ary QAM with TPC for comparison, its symbol efficiency is ηQAM = log 2(M ) · RT P C (bits/symbol). In order
to do the comparative studies under the same symbol efficiency, we select K = 6 and 64-QAM in Fig.4 as well as K = 8
and 256-QAM in Fig.5. The BER plots in both Fig.4 and Fig.5 show the significant advantage of the coded turbo-structure
OvTDM. In Fig.4, the same symbol efficiency is 4.7592 (bits/symbol) and the required Eb /N0 of 64-QAM with TPC to
achieve the BER < 10−5 is 10dB, but the turbo-structure OvTDM with TPC can achieve BER < 10−5 at 5.8dB. In Fig.5, the
same symbol efficiency is 6.3456 (bits/symbol) and the required Eb /N0 of the turbo-structure with TPC using Scheme B to
achieve the BER < 10−5 is 7.4dB less than that of 256-QAM with TPC.
Moreover, Fig.6 illustrates the comparison result of the symbol efficiency among different schemes when BER < 10−5 . The
symbol efficiency of the single structure OvTDM has been shown in [5]. Also, we plot the corresponding symbol efficiency
of the Shannon Limit [18]. In Fig.6, the turbo-structure OvTDM with TPC has a obvious improvement, compared with the
single structure OvTDM. On the other hand, with the increase of the symbol efficiency, schemes in this paper can achieve it
at a lower Eb /N0 than the Shannon Limit that represents traditional communication systems.
V. C ONCLUSION
This paper mainly focuses on structures and SISO decoding algorithms of the turbo-structure OvTDM with TPC, which
demostrate a significant improvement over the single structure OvTDM. In addition, compared with the coded QAM system of
6
8
7
bits/symbol
6
The turbo−structure OvTDM
with TPC (Scheme A)
The turbo−structure OvTDM
with TPC (Scheme B)
The single structure
Shannon Limit
5
4
3
2
1
0
2
4
6
8
Eb/No(dB)
10
12
14
16
Fig. 6. Comparison of the symbol efficiency (bits/symbol) among different schemes.
the same symbol efficiency, the BER performance of the turbo-structure OvTDM with TPC is much better. Simulation results
shows the advantage of the turbo-structure OvTDM with TPC in the communication scenario requiring high transmission rate
at a relatively low Eb /N0 .
R EFERENCES
[1] J. G. Proakis, Digital Communications, 4th ed, Macgraw Hill, New York, 2001.
[2] J. E. Mazo, “Faster-than-nyquist signaling,” Bell Syst. Tech. J., vol. 54, no. 8, pp. 1451-1462, Oct. 1975.
[3] P. Kabal and S. Pasupathy, “Partial-Response Signaling,” IEEE Trans. Commun., vol. 23, no. 9, pp. 921-934, Sep. 1975.
[4] D. Li, “A novel high spectral efficiency waveform coding-OVTDM,” International Journal of Wireless Communications and Mobile Computing, Special
Issue: 5G Wireless Communication Systems, vol. 2, no. 4-1, pp. 11-26, Dec. 2014.
[5] D. Li, Waveform Coding Theory of High Spectral Efficiency–OVTDM and Its Application, Science Press, Beijing, 2013.
[6] C. Ji and R. Liu, “Study on a High Spectrum Modulation by Introducing Intersymbol Interference,” in Proc. 2016 IEEE Int. Conf. on Signal Processing,
Communications and Computing, Hong Kong, , pp. 2904-2909, Aug. 2016.
[7] G. D. Forney, “Maximum Likelihood Sequence Estimation of Digital Sequences in the presence of intersymbol interference,” IEEE Trans. Inform.
Theory, vol 18, no. 3, pp. 363-378, May 1972.
[8] X. Dong, “Research on the performance of OvTDM and Turbo-OvTDM technology application in multi-carrier system,” M.Sc. Dissertation, Beijing
Univ. Posts Telecommun., 2013.
[9] B. Liu, “Applications of overlapped multiplexing principle in telecommunications,” Ph.D. Dissertation, Beijing Univ. Posts Telecommun., 2014.
[10] C. Berrou and A. Glavieux, “Near optimum error correcting coding and decoding: Turbo codes,” IEEE Trans. Commun., vol. 44, no. 10, pp. 1261-1271,
Oct. 1996.
[11] R. Gallager, “Low-density parity-check codes,” IRE Trans. Inform. Theory, vol. 8, no. 1, pp. 21-28, Jan. 1968.
[12] D. J. C. MacKay and R. M. Neal, “Near shannon limit performance of low density parity check codes,” Electron. Lett., vol. 33, no. 6, pp. 457-458,
Mar. 1997.
[13] R. Pyndiah, “Near optimum decoding of product codes: block turbo codes,” IEEE Trans. Commun., vol. 46, no. 8, pp. 1003-1010, Aug. 1998.
[14] J. Fang, F. Buda and E. Lemois, “Turbo Product Code: a well suitable solution to wireless packet transmission for very low error rates,” in Proc. 2nd
Int. Symp. on Turbo Codes & Related Topics, France, pp. 101-111, Sep. 2000.
[15] A. J. Viterbi, “Error bounds for convolutional codes and an asymtotically optimum decoding algorithm,” IEEE Trans. Inform. Theory, vol. 13, no. 2,
pp. 260-269, Apr. 1967.
[16] L. R. Bahl, J. Cocke, F. Jelinek, and J. Raviv, “Optimal decoding of linear codes for minimizing symbol error rate,” IEEE Trans. Inform. Theory, vol.
20, no. 2, pp. 284-287, Mar. 1974.
[17] C. Ji, “On Sequential Learning for Parameter Estimation in Particle Algorithms for State-Space Models,” International Journal of Statistics and Probability,
vol. 6, no. 1, pp. 13-23, Jan. 2017.
[18] C. E. Shannon, “A mathematical theory of communication,” Bell Syst. Tech. J., vol. 27, no. 7, pp. 379-423, 1948.
| 7cs.IT
|
Variational Walkback: Learning a Transition
Operator as a Stochastic Recurrent Net
arXiv:1711.02282v1 [stat.ML] 7 Nov 2017
Anirudh Goyal
MILA, Université de Montréal
[email protected]
Surya Ganguli
Stanford University
[email protected]
Nan Rosemary Ke
MILA, École Polytechnique de Montréal
[email protected]
Yoshua Bengio
MILA, Université de Montréal
[email protected]
Abstract
We propose a novel method to directly learn a stochastic transition operator whose
repeated application provides generated samples. Traditional undirected graphical
models approach this problem indirectly by learning a Markov chain model whose
stationary distribution obeys detailed balance with respect to a parameterized energy
function. The energy function is then modified so the model and data distributions
match, with no guarantee on the number of steps required for the Markov chain to
converge. Moreover, the detailed balance condition is highly restrictive: energy
based models corresponding to neural networks must have symmetric weights,
unlike biological neural circuits. In contrast, we develop a method for directly
learning arbitrarily parameterized transition operators capable of expressing nonequilibrium stationary distributions that violate detailed balance, thereby enabling
us to learn more biologically plausible asymmetric neural networks and more general non-energy based dynamical systems. The proposed training objective, which
we derive via principled variational methods, encourages the transition operator to
"walk back" (prefer to revert its steps) in multi-step trajectories that start at datapoints, as quickly as possible back to the original data points. We present a series
of experimental results illustrating the soundness of the proposed approach, Variational Walkback (VW), on the MNIST, CIFAR-10, SVHN and CelebA datasets,
demonstrating superior samples compared to earlier attempts to learn a transition
operator. We also show that although each rapid training trajectory is limited to a
finite but variable number of steps, our transition operator continues to generate
good samples well past the length of such trajectories, thereby demonstrating the
match of its non-equilibrium stationary distribution to the data distribution. Source
Code: http://github.com/anirudh9119/walkback_nips17
1
Introduction
A fundamental goal of unsupervised learning involves training generative models that can understand
sensory data and employ this understanding to generate, or sample new data and make new inferences.
In machine learning, the vast majority of probabilistic generative models that can learn complex probability distributions over data fall into one of two classes: (1) directed graphical models, corresponding
to a finite time feedforward generative process (e.g. variants of the Helmholtz machine (Dayan
et al., 1995) like the Variational Auto-Encoder (VAE) (Kingma and Welling, 2013; Rezende et al.,
2014)), or (2) energy function based undirected graphical models, corresponding to sampling from a
stochastic process whose equilibrium stationary distribution obeys detailed balance with respect to the
energy function (e.g. various Boltzmann machines (Salakhutdinov and Hinton, 2009)). This detailed
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
balance condition is highly restrictive: for example, energy-based undirected models corresponding
to neural networks require symmetric weight matrices and very specific computations which may not
match well with what biological neurons or analog hardware could compute.
In contrast, biological neural circuits are capable of powerful generative dynamics enabling us to
model the world and imagine new futures. Cortical computation is highly recurrent and therefore its
generative dynamics cannot simply map to the purely feed-forward, finite time generative process of
a directed model. Moreover, the recurrent connectivity of biological circuits is not symmetric, and so
their generative dynamics cannot correspond to sampling from an energy-based undirected model.
Thus, the asymmetric biological neural circuits of our brain instantiate a type of stochastic dynamics
arising from the repeated application of a transition operator∗ whose stationary distribution over
neural activity patterns is a non-equilibrium distribution that does not obey detailed balance with
respect to any energy function. Despite these fundamental properties of brain dynamics, machine
learning approaches to training generative models currently lack effective methods to model complex
data distributions through the repeated application a transition operator, that is not indirectly specified
through an energy function, but rather is directly parameterized in ways that are inconsistent with the
existence of any energy function. Indeed the lack of such methods constitutes a glaring gap in the
pantheon of machine learning methods for training probabilistic generative models.
The fundamental goal of this paper is to provide a step to filling such a gap by proposing a novel
method to learn such directly parameterized transition operators, thereby providing an empirical
method to control the stationary distributions of non-equilibrium stochastic processes that do not
obey detailed balance, and match these distributions to data. The basic idea underlying our training
approach is to start from a training example, and iteratively apply the transition operator while
gradually increasing the amount of noise being injected (i.e., temperature). This heating process
yields a trajectory that starts from the data manifold and walks away from the data due to the heating
and to the mismatch between the model and the data distribution. Similarly to the update of a
denoising autoencoder, we then modify the parameters of the transition operator so as to make the
reverse of this heated trajectory more likely under a reverse cooling schedule. This encourages the
transition operator to generate stochastic trajectories that evolve towards the data distribution, by
learning to walk back the heated trajectories starting at data points. This walkback idea had been
introduced for generative stochastic networks (GSNs) and denoising autoencoders (Bengio et al.,
2013b) as a heuristic, and without temperature annealing. Here, we derive the specific objective
function for learning the parameters through a principled variational lower bound, hence we call our
training method variational walkback (VW). Despite the fact that the training procedure involves
walking back a set of trajectories that last a finite, but variable number of time-steps, we find
empirically that this yields a transition operator that continues to generate sensible samples for many
more time-steps than are used to train, demonstrating that our finite time training procedure can sculpt
the non-equilibrium stationary distribution of the transition operator to match the data distribution.
We show how VW emerges naturally from a variational derivation, with the need for annealing
arising out of the objective of making the variational bound as tight as possible. We then describe
experimental results illustrating the soundness of the proposed approach on the MNIST, CIFAR-10,
SVHN and CelebA datasets. Intriguingly, we find that our finite time VW training process involves
modifications of variational methods for training directed graphical models, while our potentially
asymptotically infinite generative sampling process corresponds to non-equilibrium generalizations
of energy based undirected models. Thus VW goes beyond the two disparate model classes of
undirected and directed graphical models, while simultaneously incorporating good ideas from each.
2
The Variational Walkback Training Process
Our goal is to learn a stochastic transition operator pT (s0 |s) such that its repeated application yields
samples from the data manifold. Here T reflects an underlying temperature, which we will modify
during the training process. The transition operator is further specified by other parameters which
must be learned from data. When K steps are chosen to generate a sample, the generative process
QK
has joint probability p(sK
0 ) = p(sK )
t=1 pTt (st−1 |st ), where Tt is the temperature at step t. We
first give an intuitive description of our learning algorithm before deriving it via variational methods
in the next section. The basic idea, as illustrated in Fig. 1 and Algorithm 1 is to follow a walkback
∗
A transition operator maps the previous-state distribution to a next-state distribution, and is implemented by
a stochastic transformation which from the previous state of a Markov chain generates the next state
2
Figure 1: Variational WalkBack framework. The generative process is represented in the blue arrows
with the sequence of pTt (st−1 |st ) transitions. The destructive forward process starts at a datapoint
(from qT0 (s0 )) and gradually heats it through applications of qTt (st |st−1 ). Larger temperatures on
the right correspond to a flatter distribution, so the whole destructive forward process maps the data
distribution to a Gaussian and the creation process operates in reverse.
strategy similar to that introduced in Alain and Bengio (2014). In particular, imagine a destructive
process qTt+1 (st+1 |st ) (red arrows in Fig. 1), which starts from a data point s0 = x, and evolves it
QK
K
stochastically to obtain a trajectory s0 , . . . , sK ≡ sK
0 , i.e., q(s0 ) = q(s0 )
t=1 qTt (st |st−1 ), where
q(s0 ) is the data distribution. Note that the p and q chains will share the same parameters for the
transition operator (one going backwards and one forward) but they start from different priors for
their first step: q(s0 ) is the data distribution while p(s0 ) is a flat factorized prior (e.g. Gaussian).
The training procedure trains the transition operator pT to make reverse transitions of the destructive
process more likely. For this reason we index time so the destructive process operates forward in time,
while the reverse generative process operates backwards in time, with the data distribution occurring
at t = 0. In particular, we need only train the transition operator to reverse time by 1-step at each step,
making it unnecessary to solve a deep credit assignment problem by performing backpropagation
through time across multiple walk-back steps. Overall, the destructive process generates trajectories
that walk away from the data manifold, and the transition operator pT learns to walkback these
trajectories to sculpt the stationary distribution of pT at T = 1 to match the data distribution.
Because we choose qT to have the same parameters as pT , they have the same transition operator but
not the same joint over the whole sequence because of differing initial distributions for each trajectory.
We also choose to increase temperature with time in the destructive process, following a temperature
schedule T1 ≤ · · · ≤ TK . Thus the forward destructive (reverse generative) process corresponds to a
heating (cooling) protocol. This training procedure is similar in spirit to DAE’s (Vincent et al., 2008)
or NET (Sohl-Dickstein et al., 2015) but with one major difference: the destructive process in these
works corresponds to the addition of random noise which knows nothing about the current generative
process during training. To understand why tying together destruction and creation may be a good
idea, consider the special case in which pT corresponds to a stochastic process whose stationary
distribution obeys detailed balance with respect to the energy function of an undirected graphical
model. Learning any such model involves two fundamental goals: the model must place probability
mass (i.e. lower the energy function) where the data is located, and remove probability mass (i.e.
raise the energy function) elsewhere. Probability modes where there is no data are known as spurious
modes, and a fundamental goal of learning is to hunt down these spurious modes and remove them.
Making the destructive process identical to the transition operator to be learned is motivated by the
notion that the destructive process should then efficiently explore the spurious modes of the current
transition operator. The walkback training will then destroy these modes. In contrast, in DAE’s and
NET’s, since the destructive process corresponds to the addition of unstructured noise that knows
nothing about the generative process, it is not clear that such an agnostic destructive process will
efficiently seek out the spurious modes of the reverse, generative process.
We chose the annealing schedule empirically to minimize training time. The generative process
starts by sampling a state sK from a broad Gaussian p∗ (sK ), whose variance is initially equal to
2
the total data variance σmax
(but can be later adapted to match the final samples from the inference
trajectories). Then we sample from pTmax (sK−1 |sK ), where Tmax is a high enough temperature
so that the resultant injected noise can move the state across the whole domain of the data. The
injected noise used to simulate the effects of finite temperature has variance linearly proportional to
3
temperature. Thus if σ 2 is the equivalent noise injected by the transition operator pT at T = 1, we
σ2
choose Tmax = σmax
to achieve the goal of the first sample sK−1 being able to move across the entire
2
range of the data distribution. Then we successively cool the temperature as we sample “previous”
states st−1 according to pT (st−1 |st ), with T reduced by a factor of 2 at each step, followed by n
steps at temperature 1. This cooling protocol requires the number of steps to be
K = log2 Tmax + n,
(1)
in order to go from T = Tmax to T = 1 in K steps. We choose K from a random distribution.
Thus the training procedure trains pT to rapidly transition from a simple Gaussian distribution to
the data distribution in a finite but variable number of steps. Ideally, this training procedure should
then indirectly create a transition operator pT at T = 1 whose repeated iteration samples the data
distribution with a relatively rapid mixing time. Interestingly, this intuitive learning algorithm for a
recurrent dynamical system, formalized in Algorithm 1, can be derived in a principled manner from
variational methods that are usually applied to directed graphical models, as we see next.
Algorithm 1 VariationalWalkback(θ)
Train a generative model associated with a transition operator pT (s | s0 ) at temperature T (temperature 1 for sampling from the actual model), parameterized by θ. This transition operator injects noise
of variance T σ 2 at each step, where σ 2 is the noise level at temperature 1.
Require: Transition operator pT (s | s0 ) from which one can both sample and compute the gradient
of log pT (s|s0 ) with respect to parameters θ, given s and s0 .
2
Require: Precomputed σmax
, initially data variance (or squared diameter).
Require: N1 > 1 the number of initial temperature-1 steps of q trajectory (or ending a p trajectory).
repeat
Set p∗ to be a Gaussian with mean and variance of the data.
σ2
Tmax ← σmax
2
Sample n as a uniform integer between 0 and N1
K ← ceil(log2 Tmax ) + n
Sample x ∼ data (or equivalently sample a minibatch to parallelize computation and process
each element of the minibatch independently)
Let s0 = (x) and initial temperature T = 1, initialize L = 0
for t = 1 to K do
Sample st ∼ pT (s|st−1 )
Increment L ← L + log pT (st−1 |st )
(st−1 |st )
Update parameters with log likelihood gradient ∂ log pT∂θ
If t > n, increase temperature with T ← 2T
end for
Increment L ← L + log p∗ (sK )
Update mean and variance of p∗ to match the accumulated 1st and 2nd moment statistics of the
samples of sK
until convergence monitoring L on a validation set and doing early stopping
3
Variational Derivation of Walkback
The marginal probability of a data point s0 at the end of the K-step generative cooling process is
!
K
X
Y
p(s0 ) =
pT0 (s0 |s1 )
pTt (st−1 |st ) p∗ (sK )
(2)
t=2
sK
1
where sK
1 = (s1 , s2 , . . . , sK ) and v = s0 is a visible variable in our generative process, while the
cooling trajectory that lead to it can be thought of as a latent, hidden variable h = sK
1 . Recall the
decomposition of the marginal log-likelihood via a variational lower bound,
ln p(v) ≡ ln
X
p(v|h)p(h) =
X
h
p(v, h)
+DKL [q(h|v)||p(h|v)].
q(h|v)
{z
}
q(h|v) ln
h
|
L
4
(3)
Here L is the variational lower bound which motivates the proposed training procedure, and q(h|v) is
a variational approximation to p(h|v). Applying this decomposition to v = s0 and h = sK
1 , we find
ln p(s0 ) =
X
q(sk1 |s0 ) ln
sk
1
p(s0 |sk1 )p(sk1 )
+ DKL [q(sk1 |s0 ) || p(sk1 |s0 )].
q(sk1 |s0 )
(4)
Similarly to the EM algorithm, we aim to approximately maximize the log-likelihood with a 2-step
procedure. Let θp be the parameters of the generative model p and θq be the parameters of the
approximate inference procedure q. Before seeing the next example we have θq = θp . Then in the
first step we update θp towards maximizing the variational bound L, for example by a stochastic
gradient descent step. In the second step, we update θq by setting θq ← θp , with the objective to
reduce the KL term in the above decomposition. See Sec. 3.1 below regarding conditions for the
tightness of the bound, which may not be perfect, yielding a possibly biased gradient when we force
the constraint θp = θq . We continue iterating this procedure, with training examples s0 . We can
obtain an unbiased Monte-Carlo estimator of L as follows from a single trajectory:
L(s0 ) ≈
K
X
t=1
ln
pTt (st−1 |st )
+ ln p∗ (sK )
qTt (st |st−1 )
(5)
with respect to pθ , where s0 is sampled from the data distribution qT0 (s0 ), and the single sequence sK
1
is sampled from the heating process q(sK
1 |s0 ). We are making the reverse of heated trajectories more
likely under the cooling process, leading to Algorithm 1. Such variational bounds have been used
successfully in many learning algorithms in the past, such as the VAE (Kingma and Welling, 2013),
except that they use an explicitly different set of parameters for p and q. Some VAE variants (Sønderby
et al., 2016; Kingma et al., 2016) however mix the p-parameters implicitly in forming q, by using the
likelihood gradient to iteratively form the approximate posterior.
3.1
Tightness of the variational lower bound
As seen in (4), the gap between L(s0 ) and ln p(s0 ) is controlled by DKL [q(sk1 |s0 )||p(sk1 |s0 )], and is
therefore tight when the distribution of the heated trajectory, starting from a point s0 , matches the
posterior distribution of the cooled trajectory ending at s0 . Explicitly, this KL divergence is given by
K
X
p(s0 ) Y qTt (st |st−1 )
DKL =
q(sk1 |s0 ) ln ∗
.
(6)
p (sK ) t=1 pTt (st−1 |st )
k
s1
As the heating process q unfolds forward in time, while the cooling process p unfolds backwards in
time, we introduce the time reversal of the transition operator pT , denoted by pR
T , as follows. Under
repeated application of the transition operator pT , state s settles into a stationary distribution πT (s)
at temperature T . The probability of observing a transition st → st−1 under pT in its stationary state
is then pT (st−1 |st )πT (st ). The time-reversal pR
T is the transition operator that makes the reverse
transition equally likely for all state pairs, and therefore obeys
PT (st−1 |st )πT (st ) = PTR (st |st−1 )πT (st−1 )
(7)
pR
T
for all pairs of states st−1 and st . It is well known that
is a valid stochastic transition operator and
has the same stationary distribution πT (s) as pT . Furthermore, the process pT obeys detailed balance
if and only if it is invariant under time-reversal, so that pT = pR
T.
To better understand the KL divergence in (6), at each temperature Tt , we use relation (7) to replace
the cooling process PTt which occurs backwards in time with its time-reversal, unfolding forward in
time, at the expense of introducing ratios of stationary probabilities. We also exploit the fact that q
and p are the same transition operator. With these substitutions in (6), we find
DKL =
X
sk
1
q(sk1 |s0 ) ln
K
K
Y
pTt (st |st−1 ) X
p(s0 ) Y πTt (st )
k
+
q(s
|s
)
ln
.
0
1
p∗ (sK ) t=1 πTt (st−1 )
pR (s |s )
k
t=1 Tt t t−1
(8)
s1
The first term in (8) is simply the KL divergence between the distribution over heated trajectories, and
the time reversal of the cooled trajectories. Since the heating (q) and cooling (p) processes are tied,
this KL divergence is 0 if and only if pTt = pR
Tt for all t. This time-reversal invariance requirement
for vanishing KL divergence is equivalent to the transition operator pT obeying detailed balance at all
temperatures.
5
Now intuitively, the second term can be made small in the limit where K is large and the temperature
sequence is annealed slowly. To see why, note we can write the ratio of probabilities in this term as,
πT
(sK−1 ) πTK (sK )
p(s0 ) πT1 (s1 )
· · · K−1
.
(9)
πT1 (s0 ) πT2 (s1 )
πTK−1 (sK ) p∗ (sK )
which is similar in shape (but arising in a different context) to the product of probability ratios
computed for annealed importance sampling (Neal, 2001) and reverse annealed importance sampling (Burda et al., 2014). Here it is manifest that, under slow incremental annealing schedules, we
are comparing probabilities of the same state under slightly different distributions, so all ratios are
close to 1. For example, under many steps, with slow annealing, the generative process approximately
reaches its stationary distribution, p(s0 ) ≈ πT1 (s0 ).
This slow annealing to go from p∗ (sK ) to p(s0 ) corresponds to the quasistatic limit in statistical
physics, where the work required to perform the transformation is equal to the free energy difference
between states. To go faster, one must perform excess work, above and beyond the free energy difference, and this excess work is dissipated as heat into the surrounding environment. By writing the distributions in terms of energies and free energies: πTt (st ) ∝ e−E(st )/Tt , p∗ (sK ) = e−[EK (sK )−FK ] ,
and p(s0 ) = e−[E0 (s0 )−F0 ] , one can see that the second term in the KL divergence is closely related
to average heat dissipation in a finite time heating process (see e.g. (Crooks, 2000)).
This intriguing connection between the size of the gap in a variational lower bound, and the excess
heat dissipation in a finite time heating process opens the door to exploiting a wealth of work in
statistical physics for finding optimal thermodynamic paths that minimize heat dissipation (Schmiedl
and Seifert, 2007; Sivak and Crooks, 2012; Gingrich et al., 2016), which may provide new ideas
to improve variational inference. In summary, tightness of the variational bound can be achieved
if: (1) The transition operator of p approximately obeys detailed balance, and (2) the temperature
annealing is done slowly over many steps. And intriguingly, the magnitude of the looseness of the
bound is related to two physical quantities: (1) the degree of irreversiblity of the transition operator p,
as measured by the KL divergence between p and its time reversal pR , and (2) the excess physical
work, or equivalently, excess heat dissipated, in performing the heating trajectory.
To check, post-hoc, potential looseness of the variational lower bound, we can measure the degree of
irreversibility of pT by estimating the KL divergence DKL (pT (s0 |s)πT (s) || pT (s|s0 )πT (s0 )), which
is 0 if and only if pT obeys detailed balance and is therefore time-reversal invariant. This quantity
PK
pT (st+1 |st )
1
K
can be estimated by K
t=1 ln pT (st |st+1 ) , where s1 is a long sequence sampled by repeatedly
applying transition operator pT from a draw s1 ∼ πT . If this quantity is strongly positive (negative)
then forward transitions are more (less) likely than reverse transitions, and the process pT is not
time-reversal invariant. This estimated KL divergence can be normalized by the corresponding
entropy to get a relative value (with 3.6% measured on a trained model, as detailed in Appendix).
3.2 Estimating log likelihood via importance sampling
We can derive an importance sampling estimate of the negative log-likelihood by the following
procedure. For each training example x, we sample a large number of destructive paths (as in
Algorithm 1). We then use the following formulation to estimate the log-likelihood log p(x) via
Q
K
∗
pT0 (s0 = x|s1 )
t=2 pTt (st−1 |st ) p (sK )
Q
log Ex∼pD ,qT (x)qT (s1 |s0 (x,))(QK qT (st |st−1 ))
K
t=2
t
0
1
q
(s
|s
)
qT0 (x)qT1 (s1 |s0 = x)
T
t
t−1
t
t=2
(10)
3.3
VW transition operators and their convergence
The VW approach allows considerable freedom in choosing transition operators, obviating the need
for specifying them indirectly through an energy function. Here we consider Bernoulli and isotropic
Gaussian transition operators for binary and real-valued data respectively. The form of the stochastic
state update imitates a discretized version of the Langevin differential equation. The Bernoulli
(1−α)∗st−1 +α∗Fρ (st−1 )
transition operator computes the element-wise probability as ρ = sigmoid(
).
Tt
The Gaussian operator computes a conditional mean and standard deviation via µ = (1 − α) ∗ st−1 +
α ∗ Fµ (st−1 ) and σ = Tt log(1 + eFσ (st−1 ) ). Here the F functions can be arbitrary parametrized
functions, such as a neural net and Tt is the temperature at time step t.
6
A natural question is when will the finite time VW training process learn a transition operator whose
stationary distribution matches the data distribution, so that repeated sampling far beyond the training
time continues to yield data samples. To partially address this, we prove the following theorem:
Proposition 1. If p has enough capacity, training data and training time, with slow enough annealing
and a small departure from reversibility so p can match q, then at convergence of VW training, the
transition operator pT at T = 1 has the data generating distribution as its stationary distribution.
A proof can be found in the Appendix, but the essential intuition is that if the finite time generative
process converges to the data distribution at multiple different VW walkback time-steps, then it
remains on the data distribution for all future time at T = 1. We cannot always guarantee the
preconditions of this theorem but we find experimentally that its essential outcome holds in practice.
4
Related Work
A variety of learning algorithms can be cast in the framework of Fig. 1. For example, for directed
graphical models like VAEs (Kingma and Welling, 2013; Rezende et al., 2014), DBNs (Hinton et al.,
2006), and Helmholtz machines in general, q corresponds to a recognition model, transforming data
to a latent space, while p corresponds to a generative model that goes from latent to visible data in
a finite number of steps. None of these directed models are designed to learn transition operators
that can be iterated ad infinitum, as we do. Moreover, learning such models involves a complex,
deep credit assignment problem, limiting the number of unobserved latent layers that can be used to
generate data. Similar issues of limited trainable depth in a finite time feedforward generative process
apply to Generative Adversarial Networks (GANs) (Goodfellow et al., 2014), which also further
eschew the goal of specifically assigning probabilities to data points. Our method circumvents this
deep credit assignment problem by providing training targets at each time-step; in essence each past
time-step of the heated trajectory constitutes a training target for the future output of the generative
operator pT , thereby obviating the need for backpropagation across multiple steps. Similarly, unlike
VW, Generative Stochastic Networks (GSN) (Bengio et al., 2014) and the DRAW (Gregor et al.,
2015) also require training iterative operators by backpropagating across multiple computational
steps.
VW is similar in spirit to DAE (Bengio et al., 2013b), and NET approaches (Sohl-Dickstein et al.,
2015) but it retains two crucial differences. First, in each of these frameworks, q corresponds to
a very simple destruction process in which unstructured Gaussian noise is injected into the data.
This agnostic destruction process has no knowledge of underlying generative process p that is to
be learned, and therefore cannot be expected to efficiently explore spurious modes, or regions of
space, unoccupied by data, to which p assigns high probability. VW has the advantage of using a
high-temperature version of the model p itself as part of the destructive process, and so should be
better than random noise injection at finding these spurious modes. A second crucial difference is
that VW ties weights of the transition operator across time-steps, thereby enabling us to learn a bona
fide transition operator than can be iterated well beyond the training time, unlike DAEs and NET.
There’s also another related recent approach to learning a transition operator with a denoising cost,
developed in parallel, called Infusion training (Bordes et al., 2017), which tries to reconstruct the
target data in the chain, instead of the previous step in the destructive chain.
5
Experiments
VW is evaluated on four datasets: MNIST, CIFAR10 (Krizhevsky and Hinton, 2009), SVHN (Netzer
et al., 2011) and CelebA (Liu et al., 2015). The MNIST, SVHN and CIFAR10 datasets were used
as is except for uniform noise added to MNIST and CIFAR10, as per Theis et al. (2016), and the
aligned and cropped version of CelebA was scaled from 218 x 178 pixels to 78 x 64 pixels and
center-cropped at 64 x 64 pixels (Liu et al., 2015). We used the Adam optimizer (Kingma and Ba,
2014) and the Theano framework (Al-Rfou et al., 2016). More details are in Appendix and code for
training and generation is at http://github.com/anirudh9119/walkback_nips17.
Image Generation. Figure 3, 5, 6, 7, 8 (see supplementary section) show VW samples on each of
the datasets. For MNIST, real-valued views of the data are modeled. Image Inpainting. We clamped
the bottom part of CelebA test images (for each step during sampling), and ran it through the model.
Figure 1 (see Supplementary section) shows the generated conditional samples.
7
Model
bits/dim ≤
NET (Sohl-Dickstein et al., 2015)
5.40
VW(20 steps)
5.20
Deep VAE
< 4.54
VW(30 steps)
4.40
DRAW (Gregor et al., 2015)
< 4.13
ResNet VAE with IAF (Kingma et al., 2016)
3.11
Table 1: Comparisons on CIFAR10, test set average number of bits/data dimension(lower is better)
6
6.1
Discussion
Summary of results
Our main advance involves using variational inference to learn recurrent transition operators that
can rapidly approach the data distribution and then be iterated much longer than the training time
while still remaining on the data manifold. Our innovations enabling us to achieve this involved: (a)
tying weights across time, (b) tying the destruction and generation process together to efficiently
destroy spurious modes, (c) using the past of the destructive process to train the future of the creation
process, thereby circumventing issues with deep credit assignment (like NET), (d) introducing an
aggressive temperature annealing schedule to rapidly approach the data distribution (e.g. NET takes
1000 steps while VWB only takes 30 steps to do so), and (e) introducing variable trajectory lengths
during training to encourage the generator to stay on the data manifold for times longer than the
training sequence length.
Indeed, it is often difficult to sample from recurrent neural networks for many more time steps than
the duration of their training sequences, especially non-symmetric networks that could exhibit chaotic
activity. Transition operators learned by VW can be stably sampled for exceedingly long times; for
example, in experiments (see supplementary section) we trained our model on CelebA for 30 steps,
while at test time we sampled for 100000 time-steps. Overall, our method of learning a transition
operator outperforms previous attempts at learning transition operators (i.e. VAE, GSN and NET)
using a local learning rule.
Overall, we introduced a new approach to learning non-energy-based transition operators which
inherits advantages from several previous generative models, including a training objective that
requires rapidly generating the data in a finite number of steps (as in directed models), re-using the
same parameters for each step (as in undirected models), directly parametrizing the generator (as in
GANs and DAEs), and using the model itself to quickly find its own spurious modes (the walk-back
idea). We also anchor the algorithm in a variational bound and show how its analysis suggests to use
the same transition operator for the destruction or inference process, and the creation or generation
process, and to use a cooling schedule during generation, and a reverse heating schedule during
inference.
6.2
New bridges between variational inference and non-equilibrium statistical physics
We connected the variational gap to physical notions like reversibility and heat dissipation. This novel
bridge between variational inference and concepts like excess heat dissipation in non-equilbrium
statistical physics, could potentially open the door to improving variational inference by exploiting a
wealth of work in statistical physics. For example, physical methods for finding optimal thermodynamic paths that minimize heat dissipation (Schmiedl and Seifert, 2007; Sivak and Crooks, 2012;
Gingrich et al., 2016), could potentially be exploited to tighten lowerbounds in variational inference.
Moreover, motivated by the relation between the variational gap and reversibility, we verified empirically that the model converges towards an approximately reversible chain (see Appendix) making the
variational bound tighter.
6.3
Neural weight asymmetry
A fundamental aspect of our approach is that we can train stochastic processes that need not exactly
8
obey detailed balance, yielding access to a larger and potentially more powerful space of models. In
particular, this enables us to relax the weight symmetry constraint of undirected graphical models
corresponding to neural networks, yielding a more brain like iterative computation characteristic
of asymmetric biological neural circuits. Our approach thus avoids the biologically implausible
requirement of weight transport (Lillicrap et al., 2014) which arises as a consequence of imposing
weight symmetry as a hard constraint. With VW, this hard constraint is removed, although the
training procedure itself may converge towards more symmetry. Such approach towards symmetry is
consistent with both empirical observations (Vincent et al., 2010) and theoretical analysis (Arora et al.,
2015) of auto-encoders, for which symmetric weights are associated with minimizing reconstruction
error.
6.4 A connection to the neurobiology of dreams
The learning rule underlying VW, when applied to an asymmetric stochastic neural network, yields a
speculative, but intriguing connection to the neurobiology of dreams. As discussed in Bengio et al.
(2015), spike-timing dependent plasticity (STDP), a plasticity rule found in the brain (Markram
and Sakmann, 1995), corresponds to increasing the probability of configurations towards which the
network intrinsically likes to go (i.e., remembering observed configurations), while reverse-STDP
corresponds to forgetting or unlearning the states towards which the network goes (which potentially
may occur during sleep).
In the VW update applied to a neural network, the resultant learning rule does indeed strengthen
synapses for which a presynaptic neuron is active before a postsynaptic neuron in the generative
cooling process (STDP), and it weakens synapses in which a postsynaptic neuron is active before a
presynaptic neuron in the heated destructive process (reverse STDP). If, as suggested, the neurobiological function of sleep involves re-organizing memories and in particular unlearning spurious modes
through reverse-STDP, then the heating destructive process may map to sleep states, in which the
brain is hunting down and destroying spurious modes. In contrast, the cooling generative dynamics
of VW may map to awake states in which STDP reinforces neural trajectories moving towards
observed sensory data. Under this mapping, the relative incoherence of dreams compared to reality
is qualitatively consistent with the heated destructive dynamics of VW, compared to the cooled
transition operator in place during awake states.
6.5 Future work
Many questions remain open in terms of analyzing and extending VW. Of particular interest is the
incorporation of latent layers. The state at each step would now include both visible x and latent
h components. Essentially the same procedure can be run, except for the chain initialization, with
s0 = (x, h0 ) where h0 a sample from the posterior distribution of h given x.
Another interesting direction is to replace the log-likelihood objective at each step by a GAN-like
objective, thereby avoiding the need to inject noise independently on each of the pixels, during
each transition step, and allowing latent variable sampling to inject the required high-level decisions
associated with the transition. Based on the earlier results from (Bengio et al., 2013a), sampling in
the latent space rather than in the pixel space should allow for better generative models and even
better mixing between modes (Bengio et al., 2013a).
Overall, our work takes a step to filling a relatively open niche in the machine learning literature on
directly training non-energy-based iterative stochastic operators, and we hope that the many possible
extensions of this approach could lead to a rich new class of more powerful brain-like machine
learning models.
Acknowledgments
The authors would like to thank Benjamin Scellier, Ben Poole, Tim Cooijmans, Philemon Brakel,
Gaétan Marceau Caron, and Alex Lamb for their helpful feedback and discussions, as well as
NSERC, CIFAR, Google, Samsung, Nuance, IBM and Canada Research Chairs for funding, and
Compute Canada for computing resources. S.G. would like to thank the Simons, McKnight, James S.
McDonnell, and Burroughs Wellcome Foundations and the Office of Naval Research for support. Y.B
would also like to thank Geoff Hinton for an analogy which is used in this work, while discussing
contrastive divergence (personnal communication). The authors would also like to express debt of
gratitude towards those who contributed to theano over the years (as it is no longer maintained),
making it such a great tool.
9
References
Al-Rfou, R., Alain, G., Almahairi, A., and et al. (2016). Theano: A python framework for fast
computation of mathematical expressions. CoRR, abs/1605.02688.
Alain, G. and Bengio, Y. (2014). What regularized auto-encoders learn from the data-generating
distribution. J. Mach. Learn. Res., 15(1):3563–3593.
Arora, S., Liang, Y., and Ma, T. (2015). Why are deep nets reversible: a simple theory, with
implications for training. Technical report, arXiv:1511.05653.
Ba, J. L., Kiros, J. R., and Hinton, G. E. (2016).
arXiv:1607.06450.
Layer normalization.
arXiv preprint
Bengio, Y., Mesnard, T., Fischer, A., Zhang, S., and Wu, Y. (2015). An objective function for STDP.
CoRR, abs/1509.05936.
Bengio, Y., Mesnil, G., Dauphin, Y., and Rifai, S. (2013a). Better mixing via deep representations.
In Proceedings of the 30th International Conference on Machine Learning (ICML-13), pages
552–560.
Bengio, Y., Thibodeau-Laufer, E. r., Alain, G., and Yosinski, J. (2014). Deep generative stochastic networks trainable by backprop. In Proceedings of the 31st International Conference on International
Conference on Machine Learning - Volume 32, ICML’14, pages II–226–II–234. JMLR.org.
Bengio, Y., Yao, L., Alain, G., and Vincent, P. (2013b). Generalized denoising auto-encoders as
generative models. In NIPS’2013, arXiv:1305.6663.
Bordes, F., Honari, S., and Vincent, P. (2017). Learning to generate samples from noise through
infusion training. CoRR, abs/1703.06975.
Burda, Y., Grosse, R. B., and Salakhutdinov, R. (2014). Accurate and conservative estimates of MRF
log-likelihood using reverse annealing. CoRR, abs/1412.8566.
Crooks, G. E. (2000). Path-ensemble averages in systems driven far from equilibrium. Physical
review E, 61(3):2361.
Dayan, P., Hinton, G. E., Neal, R. M., and Zemel, R. S. (1995). The helmholtz machine. Neural
Comput., 7(5):889–904.
Gingrich, T. R., Rotskoff, G. M., Crooks, G. E., and Geissler, P. L. (2016). Near-optimal protocols in
complex nonequilibrium transformations. Proceedings of the National Academy of Sciences, page
201606273.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and
Bengio, Y. (2014). Generative adversarial nets. In Advances in Neural Information Processing
Systems, pages 2672–2680.
Gregor, K., Danihelka, I., Graves, A., and Wierstra, D. (2015). Draw: A recurrent neural network for
image generation. arXiv preprint arXiv:1502.04623.
Hinton, G. E., Osindero, S., and Teh, Y.-W. (2006). A fast learning algorithm for deep belief nets.
Neural Comput., 18(7):1527–1554.
Ioffe, S. and Szegedy, C. (2015). Batch normalization: Accelerating deep network training by
reducing internal covariate shift. arXiv preprint arXiv:1502.03167.
Kingma, D. and Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980.
Kingma, D. P., Salimans, T., and Welling, M. (2016). Improving variational inference with inverse
autoregressive flow. CoRR, abs/1606.04934.
Kingma, D. P. and Welling, M. (2013).
arXiv:1312.6114.
Auto-encoding variational bayes.
10
arXiv preprint
Krizhevsky, A. and Hinton, G. (2009). Learning multiple layers of features from tiny images.
Lillicrap, T. P., Cownden, D., Tweed, D. B., and Akerman, C. J. (2014). Random feedback weights
support learning in deep neural networks. arXiv:1411.0247.
Liu, Z., Luo, P., Wang, X., and Tang, X. (2015). Deep learning face attributes in the wild. In
Proceedings of the IEEE International Conference on Computer Vision, pages 3730–3738.
Markram, H. and Sakmann, B. (1995). Action potentials propagating back into dendrites triggers
changes in efficacy. Soc. Neurosci. Abs, 21.
Neal, R. M. (2001). Annealed importance sampling. Statistics and Computing, 11(2):125–139.
Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., and Ng, A. Y. (2011). Reading digits in natural
images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised
feature learning, volume 2011, page 5.
Rezende, D. J., Mohamed, S., and Wierstra, D. (2014). Stochastic backpropagation and approximate
inference in deep generative models. arXiv preprint arXiv:1401.4082.
Salakhutdinov, R. and Hinton, G. (2009). Deep boltzmann machines. In Artificial Intelligence and
Statistics.
Salimans, T., Goodfellow, I. J., Zaremba, W., Cheung, V., Radford, A., and Chen, X. (2016). Improved
techniques for training gans. CoRR, abs/1606.03498.
Schmiedl, T. and Seifert, U. (2007). Optimal finite-time processes in stochastic thermodynamics.
Physical review letters, 98(10):108301.
Sivak, D. A. and Crooks, G. E. (2012). Thermodynamic metrics and optimal paths. Physical review
letters, 108(19):190602.
Sohl-Dickstein, J., Weiss, E. A., Maheswaranathan, N., and Ganguli, S. (2015). Deep unsupervised
learning using nonequilibrium thermodynamics. CoRR, abs/1503.03585.
Sønderby, C. K., Raiko, T., Maaløe, L., Sønderby, S. K., and Winther, O. (2016). Ladder variational
autoencoders. In Advances in Neural Information Processing Systems, pages 3738–3746.
Theis, L., van den Oord, A., and Bethge, M. (2016). A note on the evaluation of generative models.
In International Conference on Learning Representations.
Vincent, P., Larochelle, H., Bengio, Y., and Manzagol, P.-A. (2008). Extracting and composing robust
features with denoising autoencoders. In Proceedings of the 25th international conference on
Machine learning, pages 1096–1103. ACM.
Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., and Manzagol, P.-A. (2010). Stacked denoising
autoencoders: Learning useful representations in a deep network with a local denoising criterion.
J. Machine Learning Res., 11.
11
Supplementary Material
A
VW transition operators and their convergence
Proposition 2. If p has enough capacity, training data and training time, with slow enough annealing
and a small departure from reversibility so p can match q, then at convergence of VW training, the
transition operator pT at T = 1 has the data generating distribution as its stationary distribution.
) match q(sK+n
), where q(s0 ) is the data distribution. It
Proof. With these conditions p(sK+n
0
0
means that p(s0 ) (the marginal at the last step of sampling) is the data distribution when running the
annealed (cooling) trajectory for K + n steps, for n any integer between 0 and N1 , where the last
n + 1 steps are at temperature 1. Since the last n steps are at temperature 1, they apply the same
transition operator. Consider any 2 consecutive sampling steps among these last n steps. Both of these
samples are coming from the same distribution (the data distribution). It means that the temperature 1
transition operator leaves the data distribution unchanged. This implies that the data distribution is an
eigenvector of the linear operator associated with the temperature 1 transition operator, or that the
data generating distribution is a stationary distribution of the temperature 1 transition operator.
B
Additional Results
Image inpainting samples from CelebA dataset are shown in Fig 2, where each top sub-figure shows
the masked image of a face (starting point of the chain), and the bottom sub-figure shows the inpainted
image. The images are drawn from the test set.
The VW samples for CelebA, CIFAR10 and SVHN are shown in Fig 4, 5, 6.
Figure 2: VW inpainting in CelebA images. Images on the left are the ground truth images corrupted
for their bottom half (which is the starting point of the chain). The goal is to fill in the bottom half of
each face image given an observed top half of an image (drawn from test set). Images on the right
show the inpainted lower halves for all these images.
C
VW on Toy Datasets
Fig. 7 and 8 shows the application of a transition operator applied on 2D datasets.
12
Figure 3: VW samples on MNIST using Gaussian noise in the transition operator. The model is
trained with 30 steps of walking away, and samples are generated using 30 annealing steps.
D
VW chains
Fig. 9, 10, 11, 12, 13, 14, 15 shows the model chains on repeated application of transition operator at
temperature = 1. This is to empirically prove the conjecture mentioned in the paper (Preposition 1)
that is, if the finite time generative process converges to the data distribution at multiple different VW
walkback time-steps, then it remains on the data distribution for all future time at T= 1
E
Architecture Details
In this section, we provide more details on the architecture that was used for each of the dataset. The
details of the hyper parameter and architecture used for each dataset can also be found in Tables 2, 3,
4 and 5. Complete specifications are available as experiment scripts at http://github.com/
anirudh9119/walkback_nips17.
E.1
MNIST
For lower bound(and IS estimates) comparisons, the network trained on MNIST is a MLP composed
of two fully connected layers with 1200 units using batch-normalization (Ioffe and Szegedy, 2015).
This network has two different final layers with a number of units corresponding to the image size (i.e
13
Figure 4: VW samples on CelebA dataset using Gaussian noise in the transition operator. Model is
trained using 30 steps to walk away and samples are generated using 30 annealing steps.
number of pixels) each corresponding to mean and variance for each pixel. We use softplus output
for the variance. We don’t share the batch-normalization parameters across different time steps.
For the real-values MNIST dataset samples, we used an encoder-decoder architecture with convolutional layers. The encoder consists of 2 convolutional layers with kernel length of 5 and stride of 2
followed by a decoder with strided convolutions. In addition, we used 5 fully connected feedforward
layers to connect the encoder and decoder. We applied batch normalization (Ioffe and Szegedy, 2015)
to the convolutional Layers, and we applied layer normalization (Ba et al., 2016) to the feedforward
layers. The network has 2 separate output layers, one corresponding the mean of the Gaussian sample,
and one corresponding to the variance of the added Gaussian noise. We use Adam (Kingma and Ba,
2014) with a learning rate of 0.0001 to optimize the network. Details of the hyper parameter and
architecture is also available in Table 2.
E.2
CIFAR10, CelebA and SVNH
We use a similar encoder-decoder architecture as we have stated above. We use 3 convolutional layers
for the encoder as well as for the decoder. We also apply batch normalization (Ioffe and Szegedy,
2015)to the convolutional layers, as well as layer normalization (Ba et al., 2016) to the feedforward
layers. Details of the hyper parameter and architecture is also available in Table 4, 5 and 3.
14
Operation
Convolution
Convolution
Fully Connected
Fully Connected
Fully Connected
Fully Connected
Fully Connected
Strided Convolution
Strided Convolution
Kernel
5x5
5x5
5x5
5x5
Strides
2
2
2
2
Feature Maps
16
32
16
1
Normalization
Batchnorm
Batchnorm
LayerNorm
LayerNorm
LayerNorm
LayerNorm
LayerNorm
Batchnorm
No
Non Linearity
Relu
Relu
Leaky Relu
Leaky Relu
Leaky Relu
Leaky Relu
Leaky Relu
Relu
None
Hidden Units
1568 * 1024
1024 * 1024
1024 * 1024
1024 * 1024
1024 * 1568
-
Table 2: Hyperparameters for MNIST experiments, for each layer of the encoder-decoder (each
row of the table). We use adam as an optimizer, learning rate of 0.0001. We model both mean and
variance of each pixel. We use reconstruction error as per-step loss function. We see improvements
using layernorm in the bottleneck, as compared to batchnorm. Using Dropout also helps, but all the
results reported in the paper are without dropout.
Operation
Convolution
Convolution
Convolution
Fully Connected
Fully Connected
Fully Connected
Fully Connected
Fully Connected
Strided Convolution
Strided Convolution
Strided Convolution
Kernel
5x5
5x5
5x5
5x5
5x5
5x5
Strides
2
2
2
2
2
2
Feature Maps
64
128
256
128
64
3
Normalization
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
No
Non Linearity
Relu
Relu
Relu
Relu
Relu
Relu
Relu
Relu
Relu
Relu
None
Hidden Units
16384 * 1024
1024 * 1024
1024 * 1024
1024 * 1024
1024 * 16384
-
Table 3: Hyperparameters for CelebA experiments, for each layer of the encoder-decoder (each
row of the table). We use adam as an optimizer, learning rate of 0.0001. We model both mean and
variance of each pixel. We use reconstruction error as per-step loss function.
Operation
Convolution
Convolution
Convolution
Fully Connected
Fully Connected
Fully Connected
Fully Connected
Fully Connected
Strided Convolution
Strided Convolution
Strided Convolution
Kernel
5x5
5x5
5x5
5x5
5x5
5x5
Strides
2
2
2
2
2
2
Feature Maps
64
128
256
128
64
3
Normalization
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
Batchnorm
No
Non Linearity
Relu
Relu
Relu
Relu
Relu
Relu
Relu
Relu
Relu
Relu
None
Table 4: Hyperparameters for Cifar experiments, for each layer of the encoder-decoder (each row of
the table). We use adam as an optimizer, learning rate of 0.0001. We model both mean and variance
of each pixel. We use reconstruction error as per-step loss function.
15
Hidden Units
4096 * 2048
2048 * 2048
2048 * 2048
2048 * 2048
2048 * 4096
-
Figure 5: VW samples on Cifar10 using Gaussian noise in the transition operator. Model is trained
using 30 steps to walk away and samples are generated using 30 annealing steps.
F
Walkback Procedure Details
The variational walkback algorithm has three unique hyperparameters. We specify the number of
Walkback steps used during training, the number of extra Walkback steps used during sampling and
also the temperature increase per step.
The most conservative setting would be to allow the model to slowly increase the temperature during
training. However, this would require a large number of steps for the model to walk to the noise, and
this would not only significantly slow down the training process, but this also means that we would
require a large number of steps used for sampling.
There may exist a dynamic approach for setting the number of Walkback steps and the temperature
schedule. In our work,
√we set this hyperparameters heuristically. We found that a heating temperature
schedule of Tt = T0 2t at step t produced good results, where T0 = 1.0 is the initial
temperature.
√
N
During sampling, we found good results using the exactly reversed schedule: Tt = √22t , where t is
the step index and N is the total number of cooling steps.
For MNIST, CIFAR, SVHN and CelelbA, we use K = 30 training steps and N = 30 sampling steps.
We also found that we could achieve better quality results if allow the model to run for a few extra
steps with a temperature of 1 during sampling. Finally, our model is able to achieve similar results
compared to the NET model(Sohl-Dickstein et al., 2015). Considering our model uses only 30 steps
for MNIST and NET (Sohl-Dickstein et al., 2015) uses 1000 steps for MNIST.
16
Figure 6: VW samples on SVHN dataset using Gaussian noise in the transition operator. Model is
trained using 30 steps to walk away and samples are generated using 30 annealing steps.
G
Higher Lower Bound: not always better samples
We have observed empirically that the variational lower bound does not necessarily correspond to
sample quality. Among trained models, higher value of the lower bound is not a clear indication
of visually better looking samples. Our MNIST samples shown in Fig 16 is an example of this
phenomenon. A model with better lower bound could give better reconstructions while not producing
better generated samples. This resonates with the finding of (Theis et al., 2016)
H
Reversibility of transition operator
We measured the degree of reversibility of pT by estimating the KL divergence
DKL (pT (s0 |s)πT (s) || pT (s|s0 )πT (s0 )), which is 0 if and only if pT obeys detailed balance and
PK
pT (st+1 |st )
1
is therefore time-reversal invariant by computing the Monte-Carlo estimator K
t=1 ln pT (st |st+1 ) ,
where sK
1 is a long sequence sampled by repeatedly applying transition operator pT from a draw
s1 ∼ πT , i.e., taking samples after a burn-in period (50 samples).
To get a sense of the magnitude of this reversibility measure, and because it corresponds to an
estimated KL divergence, we estimate the corresponding entropy (of the forward trajectory) and use
it as a normalizing denominator telling us how much we depart from reversibility in nats relative to
the number of nats of entropy. To justify this, consider that the minimal code length required to code
17
Figure 7: The proposed modeling framework trained on 2-d swiss roll data. This algorithm was
trained on 2D swiss roll for 30 annealing steps using annealing schedule increasing temperator by
1.1 each time. We have shown every 5th sample (ordering is row wise, and within each row it is
column-wise.
Figure 8: The proposed modeling framework trained on circle data. This algorithm was trained on
circle for 30 annealing time steps using annealing schedule increasing temperature by factor 1.1 each
time. We have shown every 5th sample (ordering is row wise, and within each row it is column-wise.
samples from a distribution p is the entropy H(p). But suppose we evaluate those samples from p
using q instead to code them. Then the code length is H(p) + D(p||q). So the fractional increase
in code length due to having the wrong distribution is D(p||q)/H(p), which is what we report here,
with p being the forward transition probability and q the backward transition probability.
To compute this quantity, we took our best model (in terms of best lower bound) on MNIST, and ran
it for 1000 time steps i.e (T = 1000), at a constant temperature.
We run the learned generative chain p for T time steps (after a burn in period
whose samples we ignore) getting s0 → s1 → s2 → · · · sT and computing
log p(s0 → s1 → s2 → · · · sT )/p(sT → · · · → s2 → s1) both under the same generative chain, divided by T to get the per-step average.
On the same set of runs, we compute 1/T ∗log p(s0 → s1 → s2 → · · · sT ) under the same generative
chain. This is an estimate of the entropy per unit time of the chain. This is repeated multiple times to
average over many runs and reduce the variance of the estimator.
18
Figure 9: VW sample chain (vertically, going down) starting from pure noise. Model trained using
K = 30 steps to walk away and samples are generated using 30 steps of annealing. The figure shows
every 3rd sample of the chain in each column.
The obtained ratio (nats/nats) is 3.6%, which seems fairly low but also suggests that the trained model
is not perfectly reversible.
I
Some Minor Points
• In all the image experiments, we observed that by having different batchnorm papemeters
for different steps, actually improves the result considerably. Having different batchnorm
parameters was also necessery for making it work on mixture on gaussian. The authors were
not able to make it work on MoG without different parameters. One possible way, could be
to let optimizer know that we are on different step by giving the temperature information to
the optimizer too.
• We observed better results while updating the parameters in online-mode, as compared to
batch mode. (i.e instead of accumulating gradients across different steps, we update the
parameters in an online fashion)
19
Figure 10: VW sample chain. Each coloumn above corresponds to one sampling chain. We have
shown every 10th sample. We applied the transition operator for 5000 time-steps at temperature =
1, to demonstrate that even over very long chain, the transition operator continues to generate good
samples.
J
Inception Scores on CIFAR
We computed the inception scores using 50,000 samples generated by our model. We compared the
inception scores with (Salimans et al., 2016) as the baseline model.
20
Figure 11: VW sample chain. Each column above corresponds to one sampling chain. We have
shown every 10th sample. We applied the transition operator for 5000 time-steps at temperature =
1, to demonstrate that even over very long chain, the transition operator continues to generate good
Operation
Kernel Strides Feature Maps Normalization Non Linearity
samples.
Convolution
5x5
2
64
Batchnorm
Relu
Convolution
5x5
2
128
Batchnorm
Relu
Convolution
5x5
2
256
Batchnorm
Relu
Fully Connected
Batchnorm
Relu
Fully Connected
Batchnorm
Relu
Fully Connected
Batchnorm
Relu
Fully Connected
Batchnorm
Relu
Fully Connected
Batchnorm
Relu
Strided Convolution
5x5
2
128
Batchnorm
Relu
Strided Convolution
5x5
2
64
Batchnorm
Relu
Strided Convolution
5x5
2
3
No
None
Table 5: Hyperparameters for SVHN experiments, for each layer of the encoder-decoder (each row of
the table). We use adam as an optimizer, learning rate of 0.0001. We model both mean and variance
of each pixel. We use reconstruction error as per-step loss function.
21
Hidden Units
4096 * 1024
1024 * 1024
1024 * 1024
1024 * 1024
1024 * 4096
-
Figure 12: VW sample chain. Each column above corresponds to one sampling chain. We have
shown every 10th sample. We applied the transition operator for 5000 time-steps temperature = 1.
Model
Real Data
Salimans (semi-supervised)
Salimans (unsupervised)
Salimans (supervised training without minibatch features)
VW(20 steps)
VW(30 steps)
Table 6: Inception scores on CIFAR
22
Inception Score
11.24
8.09
4.36
3.87
3.72
4.39 ± 0.2
Figure 13: VW sample chain. Each column above corresponds to one sampling chain. We have
shown every 10th sample. We applied the transition operator for 5000 time-steps at temperature =
1, to demonstrate that even over very long chain, the transition operator continues to generate good
samples.
23
Figure 14: VW sample chain. Each column above corresponds to one sampling chain. We have
shown every 10th sample. We applied the transition operator for 5000 time-steps at temperature =
1, to demonstrate that even over very long chain, the transition operator continues to generate good
samples.
24
Figure 15: VW sample chain. Each column above corresponds to one sampling chain. We have
shown every 10th sample. We applied the transition operator for 5000 time-steps at temperature =
1, to demonstrate that even over very long chain, the transition operator continues to generate good
samples.
25
Figure 16: Samples from two VW models (left and right) which have a higher lower bound than the
one whose samples are shown in Figure 5 (and comparable but slightly better importance sampling
estimators of the log-likelihood): yet, the generated samples are clearly not as good, suggesting
that either the bound is sometimes not tight enough or that the log-likelihood is not always a clear
indicator of sample quality.
26
| 9cs.NE
|
Online Buy-at-Bulk Network Design
Deeparnab Chakrabarty∗
Alina Ene†
Ravishankar Krishnaswamy‡
arXiv:1509.03212v1 [cs.DS] 10 Sep 2015
Debmalya Panigrahi§
Abstract
We present the first non-trivial online algorithms for the non-uniform, multicommodity buy-at-bulk
(MC-BB) network design problem. Our competitive ratios qualitatively match the best known approximation factors for the corresponding offline problems. In particular, we show
• A polynomial time online algorithm with a poly-logarithmic competitive ratio for the MC-BB
problem in undirected edge-weighted graphs.
• A quasi-polynomial time online algorithm with a poly-logarithmic competitive ratio for the MCBB problem in undirected node-weighted graphs.
1
• For any fixed > 0, a polynomial time online algorithm with a competitive ratio of Õ k 2 + )
(where k is the number of demands, and Õ(.) hides polylog factors) for MC-BB in directed graphs.
• Algorithms with matching competitive ratios for the prize-collecting variants of all the above problems.
Prior to our work, a logarithmic competitive ratio was known for undirected, edge-weighted graphs
only for the special case of uniform costs (Awerbuch and Azar, FOCS 1997), and a polylogarithmic
competitive ratio was known for the edge-weighted single-sink problem (Meyerson, SPAA 2004). To
the best of our knowledge, no previous online algorithm was known, even for uniform costs, in the
node-weighted and directed settings.
Our main engine for the results above is an online reduction theorem of MC-BB problems to their
single-sink (SS-BB) counterparts. We use the concept of junction-tree solutions (Chekuri et al., FOCS
2006) that play an important role in solving the offline versions of the problem via a greedy subroutine
– an inherently offline procedure. Our main technical contribution is in designing an online algorithm
using only the existence of good junction-trees to reduce an MC-BB instance to multiple SS-BB subinstances. Along the way, we also give the first non-trivial online node-weighted/directed single-sink
buy-at-bulk algorithms. In addition to the new results, our generic reduction also yields new proofs of
recent results for the online node-weighted Steiner forest and online group Steiner forest problems.
∗
Microsoft Research, 9 Lavelle Road, Bangalore, India. Email: [email protected].
Department of Computer Science and DIMAP, University of Warwick, Coventry, UK. Email: [email protected].
‡
Microsoft Research, 9 Lavelle Road, Bangalore, India. Email: [email protected].
§
Department of Computer Science, Duke University, Durham, NC, USA. Email: [email protected].
†
1
Introduction
In a typical network design problem, one has to find a minimum cost (sub) network satisfying various
connectivity and routing requirements. These are fundamental problems in combinatorial optimization,
operations research, and computer science. To model economies of scale in network design, Salman et
al. [32] proposed the buy-at-bulk framework, which has been studied extensively over the last two decades
(e.g., [6, 20, 33, 22, 29, 28, 15, 14]). In this framework, each network element is associated with a subadditive function representing the cost for a given utilization. Given a set of connectivity demands comprising k source-sink pairs, the goal is to route integral flows from the sources to the corresponding sinks
concurrently to minimize the total cost of the routing.
An important application of the problem is capacity planning in telecommunication networks or in the
Internet. As observed by Awerbuch and Azar [6], this application is inherently “online” in that terminalpairs arrive over time and need to be served without knowledge of future pairs. The authors of [6] give
a logarithmic-competitive online algorithm for the uniform case where every edge is associated with the
same cost function. However, uniformity is not always a feasible assumption, especially in heterogeneous,
dynamic networks like the Internet. Indeed recent research (e.g., [29, 28, 15]) has focused on the nonuniform setting with a different sub-additive function for every network element. In this non-uniform setting,
Meyerson [28] gives a polylogarithmic-competitive algorithm for the special case when all terminal-pairs
share the same sink. To the best of our knowledge, no non-trivial online algorithm is known for the general
multicommodity setting, which is the focus of our paper.
We consider, in increasing order of generality, undirected edge-weighted graphs, undirected nodeweighted graphs, and directed edge-weighted graphs.1 It is also convenient to classify the problems that
we study into the single-sink version where all the terminal-pairs share a common sink, and the general
multicommodity version where the sinks in the terminal-pairs may be distinct. For notational convenience,
we use the following shorthand forms for our problems: X-Y-BB where X = SS or MC (single-sink and
multicommodity, respectively) and Y = E or N or D (undirected edge-weighted, undirected node-weighted,
and the general directed case, respectively).
1.1
Our Contributions
We obtain the following new results (unless otherwise noted, our algorithms run in polynomial time):
• A poly-logarithmic competitive online algorithm for the MC-E-BB problem.
• A poly-logarithmic competitive online algorithm for MC-N-BB and SS-N-BB that runs in quasipolynomial time.
1
• An Õ(k 2 +ε )-competitive online algorithm for MC-D-BB for any constant ε > 0 with running time
nO(1/ε) , where Õ(.) hides polylogarithmic factors. For SS-D-BB, the ratio improves to Õ(k ), translating to a polylogarithmic competitive ratio in quasi-polynomial time.
• Online algorithms for prize-collecting versions of all the above problems with the same competitive
ratio.
Up to exponents in the logarithm, our online algorithms match the best known offline approximation algorithms (Chekuri et al. [15] for MC-E/N-BB and Antonakopoulos [5] for MC-D-BB); for MC-N-BB,
1
In undirected graphs, node costs can simulate edge costs; in directed graphs they are equivalent.
2
however, a polynomial time, polylogarithmic approximation is known [14], whereas our algorithm runs in
quasi-polynomial time. Furthermore, a logarithmic lower bound, even for SS-E-BB, follows from the lower
bound for the online Steiner tree problem [26], and a polylogarithmic lower bound for online SS-N-BB
follows from a matching one for set cover [3].
From a technical perspective, we derive all the multicommodity results using a generic online reduction
theorem that reduces a multicommodity instance to several single-sink instances, for which we either use
existing online algorithms or give new online algorithms. Informally, one can view this as the “online
analog” of the junction-tree approach pioneered by Chekuri et al. [15] for offline multicommodity network
design. We discuss this approach in the next subsection.
1.2
An Online Reduction to Single Sink Instances
Multicommodity network design problems, both online and offline, are typically more challenging than their
single-sink counterparts, and have historically2 required new ideas every time depending on the specific
problem at hand. The situation is no different for buy-at-bulk, both for uniform and non-uniform costs.
In the offline buy-at-bulk setting, this shortcoming is addressed by Chekuri et al. [15] (expanded to
other problems by [14, 13, 5]), who introduce a generic combinatorial framework for mapping a single
instance of a multicommodity problem to multiple instances of the corresponding single-sink problem. At
the heart of this scheme is the following observation that holds for many multicommodity problems such as
(edge/node) Steiner forest, directed Steiner network, buy-at-bulk, and set connectivity: there exists a nearoptimal3 junction-tree solution for the multicommodity problem that decomposes into solutions to multiple
single-sink problems where each single-sink problem connects some subset of the original terminal-pairs to
a particular root.
The problem now reduces to finding good junction-trees to cover all the terminal-pairs. The offline techniques [15, 13, 5] tackle this using a greedy algorithm for finding the single-sink solutions; more precisely,
in each step they find the best density (cost per terminal-pair) solution that routes a subset of terminalpairs via a single sink. A set cover style analysis then bounds the loss for repeating this procedure until
all terminal-pairs are covered. However, as the reader may have already noticed, the greedy optimization
approach is inherently offline, as finding the best-density solution requires knowledge of all terminal-pairs
upfront. Our main technical contribution in this work is an online version of the junction-tree framework.
Indeed, we show how to reduce any multicommodity buy-at-bulk instance to a collection of single-sink
instances online.
(Informal Theorem) If the junction-tree approximation factor of the MC-BB problem is α, the integrality
gap of a natural LP relaxation of the SS-BB problem is β, and there is a γ-competitive online algorithm for
the SS-BB problem, then there is an O(αβγ · polylog(n))-competitive algorithm for the MC-BB problem.
To prove the above theorem, we first write a composite-LP, which has (a) an outer-LP comprising assignment variables that fractionally assign terminal-pairs to roots, and (b) many inner-LPs which correspond to
the natural LP relaxations for the SS-BB problem for each root and the terminal-pairs fractionally assigned
to it by the outer-LP. We then apply the framework of online primal-dual algorithms (see [10] for instance)
to solve the composite-LP online. However, there are two main challenges we need to surmount.
• First, the existing framework has been mostly applied to purely covering/packing LPs4 and our inner2
For instance, compare [29] and [12] for the SS-E-BB and MC-E-BB problem, and compare Naor et al. [31] and Hajiaghayi et
al. [24] for the online node-weighted Steiner tree and Steiner forest problem.
3
We call the quality of such a solution the junction-tree approximation factor; e.g., it is O(log n) for MC-E-BB and MC-NBB [15]
4
Our current understanding of mixed packing-covering is limited [7] and does not capture the problem we want to solve.
3
LPs have both kinds of constraints, and moreover, there is an outer-LP encapsulating them. We show
nonetheless that it can be extended to solving our LP fractionally up to polylogarithmic factors. Indeed,
we use the specific flow-structure of the inner-LP, and each step of our algorithm solves many (auxiliary)
min-cost max-flow problems.
• The second difficulty is in rounding this fractional solution online. This is a hard problem, and currently we do not know how to do so even for basic network design problems such as the Steiner tree problem.
To circumvent this, we show that it suffices to only partially round the LP. More precisely, we round the
LP solution so that only the outer-LP (assignment variables) become integral, and the inner-LPs remain
fractional. This gives us an integral assignment of the terminal-pairs to different single-sink instances, with
bounded total fractional cost. Now, from the bounded integrality gap of the inner-LPs, we know that there
exist good single-sink solutions for our assignment of terminal-pairs to roots, even though we cannot find
them online5 . Using this knowledge, our final step is to run online single-sink algorithms for each root,
and send the terminal-pairs to the root as determined by the outer-LP assignment. Figure 1 summarizes our
overall approach.
initialize multiple online algorithms for the single-sink problem (one for each vertex as root).
when (si , ti ) arrives
update the fractional solution of the composite LP to satisfy the new request
round the composite LP to get an integral solution to the outer LP, which gives us an assignment
of (si , ti ) to some root r
send both si and ti to the instance of the single-sink online algorithm with root r
Figure 1: Online Framework for Multicommodity Network Design Problems
The results mentioned in Section 1.1 follow by bounding α, β, γ for the corresponding problems. For
MC-E-BB, all of these are known to be O(polylog(n)) ([15, 29, 28] respectively). For MC-N-BB, it is
known both α, β are bounded by O(polylog(n)) [14], and we bound γ in Section 5 by giving the first online
algorithms for SS-N-BB. For MC-D-BB, we need some additional work. In this case we cannot directly
bound β, since the integrality gap of the natural LP relaxation is not known to be bounded. Nevertheless,
in Section 4, we show that it suffices to work only with more structured instances for which we can bound
the integrality gap.
Finally, we illustrate the generality of our reduction theorem by noting that, when combined with existing
bounds on α, β, and γ, it immediately implies (up to polylogarithmic factors) some recent results in online
network design, such as online node-weighted Steiner forest [24], and online edge-weighted group Steiner
forest [31] – two problems for which specialized techniques were needed, even though their single-sink
counterparts were known earlier.
5
The difficulty comes from the fact that the online solution we maintain must be monotonic, i.e., the decisions are irrevocable.
4
1.3
Related Work
Buy-at-bulk network design problems have received considerable attention over the last two decades, both
in the offline and online settings. For the uniform cost model, Awerbuch and Azar [6] give an O(log n)approximation for MC-E-BB, while O(1)-approximations are known [20, 33, 22] for SS-E-BB. We also
note that O(1)-approximations have been obtained in special cases for the multicommodity problem, such
as in the rent-or-buy setting [21]. Meyerson et al. [29] give an O(log k)
√ approximation for the general
SS-E-BB, and the first non-trivial algorithm for MC-E-BB is an exp(O( log n log log n))-approximation
due to Charikar and Karagiozova [12]. This was improved to a poly-logarithmic factor by Chekuri et al.
[15] who also solve MC-N-BB [14] with similar guarantees. For directed graphs, our knowledge is much
sparser. Even for special cases like directed
√ Steiner tree and forest, the best polytime approximation factors
known are O(k ε ) [34, 11] and min(O( k, n2/3 )) [13, 18, 8] respectively, and these ideas were extended
to MC-D-BB by Antonakopoulos [5]. On the hardness side, Andrews [4] shows that even the MC-E-BB
problem is Ω log1/2−ε n)-hard, while MC-D-BB (in fact directed Steiner forest) is known to be label-cover
hard [17].
The online Steiner tree problem (a special case of online SS-E-BB) was first studied by Imase and
Waxman [26] who give an O(log k)-competitive algorithm. Berman and Coulston [9] give an O(log k)competitive algorithm for online Steiner forest, and both these results are tight, i.e., there is an Ω(log k) lower
bound. As mentioned earlier, Awerbuch and Azar’s algorithm [6] can be seen as an O(log n)-competitive
online algorithm for the uniform-cost MC-E-BB. For non-uniform buy-at-bulk, the only online algorithm
that we are aware of is Meyerson’s [28] polylog-competitive algorithm for the single-sink problem. For
online node-weighted network design, developments are much more recent. A polylogarithmic approximation for the node-weighted Steiner tree problem was first given by Naor et al. [31] and later extended to the
online node-weighted Steiner forest problem [24] and prize-collecting versions [23]. These algorithms, like
ours in this paper, utilize the online adaptation of the primal-dual and LP rounding schemas pioneered by
the work of Alon et al. [3] for the online set cover problem (see also [2] for its adaptation to network design
problems). We also note that in the node-weighted setting, the online lower bound can be strengthened to
Ω(log n log k) using online set cover lower bounds [3, 27].
2
Preliminaries and Results
We now formally state the problem, set up notation that we use throughout the paper, and state our main
theorems.
2.1
Our Problems
Buy-at-bulk Network Design. In the most general setting of the MC-D-BB problem, an instance I =
(G, X ) consists of a directed graph G = (V, E) and a collection X of terminal-pairs (si , ti ) ∈ V × V ;
each such si and ti is called a terminal. Each (si , ti ) pair also has a positive integer demand di , which we
assume to be 1 for clarity in presentation.6 Additionally, each edge e ∈ E is associated with a monotone,
sub-additive7 cost function fe : R≥0 → R≥0 . A feasible solution to the problem is a collection of paths
{P1 , . . . , Pk } where Pi is a directed path from si to ti carrying load di . Given a solution {P1 , . . . , Pk }, we
6
We can handle non-uniform demands by incurring an additional O(log D) factor in the competitive ratio and the running time
(where D is the maximum demand) by having O(log D) “unit-demand” instances, where the ith instance deals with demands
between 2i−1 and 2i .
7
That is, fe (x) ≥ fe (y) whenever x ≥ y, and fe (x + y) ≤ fe (x) + fe (y)
5
P
let load(e) = i:e∈Pi diP
denote the total load on edge e. The goal is to find a feasible solution minimizing
the objective ObjBB := e∈E fe (load(e)). In the online problem, the offline input consists of the graph G
and the cost functions fe . The pairs (si , ti ) arrive online in an unknown, possibly adversarial, order. When a
pair (si , ti ) arrives, the algorithm must select the path Pi that connects them, and this decision is irrevocable.
Reduction to the Two-metric Problem. Following previous work, throughout this paper we consider an
equivalent problem (up to constant factors) known as two-metric network design. In this problem, instead
of functions fe (.) on the edges, we are given two parameters ce and `e on each edge. One can think of
ce as a fixed buying cost, or just cost, of edge e, and `e as a per-unit flow cost, or length, of edge e. The
feasible solution
space is the
P
P same
P as for the buy-at-bulk problem, and the goal is to minimize the objective
Obj2M := e∈S Pi ce + i e∈Pi `e . The following lemma is well known (see e.g., [15]).
i
Lemma 1. Given an instance of the buy-at-bulk problem, for any ε > 0 one can find an instance of the
two-metric network design problem such that, for any feasible solution, Obj2M ≤ ObjBB ≤ (2 + ε)Obj2M .
Remark. In light of the above lemma, henceforth we abuse notation and let the buy-at-bulk problem mean
the two-metric network design problem.
2.2
Our Tools
Junction-tree solutions. Given an instance I = (G, X ) of the buy-at-bulk problem, we consider junctiontree8 solutions, a specific kind of solution to the problem introduced by [15]. In such solutions, the collection
of pairs are partitioned into groups and each group is indexed by a root vertex r ∈ V . For all terminal pairs
(si , ti ) in a group indexed by r, the path Pi from si to ti contains the root vertex r (see Figure 2).
t4
t2
s1
junction
s3
t1
t3
s4
s2
Figure 2: A group of terminal pairs routed via a junction-vertex in an undirected graph.
Formally, consider an instance I = (G, X ) of the buy-at-bulk problem, and let Opt denote the objective
value of the optimum solution. Given a partition Π := (πr1 , . . . , πrq ) of terminal pairs indexed by q different
root vertices, a junction-tree solution is one that uses single-sink solutions to connect the original terminalpairs. Indeed, for each part πr indexed by root r, consider the optimal solutions to the single-sink problem
on graph G with demands {(si , r) : (si , ti ) ∈ πr } and the single-source problem9 with pairs {(r, ti ) :
(si , ti ) ∈ πr }. Let Optr (πr ) denote the sum ofPthe objectives of the optimal solutions to the single-sink
and single-source problems, and let Opt(Π) := r∈V Optr (πr ). Let Optjunc denote the minimum Opt(Π)
8
The word tree is misleading since the final solution need not be a tree in directed graphs. Nevertheless, we continue using this
term for historical reasons. Junction trees were originally proposed for undirected graphs, where the solution is indeed a tree.
9
The single-source problem in a directed graph is identical to the single-sink problem with all the edges reversed in direction.
For undirected graphs, both problems are on the same graph.
6
over all partitions. We call this solution the optimum junction-tree solution for this instance.10 Clearly,
Optjunc ≥ Opt. The junction-tree approximation factor of I is defined to be the ratio Optjunc /Opt.
LP Relaxation. We now describe a natural flow-based LP relaxation for the single-sink buy-at-bulk problem
for an instance I = (G, T ) where T is a set of terminals that need to be connected to the root r.
minimize
X
ce xe +
e∈E
s.t
XX
i
`e fi (e)
(SS-BaB LP)
e∈E
{fi (e) : e ∈ E(G)} defines a flow from si to r of value 1
fi (e) ≤ xe
xe ≥ 0, fi (e) ≥ 0
∀si ∈ T
∀e ∈ E
∀e ∈ E
Recall that the integrality gap of (SS-BaB LP) on the instance I = (G, T ) is defined to be the ratio of Opt
to the optimal value of the LP (SS-BaB LP). Also, we define the integrality gap for the graph G to be the
worst case integrality gap (over all requests T on graph G) of the corresponding instance I = (G, T ).
2.3
Our Results
Main Technical Theorem and its Applications. Now we are ready to state our main theorem; the proof is
in Section 3. We say that an online algorithm is γ-competitive for a graph G if, for any sequence of requests
X , the online algorithm for buy-at-bulk returns a solution within a γ-factor of Opt(I), where I = (G, X ).
Theorem 2 (Reduction to Single-Sink Online Algorithms). Fix an instance I = (G, X ) of the MC-BB
problem. Suppose the following three conditions hold.
(i) The junction-tree approximation factor of I is at most α.
(ii) The integrality gap of (SS-BaB LP) on any single-sink instance on graph G is at most β.
(iii) There is a γ-competitive online SS-BB algorithm for any instance on graph G that runs in time T .
Then there is an online algorithm for I running in time poly(n, T ) whose competitive ratio is O(αβγ ·
polylog(n)).
Using this theorem, we can immediately obtain the following new results mentioned in the introduction.
Theorem 3 (Undirected Edge-weighted Buy-at-Bulk). There is a polylog(n)-competitive, polynomial time
randomized online algorithm for the MC-E-BB problem.
Proof: The theorem follows directly by combining Theorem 2 with the following results from previous
work. Chekuri et al. [15] prove that the junction-tree approximation factor for the undirected edge-weighted
buy-at-bulk problem is O(log k). Chekuri et al. [16] prove that the integrality gap of (SS-BaB LP) in undirected edge-weighted graphs is O(log k). Meyerson [28] gives a randomized polynomial time online algorithm for the single-sink buy-at-bulk problem with competitive ratio O(log4 n).
Theorem 4 (Undirected Node-weighted Buy-at-Bulk). For any constant ε > 0, there is an O(k ε polylog(n))competitive, randomized online algorithm for MC-N-BB with running time nO(1/ε) . As a corollary, this
yields a polylog(n)-competitive, quasi-polynomial time algorithm for this problem.
10
Note that copies of the same edge appearing in multiple single-sink solutions are treated as distinct edges in the junction-tree
solution. Hence, decomposing the optimal multicommodity solution into its constituent paths does not yield Optjunc = Opt.
7
Theorem 5 (Directed Buy-at-Bulk). For any constant ε > 0, there is an O k 1/2+ε polylog(n)) -competitive,
polynomial time online algorithm for the MC-D-BB.
We again use Theorem 2 to prove the above theorems. However, unlike for MC-E-BB, we are not aware of
any online algorithms for the SS-N-BB and SS-D-BB problems. We therefore first give online algorithms
for these problems, and then use Theorem 2; the details appear in Sections 5 and 4.
Finally, we can almost directly use Theorem 2 to also obtain matching results for prize-collecting versions
of the above problems. Recall that in a prize-collecting problem, every terminal-pair also comes with a
penalty qi , and the algorithm can opt to not satisfy the request by incurring this value in the objective. We
give the extension of our results to the corresponding prize-collecting problems in Section 6.
Theorem 6. For each of the above problems, there is an online algorithm with matching running time and
competitive ratio for the corresponding prize-collecting version.
In addition to the new results mentioned above, we can also use Theorem 2 to give alternative proofs
(with slightly worse polylog factors) of some recent results in online network design. By combining Theorem 2 with the polylog(n)-competitive algorithm for online group Steiner Tree due to Alon et al. [2], we obtain a polylog(n)-competitive online algorithm for the group Steiner forest problem – a result shown earlier
by Naor et al. [31]. Similarly, by combining Theorem 2 with the polylog(n)-competitive online algorithm
for the node-weighted Steiner tree problem due to Naor et al. [31], we obtain a polylog(n)-competitive
online algorithm for the node-weighted Steiner forest problem – a result shown earlier by Hajiaghayi et al.
[24].
Height Reduction Theorem. One of the technical tools that we use repeatedly in this paper is the following
result, which builds on the work of Helvig et al. [25]. We give the proof in Appendix A.
Theorem 7. Given a directed graph G = (V, E) with edge costs ce and lengths `e , for all h > 0, we can
efficiently find an upward directed, layered graph Gup
h on (h + 1) levels and edges (with new costs and
lengths) only between successive levels going from bottom (level h) to top (level 0), such that each layer has
n vertices corresponding to the vertices of G, and, for any set of terminals X and any root vertex r,
(i) the optimal objective value of the single-sink buy-at-bulk problem to connect X (at level h) with
1/h )φ, where φ is the objective value of an optimal
r (at level 0) on the graph Gup
h is at most O(hk
solution of the same instance on the original graph G;
(ii) given a integral (resp. fractional solution) of objective value φ for the single-sink buy-at-bulk
problem to connect X with r on the graph Gup
h , we can efficiently recover an integral (resp. fractional
solution) of objective value at most φ for the problem on the original graph G.
Likewise, we can obtain a downward directed, layered graph Gdown
on (h + 1)-levels with edges going from
h
top to bottom, with the same properties as above except for single-source instances instead.
3
Proof of Theorem 2 (Online Reduction to Single-Sink Instances)
There are three main steps in the proof. In Section 3.1, we describe the composite LP which is a relaxation
of optimal junction-tree solutions (for technical reasons, we first need to pre-process the graph). Next, in
Section 3.2, we show how to fractionally solve the LP online. Third, in Section 3.3, we show how to partially
round the LP online. The resulting solution then decomposes as fractional solutions to different single-sink
instances. Finally, we use the bounded integrality gap and the online algorithm for SS-BB to wrap up the
proof in Section 3.4.
8
3.1
The Composite-LP Relaxation: MC-BaB LP
We first apply Theorem 7 with h = Θ(log n) to obtain layered graphs Gup (resp., Gdown ) of height O(log n)
where all the edges are directed upward (resp. downward); see Figure 3 for an illustration. The reason for
this preprocessing is that the length of the (si , ti ) paths appear as a factor in our final competitive ratio and
the above step bounds it to a logarithmic factor. Recall that the graph Gup (resp., Gdown ) approximately
preserves the single-sink (resp., single-source) solutions for any set of terminals and any root. After this
step, we can imagine that all the roots (of the single-sink instances we will solve) are vertices in level 0,
and all the terminals will be vertices in level h = Θ(log n). For clarity of presentation, we refer to the root
and terminal vertices by the same name in both Gup and Gdown (even though the graphs are completely
disjoint). Overloading notation, let V denote the vertices in level 0 in both Gup and Gdown , and let E be the
union of the edge sets of Gup and Gdown . Furthermore, the cost ce and length `e of these edges are inherited
from Theorem 7.
v1,0
v2,0 v3,0 v4,0 v5,0 v6,0
(roots are here)
v1,0
level 0
v2,0 v3,0 v4,0 v5,0 v6,0
V
v1,3 v2,3 v3,3 v4,3 v5,3 v6,3
Gup
level 3
v1,3 v2,3 v3,3 v4,3 v5,3 v6,3
(terminals arrive here)
Gdown
Figure 3: The graphs Gup and Gdown with h = 3, where the original graph G has 6 vertices {v1 , v2 , . . . , v6 }.
Now, using the junction-tree decomposition with approximation factor α, we get the following lemma.
Lemma 8. There exists a set R∗ ⊆ V of root vertices, a partition Π∗ := {πr : r ∈ R∗ } of the terminalpairs in X , a collection of in-trees {Trup : r ∈ R∗ } rooted at r in Gup , and a collection of out-trees
{Trdown : r ∈ R∗ } rooted at r in Gdown such that
(i) Each (si , ti ) ∈ X belongs to πr for some r ∈ R∗ .
(ii) For each r ∈ R∗ , the in-tree Trup is a feasible solution to the single-sink buy-at-bulk problem connecting {si : (si , ti ) ∈ πr } to r in Gup ; likewise, the out-tree Trdown is a feasible solution to the
single-source buy-at-bulk problem connecting r to {ti : (si , ti ) ∈ πr } in Gdown .
(iii) The sum of objective values of the single-source and single-sink solutions is at most O(α log n) ·
Opt(I).
Proof: Since the junction-tree approximation ratio of the given instance is α, there exists a junction-tree
solution given by a set of roots R∗ and a partition Π∗ := {πr : r ∈ R∗ } such that the total objective value
of all the single-sink and single-source junction-trees is at most αOpt(I). Moreover, by Theorem 7, because
we choose the height h = Θ(log n), the objective value of each single-sink and single-source solution in G
increases by a factor of O(log n) in Gup and Gdown respectively. Thus, the overall objective value of the
resulting junction-trees in the graphs Gup and Gdown is O(α log n) · Opt(I).
9
The above lemma motivates the LP relaxation given in Fig. 4 which seeks to assign each (si , ti ) pair to
some rooted instance, and then minimizes the total fractional objective value of the rooted instances. Each
individual rooted instance is represented by an inner-LP (see the boxed constraints in Fig. 4).
minimize
XX
ce xre +
r∈V e∈E
s.t
X
r∈V
X
XX
r
r
+
f
`e f(e,s
(e,ti )
i)
(MC-BaB LP)
(si ,ti )∈X r∈R e∈E
zir ≥ 1
∀i
zir ≥ 0
(2b)
r
} define a flow from si to r of value zir in Gup
{f(e,s
i)
r
{f(e,t
} define a flow from r to ti of value zir in Gdown
i)
r
f(e,s
≤ xre
i)
r
f(e,t
≤ xre
i)
r
xre ≥ 0, f(·)
(2a)
∀i, ∀r ∈ V
(2c)
∀i, ∀r ∈ V
(2d)
∀i, ∀e, ∀r ∈ V
∀i, ∀e, ∀r ∈ V
≥0
(2e)
(2f)
(2g)
Figure 4: Composite LP for MC-BB. Equations (2a) and (2b) form the outer-LP; (2c)-(2g) form the innerLPs.
In the LP, zir denotes the extent to which the pair (si , ti ) chooses root r to route its flow. Within each
inner-LP corresponding to a root r, {xre } are the variables which denotes whether edge e is used to route flow
r
r
in the corresponding rooted instance, and f(e,s
(resp. f(e,t
) denotes the amount of flow si sends (resp.,
i)
i)
ti receives) along e to (resp., from) root r. Observe that if the zir variables are integral, then the inner-LP
corresponding to every root r constitutes a feasible fractional solution to (SS-BaB LP) for the single-sink
instance I 0 = (Gup , X 0 ) where X 0 = {(si , r) : zir = 1} and the single-source instance I 00 = (Gdown , X 00 )
where X 00 = {(r, ti ) : zir = 1}. The next lemma, which bounds the optimal value of the MC-BaB LP,
follows directly from Lemma 8.
Lemma 9. The optimum value of (MC-BaB LP) is O(α log n) · Opt(I).
3.2
An online fractional algorithm for the MC-BaB LP
Theorem 10. There is a randomized, polynomial-time online algorithm that returns a feasible fractional
solution for (MC-BaB LP) of value at most O(α log3 n) · Opt(I).
In the remainder of the subsection, we prove the above theorem. We remark that the overall reduction
uses Theorem 10 as a black-box and the time-constrained reader can skip the proof and move to Section 3.3.
To simplify the exposition, we assume that we know the cost of an optimal solution Opt(I) up to a
constant factor, using a standard doubling trick11 . Once we know Opt(I), by re-scaling all the parameters
Suppose our online algorithm has a competitive ratio of α, and the true cost of an optimal solution is c∗ . Then, we begin
with an initial guess for the optimal cost, and run the online algorithm assuming this guess is the correct estimate for c∗ . If our
online algorithm fails to find a feasible solution of cost at most α times the current guess, we double our guess and run the online
algorithm again. Eventually, our guess will exceed the optimal cost c∗ by at most a factor of two, and for this guess, the algorithm
11
10
in the problem, we may assume that it equals 1. Next, we delete any edge in Gup or Gdown that has cost
ce or length `e larger than 1 as such edges cannot participate in any optimal solution. Subsequently, we
initialize all xre variables to 1/n5 . Likewise, we initialize all zir variables to 1/n5 and also send an initial
flow of 1/n5 from each si to r in Gup on an arbitrary flow path from si to r and likewise from r to ti in
Gdown . This setting ensures that the cost of the initial solution is o(1).
In the following, we partition the edge set E into disjoint sets {Ej : 0 ≤ j ≤ h − 1}, where Ej denotes
the set of edges in E between levels j and j + 1. Furthermore, for clarity of exposition, we describe a
‘continuous-time’ version of the algorithm where we increase the variables as a function of time. We note
that this algorithm can easily be discretized for a polynomial-time12 implementation. The algorithm is given
as Algorithm 1.
Algorithm 1 Online Fractional Algorithm for (MC-BaB LP)
When a terminal-pair (si , ti ) arrives, we update the LP solution using the following steps:
(1) Let Ri denote the set of roots r in level 0 such that si is connected to r in Gup and r is connected to ti
in Gdown . For each r ∈ Ri , initialize a flow of value 1/n5 using any arbitrary flow path from si to r in
Gup and likewise from r to ti in Gdown . Also set zir = 1/n5 for these roots.
P
(2) Repeat the following while r∈Ri zir < 1:
r
r
(a) Call an edge e ∈ E tight for root r if xre = f(e,s
or xre = f(e,t
.
i)
i)
(b) Edge Update: For all tight edges e ∈ E, update xre at the rate
dxre
dt
:=
xre
ce .
(c) Flow Update: Solve the following min-cost max-flow problem for each r ∈ Ri : maximize ∆ such
that
r
- there exists a flow {g(e,s
} sending ∆ units of flow from si to r in Gup ,
i)
r
- there exists a flow {g(e,t
} sending ∆ units of flow from r to ti in Gdown ,
i)
r
r
- Capacity constraints: g(e,s
≤ TheAlgorithm(Algorithm 1).xre /ce and g(e,t
≤ xre /ce
i)
i)
for all tight edges e,
P
P
r
r
≤ zir .
≤ zir and e `e · g(e,t
- Cost constraint: e `e · g(e,s
i)
i)
r
(d) Update f(e,s
at the rate
i)
zir at the rate
dzir
dt
r
df(e,s
dt
i)
r
r
:= g(e,s
, and f(e,t
at the rate
i)
i)
r
df(e,t
i)
dt
r
= g(e,t
for all e, and update
i)
= ∆.
We increase the x variables on tight edges at a rate inversely proportional to their cost, similar to the wellknown online set cover algorithm [3]. However, the “flow constraints” are not pure packing (or covering)
constraints and there is no general-purpose way of handling them. Indeed, we determine the rate of increase
of the flow variables by solving an auxiliary min-cost max-flow subroutine which routes incremental flows of
equal value from si to r in Gup and from r to ti in Gdown respecting capacity constraints (i.e., for edges that
are tight, the incremental flow is at most the rate of increase of x). This maintains feasibility in the inner LP.
Moreover, to bound the rate of increase in objective, we enforce that the total length of the incremental flow
will compute a feasible solution of cost at most 2αc∗ . Moreover, since our guesses double every time, the total cost of the edges
bought by the online algorithm over all the runs across different guesses is at most Θ(αc∗ ).
12
The polynomial is in the size of the input to this algorithm, which for some of our algorithms/results is quasi-polynomial in the
size of the actual problem instance as stated in the introduction.
11
is at most zir (this is the “cost” constraint in the min-cost max-flow problem). We stress that the incremental
flows from the auxiliary problem dictate the rate at which we increase the original flow variables in the LP.
The final solution is feasible since the algorithm runs until the outer-LP constraint is satisfied.
First, note that the total cost of initialization is o(1) over all the edge and flow variables. So it suffices
to bound the cost of the updates. The next lemma relates the total cost of the updates to the total time τ for
which the algorithm runs, and the subsequent lemma bounds τ in terms of Opt(I).
Lemma 11. The LP objective value at the end of the above algorithm is O(log n) · τ , where τ is the
(continuous) time for which the algorithm runs.
Proof: We show that at any time t, the rate of the increase of the LP objective value in the algorithm is at
most O(log n); this proves the lemma.13 The objective increases because of increase in x variables and flow
variables f ; we bound these separately.
We first upper bound the objective increase due to the changes
P in the x variables. Fix
P a level j and let
Ejtgt denote the set of tight edges in Ej at time t. By definition, e∈E tgt ∩Gup xre (resp., e∈E tgt ∩Gdown xre )
j
j
equals the total flow on these edges for the pair (si , ti ).
j form
P a cut separating si from
PSince the edges in EP
ti , the total flow across this cut is at most zir . Since r zir < 1, we have r e∈E tgt xre < 2. Now, the
j
rate of increase of each such tight edge is precisely xre /ce , which implies that the total rate of increase of the
LP value due to the increase of x is at most
X X
r e∈E tgt
j
ce
X X
dxre
≤
xre ≤ 2.
dt
tgt
r
e∈Ej
Summing over all levels gives the desired O(log n) bound.
Next, we upper bound the objective increase due to the changes in the f variables. When these variables
are updated, the total rate of increase of the objective due to the lengths of the (si , r) and (r, ti ) flow paths
is at most zir — this is precisely the “cost” constraint in the auxiliary flow problem. Hence the total rate of
increase of flow lengths is at most 2, completing the proof.
Given the above lemma, we are left to relate τ to Opt(I) in order to complete the proof of Theorem 10.
Lemma 12. The time duration τ of the above algorithm satisfies τ = O(α log2 n) · Opt(I).
We will need several new definitions and auxiliary lemmas in order to prove Lemma 12. Recall from Lemma 8
that we can assume that the solution that we are comparing against is the set of junction-trees defined by Trup
and Trdown for r ∈ R∗ . Also, recall that the terminal-pairs are partitioned by the groups Π∗ = {πr : r ∈ R∗ }.
For every (si , ti ) ∈ X , let Ps∗i denote the path from si to the
root r in Trup such that (si , ti ) ∈ πr . Similarly,
P
let Pt∗i denote the path from r to ti in Trdown . Let `(P ) = e∈P `e for any path P . Lemma 8 asserts that
X
X
X
ce +
`(Ps∗i ) + `(Pt∗i ) = O(α log n) · Opt(I)
(3)
r∈R∗
e∈Trup ∪Trdown
(si ,ti )∈X
To bound τ against the optimal junction-tree solution, we use two sets of charging clocks:
13
We remark that the “log n” corresponds to the number of levels in Gh justifying the preprocessing step before the LP description.
12
• We maintain an edge clock on every (e, r) pair such that e ∈ Trup or e ∈ Trdown , i.e., if e is used by
the optimal junction-tree solution in the single-source (or single-sink) instance corresponding to r. In
particular, note that if an edge e is in multiple junction-trees, then it has a separate clock for each such
tree.
• We maintain a terminal clock on every terminal-pair (si , ti ) ∈ X .
The crucial invariant that we maintain is the following: at any time instant t, at least one clock “ticks,” i.e.,
augments its counter at unit rate. The overall goal would then be to bound the total time for which all the
charging clocks can cumulatively tick.
First, we describe the rule for the ticking of the clocks. Fix a time t, and let the terminal-pair (si , ti ) be
the pair that is active at time t. Let r denote the root vertex which (si , ti ) has been assigned to in the optimal
junction-tree solution from Lemma 8, i.e., (si , ti ) ∈ πr . Now, consider the flow-paths Ps∗i in Gup and Pt∗i in
Gdown . We can have one of two situations:
- If any variable xre is tight for any edge e ∈ Ps∗i ∪ Pt∗i at time t, then the edge clock on the pair (e, r)
ticks at time t. If there are multiple such edges, then all the corresponding clocks tick.
- Otherwise, both paths are free of tight edges. In this case, the terminal clock for (si , ti ) ticks at time
t.
Lemma 13. For any pair (e, r) such that e ∈ Trup ∪ Trdown , its edge clock ticks for O(ce log n) time.
Proof: Notice that xre is initialized to 1/n5 for all roots r, and increases at the rate
dxre
xr
= e
dt
ce
(4)
at all times when the edge clock on (e, r) ticks. To see why, consider a time t when the clock on (e, r) ticks,
and let (si , ti ) denote the active terminal-pair at time t. It must be that (i) (si , ti ) has been assigned to root
r
r
or xre = f(e,t
. But in this case, we increase such variables at rate xre /ce
r in Π∗ , and (ii) either xre = f(e,s
i)
i)
in our algorithm (Step (2a)). Therefore, we can infer that the value of xre would be 1 after the edge clock
on e has ticked for time O(ce log n). But clearly, e cannot be a tight edge for any subsequent terminal-pair
(si , ti ) once xe reaches 1; therefore, the edge clock on (e, r) ticks for O(ce log n) time overall.
Lemma 14. For every terminal-pair (si , ti ) connected by the optimal junction-tree solution through the root
vertex r, the total time for which the terminal clock ticks is at most O(log n) · max(`(Ps∗i ), `(Pt∗i )).
Proof: Recall that if the terminal clock for (si , ti ) is ticking at time t, then it must mean that no edge is
tight on either path Ps∗i or Pt∗i . In this case, we show that the variable zir increases at a fast enough rate,
where r is the root (si , ti ) is assigned to in the optimal junction-tree, i.e., (si , ti ) ∈ πr . We show this by
exhibiting a feasible solution to the auxiliary LP considered in Step (2b) of the algorithm for root r. Indeed,
send the flow from si to r along Ps∗i , and likewise from r to ti along Pt∗i . Also set the value of ∆ to be
zir / max(`(Ps∗i ), `(Pt∗i )). Clearly, on the edges of these flow paths, we do not have any capacity constraints
since no edge is tight. So, the only constraints are the cost constraints which are satisfied by the choice of
∆. Hence, the rate of increase of zir is at least
dzir
zir
≥
dt
max(`(Ps∗i ), `(Pt∗i ))
13
(5)
at all times when the terminal clock on (si , ti ) ticks. This proves the claim, for otherwise the variable zir
would have reached 1, and the algorithm would have completed processing (si , ti ).
Since at least one clock ticks at all times, the total time clocked is at least τ , the duration of the algorithm.
Lemma 13 and Lemma 14 imply that
X
X
X
ce +
`(Ps∗i ) + `(Pt∗i )
τ ≤ O(log n)
e∈Trup ∪Trdown
r∈R∗
(si ,ti )∈X
which together with (3) completes the proof of Lemma 12. Theorem 10 follows from Lemma 11 and
Lemma 12.
3.3
Partial Online LP Rounding
We partially round the fractional solution returned by Theorem 10 to obtain integral values for only the
outer-LP variables zir , i.e., each (si , ti ) pair is integrally assigned to a root. The inner-LP variables x and
f continue to be fractional but represent unit fractional flow from si to r and r to ti for the (si , ti ) pairs
assigned to r. The partial rounding algorithm is given as Algorithm 2.
Algorithm 2 Online Partial Rounding Algorithm
(1) Initialization: Each root chooses a threshold τr ∈ [1/2n, 1/(3 log n)] uniformly at random.
r
=
(2) Partial Rounding: At each time, maintain the scaled solution x̃re = min (1, xre /τr ), f˜(·)
r
min 1, f(·) /τr . Also set z̃ir = 1 if zir ≥ τr .
Theorem 15. The scaled solution (x̃, f˜) component-wise dominates a feasible solution to the outer-LP, and
the expected objective value of the scaled solution (x̃, f˜) is at most O(α log5 n) · Opt(I). Moreover, for
each (si , ti ), there exists at least one root r such that z̃ir ≥ 1 with probability at least 1 − 1/n3 .
Proof: Since each root r chooses its threshold τr independently and uniformly at random from [1/2n, 1/ log n],
the probability that z̃ir = 1 is at least zir log n (since z̃ir = 1 if and only if τr ≤ zir ). Since this is independent for different roots, a standard Chernoff-Hoeffding bound application (see, e.g., [30]) shows that each
(si , ti ) pair has z̃ir = 1 for some root r with probability at least 1 − 1/n3 . Moreover, the expected value of
any variable xre is given by
E [x̃re ]
≤
Z
log n
τr =1/2n
xre
log ndτr ≤ O(log2 n)xre .
τr
A similar argument shows that the expected values of scaled flow variables are also bounded by O(log2 n)
times their values in the fractional solution. This shows that the expected objective value of the (x̃, f˜) solution is at most O(log2 n) times the value of (x, f ); by Theorem 10, the latter is at most O(α log3 n)Opt(I).
Combining these facts gives us the desired bound on the value of the scaled solution.
It remains to show that the scaled solution dominates a feasible solution to the LP. To this end, fix
some root r and let Xr denote the set of (si , ti ) pairs for which zir = 1. We need to show that installing
r
capacities of {f˜(e,s
} on the edges can support unit flow from si to r in Gup for all (si , ti ) ∈ Xr . Suppose
i)
14
for contradiction that there is is a cut Q separating si from r of capacity strictly smaller than 1. This
r
r
implies that every edge e ∈ Q must have f(e,s
≤ τr ; otherwise, we would have an edge e with f˜(e,s
=
i)
i)
1,
which
contradicts
our
assumption
on
the
cut
capacity.
But
then
the
value
of
the
min-cut
is
precisely
P
r
f
e∈Q (e,si ) /τr , which must be at least 1 because of the following two observations: (i) we know that
P
r
r
{f(e,si ) } is a feasible flow from si to r of value zir and hence it must be that e∈Q f(e,s
≥ zir , and (ii)
i)
since z̃ir = 1, it must be that zir ≥ τr . This contradicts the assumption that the cut capacity is strictly
r
smaller than 1. A similar argument shows that the variables {f˜(e,t
} can support unit flow from r to ti for
i)
every (si , ti ) with z̃ir = 1.
3.4
Wrapping up: Invoking the Single-Sink Online Algorithm
We are now ready to put all the pieces together and present our overall online multicommodity buy-at-bulk
algorithm as Algorithm 3. SingleSinkAlg is the online algorithm for SS-BB alluded to in point (iii) of the
statement of Theorem 2.
Algorithm 3 Online Multicommodity Buy-at-Bulk Algorithm
when (si , ti ) arrives
(1) update the fractional solution of the composite LP using the algorithm (Algorithm 1, Section 3.2).
(2) partially round the solution using algorithm in (Fig. 2, Section 3.3).
(3) if(∃r : zir ≥ 1): send both si and ti to the instance of SingleSinkAlg with root r.
(4) else: buy a trivial shortest path between si and ti on the metric (c + `) and route along this path
Clearly the algorithm produces a feasible solution; so we now argue about the expected objective value.
Fix an (si , ti ) pair. Since the probability that a terminal-pair is not assigned to a root is ≤ 1/n3 (by Theorem 15), the expected total contribution of such unassigned terminal-pairs is ≤ 1 = Opt(I). For a root
r, let πr be the terminal-pairs assigned to r. We know that (x̃, f˜) restricted to πr dominates a feasible
solutionP
in (SS-BaB LP). Letting LPr denote the contribution of this restriction to the overall LP value,
we get r LPr = O(α log5 n) · Opt(I). By the integrality gap condition, we get that Optr , i.e. the integral optimum objective value of the instance generated by r and πr , is at most β · LPr . (Here we are
using the fact from Theorem 7 that moving to the layered instance does not increase the integrality gap.)
The objective value of the solution produced by SingleSinkAlg is at most γ · Optr , where γ is the competitive ratio of SingleSinkAlg. Putting these observations together, we conclude that the overall objective
value of the solution returned by the online algorithm is O(αβγ log5 n) · Opt(I). This completes the proof
of Theorem 2.
4
Online Directed Buy-at-Bulk
In this section, we prove Theorem 5. A natural approach is to use the reduction given by Theorem 2. To this
end, we need to establish the following: the existence of a junction-tree scheme with a good approximation;
a good upper bound on the integrality gap for single-sink instances of the LP given in Section 2; and an
online algorithm for single-sink instances with a good competitive ratio.
15
Extending the work of Chekuri
√ et al. [13], Antonakopoulos [5] shows the existence of a junction-tree
scheme with approximation O( k). Unfortunately, the integrality
gap of the LP relaxation is not very well
√
understood even for Steiner tree instances; [35] gives an Ω( k) lower bound14 on the integrality gap for the
Steiner tree problem and no suitable upper bound is known. We overcome this difficulty as follows. Instead
of working with general graphs, we pre-process the instance and obtain a tree-like graph for which we can
show that the LP has a good integrality gap. Finally, we give the first non-trivial online algorithm for the
directed single-sink buy-at-bulk problem. These results, together with our reduction (Theorem 2), imply the
online algorithm for MC-D-BB.
We devote the rest of this section to the proof of Theorem 5; to aid the reader, we restate the theorem
below.
1
Theorem 16. For any constant ε > 0, there is a O(k 2 +ε polylog(n))-competitive, polynomial time randomized online algorithm for the general buy-at-bulk problem.
Pre-processing step. We first give our reduction from general instances of the directed buy-at-bulk problem
1
to much more structured instances; the reduction loses a factor of O(k 2 +ε ) in the approximation ratio.
Let h = d1/εe. Given an instance I = (G, X ) of the directed buy-at-bulk problem, we map it to a
tree-like instance J = (H, X ) as follows. We start by applying Theorem 7 to G to obtain the graphs Gup
and Gdown ; recall that these graphs are layered (h + 1)-level graphs with n vertices (corresponding to the
vertices in G) in each level, and the levels are numbered 0, 1, . . . , h with 0 being called the root level. The
graph Gup has edges directed from higher numbered levels to lower numbered levels, and Gdown has edges
in the opposite direction. To facilitate the construction of the graph H, we now create n trees from Gup and
n trees from Gdown as follows.
For every “root vertex” r at level 0 in Gup (resp. Gdown ), the tree Trup (resp. Trdown ) is constructed as
follows:
• The 0th layer of Trup has just one vertex – the root r.
• For each i such that 1 ≤ i ≤ h, the i-th layer of Trup contains all (i + 1)-length tuples (r, v1 , . . . , vi )
where vj is a vertex present in the j-th layer of Gup .
• For every edge e = (vi , vi−1 ) ∈ Gup , there is an arc from (r, v1 , . . . , vi−1 , vi ) to (r, v1 , . . . , vi−1 )
inheriting the same cost ce and length `e .
Therefore each tree Trup is an in-arborescence, with all edges directed towards the root. The tree Trdown is
constructed analogously except all edges are directed away from the root. In the following, we use the term
leaves to refer to the vertices on layer h of these trees.
down , we have 2n
After performing the above operation for every root vertex r in level 0 of Gup
r and Gr
trees. Then the final graph H is obtained as follows (see Figure 5). For each root r ∈ V , we first add an arc
from the root of Trup to Trdown of zero cost and length. Finally, for every (si , ti ) pair, we add the vertices si
and ti to H and the following arcs connecting them to the trees: for each tree Trup , we add an arc from si
to each leaf of Trup of the form (r, v1 , . . . , vh ) with vh = si ; for each tree Trdown , we add an arc to ti from
each leaf of Trup of the form (r, v1 , . . . , vh ) with vh = ti . These new arcs have zero cost and length (i.e.,
ce = `e = 0).
This completes the construction of H. Note that the graph H has nO(h) vertices and a similar number
of edges. Our new instance is J = (H, X ) and we will apply Theorem 2 to this instance.
We first relate the objective values of J and I.
14
However, in these instances, n is exponentially large in k. So, they do not rule out a polylog(n) upper bound.
16
r1
r2
Trup
1
s1
s2
···
Trdown
1
t1
sk
t2
···
Trup
2
Trdown
2
r2
r2
tk
Figure 5: Construction of graph H.
Lemma 17. Every feasible solution for J is a junction-tree solution.
Proof: Note that any (si , ti ) path in H has the following structure: si connects to a leaf node of Trup for
some r ∈ V , then continues to the root r, then traverse the edge to the root of Trdown , then goes down to
a leaf of Trdown and finally connects to ti . Thus, for any feasible solution for J , the (si , ti ) pairs can be
partitioned based on the root r through which they connect.
Lemma 18. Any feasible solution for J can be mapped to a feasible solution — in fact, a junction-tree
solution — for I of equal or smaller objective value.
Proof: Note that from the previous lemma, any feasible solution S in J is a junction-tree solution. Therefore, there is a partition ΠS of the (si , ti ) pairs depending on which root vertex they are using to connect.
Moreover, it follows from our construction of the trees in H that any edge in Trup (resp. Trdown ) corresponds
to an edge in Gup (resp. Gdown ). Therefore, if we map each edge appearing in solution S to its corresponding edge in Gup or Gdown , we obtain a mapping from each junction tree of S rooted at r to a junction tree in
Gup ∪ Gdown rooted at r that is connecting the same subset of pairs. Finally, by Theorem 7, each junction
tree in Gup ∪ Gdown rooted at r can be mapped, without increasing the objective value, to a junction tree in
G rooted at r that is connecting the same subset of pairs. This completes the proof of the lemma.
Lemma 19. Opt(J ) ≤ O(hk 1/h )Optjunc (I), where Optjunc (I) is the objective value of an optimal junctiontree solution for I.
Proof: Consider the optimal junction tree solution for I. Let the optimum partition be Π = (πr1 , . . . , πrq )
where R∗ = {r1 , r2 , . . . , rq } is the set of roots of the junction trees. For each r ∈ R∗ , let Xr = {si :
(si , ti ) ∈ πr } be the set of sources of πr and Yr be the corresponding sinks. From Theorem 7, we know
that the optimum objective value of any one single-sink problem connecting Xr to r in Gup is at most
O(hk 1/h ) times the objective value of the optimum solution connecting each source in Xr to r. An analogous upper bound holds for every optimal single-source solution connecting r to each sink in Yr . Therefore,
we get that the total sum of objective values of each of the junction trees in Gup and Gdown is at most
17
O(hk 1/h )Optjunc (I). Now notice that any solution SG for (Gup , Xr ) can easily be “simulated” by a solution ST in the tree Trup : indeed, for every root-vertex path (r, v1 , v2 , . . . , vi ) in the solution SG , include the
edge from (r, v1 , v2 , . . . , vi ) to (r, v1 , v2 , . . . , vi−1 ) in ST (recall the vertices in Trup exactly correspond to
such root-vertex paths). It is easy to see that the objective value of the solution ST in Trup is the same as that
of SG . Similarly, any solution for (Gdown , Yr ) can be simulated in Trdown with the same objective value. It
follows that there is a feasible solution in J of objective value at most O(hk 1/h )Optjunc (I).
1
Corollary 20. Opt(J ) ≤ O(k 2 +ε )Opt(I).
√
Proof Sketch: Antonakopoulos [5] shows that there exists a junction-tree solution of cost at most O( k)Opt(I).
The corollary follows from this work and the fact that we set h = Θ(1/ε).
Now we are ready to show that the new instance J has the properties required by the reduction, i.e.,
Theorem 2 can be applied. In the following lemma, a single-source (resp. single-sink) sub-instance
J 0 = (H, T , v) of J = (H, X ) is a single-source (resp. single-sink) instance of the following form:
the graph is the same as in J , namely H; the set of terminals T is a subset of the sources (resp. sinks) of
X ; the terminals T need to be connected to a root vertex v ∈ V (H) \ {si , ti : i ∈ [k]}.
Lemma 21. Let J be the instance described above. Let α, β, γ be as in the statement of Theorem 2. We
have
(i) The junction-tree approximation factor of J is 1 (i.e., α = 1).
(ii) The integrality gap of (SS-BaB LP) for any single-sink/source sub-instance J 0 = (H, T , v) is O(h log n log k)
(i.e., β = O(h log n log k)).
(iii) There is a O(h2 log2 n log k)-competitive algorithm for any single-sink/source sub-instance J 0 =
(H, T , v) (i.e., γ = O(h2 log2 n log k)).
Proof: Property (i) follows from Lemma 17. Thus we focus on proving (ii) and (iii). In the following, we
assume that we are working with a single-sink sub-instance J 0 of J ; the proof is very similar for singlesource sub-instances and we omit it.
In order to show (ii) and (iii), we will map the sub-instance J 0 = (H, T , v) to an instance of the group
Steiner tree problem on a tree as follows. Since v ∈ V (H) \ {si , ti : i ∈ [k]}, we have v ∈ Trup ∪ Trdown
for some r. We first consider the case when v ∈ Trup . In order to define the tree of the group Steiner tree
instance, we start with the subtree Tv of Trup rooted at v. We add the following ‘dangling’ edges to Tv for
each source si ∈ T : for each leaf vertex u ∈ Trup such that si has an edge in Trup to u, we add a new vertex
u0 to Tv and connect it to u. Let T be the resulting tree. We assign weights to the edges of T as follows. Each
of the old edges e ∈ Tv receives a weight equal to ce . Each of the new edges uu0 ∈ E(T ) \ E(Tv ) receives a
weight equal to `(Pu ), where Pu is the path of Trup from u to v. Finally, we define the following groups: for
each source si ∈ T , we introduce a group Si consisting of all the new vertices u0 ∈ V (T ) \ V (Tv ) such that
si is connected to its partner u in Trup (that is, Trup has an edge from si to u). In the resulting group Steiner
tree instance, the goal is to connect all of the groups S = {Si : si ∈ T } to the root v using a minimum
weight subtree of T ; we let (T, S, v) denote this instance.
Now the key claim is that the feasible solutions to the single-sink buy-at-bulk (H, T , v) are in a oneto-one correspondence with feasible solution to the group Steiner tree instance (T, S, v); moreover, the
objective value of a solution to the former is equal to the weight of the solution to the latter. To see this,
consider a feasible solution P for the single-sink buy-at-bulk instance. Note that, for each source si ∈ T , P
has a path connecting si to v; it follows from our construction of H that this path consists of an edge from
18
si to a leaf u of Trup followed by the unique path in Trup from u to v. Thus we can construct a feasible group
Steiner tree solution by connecting each group Si using the path of T from u0 to v, where u0 is the partner of
the leaf u of Trup through which si connects to the root in the buy-at-bulk solution. The weight of the edge
u0 u captures the `-cost of si ’s path and the weight of the path from u to v captures the c-cost of si ’s path.
Moreover, we can apply the same argument to fractional solutions to the two problems and show that
there is a bijection between feasible fractional solutions to (SS-BaB LP) and feasible fractional solutions
to the LP relaxation for group Steiner tree of Garg, Konjevod, and Ravi [19]; as before, these corresponding solutions have the same objective values. Now the desired upper bound on the integrality gap
of (SS-BaB LP) follows from the work of [19] who showed that the integrality gap of the group Steiner
tree LP is O(log N log K) where N = maxi |Si | is the maximum size of a group and K is the number of
groups. In our setting, K ≤ |X | ≤ k and N ≤ nh . Therefore the integrality gap is O(h log n log k), which
establishes property (ii).
Moreover, notice that the above reduction can also be used to obtain an online algorithm for the singlesink (and single-source) sub-instances. Indeed, we simply use the online group Steiner tree algorithm of
Alon et al. [1] which has a competitive ratio O(log2 N log K) = O(h2 log2 n log k). This proves property
(iii).
Now we are ready to complete the proof of Theorem 16.
Proof of Theorem 16: Given the instance I = (G, X ), we construct the graph H as described above; the
time taken to do so is nO(h) . We pass the instance (H, X ) to Theorem 2 and, using Lemma 21, we obtain
an online algorithm that, for any collection of pairs X and any adversarial ordering of X , returns a solution
of cost O(polylog(n)) · Opt(J ). By Lemma 19 and Corollary 20, we can map solutions for (H, X ) to
solutions for (G, X ).
We note that the approach described above also gives us new online algorithms for the single-sink buy-atbulk problem on directed graphs.
For single-sink instances, the junction-tree approximation is equal to 1
√
and thus we save a factor of k.
Corollary 22. There is a polynomial time O(k ε · polylog(n))-competitive online algorithm for the singlesink (or single-source) buy-at-bulk problem on directed graphs. The competitive ratio can be improved to
polylog(n) if the running time can be quasi-polynomial in n.
5
Online Single-sink, Undirected, Node-weighted Buy-at-Bulk
In this section, we prove Theorem 4. Again, we use our main theorem (Theorem 2) and reduce the multicommodity buy-at-bulk problem to the single-sink version of the problem. We combine the reduction theorem with the following results from previous work. Chekuri et al. [13] show the existence of a junction-tree
scheme with approximation factor O(log k). Moreover, the natural LP relaxation for the single-sink buyat-bulk problem on graphs with node costs is also O(log k) [15]. For the single-source online algorithm,
we resort to the algorithms from the previous section for the more general directed single-sink buy-at-bulk
problem (see Corollary 22). We now obtain the desired result by the following parameter settings:
α = O(log k)
β = O(log k)
γ = O(polylog(n))
T
= nO(log n) .
19
6
Online Prize-Collecting Buy-at-Bulk
In the prize-collecting version of the buy-at-bulk problem, each terminal pair (si , ti ) also comes with a
penalty qi and the algorithm may choose not to serve this request and incur the penalty in the total cost. We
show that our online reduction framework (Theorem 2) can be easily modified to handle prize-collecting
versions as follows.
Theorem 23. Let I be a buy-at-bulk instance and suppose the three conditions of Theorem 2 hold. Then
there is an O(αβγ · polylog(n))-competitive online algorithm for the online, prize-collecting buy-at-bulk
problem on I with arbitrary penalties.
Proof: We closely follow the proof of Theorem 2. The first difference is in the LP-formulation. Now, for
each (si , ti ) pair we have an extra variable zi,0 which indicates whether we choose to discard this pair (and
pay the corresponding penalty) or not. We point out the differences with (MC-BaB LP). The new objective
function is
XX
X XX
X
r
r
minimize
ce xre +
`e f(e,s
+
f
+
qi zi,0
(e,ti )
i)
r∈V e∈E
(si ,ti )∈X r∈R e∈E
(si ,ti )∈X
and (2a) is replaced by
X
r∈V
zir + zi,0 ≥ 1
∀i
Observe that the optimum value of this modified LP is at most O(α log n) times the optimum: set zi,0 = 1
for the pairs the integral optimum solution does not connect, and for the rest apply Lemma 9. Also observe
that the modified LP can be thought of the old LP on a modified instance where the graphs Gup and Gdown
(obtained from Theorem 7) have another vertex “0” at the root level, and each si has a direct path from si to
“0” in Gup with total length qi /2 and no fixed cost, and similarly, each ti has a path from “0” to ti in Gdown
with total length qi /2 and no fixed cost. The rest of the proof now follows exactly as in Section 3, by also
including the special vertex as a possible root while rounding to make the outer LP variables integral.
7
Conclusion
In this paper, we gave the first polylogarithmic-competitive online algorithms for the non-uniform multicommodity buy-at-bulk problem. Our result is a corollary of a generic online reduction technique that we
proposed in this paper for converting a multicommodity instance into several single-sink instances, which
are often easier to design algorithms for. We believe that this reduction will have other applications beyond the buy-at-bulk framework, and illustrate this by showing that recent results on online node-weighted
Steiner forest and online generalized connectivity directly follow from our reduction theorem. Our work
also opens up new directions for future research. For instance, our algorithm for the node-weighted problem
runs in quasi-polynomial time, and a concrete open question is to get a polynomial-time polylogarithmiccompetitive algorithm for the SS-N-BB problem (this suffices for MC-N-BB as well by our main theorem).
Another technical question concerns non-uniform demands. While our algorithm can be extended to the
case of non-uniform demands, the approximation ratio incurs an additional O(log D) factor, where D is
the ratio of the largest to the smallest demand. It would be interesting to eliminate this dependence on D
since the corresponding offline results do not have this dependence. More generally, a broader question is
to investigate other mixed packing-covering LPs that can be solved and rounded online.
20
Acknowledgements
D. Panigrahi is supported in part by NSF Award CCF-1527084, a Google Faculty Research Award, and a
Yahoo FREP Award.
References
[1] N. Alon, B. Awerbuch, Y. Azar, N. Buchbinder, and J. Naor. A general approach to online network optimization problems. In Proceedings, ACM-SIAM Symposium on Discrete Algorithms (SODA), pages
577–586, 2004.
[2] N. Alon, B. Awerbuch, Y. Azar, N. Buchbinder, and J. Naor. A general approach to online network
optimization problems. ACM Trans. on Alg., 2(4):640–660, 2006.
[3] N. Alon, B. Awerbuch, Y. Azar, N. Buchbinder, and J. Naor. The online set cover problem. SIAM J.
Comput., 39(2):361–370, 2009.
[4] M. Andrews. Hardness of buy-at-bulk network design. In Proceedings, IEEE Symposium on Foundations of Computer Science (FOCS), pages 115–124, 2004.
[5] S. Antonakopoulos. Approximating directed buy-at-bulk network design. In Proceedings, Workshop
on Approximation and Online Algorithms (WAOA), pages 13–24, 2010.
[6] B. Awerbuch and Y. Azar. Buy-at-bulk network design. In Proceedings, IEEE Symposium on Foundations of Computer Science (FOCS), pages 542–547, 1997.
[7] Y. Azar, U. Bhaskar, L. Fleischer, and D. Panigrahi. Online mixed packing and covering. In Proceedings, ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 85–100, 2013.
[8] P. Berman, A. Bhattacharyya, K. Makarychev, S. Raskhodnikova, and G. Yaroslavtsev. Approximation
algorithms for spanner problems and directed steiner forest. Inform. and Comput., 222:93–107, 2013.
[9] P. Berman and C. Coulston. On-line algorithms for steiner tree problems (extended abstract). In
Proceedings, ACM Symp. on Theory of Computing (STOC), pages 344–353, 1997.
[10] N. Buchbinder and J. Naor. Online primal-dual algorithms for covering and packing. Math. Oper. Res.,
34(2):270–286, 2009.
[11] M. Charikar, C. Chekuri, T. Cheung, Z. Dai, A. Goel, S. Guha, and M. Li. Approximation algorithms
for directed steiner problems. J. Algorithms, 33(1):73–91, 1999.
[12] M. Charikar and A. Karagiozova. On non-uniform multicommodity buy-at-bulk network design. In
Proceedings, ACM Symp. on Theory of Computing (STOC), pages 176–182, 2005.
[13] C. Chekuri, G. Even, A. Gupta, and D. Segev. Set connectivity problems in undirected graphs and the
directed steiner network problem. ACM Transactions on Algorithms, 7(2):18, 2011.
[14] C. Chekuri, M. T. Hajiaghayi, G. Kortsarz, and M. R. Salavatipour. Approximation algorithms for
node-weighted buy-at-bulk network design. In Proceedings, ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1265–1274, 2007.
21
[15] C. Chekuri, M. T. Hajiaghayi, G. Kortsarz, and M. R. Salavatipour. Approximation algorithms for
nonuniform buy-at-bulk network design. SIAM J. Comput., 39(5):1772–1798, 2010.
[16] C. Chekuri, S. Khanna, and J. Naor. A deterministic algorithm for the cost-distance problem. In Proceedings, ACM-SIAM Symposium on Discrete Algorithms (SODA), volume 7, pages 232–233, 2001.
[17] Y. Dodis and S. Khanna. Design networks with bounded pairwise distance. Proceedings, ACM Symp.
on Theory of Computing (STOC), pages 750 – 759, 1999.
[18] M. Feldman, G. Kortsarz, and Z. Nutov. Improved approximation algorithms for directed steiner forest.
J. Comput. System Sci., 78(1):279–292, 2012.
[19] N. Garg, G. Konjevod, and R. Ravi. A polylogarithmic approximation algorithm for the group steiner
tree problem. J. Algorithms, 37(1):66–84, 2000.
[20] S. Guha, A. Meyerson, and K. Munagala. A constant factor approximation for the single sink edge
installation problem. SIAM J. Comput., 38(6):2426–2442, 2009.
[21] A. Gupta, A. Kumar, M. Pál, and T. Roughgarden. Approximation via cost-sharing: A simple approximation algorithm for the multicommodity rent-or-buy problem. In Proceedings, IEEE Symposium on
Foundations of Computer Science (FOCS), pages 606–615, 2003.
[22] A. Gupta, A. Kumar, and T. Roughgarden. Simpler and better approximation algorithms for network
design. In Proceedings, ACM Symp. on Theory of Computing (STOC), pages 365–372, 2003.
[23] M. Hajiaghayi, V. Liaghat, and D. Panigrahi. Near-optimal online algorithms for prize-collecting
steiner problems. In Proceedings, International Colloquium on Automata, Languages and Processing
(ICALP), pages 576–587, 2014.
[24] M. T. Hajiaghayi, V. Liaghat, and D. Panigrahi. Online node-weighted steiner forest and extensions
via disk paintings. In Proceedings, IEEE Symposium on Foundations of Computer Science (FOCS),
pages 558–567, 2013.
[25] C. S. Helvig, G. Robins, and A. Zelikovsky. An improved approximation scheme for the group steiner
problem. Networks, 37(1):8–20, 2001.
[26] M. Imase and B. M. Waxman. Dynamic steiner tree problem. SIAM J. Discrete Math., 4(3):369–384,
1991.
[27] S. Korman. On the use of randomization in the online set cover problem. M.S. thesis, Weizmann
Institute of Science, 2005.
[28] A. Meyerson. Online algorithms for network design. In Proceedings, ACM Symposium on Parallelism
in Algorithms and Architectures (SPAA), pages 275–280, 2004.
[29] A. Meyerson, K. Munagala, and S. A. Plotkin. Cost-distance: Two metric network design. SIAM J.
Comput., 38(4):1648–1659, 2008.
[30] R. Motwani and P. Raghavan. Randomized Algorithms. Cambridge University Press, 1997.
[31] J. Naor, D. Panigrahi, and M. Singh. Online node-weighted steiner tree and related problems. In
Proceedings, IEEE Symposium on Foundations of Computer Science (FOCS), pages 210–219, 2011.
22
[32] F. S. Salman, J. Cheriyan, R. Ravi, and S. Subramanian. Approximating the single-sink linkinstallation problem in network design. SIAM J. Optimization, 11(3):595–610, 2001.
[33] K. Talwar. The single-sink buy-at-bulk LP has constant integrality gap. In Proceedings, MPS Conference on Integer Programming and Combinatorial Optimization (IPCO), pages 475–486, 2002.
[34] A. Zelikovsky. A series of approximation algorithms for the acyclic directed steiner tree problem.
Algorithmica, 18(1):99–110, 1997.
[35] L. Zosin and S. Khuller. On directed steiner trees. In Proceedings, ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 59–63, 2002.
A
Reduction to Layered Instances (Proof of Theorem 7)
In this section, we prove Theorem 7, which is an extension of Zelikovsky’s ‘height reduction lemma’ for the
buy-at-bulk problem; Zelikovsky’s original Lemma was for a single metric, whereas in our setting there is
both a cost and a length metric.
We prove the up-ward case; the down-ward case follows analogously. In order to simplify the notation,
we remove the superscript up . For this reduction, we will adapt the notion of layered expansion of a graph,
which has been in the folklore for many years and has been used recently by several papers (see, e.g.,[13,
31]). The h-level layered expansion of G is a layered DAG Gh of h+1 levels (we index the level 0, 1, . . . , h)
defined as follows:
(i) For each i such that 0 ≤ i ≤ h, the vertices in level i are copies of the vertices of G; we let vi to
denote the copy of vertex v ∈ V at level i.
(ii) For each i such that 1 ≤ i ≤ h, there is a directed edge from every vertex in level i to every vertex in
i from
level i − 1. The fixed cost of an edge (ui , vi−1 ) is given by that of the shortest directed path Puv
1−i/h
ui to vi−1 in G according to the metric ce + k
`e . The length of this edge is set to be the length
i in the ` metric.
of the path Puv
We now relate the optimal objective values for the two instances. One of the directions of the reduction
is straightforward.
Lemma 24. For any root r and any set of terminals X, if there is a feasible integral/fractional solution
of objective/LP value φ for the single-sink buy-at-bulk problem connecting X to r on the h-level layered
expansion Gh , then there is a feasible integral/fractional solution of objective/LP at most φ for the same
problem in G.
Proof: Note that for every edge in Gh , there is a corresponding path in G with the property that the sum of
edge costs and lengths on the path is at most the cost and length of the edge in G. Therefore, replacing the
edges in the solution for the layered graph by the corresponding paths in G yields a feasible solution in G
without increasing the overall cost and length. Notice that the same “embedding” of edges in Gh to paths in
G can be applied to the fractional solution on Gh as well. This shows property (ii) of the theorem statement.
The more interesting direction is to show that the optimal objective value on the layered graph Gh can be
bounded in terms of the optimal objective value on the original graph G. To show this, we will re-purpose
the so-called “height reduction” lemma of Helvig, Robins, and Zelikovsky [25]. We restate the lemma in a
form that will be useful for us.
23
Lemma 25. For any in-tree T defined on the edges of G that is rooted at r and contains all the terminals in
X, and for any integer h ≥ 1, there is an in-tree T 0 (on the same vertices as G but over a different edge set)
that is also rooted at r and contains all the terminals, and has the following properties:
(i) T 0 contains h + 1 levels of vertices, i.e., has height h.
(ii) T 0 is an k 1/h -ary tree, i.e., each non-leaf vertex has k 1/h children.
(iii) Each edge e0 = (u0 , v 0 ) in tree T 0 corresponds to the unique directed path pe0 in T from u0 to v 0 .
Moreover, the number of terminals in the subtree of T rooted at u0 is exactly k 1−i/h , where e0 is an
edge between levels i − 1 and i of T .
(iv) Each edge in T is in at most 2hk 1/h such paths pe0 for edges e0 ∈ T 0 .
For an edge e0 ∈ T 0 , suppose we define its cost to be the cost of the path pe0 , and its length to be the
length of the path pe0 . Then it is easy to see that the overall cost φT 0 of tree T 0 is O(hk 1/h ) times that of tree
T ; this is due to the following implications of the above lemma: (a) the total (buying) cost of all edges in T 0
is at most 2hk 1/h times that of T since each edge is reused at most 2hk 1/h times, and (b) for any terminal
x ∈ X, the edges on its path to the root in T 0 correspond to disjoint sub-paths in the unique path between x
and r in T , and hence the total length cost in T 0 is at most that in T .
Using this lemma, we can now complete the reduction by “embedding” the tree T 0 in the layered graph
Gh .
Lemma 26. If there is a feasible solution of overall cost φ for the single-sink buy-at-bulk problem on G, then
there is a feasible solution of overall cost O(hk 1/h φ) for the same problem on the h-level layered extension
Gh .
Proof: Let T be the union of the paths in the optimum solution on the graph G. It’s easy to see that T
is a directed in-tree. First, we use Lemma 25 to transform T to tree T 0 of height h. As noted earlier, the
overall cost φT 0 of T 0 is O(hk 1/h φ). Now, we construct a feasible tree Th in Gh using this solution T 0 as
follows: consider each edge (u, v) in T 0 where u is at level i and v is at level (i − 1). Then, include the edge
(ui , vi−1 ) in Th . Clearly, since T 0 connects all the terminals to the root, so does Th . Moreover, notice that
there is a 1-to-1 mapping between edges in T 0 and edges in Th .
To bound the objective value of the subtree, we relate the objective value for each edge of the subtree in
Gh to its corresponding mapped edge in T 0 . First, note that the overall contribution of an edge e0 = (u0 , v 0 )
between layers i and i + 1 towards φT 0 is equal to the sum of costs and k 1−i/h times the lengths of the
edges on the associated path pe0 from u0 to v 0 . This is because, by property (iii) of Lemma 25, the number
of demands in the subtree rooted at u0 is exactly k 1−i/h and all of them traverse this edge to reach r. Next,
0 ) between layers i and i + 1 in T is equal to the
we note that, by definition, the cost of the edge (u0i , vi+1
h
0
0
1−i/h
shortest directed path from u to v in G according to the metric ce + k
`e . Since we chose the shortest
0 ) ∈ T is at most the contribution of (u0 , v 0 ) towards
path, we get that the buying cost of edge (u0i , vi+1
h
φT 0 . Moreover, the total length cost is at most k 1−i/h times the length of the shortest path, which is at most
0 ) ∈ T (again, this uses the fact that there are exactly k 1−i/h terminals which
the fixed cost of (u0i , vi+1
h
route through this edge in Th also). It therefore follows that the overall cost of the solution that we have
inductively constructed in Gh is at most twice the overall cost of T 0 , which is at most O(hk 1/h )φ.
Theorem 7 follows from Lemmas 24 and 26.
24
| 8cs.DS
|
Counterfactual Multi-Agent Policy Gradients
Jakob N. Foerster1,†
Gregory Farquhar1,†
[email protected]
[email protected]
Triantafyllos Afouras1
Nantas Nardelli1
Shimon Whiteson1
[email protected]
[email protected]
[email protected]
arXiv:1705.08926v2 [cs.AI] 14 Dec 2017
1
University of Oxford, United Kingdom
Abstract
Many real-world problems, such as network packet routing
and the coordination of autonomous vehicles, are naturally
modelled as cooperative multi-agent systems. There is a great
need for new reinforcement learning methods that can efficiently learn decentralised policies for such systems. To
this end, we propose a new multi-agent actor-critic method
called counterfactual multi-agent (COMA) policy gradients.
COMA uses a centralised critic to estimate the Q-function
and decentralised actors to optimise the agents’ policies. In
addition, to address the challenges of multi-agent credit assignment, it uses a counterfactual baseline that marginalises
out a single agent’s action, while keeping the other agents’
actions fixed. COMA also uses a critic representation that allows the counterfactual baseline to be computed efficiently in
a single forward pass. We evaluate COMA in the testbed of
StarCraft unit micromanagement, using a decentralised variant with significant partial observability. COMA significantly
improves average performance over other multi-agent actorcritic methods in this setting, and the best performing agents
are competitive with state-of-the-art centralised controllers
that get access to the full state.
1
Introduction
Many complex reinforcement learning (RL) problems such
as the coordination of autonomous vehicles (Cao et al.
2013), network packet delivery (Ye, Zhang, and Yang 2015),
and distributed logistics (Ying and Dayong 2005) are naturally modelled as cooperative multi-agent systems. However, RL methods designed for single agents typically fare
poorly on such tasks, since the joint action space of the
agents grows exponentially with the number of agents.
To cope with such complexity, it is often necessary to resort to decentralised policies, in which each agent selects its
own action conditioned only on its local action-observation
history. Furthermore, partial observability and communication constraints during execution may necessitate the use of
decentralised policies even when the joint action space is not
prohibitively large.
Hence, there is a great need for new RL methods that
can efficiently learn decentralised policies. In some settings,
the learning itself may also need to be decentralised. However, in many cases, learning can take place in a simulator
or a laboratory in which extra state information is available and agents can communicate freely. This centralised
†
Equal contribution
training of decentralised policies is a standard paradigm for
multi-agent planning (Oliehoek, Spaan, and Vlassis 2008;
Kraemer and Banerjee 2016) and has recently been picked
up by the deep RL community (Foerster et al. 2016; Jorge,
Kågebäck, and Gustavsson 2016). However, the question of
how best to exploit the opportunity for centralised learning
remains open.
Another crucial challenge is multi-agent credit assignment (Chang, Ho, and Kaelbling 2003): in cooperative settings, joint actions typically generate only global rewards,
making it difficult for each agent to deduce its own contribution to the team’s success. Sometimes it is possible to
design individual reward functions for each agent. However,
these rewards are not generally available in cooperative settings and often fail to encourage individual agents to sacrifice for the greater good. This often substantially impedes
multi-agent learning in challenging tasks, even with relatively small numbers of agents.
In this paper, we propose a new multi-agent RL method
called counterfactual multi-agent (COMA) policy gradients,
in order to address these issues. COMA takes an actor-critic
(Konda and Tsitsiklis 2000) approach, in which the actor,
i.e., the policy, is trained by following a gradient estimated
by a critic. COMA is based on three main ideas.
First, COMA uses a centralised critic. The critic is only
used during learning, while only the actor is needed during
execution. Since learning is centralised, we can therefore use
a centralised critic that conditions on the joint action and all
available state information, while each agent’s policy conditions only on its own action-observation history.
Second, COMA uses a counterfactual baseline. The idea
is inspired by difference rewards (Wolpert and Tumer 2002;
Tumer and Agogino 2007), in which each agent learns from
a shaped reward that compares the global reward to the reward received when that agent’s action is replaced with a
default action. While difference rewards are a powerful way
to perform multi-agent credit assignment, they require access to a simulator or estimated reward function, and in general it is unclear how to choose the default action. COMA
addresses this by using the centralised critic to compute an
agent-specific advantage function that compares the estimated return for the current joint action to a counterfactual
baseline that marginalises out a single agent’s action, while
keeping the other agents’ actions fixed. This is similar to cal-
culating an aristocrat utility (Wolpert and Tumer 2002), but
avoids the problem of a recursive interdependence between
the policy and utility function because the expected contribution of the counterfactual baseline to the policy gradient
is zero. Hence, instead of relying on extra simulations, approximations, or assumptions regarding appropriate default
actions, COMA computes a separate baseline for each agent
that relies on the centralised critic to reason about counterfactuals in which only that agent’s action changes.
Third, COMA uses a critic representation that allows the
counterfactual baseline to be computed efficiently. In a single forward pass, it computes the Q-values for all the different actions of a given agent, conditioned on the actions of all
the other agents. Because a single centralised critic is used
for all agents, all Q-values for all agents can be computed in
a single batched forward pass.
We evaluate COMA in the testbed of StarCraft unit micromanagement1 , which has recently emerged as a challenging RL benchmark task with high stochasticity, a large stateaction space, and delayed rewards. Previous works (Usunier
et al. 2016; Peng et al. 2017) have made use of a centralised
control policy that conditions on the entire state and can use
powerful macro-actions, using StarCraft’s built-in planner,
that combine movement and attack actions. To produce a
meaningfully decentralised benchmark that proves challenging for scenarios with even relatively few agents, we propose
a variant that massively reduces each agent’s field-of-view
and removes access to these macro-actions.
Our empirical results on this new benchmark show that
COMA can significantly improve performance over other
multi-agent actor-critic methods, as well as ablated versions
of COMA itself. In addition, COMA’s best agents are competitive with state-of-the-art centralised controllers that are
given access to full state information and macro-actions.
2
Related Work
Although multi-agent RL has been applied in a variety of
settings (Busoniu, Babuska, and De Schutter 2008; Yang and
Gu 2004), it has often been restricted to tabular methods
and simple environments. One exception is recent work in
deep multi-agent RL, which can scale to high dimensional
input and action spaces. Tampuu et al. (2015) use a combination of DQN with independent Q-learning (Tan 1993;
Shoham and Leyton-Brown 2009) to learn how to play twoplayer pong. More recently the same method has been used
by Leibo et al. (2017) to study the emergence of collaboration and defection in sequential social dilemmas.
Also related is work on the emergence of communication
between agents, learned by gradient descent (Das et al. 2017;
Mordatch and Abbeel 2017; Lazaridou, Peysakhovich, and
Baroni 2016; Foerster et al. 2016; Sukhbaatar, Fergus, and
others 2016). In this line of work, passing gradients between
agents during training and sharing parameters are two common ways to take advantage of centralised training. However, these methods do not allow for extra state information
1
StarCraft and its expansion StarCraft: Brood War are trademarks of Blizzard EntertainmentTM .
to be used during learning and do not address the multi-agent
credit assignment problem.
Gupta, Egorov, and Kochenderfer (2017) investigate
actor-critic methods for decentralised execution with centralised training. However, in their methods both the actors
and the critic condition on local, per-agent, observations and
actions, and multi-agent credit assignment is addressed only
with hand-crafted local rewards.
Most previous applications of RL to StarCraft micromanagement use a centralised controller, with access to the full
state, and control of all units, although the architecture of
the controllers exploits the multi-agent nature of the problem. Usunier et al. (2016) use a greedy MDP, which at each
timestep sequentially chooses actions for agents given all
previous actions, in combination with zero-order optimisation, while Peng et al. (2017) use an actor-critic method that
relies on RNNs to exchange information between the agents.
The closest to our problem setting is that of Foerster et
al. (2017), who also use a multi-agent representation and decentralised policies. However, they focus on stabilising experience replay while using DQN and do not make full use
of the centralised training regime. As they do not report on
absolute win-rates we do not compare performance directly.
However, Usunier et al. (2016) address similar scenarios
to our experiments and implement a DQN baseline in a
fully observable setting. In Section 6 we therefore report our
competitive performance against these state-of-the-art baselines, while maintaining decentralised control. Omidshafiei
et al. (2017) also address the stability of experience replay in
multi-agent settings, but assume a fully decentralised training regime.
(Lowe et al. 2017) concurrently propose a multi-agent
policy-gradient algorithm using centralised critics. Their approach does not address multi-agent credit assignment. Unlike our work, it learns a separate centralised critic for each
agent and is applied to competitive environments with continuous action spaces.
Our work builds directly off of the idea of difference
rewards (Wolpert and Tumer 2002). The relationship of
COMA to this line of work is discussed in Section 4.
3
Background
We consider a fully cooperative multi-agent task that can
be described as a stochastic game G, defined by a tuple
G = hS, U, P, r, Z, O, n, γi, in which n agents identified
by a ∈ A ≡ {1, ..., n} choose sequential actions. The environment has a true state s ∈ S. At each time step, each
agent simultaneously chooses an action ua ∈ U , forming
a joint action u ∈ U ≡ U n which induces a transition in
the environment according to the state transition function
P (s0 |s, u) : S × U × S → [0, 1]. The agents all share the
same reward function r(s, u) : S × U → R and γ ∈ [0, 1)
is a discount factor.
We consider a partially observable setting, in which
agents draw observations z ∈ Z according to the observation function O(s, a) : S × A → Z. Each agent has an
action-observation history τ a ∈ T ≡ (Z × U )∗ , on which it
conditions a stochastic policy π a (ua |τ a ) : T × U → [0, 1].
We denote joint quantities over agents in bold, and joint
quantities over agents other than a given agent a with the
superscript −a.
P∞
The discounted return is Rt = l=0 γ l rt+l . The agents’
joint policy induces a value function, i.e., an expectation
over Rt , V π (st ) = Est+1:∞ ,ut:∞ [Rt |st ], and an actionvalue function Qπ (st , ut ) = Est+1:∞ ,ut+1:∞ [Rt |st , ut ].
The advantage function is given by Aπ (st , ut ) =
Qπ (st , ut ) − V π (st ).
Following previous work (Oliehoek, Spaan, and Vlassis
2008; Kraemer and Banerjee 2016; Foerster et al. 2016;
Jorge, Kågebäck, and Gustavsson 2016), our problem setting
allows centralised training but requires decentralised execution. This is a natural paradigm for a large set of multi-agent
problems where training is carried out using a simulator with
additional state information, but the agents must rely on local action-observation histories during execution. To condition on this full history, a deep RL agent may make use of a
recurrent neural network (Hausknecht and Stone 2015), typically with a gated model such as LSTM (Hochreiter and
Schmidhuber 1997) or GRU (Cho et al. 2014).
In Section 4, we develop a new multi-agent policy gradient method for tackling this setting. In the remainder of
this section, we provide some background on single-agent
policy gradient methods (Sutton et al. 1999). Such methods
optimise a single agent’s policy, parameterised by θπ , by
performing gradient ascent on an estimate of the expected
discounted total reward J = Eπ [R0 ]. Perhaps the simplest
form of policy gradient is REINFORCE (Williams 1992), in
which the gradient is:
" T
#
X
g = Es0:∞ ,u0:∞
Rt ∇θπ log π(ut |st ) .
(1)
t=0
In actor-critic approaches (Sutton et al. 1999; Konda and
Tsitsiklis 2000; Schulman et al. 2015), the actor, i.e., the
policy, is trained by following a gradient that depends on
a critic, which usually estimates a value function. In particular, Rt is replaced by any expression equivalent to
Q(st , ut ) − b(st ), where b(st ) is a baseline designed to reduce variance (Weaver and Tao 2001). A common choice is
b(st ) = V (st ), in which case Rt is replaced by A(st , ut ).
Another option is to replace Rt with the temporal difference
(TD) error rt + γV (st+1 ) − V (s), which is an unbiased estimate of A(st , ut ). In practice, the gradient must be estimated from trajectories sampled from the environment, and
the (action-)value functions must be estimated with function
approximators. Consequently, the bias and variance of the
gradient estimate depends strongly on the exact choice of
estimator (Konda and Tsitsiklis 2000).
In this paper, we train critics f c (·, θc ) on-policy to estimate either Q or V , using a variant of TD(λ) (Sutton
1988) adapted for use with deep neural networks. TD(λ)
Pn
(n)
uses a mixture of n-step returns Gt = l=1 γ l−1 rt+l +
γ n f c (·t+n , θc ). In particular, the critic parameters θc are updated by minibatch gradient descent to minimise the following loss:
Lt (θc ) = (y (λ) − f c (·t , θc ))2 ,
(2)
P∞
(n)
where y (λ) = (1 − λ) n=1 λn−1 Gt , and the n-step
(n)
returns Gt are calculated with bootstrapped values estimated by a target network (Mnih et al. 2015) with parameters copied periodically from θc .
4
Methods
In this section, we describe approaches for extending policy
gradients to our multi-agent setting.
Independent Actor-Critic
The simplest way to apply policy gradients to multiple
agents is to have each agent learn independently, with its
own actor and critic, from its own action-observation history. This is essentially the idea behind independent Qlearning (Tan 1993), which is perhaps the most popular
multi-agent learning algorithm, but with actor-critic in place
of Q-learning. Hence, we call this approach independent
actor-critic (IAC).
In our implementation of IAC, we speed learning by sharing parameters among the agents, i.e., we learn only one actor and one critic, which are used by all agents. The agents
can still behave differently because they receive different observations, including an agent-specific ID, and thus evolve
different hidden states. Learning remains independent in the
sense that each agent’s critic estimates only a local value
function, i.e., one that conditions on ua , not u. Though we
are not aware of previous applications of this specific algorithm, we do not consider it a significant contribution but
instead merely a baseline algorithm.
We consider two variants of IAC. In the first, each agent’s
critic estimates V (τ a ) and follows a gradient based on
the TD error, as described in Section 3. In the second,
each agent’s critic estimates Q(τ a , ua ) and follows a gradient based on the advantage:
A(τ a , ua ) = Q(τ a , ua ) −
P
V (τ a ), where V (τ a ) = ua π(ua |τ a )Q(τ a , ua ). Independent learning is straightforward, but the lack of information
sharing at training time makes it difficult to learn coordinated strategies that depend on interactions between multiple agents, or for an individual agent to estimate the contribution of its actions to the team’s reward.
Counterfactual Multi-Agent Policy Gradients
The difficulties discussed above arise because, beyond parameter sharing, IAC fails to exploit the fact that learning
is centralised in our setting. In this section, we propose
counterfactual multi-agent (COMA) policy gradients, which
overcome this limitation. Three main ideas underly COMA:
1) centralisation of the critic, 2) use of a counterfactual baseline, and 3) use of a critic representation that allows efficient
evaluation of the baseline. The remainder of this section describes these ideas.
First, COMA uses a centralised critic. Note that in IAC,
each actor π(ua |τ a ) and each critic Q(τ a , ua ) or V (τ a )
conditions only on the agent’s own action-observation history τ a . However, the critic is used only during learning and
only the actor is needed during execution. Since learning is
centralised, we can therefore use a centralised critic that conditions on the true global state s, if it is available, or the joint
action-observation histories τ otherwise. Each actor conditions on its own action-observation histories τ a , with parameter sharing, as in IAC. Figure 1a illustrates this setup.
A naive way to use this centralised critic would be for
each actor to follow a gradient based on the TD error estimated from this critic:
g = ∇θπ log π(u|τta ) (r + γV (st+1 ) − V (st )) .
(3)
However, such an approach fails to address a key credit
assignment problem. Because the TD error considers only
global rewards, the gradient computed for each actor does
not explicitly reason about how that particular agent’s actions contribute to that global reward. Since the other agents
may be exploring, the gradient for that agent becomes very
noisy, particularly when there are many agents.
Therefore, COMA uses a counterfactual baseline. The
idea is inspired by difference rewards (Wolpert and Tumer
2002), in which each agent learns from a shaped reward
Da = r(s, u) − r(s, (u−a , ca )) that compares the global
reward to the reward received when the action of agent a is
replaced with a default action ca . Any action by agent a that
improves Da also improves the true global reward r(s, u),
because r(s, (u−a , ca )) does not depend on agent a’s actions.
Difference rewards are a powerful way to perform multiagent credit assignment. However, they typically require
access to a simulator in order to estimate r(s, (u−a , ca )).
When a simulator is already being used for learning, difference rewards increase the number of simulations that
must be conducted, since each agent’s difference reward
requires a separate counterfactual simulation. Proper and
Tumer (2012) and Colby, Curran, and Tumer (2015) propose estimating difference rewards using function approximation rather than a simulator. However, this still requires a
user-specified default action ca that can be difficult to choose
in many applications. In an actor-critic architecture, this approach would also introduce an additional source of approximation error.
A key insight underlying COMA is that a centralised critic
can be used to implement difference rewards in a way that
avoids these problems. COMA learns a centralised critic,
Q(s, u) that estimates Q-values for the joint action u conditioned on the central state s. For each agent a we can then
compute an advantage function that compares the Q-value
for the current action ua to a counterfactual baseline that
marginalises out ua , while keeping the other agents’ actions
u−a fixed:
X
Aa (s, u) = Q(s, u) −
π a (u0a |τ a )Q(s, (u−a , u0a )).
u0a
(4)
Hence, Aa (s, ua ) computes a separate baseline for each
agent that uses the centralised critic to reason about counterfactuals in which only a’s action changes, learned directly
from agents’ experiences instead of relying on extra simulations, a reward model, or a user-designed default action.
This advantage has the same form as the aristocrat utility (Wolpert and Tumer 2002). However, optimising for an
aristocrat utility using value-based methods creates a selfconsistency problem because the policy and utility function
depend recursively on each other. As a result, prior work
focused on difference evaluations using default states and
actions. COMA is different because the counterfactual baseline’s expected contribution to the gradient, as with other
policy gradient baselines, is zero. Thus, while the baseline
does depend on the policy, its expectation does not. Consequently, COMA can use this form of the advantage without
creating a self-consistency problem.
While COMA’s advantage function replaces potential extra simulations with evaluations of the critic, those evaluations may themselves be expensive if the critic is a deep
neural network. Furthermore, in a typical representation, the
number of output nodes of such a network would equal |U |n ,
the size of the joint action space, making it impractical to
train. To address both these issues, COMA uses a critic representation that allows for efficient evaluation of the baseline. In particular, the actions of the other agents, u−a
t , are
part of the input to the network, which outputs a Q-value
for each of agent a’s actions, as shown in Figure 1c. Consequently, the counterfactual advantage can be calculated efficiently by a single forward pass of the actor and critic, for
each agent. Furthermore, the number of outputs is only |U |
instead of (|U |n ). While the network has a large input space
that scales linearly in the number of agents and actions, deep
neural networks can generalise well across such spaces.
In this paper, we focus on settings with discrete actions.
However, COMA can be easily extended to continuous actions spaces by estimating the expectation in (4) with Monte
Carlo samples or using functional forms that render it analytical, e.g., Gaussian policies and critic.
The following lemma establishes the convergence of
COMA to a locally optimal policy. The proof follows directly from the convergence of single-agent actor-critic algorithms (Sutton et al. 1999; Konda and Tsitsiklis 2000),
and is subject to the same assumptions.
Lemma 1. For an actor-critic algorithm with a compatible
TD(1) critic following a COMA policy gradient
"
#
X
a a a
a
gk = Eπ
∇θk log π (u |τ )A (s, u)
(5)
a
at each iteration k,
lim inf ||∇J|| = 0 w.p. 1.
(6)
Proof. The COMA gradient is given by
"
#
X
a a a
a
g = Eπ
∇θ log π (u |τ )A (s, u) ,
(7)
k
a
Aa (s, u) = Q(s, u) − b(s, u−a ),
(8)
where θ are the parameters of all actor policies, e.g. θ =
{θ1 , . . . , θ|A| }, and b(s, u−a ) is the counterfactual baseline
defined in equation 4.
First consider the expected contribution of the this baseline b(s, u−a ):
"
#
X
a a a
−a
gb = −Eπ
∇θ log π (u |τ )b(s, u ) , (9)
a
A1 t
A2 t
Critic
1
(h , )
h
1
t
(h , )
u
1
= (hat, )
Aa t
u
t
h
t
2
h at
Actor 2
st
u1 t
rt
u2 t
(uat,
( )
2
Actor 1
o 1t
a
2
(hat-1)
GRU
a
t
)
COMA
{Q(ua=1, u-at,..),. .,Q(ua=|U|, u-at,..)}
(hat)
o 2t
Environment
(oat, a, uat-1)
(u-at, st, oat, a, ut-1)
(a)
(b)
(c)
Figure 1: In (a), information flow between the decentralised actors, the environment and the centralised critic in COMA; red
arrows and components are only required during centralised learning. In (b) and (c), architectures of the actor and critic.
where the expectation Eπ is with respect to the state-action
distribution induced by the joint policy π. Now let dπ (s) be
the discounted ergodic state distribution as defined by Sutton
et al. (1999):
X
XX
gb = −
dπ (s)
π(u−a |τ −a) ·
s
a u−a
X
π (u |τ a )∇θ log π a (ua |τ a )b(s, u−a )
a
a
(10)
ua
=−
X
dπ (s)
XX
s
π(u−a |τ −a) ·
a u−a
X
=−
dπ (s)
XX
s
a
(11)
π(u−a |τ −a)b(s, u−a )∇θ 1
u−a
= 0.
(12)
Clearly, the per-agent baseline, although it reduces variance,
does not change the expected gradient, and therefore does
not affect the convergence of COMA.
The remainder of the expected policy gradient is given by:
"
#
X
a a a
g = Eπ
∇θ log π (u |τ )Q(s, u)
(13)
a
"
= Eπ ∇θ log
#
Y
a
a
a
π (u |τ )Q(s, u) .
(14)
a
Writing the joint policy as a product of the independent actors:
Y
π(u|s) =
π a (ua |τ a ),
(15)
a
yields the standard single-agent actor-critic policy gradient:
g = Eπ [∇θ log π(u|s)Q(s, u)] .
(16)
Konda and Tsitsiklis (2000) prove that an actor-critic following this gradient converges to a local maximum of the
expected return J π , given that:
1. the policy π is differentiable,
amongst several further assumptions. The parameterisation
of the policy (i.e., the single-agent joint-action learner is decomposed into independent actors) is immaterial to convergence, as long as it remains differentiable. Note however
that COMA’s centralised critic is essential for this proof to
hold.
5
∇θ π (ua |τ a )b(s, u−a )
a
ua
X
2. the update timescales for Q and π are sufficiently slow,
and that π is updated sufficiently slower than Q, and
3. Q uses a representation compatible with π,
Experimental Setup
In this section, we describe the StarCraft problem to which
we apply COMA, as well as details of the state features,
network architectures, training regimes, and ablations.
Decentralised StarCraft Micromanagement. StarCraft
is a rich environment with stochastic dynamics that cannot
be easily emulated. Many simpler multi-agent settings, such
as Predator-Prey (Tan 1993) or Packet World (Weyns, Helleboogh, and Holvoet 2005), by contrast, have full simulators
with controlled randomness that can be freely set to any state
in order to perfectly replay experiences. This makes it possible, though computationally expensive, to compute difference rewards via extra simulations. In StarCraft, as in the
real world, this is not possible.
In this paper, we focus on the problem of micromanagement in StarCraft, which refers to the low-level control of individual units’ positioning and attack commands
as they fight enemies. This task is naturally represented as
a multi-agent system, where each StarCraft unit is replaced
by a decentralised controller. We consider several scenarios
with symmetric teams formed of: 3 marines (3m), 5 marines
(5m), 5 wraiths (5w), or 2 dragoons with 3 zealots (2d 3z).
The enemy team is controlled by the StarCraft AI, which
uses reasonable but suboptimal hand-crafted heuristics.
We allow the agents to choose from a set of discrete
actions: move[direction], attack[enemy id],
stop, and noop. In the StarCraft game, when a unit selects
an attack action, it first moves into attack range before
firing, using the game’s built-in pathfinding to choose a
route. These powerful attack-move macro-actions make the
control problem considerably easier.
To create a more challenging benchmark that is meaningfully decentralised, we impose a restricted field of view
on the agents, equal to the firing range of ranged units’
weapons, shown in Figure 2. This departure from the standard setup for centralised StarCraft control has three effects.
x
Figure 2: Starting position with example local field of view
for the 2d 3z map.
First, it introduces significant partial observability. Second, it means units can only attack when they are in range
of enemies, removing access to the StarCraft macro-actions.
Third, agents cannot distinguish between enemies who are
dead and those who are out of range and so can issue invalid attack commands at such enemies, which results in no
action being taken. This substantially increases the average
size of the action space, which in turn increases the difficulty
of both exploration and control.
Under these difficult conditions, scenarios with even relatively small numbers of units become much harder to solve.
As seen in Table 1, we compare against a simple hand-coded
heuristic that instructs the agents to run forwards into range
and then focus their fire, attacking each enemy in turn until
it dies. This heuristic achieves a 98% win rate on 5m with
a full field of view, but only 66% in our setting. To perform
well in this task, the agents must learn to cooperate by positioning properly and focussing their fire, while remembering
which enemy and ally units are alive or out of view.
All agents receive the same global reward at each time
step, equal to the sum of damage inflicted on the opponent
units minus half the damage taken. Killing an opponent generates a reward of 10 points, and winning the game generates
a reward equal to the team’s remaining total health plus 200.
This damage-based reward signal is comparable to that used
by Usunier et al. (2016). Unlike (Peng et al. 2017), our approach does not require estimating local rewards.
State Features. The actor and critic receive different input features, corresponding to local observations and global
state, respectively. Both include features for allies and enemies. Units can be either allies or enemies, while agents are
the decentralised controllers that command ally units.
The local observations for every agent are drawn only
from a circular subset of the map centred on the unit it
controls and include for each unit within this field of view:
distance, relative x, relative y, unit type
and shield.2 All features are normalised by their maximum values. We do not include any information about the
units’ current target.
The global state representation consists of similar features, but for all units on the map regardless of fields of
view. Absolute distance is not included, and x-y locations
are given relative to the centre of the map rather than to
a particular agent. The global state also includes health
points and cooldown for all agents. The representation fed to the centralised Q-function critic is the concatenation of the global state representation with the local observation of the agent whose actions are being evaluated.
Our centralised critic that estimates V (s), and is therefore
agent-agnostic, receives the global state concatenated with
all agents’ observations. The observations contain no new
information but include the egocentric distances relative to
that agent.
Architecture & Training. The actor consists of 128-bit
gated recurrent units (GRUs) (Cho et al. 2014) that use
fully connected layers both to process the input and to produce the output values from the hidden state, hat . The IAC
critics use extra output heads appended to the last layer of
the actor network. Action probabilities are produced from
the final layer, z, via a bounded softmax distribution that
lower-bounds the probability of any given action by /|U |:
P (u) = (1 − )softmax(z)u + /|U |). We anneal linearly from 0.5 to 0.02 across 750 training episodes. The centralised critic is a feedforward network with multiple ReLU
layers combined with fully connected layers. Hyperparameters were coarsely tuned on the 5m scenario and then used
for all other maps. We found that the most sensitive parameter was TD(λ), but settled on λ = 0.8, which worked best
for both COMA and our baselines. Our implementation uses
TorchCraft (Synnaeve et al. 2016) and Torch 7 (Collobert,
Kavukcuoglu, and Farabet 2011). Pseudocode and further
details on the training procedure are in the appendix.
We experimented with critic architectures that are factored at the agent level and further exploit internal parameter
sharing. However, we found that the bottleneck for scalability was not the centralisation of the critic, but rather the difficulty of multi-agent exploration. Hence, we defer further
investigation of factored COMA critics to future work.
Ablations. We perform ablation experiments to validate
three key elements of COMA. First, we test the importance
of centralising the critic by comparing against two IAC variants, IAC-Q and IAC-V . These critics take the same decentralised input as the actor, and share parameters with the actor network up to the final layer. IAC-Q then outputs |U |
Q-values, one for each action, while IAC-V outputs a single state-value. Note that we still share parameters between
agents, using the egocentric observations and ID’s as part of
the input to allow different behaviours to emerge. The cooperative reward function is still shared by all agents.
Second, we test the significance of learning Q instead of
2
After firing, a unit’s cooldown is reset, and it must drop
before firing again. Shields absorb damage until they break, after
which units start losing health. Dragoons and zealots have shields
but marines do not.
90
80
70
60
50
40
30
20
10
0
IAC-V
c e ntral-QV
he uris tic
Ave rag e Win %
Ave rag e Win %
COMA
c e ntral-V
20k 40k 60k 80k 100k 120k 140k
# Epis o de s
90
80
70
60
50
40
30
20
10
0
90
80
70
60
50
40
30
20
10
0
10k 20k 30k 40k 50k 60k 70k
# Epis o de s
(b) 5m
70
Ave rag e Win %
Ave rag e Win %
(a) 3m
IAC-Q
5k
10k 15k 20k 25k 30k 35k
# Epis o de s
(c) 5w
60
50
40
30
20
10
0
5k 10k 15k 20k 25k 30k 35k 40k
# Epis o de s
(d) 2d 3z
Figure 3: Win rates for COMA and competing algorithms on four different scenarios. COMA outperforms all baseline methods.
Centralised critics also clearly outperform their decentralised counterparts. The legend at the top applies across all plots.
V . The method central-V still uses a central state for the
critic, but learns V (s), and uses the TD error to estimate the
advantage for policy gradient updates.
Third, we test the utility of our counterfactual baseline.
The method central-QV learns both Q and V simultaneously and estimates the advantage as Q − V , replacing
COMA’s counterfactual baseline with V . All methods use
the same architecture and training scheme for the actors, and
all critics are trained with TD(λ).
6
Results
Figure 3 shows average win rates as a function of episode for
each method and each StarCraft scenario. For each method,
we conducted 35 independent trials and froze learning every 100 training episodes to evaluate the learned policies
across 200 episodes per method, plotting the average across
episodes and trials. Also shown is one standard deviation in
performance.
The results show that COMA is superior to the IAC baselines in all scenarios. Interestingly, the IAC methods also
eventually learn reasonable policies in 5m, although they
need substantially more episodes to do so. This may seem
counterintuitive since in the IAC methods, the actor and
critic networks share parameters in their early layers (see
Section 5), which could be expected to speed learning. However, these results suggest that the improved accuracy of policy evaluation made possible by conditioning on the global
state outweighs the overhead of training a separate network.
Furthermore, COMA strictly dominates central-QV , both
in training speed and in final performance across all settings.
This is a strong indicator that our counterfactual baseline is
crucial when using a central Q-critic to train decentralised
policies.
Learning a state-value function has the obvious advantage
of not conditioning on the joint action. Still, we find that
COMA outperforms the central-V baseline in final performance. Furthermore, COMA typically achieves good policies faster, which is expected as COMA provides a shaped
training signal. Training is also more stable than central-V ,
which is a consequence of the COMA gradient tending to
zero as the policy becomes greedy. Overall, COMA is the
best performing and most consistent method.
Usunier et al. (2016) report the performance of their best
agents trained with their state-of-the-art centralised controller labelled GMEZO (greedy-MDP with episodic zeroorder optimisation), and for a centralised DQN controller,
both given a full field of view and access to attack-move
macro-actions. These results are compared in Table 1 against
the best agents trained with COMA for each map. Clearly,
Local Field of View (FoV)
map
heur.
IAC-V
IAC-Q
cnt-V
cnt-QV
3m
5m
5w
2d 3z
35
66
70
63
47 (3)
63 (2)
18 (5)
27 (9)
56 (6)
58 (3)
57 (5)
19 (21)
83 (3)
67 (5)
65 (3)
36 (6)
83 (5)
71 (9)
76 (1)
39 (5)
Full FoV, Central Control
COMA
mean best
87 (3)
81 (5)
82 (3)
47 (5)
98
95
98
65
heur.
DQN
GMEZO
74
98
82
68
99
70
61
100
743
90
Table 1: Mean win percentage averaged across final 1000 evaluation episodes for the different maps, for all methods and the
hand-coded heuristic in the decentralised setting with a limited field of view. The highest mean performances are in bold,
while the values in parentheses denote the 95% confidence interval, for example 87(3) = 87 ± 3. Also shown, maximum
win percentages for COMA (decentralised), in comparison to the heuristic and published results (evaluated in the centralised
setting).
in most settings these agents achieve performance comparable to the best published win rates despite being restricted to
decentralised policies and local fields of view.
7
Conclusions & Future Work
This paper presented COMA policy gradients, a method
that uses a centralised critic in order to estimate a counterfactual advantage for decentralised policies in mutliagent RL. COMA addresses the challenges of multi-agent
credit assignment by using a counterfactual baseline that
marginalises out a single agent’s action, while keeping
the other agents’ actions fixed. Our results in a decentralised StarCraft unit micromanagement benchmark show
that COMA significantly improves final performance and
training speed over other multi-agent actor-critic methods
and remains competitive with state-of-the-art centralised
controllers under best-performance reporting. Future work
will extend COMA to tackle scenarios with large numbers
of agents, where centralised critics are more difficult to train
and exploration is harder to coordinate. We also aim to develop more sample-efficient variants that are practical for
real-world applications such as self-driving cars.
Acknowledgements
This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement number 637713). It was also supported by the OxfordGoogle DeepMind Graduate Scholarship, the UK EPSRC
CDT in Autonomous Intelligent Machines and Systems, and
a generous grant from Microsoft for their Azure cloud computing services. We would like to thank Nando de Freitas,
Yannis Assael, and Brendan Shillingford for helpful comments and discussion. We also thank Gabriel Synnaeve,
Zeming Lin, and the rest of the TorchCraft team at FAIR
for their work on the interface.
3
5w DQN and GMEZO benchmark performances are of a policy trained on a larger map and tested on 5w
References
Busoniu, L.; Babuska, R.; and De Schutter, B. 2008. A
comprehensive survey of multiagent reinforcement learning.
IEEE Transactions on Systems Man and Cybernetics Part C
Applications and Reviews 38(2):156.
Cao, Y.; Yu, W.; Ren, W.; and Chen, G. 2013. An overview
of recent progress in the study of distributed multi-agent
coordination. IEEE Transactions on Industrial informatics
9(1):427–438.
Chang, Y.-H.; Ho, T.; and Kaelbling, L. P. 2003. All learning
is local: Multi-agent learning in global reward games. In
NIPS, 807–814.
Cho, K.; van Merriënboer, B.; Bahdanau, D.; and Bengio, Y. 2014. On the properties of neural machine
translation: Encoder-decoder approaches. arXiv preprint
arXiv:1409.1259.
Colby, M. K.; Curran, W.; and Tumer, K. 2015. Approximating difference evaluations with local information. In
Proceedings of the 2015 International Conference on Autonomous Agents and Multiagent Systems, 1659–1660. International Foundation for Autonomous Agents and Multiagent Systems.
Collobert, R.; Kavukcuoglu, K.; and Farabet, C. 2011.
Torch7: A matlab-like environment for machine learning. In
BigLearn, NIPS Workshop.
Das, A.; Kottur, S.; Moura, J. M.; Lee, S.; and Batra, D.
2017. Learning cooperative visual dialog agents with deep
reinforcement learning. arXiv preprint arXiv:1703.06585.
Foerster, J.; Assael, Y. M.; de Freitas, N.; and Whiteson, S.
2016. Learning to communicate with deep multi-agent reinforcement learning. In Advances in Neural Information
Processing Systems, 2137–2145.
Foerster, J.; Nardelli, N.; Farquhar, G.; Torr, P.; Kohli, P.;
Whiteson, S.; et al. 2017. Stabilising experience replay for
deep multi-agent reinforcement learning. In Proceedings of
The 34th International Conference on Machine Learning.
Gupta, J. K.; Egorov, M.; and Kochenderfer, M. 2017. Coop-
erative multi-agent control using deep reinforcement learning.
Hausknecht, M., and Stone, P. 2015. Deep recurrent
q-learning for partially observable mdps. arXiv preprint
arXiv:1507.06527.
Hochreiter, S., and Schmidhuber, J. 1997. Long short-term
memory. Neural computation 9(8):1735–1780.
Jorge, E.; Kågebäck, M.; and Gustavsson, E. 2016. Learning
to play guess who? and inventing a grounded language as a
consequence. arXiv preprint arXiv:1611.03218.
Konda, V. R., and Tsitsiklis, J. N. 2000. Actor-critic algorithms. In Advances in neural information processing systems, 1008–1014.
Kraemer, L., and Banerjee, B. 2016. Multi-agent reinforcement learning as a rehearsal for decentralized planning. Neurocomputing 190:82–94.
Lazaridou, A.; Peysakhovich, A.; and Baroni, M. 2016.
Multi-agent cooperation and the emergence of (natural) language. arXiv preprint arXiv:1612.07182.
Leibo, J. Z.; Zambaldi, V.; Lanctot, M.; Marecki, J.; and
Graepel, T. 2017. Multi-agent reinforcement learning in sequential social dilemmas. arXiv preprint arXiv:1702.03037.
Lowe, R.; Wu, Y.; Tamar, A.; Harb, J.; Abbeel, P.; and
Mordatch, I. 2017. Multi-agent actor-critic for mixed
cooperative-competitive environments.
arXiv preprint
arXiv:1706.02275.
Mnih, V.; Kavukcuoglu, K.; Silver, D.; Rusu, A. A.; Veness, J.; Bellemare, M. G.; Graves, A.; Riedmiller, M.;
Fidjeland, A. K.; Ostrovski, G.; et al. 2015. Humanlevel control through deep reinforcement learning. Nature
518(7540):529–533.
Mordatch, I., and Abbeel, P. 2017. Emergence of grounded
compositional language in multi-agent populations. arXiv
preprint arXiv:1703.04908.
Oliehoek, F. A.; Spaan, M. T. J.; and Vlassis, N. 2008. Optimal and approximate Q-value functions for decentralized
POMDPs. 32:289–353.
Omidshafiei, S.; Pazis, J.; Amato, C.; How, J. P.; and Vian,
J. 2017. Deep decentralized multi-task multi-agent rl under
partial observability. arXiv preprint arXiv:1703.06182.
Peng, P.; Yuan, Q.; Wen, Y.; Yang, Y.; Tang, Z.; Long, H.;
and Wang, J. 2017. Multiagent bidirectionally-coordinated
nets for learning to play starcraft combat games. arXiv
preprint arXiv:1703.10069.
Proper, S., and Tumer, K. 2012. Modeling difference rewards for multiagent learning. In Proceedings of the 11th
International Conference on Autonomous Agents and Multiagent Systems-Volume 3, 1397–1398. International Foundation for Autonomous Agents and Multiagent Systems.
Schulman, J.; Moritz, P.; Levine, S.; Jordan, M. I.; and
Abbeel, P. 2015. High-dimensional continuous control using
generalized advantage estimation. CoRR abs/1506.02438.
Shoham, Y., and Leyton-Brown, K. 2009. Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foundations. New York: Cambridge University Press.
Sukhbaatar, S.; Fergus, R.; et al. 2016. Learning multiagent communication with backpropagation. In Advances in
Neural Information Processing Systems, 2244–2252.
Sutton, R. S.; McAllester, D. A.; Singh, S. P.; Mansour, Y.;
et al. 1999. Policy gradient methods for reinforcement
learning with function approximation. In NIPS, volume 99,
1057–1063.
Sutton, R. S. 1988. Learning to predict by the methods of
temporal differences. Machine learning 3(1):9–44.
Synnaeve, G.; Nardelli, N.; Auvolat, A.; Chintala, S.;
Lacroix, T.; Lin, Z.; Richoux, F.; and Usunier, N. 2016.
Torchcraft: a library for machine learning research on realtime strategy games. arXiv preprint arXiv:1611.00625.
Tampuu, A.; Matiisen, T.; Kodelja, D.; Kuzovkin, I.; Korjus,
K.; Aru, J.; Aru, J.; and Vicente, R. 2015. Multiagent cooperation and competition with deep reinforcement learning.
arXiv preprint arXiv:1511.08779.
Tan, M. 1993. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth
international conference on machine learning, 330–337.
Tumer, K., and Agogino, A. 2007. Distributed agent-based
air traffic flow management. In Proceedings of the 6th international joint conference on Autonomous agents and multiagent systems, 255. ACM.
Usunier, N.; Synnaeve, G.; Lin, Z.; and Chintala, S. 2016.
Episodic exploration for deep deterministic policies: An application to starcraft micromanagement tasks. arXiv preprint
arXiv:1609.02993.
Weaver, L., and Tao, N. 2001. The optimal reward baseline
for gradient-based reinforcement learning. In Proceedings
of the Seventeenth conference on Uncertainty in artificial intelligence, 538–545. Morgan Kaufmann Publishers Inc.
Weyns, D.; Helleboogh, A.; and Holvoet, T. 2005. The
packet-world: A test bed for investigating situated multiagent systems. In Software Agent-Based Applications, Platforms and Development Kits. Springer. 383–408.
Williams, R. J. 1992. Simple statistical gradient-following
algorithms for connectionist reinforcement learning. Machine learning 8(3-4):229–256.
Wolpert, D. H., and Tumer, K. 2002. Optimal payoff functions for members of collectives. In Modeling complexity in
economic and social systems. World Scientific. 355–369.
Yang, E., and Gu, D. 2004. Multiagent reinforcement learning for multi-robot systems: A survey. Technical report,
tech. rep.
Ye, D.; Zhang, M.; and Yang, Y. 2015. A multi-agent framework for packet routing in wireless sensor networks. sensors
15(5):10026–10047.
Ying, W., and Dayong, S. 2005. Multi-agent framework
for third party logistics in e-commerce. Expert Systems with
Applications 29(2):431–436.
Appendix
Training Details and Hyperparameters
Training is performed in batch mode, with a batch size of 30. Due to parameter sharing, all agents can be processed in parallel,
with each agent for each episode and time step occupying one batch entry. The training cycle progresses in three steps (completion of all three steps constitutes as one episode in our graphs): 1) collect data: collect 30
n episodes; 2) train critic: for each
time step, apply a gradient step to the feed-forward critic, starting at the end of the episode; and 3) train actor: fully unroll the
recurrent part of the actor, aggregate gradients in the backward pass across all time steps, and apply a gradient update.
We use a target network for the critic, which updates every 150 training steps for the feed-forward centralised critics and
every 50 steps for the recurrent IAC critics. The feed-forward critic receives more learning steps, since it performs a parameter
update for each timestep. Both the actor and the critic networks are trained using RMS-prop with learning rate 0.0005 and alpha
0.99, without weight decay. We set gamma to 0.99 for all maps.
Although tuning the skip-frame in StarCraft can improve absolute performance (Peng et al. 2017), we use a default value of
7, since the main focus is a relative evaluation between COMA and the baselines.
Algorithm
Algorithm 1 Counterfactual Multi-Agent (COMA) Policy Gradients
Initialise θ1c , θ̂1c , θπ
for each training episode e do
Empty buffer
do
for ec = 1 to BatchSize
n
s1 = initial state, t = 0, ha0 = 0 for each agent a
while st 6= terminal and t < T do
t=t+1
for each agent a do
hat = Actor oat , hat−1 , uat−1 , a, u; θi
Sample uat from π(hat , (e))
Get reward rt and next state st+1
Add episode to buffer
Collate episodes in buffer into single batch
for t = 1 to T do // from now processing all agents in parallel via single batch
Batch unroll RNN using states, actions and rewards
Calculate TD(λ) targets yta using θ̂ic
for t = T down to 1 do
∆Qat = yta − Q saj , u
∆θc = ∇θc (∆Qat )2 // calculate critic gradient
c
θi+1
= θic − α∆θc // update critic weights
Every C steps reset θ̂ic = θic
for t = T down to 1 do
P
Aa (sat , u) = Q(sat , u) − u Q(sat , u, u−a )π(u|hat ) // calculate COMA
∆θπ = ∆θπ + ∇θπ log π(u|hat )Aa (sat , u) // accumulate actor gradients
π
θi+1 = θiπ + α∆θπ // update actor weights
| 2cs.AI
|
An economic approach to vehicle dispatching for ride sharing
Mengjing Chen, Weiran Shen, Pingzhong Tang, and Song Zuo
IIIS, Tsinghua University ∗
arXiv:1707.01625v2 [cs.SY] 1 Mar 2018
March 2, 2018
Abstract
Over the past few years, ride-sharing has emerged as an effective way to relieve traffic congestion. A key problem
for these platforms is to come up with a revenue-optimal (or GMV-optimal) pricing scheme and an induced vehicle
dispatching policy that incorporate geographic and temporal information. In this paper, we aim to tackle this problem
via an economic approach.
Modeled naively, the underlying optimization problem may be non-convex and thus hard to compute. To this
end, we use a so-called “ironing” technique to convert the problem into an equivalent convex optimization one via a
clean Markov decision process (MDP) formulation, where the states are the driver distributions and the decision
variables are the prices for each pair of locations. Our main finding is an efficient algorithm that computes the exact
revenue-optimal (or GMV-optimal) randomized pricing schemes. We characterize the optimal solution of the MDP
by a primal-dual analysis of a corresponding convex program. We also conduct empirical evaluations of our solution
through real data of a major ride-sharing platform and show its advantages over fixed pricing schemes as well as
several prevalent surge-based pricing schemes.
1
Introduction
The recently established applications of shared mobility, such as ride-sharing, bike-sharing, and car-sharing, have been
proven to be an effective way to utilize redundant transportation resources and to optimize social efficiency (Cramer
and Krueger, 2016). Over the past few years, intensive researches have been done on topics related to the economic
aspects of shared mobility (Crawford and Meng, 2011; Kostiuk, 1990; Oettinger, 1999).
Despite these researches, the problem of how to design revenue optimal prices and vehicle dispatching schemes
has been largely open and one of the main research agendas in sharing economics. There are at least two challenges
when one wants to tackle this problem in the real-world applications. First of all, due to the nature of transportation,
the price and dispatch scheme must be geographically dependent. Secondly, the price and dispatch scheme must take
into consideration the fact that supplies and demands in these environments may change over time. As a result, it
may be difficult to compute, or even to represent a price and dispatch scheme for such complex environments.
Traditional price and dispatch schemes for taxis (Laporte, 1992; Gendreau et al., 1994; Ghiani et al., 2003) and
airplanes (Gale and Holmes, 1993; Stavins, 2001; McAfee and Te Velde, 2006) do not capture the dynamic aspects of
the environments: taxi fees are normally calculated by a fixed rate of distance and time and the prices of flight tickets
are sold via relatively long booking periods, while in contrast, the customers of shared vehicles make their decisions
instantly.
The dynamic ride-sharing market studied in this paper is also known to have imbalanced supply and demand,
either globally in a city or locally in a particular time and location. Such imbalance in supply and demand is known to
cause severe consequences on revenues (e.g, the so-called wild goose chase phenomenon (Castillo et al., 2017)). Surging
price is a way to balance dynamic supply and demand (Chen and Sheldon, 2015) but there is no known guarantee that
surge based pricing can dispatch vehicles efficiently and solve the imbalanced supplies and demands. Traditional
dispatch schemes (Laporte, 1992; Gendreau et al., 1994; Ghiani et al., 2003) focus more on the algorithmic aspect of
static vehicle routing, without consider pricing. However, vehicle dispatching and pricing problem are tightly related,
since a new price scheme will surely induces a change on supply and demand since the drivers and passengers are
strategic. In this paper, we aim to come up with price schemes with desirable induced supplies and demands.
∗ Contacts:
[email protected], {emersonswr, kenshinping, songzuo.z}@gmail.com
1
1.1
Our contribution
In this paper, we propose a graph model to analyze the vehicle pricing and dispatching problem mentioned above. In
the graph, each node refers to a region in the city and each edge refers to a possible trip that includes a pair of origin
and destination as well as a cost associated with the trip on this edge. The design problem is, for the platform, to set a
price and specify the vehicle dispatch for each edge at each time step. Drivers are considered to be non-strategic in
our model, meaning that they will accept whatever offer assigned to them. The objective of the platform can either be
its revenue or the GMV or any convex combination between them.
Our model naturally induces a Markov Decision Process (MDP) with the driver distributions on each node as
states, the price and dispatch along each edge as actions, and the revenue as immediately reward. Although the
corresponding mathematical program is not convex (thus computationally hard to compute) in general, we show that
it can be reduced to a convex one without loss of generality. In particular, in the resulting convex program where the
throughput along each source and destination pair in each time period are the variables, all the constraints are linear
and hence the exact optimal solutions can be efficiently computed (Theorem 3.1).
We further characterize the optimal solution via primal-dual analysis. In particular, a pricing scheme is optimal
if and only if the marginal contribution of the throughput along each edge equals to the system-wise marginal
contribution of additional supply minus the difference of the long term contributions of unit supply at the origin and
the destination (see Section 5).
We also perform extensive empirical analysis based on a public dataset with more than 8.5 million orders. We
compare our policy with other intensively studied policies such as surge pricing (Chen and Sheldon, 2015; Cachon
et al., 2016; Castillo et al., 2017). Our simulations show that, in both the static and the dynamic environment, our
optimal pricing and dispatching scheme outperforms surge pricing by 17% and 33%. Interestingly, our simulations
show that our optimal policy has much stronger ability in dispatching the vehicles than other policies, which results
directly in its performance boost (see Section 6).
1.2
Related work
Driven by real-life applications, a large number of researches have been done on ride-share markets. Some of them
employ queuing networks to model the markets (Iglesias et al., 2016; Banerjee et al., 2015; Tang et al., 2016). Iglesias
et al. (2016) describe the market as a closed, multi-class BCMP queuing network which captures the randomness of
customer arrivals. They assume that the number of customers is fixed, since customers only change their locations
but don’t leave the network. In contrast, the number of customer are dynamic in our model and we only consider
the one who asks for a ride (or sends a request to the platform). Banerjee et al. (2015) also use a queuing theoretic
approach to analyze the ride-share markets and mainly focus on the behaviors of drivers and customers. They assume
that the drivers enter or leave the market with certain possibilities. Bimpikis et al. (2016) take account for the spatial
dimension of pricing schemes in ride-share markets. They price for each region and their goal is to rebalance the
supply and demand of the whole market. However, we price for each routing and aim to maximize the total revenue
or social welfare of the platform. We also refer the readers to the line of researches initiated by (Ma et al., 2013) for
the problems about the car-pooling in the ride-sharing systems (Alonso-Mora et al., 2017; Zhao et al., 2014; Chan and
Shaheen, 2012).
Many works on ride-sharing consider both the customers and the drivers to be strategic, where the drivers
may reject the requests or leave the system if the prices are too low (Banerjee et al., 2015; Fang et al., 2017). As we
mentioned, if the revenue sharing ratios between the platform and the drivers can be dynamic, then the pricing
problem and the revenue sharing problem could be independent and hence the drivers are non-strategic in the pricing
problem. In addition, the platform can also increase the profit by adopting dynamic revenue sharing schemes (Balseiro
et al., 2017).
Another work closely related to ours is by Banerjee et al. (2017). Their work is concurrent and has been developed
independently from ours. In particular, the customers arrive according to a queuing model and their pricing policy is
state-independent and depends on the transition volume. Both their and our models are built upon the underlying
Markovian transitions between the states (the distribution of drivers over the graph). The major differences are: (i) our
model is built for the dynamic environments with a very large number of customers (each of them is non-atomic) to
meet the practical situations, while theirs adopts discrete agent settings; (ii) they overcome the non-convexity of the
problem by relaxation and focus only on concave objectives, which makes this work hard to use for real applications,
while we solve the problem via randomized pricing and transform the problem to a convex program; (iii) they prove
2
approximation bounds of the relaxation problem, while we give exact optimal solutions of the problem by efficiently
solving the convex program.
2
Model
A passenger (she) enters the ride-sharing platform and sends a request including her origin and destination to the
platform. The platform receives the request and determines a price for it. If user accepts the price, then the platform
may decide whether to send a driver (he) to pick her up. The platform is also able to dispatch drivers from one place to
another even there is no request to be served. By the pricing and dispatching methods above, the goal of maximizing
revenue or social welfare of the entire platform can be achieved. Our model incorporates the two methods into a
simple pricing problem. In this section, we define basic components of our model and consider two settings: dynamic
environments with a finite time horizon and static environments with an infinite time horizon. Finally we reduce the
action space of the problem and give a simple formulation.
Requests We use a strongly connected digraph G = (V, E) to model the geographical information of a city.
Passengers can only take rides from nodes to nodes on the graph. When a passenger enters the platform, she expects
to get a ride from node s to node t, and is willing to pay at most x ≥ 0 for the ride. She then sends to the platform a
request, which is associated with the tuple e = (s, t). Upon receiving the request, The platform sets a price p for it. If
the price is accepted by the passenger (i.e., x ≥ p), then the platform tries to send a driver to pick her up. We say that
the platform rejects the request, if no driver is available.
A request is said to be accepted if both the passenger accepts the price p and there are available drivers. Otherwise,
the request is considered to end immediately.
Drivers Clearly, within each time period, the total number of accepted requests starting from s cannot be more
than the number of drivers available at s. Formally, let q(e) denote the total number of accepted request along edge e,
then:
Õ
q(e) ≤ w(v), ∀v ∈ V,
(2.1)
e∈OUT(v)
where OUT(v) is the set of edges starting from v and w(v) is the number of currently available drivers at node v.
In particular, we assume that both the total number of drivers and the number of requests are very large, which is
often the case in practice, and consider each driver and each request to be non-atomic. For simplicity, we normalize
the total amount of drivers on the graph to be 1, thus w(v) is a real number in [0, 1]. We also normalize the number of
requests on each edge with the total number of drivers. Note that the amount of requests on an edge e can be more
than 1, if there are more requests on e than the total drivers on the graph.
Geographic Status For each accepted request on edge e, the platform will have to cover a transportation cost cτ (e)
for the driver. In the meanwhile, the assigned driver, who currently at node s, will not be available until he arrives
the destination t. Let ∆τ(e) be the traveling time from s to t and τ be the timestep of the driver leaving s. He will be
available again at timestep τ + ∆τ(e) on node t. Formally, the amount of available drivers on any v ∈ V is evolving
according to the following equations:
Õ
Õ
wτ+1 (v) = wτ (v) −
qτ (e) +
qτ+1−∆τ(e) (e),
(2.2)
e ∈OUT(v)
e∈IN(v)
where IN(v) is the set of edges ending at v. Here we add subscripts to emphasize the timestamp for each quantity. In
particular, throughout this paper, we focus on the discrete time step setting, i.e., τ ∈ N.
Demand Function As we mentioned, the platform could set different prices for the requests. Such prices may
vary with the request edge e, time step t, and the driver distribution but must be independent of the passenger’s
private value x as it is not observable. Formally, let Dτ (·|e) : R+ → R+ be the demand function of edge e, i.e., Dτ (p|e)
is the amount of requests on edge e with private value x ≥ p in time step τ.1 Then the amount of accepted requests
qτ (e) ≤ E[Dτ (pτ (e)|e)], where the expectation is taken over the potential randomness of the pricing rule pτ (e).2
1 In
practice, such a demand function can be predicted from historical data Tong et al. (2017); Moreira-Matias et al. (2013).
randomized pricing rule may set different prices for the requests on the same edge e.
2 The
3
Design Objectives In this paper, we consider a class of state-irrelevant objective functions. A function is stateirrelevant if its value only depends on the amount of accepted request on each edge q(e) but not the driver distribution
of the system w(v). Note that a wide range of objectives are included in our class of objectives, such as the revenue of
the platform:
Õ
REVENUE(p, q) =
E[(pτ (e) − cτ (e)) · qτ (e)],
e,τ
and the social welfare of the entire system:
WELFARE(p, q) =
Õ
E[(x − cτ (e)) · qτ (e)].
e,τ
In general, our techniques work for any state-irrelevant objectives. Let g(p, q) denote the general objective
function and the dispatching and pricing problem can be formulated as follows:
Õ
maximize
g(pτ (e), qτ (e)|e)
(2.3)
e,τ
subject to
(2.1) and (2.2).
Static and Dynamic Environment In general, our model is defined for a dynamic environment in the sense that
the demand function Dτ and the transportation cost cτ could be different for each time step τ. In particular, we study
the problem (2.3) in general dynamic environments with finite time horizon from τ = 1 to T, where the initial driver
distribution w1 (v) is given as input.
In addition, we also study the special case with static environment and infinite time horizon, where Dτ ≡ D and
cτ ≡ c are consistent across each time step.
2.1
Reducing the action space
In this section, we rewrite the problem to an equivalent reduced form by incorporating the action of dispatching
into pricing, i.e., using p to express q. The idea is straightforward: (i) for the requests rejected by the platform, the
platform could equivalently set an infinitely large price; (ii) if the platform is dispatching available drivers (without
requests) from node s to t, we can create virtual requests from s to t with 0 value and let the platform sets price 0 for
these virtual requests. In fact, we can assume without loss of generality that D(0|e) ≡ 1, the total amount of drivers,
because one can always add enough virtual requests for the edges with maximum demand less than 1 or remove the
requests with low values for the edges with maximum demand exceeds the total driver supply, 1.
As a result, we may conclude that q(e) ≤ D(p|e). Since our goal is to maximize the objective g(p, q), raising prices
to achieve the same amount of flow q(e) (such that E[D(p|e)] = q(e)) never eliminates the optimal solution. In other
words,
Observation 2.1. The original problem is equivalent to the following reduced problem, where the flow variables qτ (e)
are uniquely determined by the price variables pτ (e):
Õ
maximize
g(pτ (e), Dτ (pτ (e)|e))
e,τ
subject to
qτ (e) = E[D(pτ (e)|e)]
(2.4)
(2.1) and (2.2).
3
Problem Analysis
In this section, we demonstrate how the original problem (2.4) can be equivalently rewritten as a Markov decision
process with a convex objective function. Formally,
Theorem 3.1. The original problem (2.4) of the instance hG, D, g, ∆τi is equivalent to a Markov decision process problem
of another instance hG 0, D 0, g 0, ∆τ 0i with g 0 being convex.
The proof of Theorem 3.1 will be immediate after Lemma 3.2 and 3.4. The equivalent Markov decision process
problem could be formulated as a convex program, and hence can be solved efficiently.
4
3.1
Unifying travel time
Note that the original problem (2.4), in general, is not a MDP by itself, because the current state wτ+1 (v) may depend
on the action qτ+1−∆τ(e) in (2.2). Hence our first step is to equivalently map the original instance to another instance
with traveling time is always 1, i.e., ∆τ(e) ≡ 1:
Lemma 3.2 (Unifying travel time). The original problem (2.4) of an general instance hG, D, g, ∆τi is equivalent to the
problem of a 1-travel time instance hG 0, D 0, g 0, ∆τ 0i, where ∆τ 0(·) ≡ 1.
Intuitively, we tackle this problem by adding virtual nodes into the graph to replace the original edges. This
operation splits the entire trip into smaller ones, and at each time step, all drivers become available.
Proof. For edges with traveling time ∆τ(e) = 1, we are done.
e
For edges with traveling time ∆τ(e) > 1, we add ∆τ(e) − 1 virtual nodes into the graph, i.e., v1e, . . . , v∆τ(e)−1
, and
the directed edges connecting them to replace the original edge e, i.e.,
e
e
e
E 0(e) = {(s, v1e ), (v1e, v2e ), . . . , (v∆τ(e)−2
, v∆τ(e)−1
), (v∆τ(e)−1
, t)},
Ø
Ø
e
E0 =
E 0(e), V 0 =
{v1e, . . . , v∆τ(e)−1
} ∪ V.
e ∈E
e ∈E
We set the demand function of each new edge e 0 ∈ E 0(e) to be identical to those of the original edge e: D 0(·|e 0) ≡ D(·|e).
An important but natural constraint is that if a driver handles a request on edge e of the original graph, then he
must go along all edges in E 0(e) of the new graph, because he cannot leave the passenger halfway. To guarantee this,
we only need to guarantee that all edges in E 0(e) have the same price. Also, we need to split the objective of traveling
along e into the new edges, i.e., each new edge has objective function
g 0(p, q|e 0) = g(p, q|e)/∆τ(e), ∀e 0 ∈ E 0(e).
One can easily verify that the above operations increase the graph size to at most maxe ∈E ∆τ(e)∗ times of that of
the original one. In particular, there is a straightforward bijection between the dispatching behaviors of the original
G = (V, E) and the new graph G 0 = (V 0, E 0). Hence we can always recover the solution to the original problem.
3.2
Flow formulation and randomized pricing
By Lemma 3.2, the original problem (2.4) can be formulated as an MDP:
Definition 3.3 (Markov Decision Process). The vehicle pricing and dispatching problem is a Markov decision process,
denoted by a tuple (G, D, g, S, A, W), where G = (V, E) is the given graph, D is the demand function, objective g is the
reward function, S = ∆(V) is the state space including all possible driver distributions over the nodes, A is the action
space, and W is the state transition rule:
Õ
Õ
wτ+1 (v) = wτ (v) −
qτ (e) +
qτ (e).
(3.1)
e∈OUT(v)
e ∈IN(v)
However, by naïvely using the pricing functions pτ (e) as the actions, the induced flow qτ (e) = E[Dτ (pτ (e)|e)],
in general, is neither convex nor concave. In other words, both the reward g and the state transition W of the
corresponding MDP is non-convex. As a result, it is hard to solve the MDP efficiently.
In this section, we show that by formulating the MDP with the flows qτ (e) as actions, the corresponding MDP is
convex.
Lemma 3.4 (Flow-based MDP). In the MDP (G, D, g, S, A, W) with all possible flows as the action set A, i.e., A = [0, 1] |E | ,
the state transition rules are linear functions of the flows and the reward functions g are convex functions of the flows.
Proof. To do this, we first need to rewrite the prices pτ (e) as functions of the flows qτ (e). In general, since the prices
could be randomized, the inverse function of qτ (e) = E[Dτ (pτ (e)|e)] is not unique.
5
Note that conditional on fixed flows qτ (e), the state transition of the MDP is also fixed. In this case, different
prices yielding such specific flows only differs in the rewards. In other words, it is without loss of generality to let the
inverse function of prices be as follows:
pτ (e) = arg max gτ (pτ (e), qτ (e)|e), s.t. qτ (e) = E[Dτ (pτ (e)|e)].
p
In particular, since the objective function g we studied in this paper is linear and weakly increasing in the prices
p and the demand function D(p|e) is decreasing in p, the inversed price function could be defined as follows:
• Let gτ (q|e) = gτ (Dτ−1 (q|e), q|e), i.e., the objective obtained by setting the maximum fixed price p = Dτ−1 (q|e)
such that the induced flow is exactly q;
• Let ĝτ (q|e) be the ironed objective function, i.e., the smallest concave function that upper-bounds gτ (q|e) (see
Figure 1);
• For any given qτ (e), the maximum objective on edge e is ĝτ (qτ (e)|e) and could be achieve by setting the price
to be randomized over Dτ−1 (q 0 |e) and Dτ−1 (q 00 |e).
Figure 1: Ironed objective function
Finally, we prove the above claim to complete the proof of Lemma 3.4.
By the definition of ĝτ (q|e), for any randomized price p,
E[gτ (Dτ (p|e)|e)] ≤ E[ĝτ (Dτ (p|e)|e)].
p
p
Since ĝ is concave, applying Jensen’s inequality yields:
E[ĝτ (Dτ (p|e)|e)] ≤ ĝτ E[Dτ (p|e)] e = ĝτ (q̄|e)
p
p
Now it suffices to show that the upper bound ĝτ (q̄|e) is attainable.
If ĝτ (q̄|e) = gτ (q̄|e), then the right-hand-side could be achieved by letting pτ (e) be the deterministic price
Dτ−1 (q̄|e).
Otherwise, let I = (q 0, q 00) be the ironed interval (where ĝτ (q|e) > gτ (q|e), ∀q ∈ I but ĝτ (q 0 |e) = gτ (q 0 |e) and
ĝτ (q 00 |e) = gτ (q 00 |e)) containing q̄. Thus q̄ can be written as a convex combination of the end points q 0 and q 00:
q̄ = λq 0 + (1 − λ)q 00. Note that the function ĝτ is linear within the interval I. Therefore
λgτ (q 0 |e) + (1 − λ)gτ (q 00 |e) = λĝτ (q 0 |e) + (1 − λ)ĝτ (q 00 |e) = ĝτ (λq 0 + (1 − λ)q 00 |e) = ĝτ (q̄|e).
In other words, the upper bound ĝτ (q̄|e) could be achieved by setting the price to be q 0 with probability λ and q 00
with probability 1 − λ. In the meanwhile, the flow qτ (e) would retain the same.
Proof of Theorem 3.1. The theorem is implied by Lemma 3.2 and Lemma 3.4. In particular, the reward function is the
ironed objective function ĝ.
In the rest of the paper, we will focus on the following equivalent problem:
Õ
maximize
ĝτ (qτ (e)|e)
e,τ
subject to
(2.1) and (3.1).
6
(3.2)
4
Optimal Solution in Static Environment
In this setting, we restrict our attention to the case where the environment is static, hence the objective function does
not change over time, i.e., ∀τ ∈ [T], ĝτ (q|e) ≡ ĝ(q|e). We aim to find the optimal stationary policy that maximizes the
objective function, i.e., the decisions qτ depends only on the current state wτ .
In this section, we discretize the MDP problem and focus on stable policies. With the introduction of the ironed
objective function ĝτ , we show that for any discretization scheme, the optimal stationary policy of the induced
discretized MDP is dominated by a stable dispatching scheme. Then we formulate the stable dispatching scheme as a
convex problem, which means the optimal stationary policy can be found in polynomial time.
Definition 4.1. A stable dispatching scheme is a pair of state and policy (wτ , π), such that if policy π is applied, the
distribution of available drivers does not change over time, i.e., wτ+1 (v) = wτ (v).
In particular, under a stable dispatching scheme, the state transition rule (3.1) is equivalent to the following form:
Õ
Õ
q(e) =
q(e).
(4.1)
e ∈OUT(v)
e ∈IN(v)
Definition 4.2. Let M = (G, D, ĝ, S, A, W) be the original MDP problem. A discretized MDP DM with respect to M is
a tuple (G d, Dd, ĝd, Sd, Ad, Wd ), where G d = G, Dd = D, ĝd = ĝ, Wd = W, Sd is a finite subset of S, and Ad is a finite
subset of A that contains all feasible transition flows between every two states in Sd .
Theorem 4.3. Let DM and M be a discretized MDP and the corresponding original MDP. Let πd : Sd → Ad be an
optimal stationary policy of DM. Then there exists a stable dispatching scheme (w, π), such that the time-average
objective of π in M is no less than that of πd in DM.
Proof. Consider policy πd in DM. Starting from any state in Sd with policy πd , let {wτ }0∞ be the subsequent state
sequence. Since DM has finitely many states and policy πd is a stationary policy, there must be an integer n, such
that wn = wm for some m < n and from time step m on, the state sequence become a periodic sequence. Define
w̄ =
n−1
1 Õ
wk ,
n − m k=m
q̄ =
n−1
1 Õ
πd (wk )
n − m k=m
Denote by πd (wk |e) or qd (e) the flow at edge e of the decision πd (wk ). Sum the transition equations for all the time
steps m ≤ k < n, and we get:
n−1
Õ
wk+1 (v) −
k=m
n−1
Õ
k=m
wk (v) =
n−1 Õ
n−1
Õ
©
ª Õ© Õ
ª
πd (wk |e)® −
πd (wk |e)®
k=m «IN(v)
¬ k=m «OUT(v)
¬
© Õ
ª ©Õ
ª
w̄(v) = w̄(v) −
q̄(e)® +
q̄(e)®
«OUT(v)
¬ «IN(v)
¬
Also, policy πd is a valid policy, so ∀v ∈ V and ∀m ≤ k < n:
Õ
qk (e) ≤ wk (v)
OUT(v)
Summing over k, we have:
Õ
q̄(e) ≤ w̄(v)
OUT(v)
Now consider the original problem M. Let w = w̄ and π be any stationary policy such that:
• π(w) = q̄;
7
• starting from any state w 0 , w, policy π leads to state w within finitely many steps.
Note that the second condition can be easily satisfied since the graph G is strongly connected.
With the above definitions, we know that (w, π) is a stable dispatching scheme. Now we compare the objectives of
the two policies πd and π. The time-average objective function is not sensitive about the first finitely many immediate
objectives. And since the state sequences of both policies πd and π are periodic, Their time-average objectives can be
written as:
n−1
1 ÕÕ
ĝ(qd (e)|e)
n − m k=m e ∈E
Õ
OBJ(π) =
ĝ(q̄(e)|e)
OBJ(πd ) =
e ∈E
By Jensen’s inequality, we have:
n−1
Õ
1 ÕÕ
OBJ(πd ) =
ĝ(qd (e)|e) ≤
ĝ
n − m k=m e∈E
e∈E
"
! #
n
Õ
1 Õ
ĝ(q̄(e)|e) = OBJ(π)
qd (e) e =
n − m k=m
e ∈E
With Theorem 4.3, we know there exists a stable dispatching scheme that dominates the optimal stationary policy
of the our discretized MDP. Thus we now only focus on stable dispatching schemes. The problem of finding an
optimal stable dispatching scheme can be formulated as a convex program with linear constraints:
Õ
maximize
ĝ(q|e)
e∈E
(4.2)
subject to (2.1) and (4.1).
Because ĝ(q|e) is concave, the program is convex. Since all convex programs can be solved in polynomial time,
our algorithm for finding optimal stationary policy of maximizing the objective functions is efficient.
5
Characterization of optimality
In this section, we characterize the optimal solution via dual analysis. For the ease of presentation, we consider
Program 4.2 in the static environment with infinite horizon. Our characterization directly extends to the dynamic
environment.
The Lagrangian is defined to be
!
Õ
Õ
Õ © Õ
Õ
ª
L(q, λ, µ) = −
ĝ(q|e) + λ
q(e) − 1 +
µv
q(e) −
q(e)®
e ∈E
e∈E
v ∈V
IN(v)
«OUT(v)
¬
Õ
=−λ+
[−ĝ(q|e) + (λ + µs − µt )q(e)] ,
e ∈E
where s and t are the origin and destination of e, i.e., e = (s, t), and λ and µ are Lagrangian multipliers
Í with λ ≥ 0.
Note that we implicitly transform program 4.2 to the standard form that minimizes the objective − e∈E ĝ(q∗ |e).
The Lagrangian dual function is
Õ
h(λ, µ) = inf L(q, λ, µ) =
[−ĝ(q̃|e) + (λ + µs − µt )q̃(e)] ,
q
e ∈E
where q̃(e) is a function of λ and µ such that λ + µs − µt = ĝ 0(q̃|e), where ĝ 0(q̃|e) is the derivative of the objective
function with respect to flow q. The dual program corresponding to Program 4.2 is
maximize
h(λ, µ)
subject to
λ≥0
According to the KKT conditions, we have the following characterization for optimal solutions.
8
(5.1)
∗ ∗
Theorem 5.1. Let q∗ (e) be a feasible solution to the primal program 4.2 and (λ
to the dual
Í , µ ) be∗a feasible ∗solution
∗
∗
∗
program 5.1. Then both q (e) and (λ , µ ) are primal and dual optimal with − e∈E ĝ(q |e) = h(λ , µ∗ ), if and only if
!
Õ
λ∗
q∗ (e) − 1 = 0
(5.2)
e∈E
ĝ (q |e) = λ∗ + µ∗s − µ∗t , ∀v ∈ V
0
∗
(5.3)
Proof. According to the definition of h(λ, µ), we have h(λ∗, µ∗ ) = inf q L(q, λ∗, µ∗ ). Since ĝ(q|e) are concave functions,
Equation 5.3 is equivalent to the fact that q∗ (e) minimizes the function L(q, λ∗, µ∗ ).
h(λ∗, µ∗ ) = inf L(q, λ∗, µ∗ )
q
=L(q∗, λ∗, µ∗ )
!
=−
Õ
ĝ(q |e) + λ
∗
e ∈E
=−
Õ
∗
Õ
q (e) − 1 +
∗
Õ
v ∈V
e∈E
∗
Õ
© Õ ∗
ª
µ∗v
q (e) −
q∗ (e)®
IN(v)
«OUT(v)
¬
ĝ(q |e),
e∈E
where the last equation uses the Equation 5.2 and the fact that q∗ (e) is feasible.
Continuing with Theorem 5.1, we will analyze the dual variables from the economics angle and some interesting
insights into this problem for real applications.
5.1
Economic interpretation
The dual variables have useful economic interpretations (see (Boyd and Vandenberghe, 2004, Chapter 5.6)). λ∗ is the
system-wise marginal contribution of the drivers (i.e. the increase in the objective function when a small amount of
drivers are added to the system). Note that by the complementary slackness (Equation 5.2), if λ∗ > 0, the sum of the
total flow must be 1, meaning that all drivers are busy, and more requests can be accepted (hence increase revenue) if
more drivers are added to the system. Otherwise, there must be some idle drivers, and adding more drivers cannot
increase the revenue.
µ∗v is the marginal contribution of the drivers at node v. If we allow the outgoing flow from node v to be slightly
more than the incoming flow to node v, then µv is the revenue gain from adding more drivers at node v.
5.2
Insights for applications
The way we formulate and solve the problem, in fact, naturally leads to two interesting insights into this problem,
which are potentially useful for real applications.
1. Scalability In our model, the size of the convex program increases linearly in the number of edges, hence
quadratically in the number of regions. This could be one hidden feature that is potentially an obstacle to real
applications, where the number of regions in a city might be quite large.
A key observation to the issue is that any dispatching policy induced by a real system is a feasible solution of our
convex program and any improvement (maybe via gradient descent) from such policy in fact leads to a better solution
for this system. In other words, it might be hard to find the exact optimal or nearly optimal solutions, but it is easy to
improve from the current state. Therefore, in practice, the platform can keep running the optimization in background
and apply the most recent policy to gain more revenue (or achieve a higher value of some other objectives).
2. Alternative solution As suggested by the characterization and its economic interpretation, instead of solving the
convex programs directly, we also have an alternative way to find the optimal policy by solving the dual program.
The optimal policy can be easily recovered from dual optimal solutions. In particular, according to the economic
interpretation of dual variables, we need to estimate the marginal contributions of drivers.
More importantly, the number of dual variables (= the number of regions) is much smaller than the number of
primal variables (= the number of edges ≈ square of the former). So solving the dual program may be more efficient
when applied to real systems, and is also of independent interest of this paper.
9
6
Empirical Analysis
We design experiments to demonstrate the good performance of our algorithms for real applications. In this section,
we first describe the dataset and then introduce how to extract useful information for our model from the dataset.
Two benchmark policies, FIXED and SURGE, are compared with our pricing policy. The result analysis includes
demand-supply balance and instantaneous revenue in both static and dynamic environments.
6.1
Dataset
We perform our empirical analysis based on a public dataset from a major ride-sharing company. The dataset includes
the orders in a city for three consecutive weeks and the total number of orders is more than 8.5 million. An order is
created when a passenger send a ride request to the platform.
Each order consists of a unique order ID, a passenger ID, a driver ID, an origin, a destination, and an estimated
price, and the timestamp when the order is created (see Table 1 for example). The driver ID might be empty if no
driver was assigned to pick up the passenger. There are 66 major regions of the city and the origins and destinations
in the dataset are given as the region IDs. We say a request is related to a region if the region is either the origin or
the destination of the request. And the popularity of a region is defined as the number of related requests. Since some
of the regions in the dataset have very low popularity values, we only consider the most popular 21 or 5 regions in
the two settings (see Section 6.4 and Section 6.5 respectively for details). The related requests of the most popular 21
(or 5) regions cover about 90% (or 50%) of the total requests in the original dataset.
For ease of presentation, we relabel the region IDs in descending order of their popularities (so region #1 is the
most popular region). Figure 2 illustrates the frequencies of requests on different origin-destination pairs. From the
figure, one can see that the frequency matrix is almost symmetric and the destination of a request is most likely to be
in the same region as the origin.
order
hash
driver
hash
user
hash
origin
hash
dest
hash
price
37.5
timestamp
01-15 00:35:11
Table 1: An example of a row in the dataset, where “hash” stands for some hash strings of the IDs that we didn’t show
the exact value here.
Heatmap of Routes
Log Frequency
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
12.5
Origin IDs
10.0
7.5
5.0
2.5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
0.0
Destination IDs
Figure 2: The logarithmic frequencies of request routes.
6.2
Data preparation
The time consumptions from nodes to nodes and demand curves for edges are known in our model. However, the
dataset doesn’t provide such information directly. We filter out "abnormal" requests and apply a linear regression to
10
Log Frequency
8
6
30
0
0
8
6
4
2
0
15
2
15
30
4
Time & Price Filtered
Log Frequency
Time
in Mins
45 60 75 90 105 120 135 150
Time
in Mins
45 60 75 90 105 120 135 150
Time & Price
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80
Price in Local Currency
0
Price in Local Currency
(a) Time & price without filtering
(b) Time & price with filtering
Figure 3: The logarithmic frequencies of (time, price) pairs, with or without filtering the “abnormal” requests.
0
Density
Request value from region 6 to 2
Density
Request value from region 9 to 11
5
10
15
Value
20
25
0
5
10
15
20
Value
25
30
Figure 4: Fitting request values to lognormal distributions.
get the relationship of the travel time and the price. It makes possible to infer the travel time from the order price.
For the demand curves, we observe the values of each edge and fit them to lognormal distributions.
Distance and travel time The distance (or equivalently the travel time) from one region to another is required to
perform our simulation. We approximate the travel time by the time interval of two consecutive requests assigned to
the same driver. In Figure 3(a), we plot the frequencies of requests with certain (time, price) pairs. We cannot see
clear relationship between time and price, which are supposed to be roughly linearly related in this figure.3 We think
that this is due to the existence of two types of “abnormal” requests:
• Cancelled requests, usually with very short completion time but not necessarily low prices (appeared in the
right-bottom part of the plot);
• The last request of a working period, after which the driver might go home or have a rest. These requests usually
have very long completion time but not necessarily high enough prices (appeared in the left-top part of the
plot).
With the observations above, we filter out the requests with significantly longer or shorter travel time compared
with most of the requests with the same origin and destination. Figure 3(b) illustrates the frequencies of requests
after such filtering. As expected, the brightest region roughly surrounds the 30◦ line in the figure. By applying a
standard linear regression, the slope turns out to be approximately 0.5117 CNY per minute. One may also notice
some “right-shifting shadows” of the brightest region, which are caused by the surge-pricing policy with different
multipliers.
3 The price of a ride is the maximum of a two-dimension linear function of the traveled distance and spent time and a minimal price (which is 7
CNY as one can see the vertical bright line at price = 7 in Figure 3). Since the traveled distance is almost linearly related to the spent time, the
price, if larger than the minimal price, should also be almost linearly related to the traveling time. Readers may notice that from the figures, there
are many requests with price less than 7 (even as low as 0). This is because there are many coupons given to passengers to stimulate their demand
for riding and the prices given in the dataset are after applying the coupons.
11
Revenue per Minute
Revenue per Minute
1600
1400
1200
1000
800
600
400
200
0
0
20
40
60
80
# of Iterations
100
400
300
200
100
0
120
(a) Static environment
0
100
200
300
400
500
# of Iterations
600
700
800
(b) Dynamic environment
Figure 5: Convergence of revenue.
Estimation of demand curves To estimate the demand curves, we first gather all the requests along the same edge
(also within the same time period for dynamic environment, see Section 6.5) and take the prices associated with the
requests as the values of the passengers. Then, we fit the values of each edge (and each time period for dynamic
environment) to a lognormal distribution. The reason that we choose the lognormal distribution is two-fold: (i) the
data fits lognormal distributions quite well (see Figure 4 as examples); (ii) lognormal distributions are commonly used
in some related literatures Ostrovsky and Schwarz (2011); Lahaie and Pennock (2007); Roberts et al. (2016); Shen and
Tang (2017).
We set the cost of traveling to be zero, because we do not have enough information from the dataset to infer the
cost.
6.3
Benchmarks
We consider two benchmark policies:
• FIXED: fixed per-minute pricing, i.e., the price of a ride equals to the estimated traveling time from the origin
to the destination of this ride multiplied by a per-minute price α, where α is a constant across the platform.
• SURGE: based on FIXED policy, using surge pricing to clear the local market when supply is not enough.
In other words, the price of a ride equals to the estimated traveling time multiplied by αβ, where α is the
fixed per-minute price and β ≥ 1 is the surge multiplier. Note that β is dynamic and can be different for
requests initiated at different regions, while the requests initiated at the same regions will share the same surge
multipliers.
In the rest of this section, we evaluate and compare our dynamic pricing policy DYNAM with these two benchmarks
in both static and dynamic environments.
6.4
Static environment
We first present the empirical analysis for the static environment, which is simpler than the dynamic environment
that we will consider next, hence easier to begin with.
In the static environment, we use the average of the statistics of all 21 days as the inputs to our model. For
example, the demand function D(p|e) is estimated based on the frequencies and prices of the requests along edge
e averaged over time. Similarly, the total supply of drivers is estimated based on the total durations of completed
requests.
With the static environment, we can instantiate the convex program (4.2) and solve via standard gradient descent
algorithms. In our case, we simply use the MATLAB function fmincon to solve the convex program on a PC with
Intel i5-3470 CPU. We did not apply any additional techniques to speed-up the computation as the optimization
of running time is not the main focus of this paper. Figure 5(a) illustrates the convergence of the objective value
(revenue) with increasing number of iterations, where each iteration roughly takes 0.2 second.
12
Dynamic Environment
1550
Revenue per Minute
Revenue per Minute
Static Environment
DYNAM
FIXED
SURGE
1500
1450
1400
1350
1300
1250
1200
0
4
8
12
Hours
16
20
DYNAM
FIXED
1000
800
600
400
200
0
24
SURGE
0
4
(a) Static environment
8
12
16
Hours
20
24
(b) Dynamic environment
Figure 6: Instantaneous revenue in different environments.
300%
200%
100%
0%
500%
400%
300%
200%
100%
0
4
8
12
Hours
16
20
24
0%
500%
400%
300%
200%
100%
0
4
8
12
Hours
16
20
24
0%
600%
4
8
12
Hours
16
20
24
500%
400%
300%
200%
0%
DYNAM
FIXED
SURGE
700%
100%
0
Region #5
800%
DYNAM
FIXED
SURGE
700%
Supply/Demand
400%
600%
Region #4
800%
DYNAM
FIXED
SURGE
700%
Supply/Demand
500%
600%
Region #3
800%
DYNAM
FIXED
SURGE
700%
Supply/Demand
Supply/Demand
600%
Region #2
800%
DYNAM
FIXED
SURGE
700%
Supply/Demand
Region #1
800%
600%
500%
400%
300%
200%
100%
0
4
8
12
Hours
16
20
24
0%
0
4
8
12
Hours
16
20
24
Figure 7: Instantaneous supply ratios for different regions.
To compare the performance of policy DYNAM with the benchmark policies FIXED and SURGE, we also simulates
them under the same static environment. In particular, the length of each timestep is set to be 15 minutes and the
number of steps in simulation is 96 (so 24 hours in total). For both FIXED and SURGE, we use the per-minute price
fitted from data as the base price, α = 0.5117, and allow the surge ratio β to be in [1.0, 5.0]. To make the evaluations
comparable, we use the distribution of drivers under the stationary solution of our convex program as the initial
driver distributions for FIXED and SURGE. Figure 6(a) shows how the instantaneous revenues evolve as the time
goes by, where DYNAM on average outperforms FIXED and SURGE by roughly 24% and 17%, respectively.
Note that our policy DYNAM is stationary under the static environment, the instantaneous revenue is constant
(the red horizontal line). Interestingly, the instantaneous revenue curves of both FIXED and SURGE are decreasing
and the one of FIXED is decreasing much faster. The observation reflects that both FIXED and SURGE are not
doing well in dispatching the vehicles: FIXED simply never balances the supply and demand, while SURGE shows
better control in the balance of supply and demand because the policy seeks to balance the demand with local supply
when supply can not meet the demand. However, neither of them really balance the global supply and demand, so
the instantaneous revenue decrease as the supply and demand become more unbalanced.
In other words, the empirical analysis supports our insight about the importance of vehicle dispatching in
ride-sharing platforms.
6.5
Dynamic environment
In the dynamic environment, the parameters (i.e., the demand functions and the total number of requests) are estimated
based on the statistics of each hour but averaged over different days. For example, the demand functions Dh (p|e)
are defined for each edge e and each of the 24 hours, h ∈ {0, . . . , 23}. In particular, we only use the data from the
weekdays (14 days in total)4 among the most popular 5 regions for the estimation.
Again, we instantiate the convex program (3.2) for the dynamic environment and solve via the fmincon function
on the same PC that we used for the static case. Figure 5(b) shows the convergence of the objective value with
increasing number of iterations, where each iteration takes less than 1 minute.
4 The reason that we only use data from weekdays is that the dynamics of demands and supplies in weekdays do have similar patterns but quite
different from the patterns of weekends.
13
We setup FIXED and SURGE in exactly the same way as we did for the static environment, except that the initial
driver distribution is from the solution of the convex program for dynamic environment.
Figure 6(b) shows the instantaneous revenue along the simulation. In particular, the relationship DYNAM
SURGE FIXED holds almost surely. Moreover, the advantages of DYNAM over the other two policies are more
significant at the high-demand “peak times”. For example, at 8 a.m., DYNAM (∼800) outperforms SURGE (∼600) and
FIXED (∼500) by roughly 33% and 60%, respectively.
Demand-supply balance Balancing the demand and supply is not the goal of our dispatching policy. However, a
policy without such balancing abilities are unlikely to perform well. In Figure 7, we plot the supply ratios (defined as
the local instantaneous supply divided by the local instantaneous demand) for all the 5 regions during the 24 hours of
the simulation.
From the figures, we can easily check that comparing with the other two lines, the red line (the supply ratio of
DYNAM) tightly surrounds the “balance” line of 100%, which means that the number of available drivers at any time
and at each region is close to the number of requests sent from that region at that time. The lines of other two policies
sometimes could be very far from the “balance” line, that is, the drivers under policy FIXED and SURGE are not in
the location where many passengers need the service.
As a result, our policy DYNAM shows much stronger power in vehicle dispatching and balancing demand and
supply in dynamic ride-sharing systems. Such advanced techniques in dispatching can in turn help the platform to
gain higher revenue through serving more passengers.
References
Javier Alonso-Mora, Samitha Samaranayake, Alex Wallar, Emilio Frazzoli, and Daniela Rus. 2017. On-demand
high-capacity ride-sharing via dynamic trip-vehicle assignment. PNAS (2017), 201611675.
Santiago Balseiro, Max Lin, Vahab Mirrokni, Renato Paes Leme, and Song Zuo. 2017. Dynamic revenue sharing. In
NIPS 2017.
Siddhartha Banerjee, Daniel Freund, and Thodoris Lykouris. 2017. Pricing and Optimization in Shared Vehicle Systems:
An Approximation Framework. In EC 2017.
Siddhartha Banerjee, Carlos Riquelme, and Ramesh Johari. 2015. Pricing in Ride-share Platforms: A QueueingTheoretic Approach. (2015).
Kostas Bimpikis, Ozan Candogan, and Saban Daniela. 2016. Spatial Pricing in Ride-Sharing Networks. (2016).
Stephen Boyd and Lieven Vandenberghe. 2004. Convex optimization. Cambridge university press.
Gerard P Cachon, Kaitlin M Daniels, and Ruben Lobel. 2016. The role of surge pricing on a service platform with
self-scheduling capacity. (2016).
Juan Camilo Castillo, Dan Knoepfle, and Glen Weyl. 2017. Surge pricing solves the wild goose chase. In EC 2017. ACM,
241–242.
Nelson D Chan and Susan A Shaheen. 2012. Ridesharing in north america: Past, present, and future. Transport Reviews
32, 1 (2012), 93–112.
M Keith Chen and Michael Sheldon. 2015. Dynamic pricing in a labor market: Surge pricing and flexible work on the
Uber platform. Technical Report. Mimeo, UCLA.
Judd Cramer and Alan B Krueger. 2016. Disruptive change in the taxi business: The case of Uber. The American
Economic Review 106, 5 (2016), 177–182.
Vincent P Crawford and Juanjuan Meng. 2011. New york city cab drivers’ labor supply revisited: Reference-dependent
preferences with rationalexpectations targets for hours and income. AER 101, 5 (2011), 1912–1932.
Zhixuan Fang, Longbo Huang, and Adam Wierman. 2017. Prices and subsidies in the sharing economy. In Proceedings
of the 26th International Conference on World Wide Web. WWW 2017, 53–62.
14
Ian L Gale and Thomas J Holmes. 1993. Advance-purchase discounts and monopoly allocation of capacity. The
American Economic Review (1993), 135–146.
Michel Gendreau, Alain Hertz, and Gilbert Laporte. 1994. A tabu search heuristic for the vehicle routing problem.
Management science 40, 10 (1994), 1276–1290.
Gianpaolo Ghiani, Francesca Guerriero, Gilbert Laporte, and Roberto Musmanno. 2003. Real-time vehicle routing:
Solution concepts, algorithms and parallel computing strategies. European Journal of Operational Research 151, 1
(2003).
Ramon Iglesias, Federico Rossi, Rick Zhang, and Marco Pavone. 2016. A BCMP Network Approach to Modeling and
Controlling Autonomous Mobility-on-Demand Systems. arXiv preprint arXiv:1607.04357 (2016).
Peter F Kostiuk. 1990. Compensating differentials for shift work. Journal of political Economy 98, 5, Part 1 (1990),
1054–1075.
Sébastien Lahaie and David M Pennock. 2007. Revenue analysis of a family of ranking rules for keyword auctions. In
EC 2007. ACM, 50–56.
Gilbert Laporte. 1992. The vehicle routing problem: An overview of exact and approximate algorithms. European
journal of operational research 59, 3 (1992).
Shuo Ma, Yu Zheng, and Ouri Wolfson. 2013. T-share: A large-scale dynamic taxi ridesharing service. In ICDE. IEEE,
410–421.
R Preston McAfee and Vera Te Velde. 2006. Dynamic pricing in the airline industry. forthcoming in Handbook on
Economics and Information Systems, Ed: TJ Hendershott, Elsevier (2006).
Luis Moreira-Matias, Joao Gama, Michel Ferreira, Joao Mendes-Moreira, and Luis Damas. 2013. Predicting taxi–
passenger demand using streaming data. IEEE Transactions on Intelligent Transportation Systems 14, 3 (2013),
1393–1402.
Gerald S Oettinger. 1999. An empirical analysis of the daily labor supply of stadium venors. Journal of political
Economy 107, 2 (1999), 360–392.
Michael Ostrovsky and Michael Schwarz. 2011. Reserve prices in internet advertising auctions: A field experiment. In
EC 2011Practical. ACM, 59–60.
Ben Roberts, Dinan Gunawardena, Ian A Kash, and Peter Key. 2016. Ranking and tradeoffs in sponsored search
auctions. ACM Transactions on Economics and Computation 4, 3 (2016), 17.
Weiran Shen and Pingzhong Tang. 2017. Practical versus Optimal Mechanisms. In AAMAS. 78–86.
Joanna Stavins. 2001. Price discrimination in the airline market: The effect of market concentration. Review of
Economics and Statistics 83, 1 (2001), 200–202.
Christopher S Tang, Jiaru Bai, Kut C So, Xiqun Michael Chen, and Hai Wang. 2016. Coordinating supply and demand
on an on-demand platform: Price, wage, and payout ratio. (2016).
Yongxin Tong, Yuqiang Chen, Zimu Zhou, Lei Chen, Jie Wang, Qiang Yang, Jieping Ye, and Weifeng Lv. 2017. The
simpler the better: a unified approach to predicting original taxi demands based on large-scale online platforms. In
KDD 2017. ACM, 1653–1662.
Dengji Zhao, Dongmo Zhang, Enrico H Gerding, Yuko Sakurai, and Makoto Yokoo. 2014. Incentives in ridesharing
with deficit control. In AAMAS 2014.
15
| 3cs.SY
|
Mean Square Capacity of Power Constrained Fading Channels with
Causal Encoders and Decoders*
arXiv:1509.04784v1 [math.OC] 16 Sep 2015
Liang Xu1 , Lihua Xie1 and Nan Xiao2
Abstract— This paper is concerned with the mean square
stabilization problem of discrete-time LTI systems over a power
constrained fading channel. Different from existing research
works, the channel considered in this paper suffers from both
fading and additive noises. We allow any form of causal channel
encoders/decoders, unlike linear encoders/decoders commonly
studied in the literature. Sufficient conditions and necessary
conditions for the mean square stabilizability are given in
terms of channel parameters such as transmission power and
fading and additive noise statistics in relation to the unstable
eigenvalues of the open-loop system matrix. The corresponding
mean square capacity of the power constrained fading channel
under causal encoders/decoders is given. It is proved that
this mean square capacity is smaller than the corresponding
Shannon channel capacity. In the end, numerical examples are
presented, which demonstrate that the causal encoders/decoders
render less restrictive stabilizability conditions than those under
linear encoders/decoders studied in the existing works.
I. I NTRODUCTION
Control over communication networks has been a hot
research topic in the past decade [1]. This is mainly motivated by the rapid development of wireless communication
technology that enables the connection of geographically
distributed systems and devices. However, the insertion of
wireless communication networks also poses challenges in
analysis and design of control systems due to constraints
and uncertainties in communications. One must take the
communication networks into consideration and analyze how
they affect the stability and performance of the closed-loop
control systems.
Until now, there have been plentiful results that reveal
requirements on communication channels to ensure the stabilizability. For noiseless digital channels, the celebrated data
rate theorem is given in [2]. For noisy channels, the problem
is complicated by the fact that different channel capacities
are required under different stability definitions. For almost
sure stability, [3] shows that the Shannon capacity in relation
to unstable dynamics of a system constitutes the critical
condition for its stabilizability. While for moment stability,
[4] shows that the Shannon capacity is too optimistic while
the zero-error capacity is too pessimistic, and the anytime
*This work was supported by the National Research Foundation of Singapore under Grant NRF2011NRF-CRP001-090 and the National Natural
Science Foundation of China under Grant 61304044.
1 Liang Xu and Lihua Xie are with EXQUISITUS, Centre for E-City,
School of Electrical and Electronic Engineering, Nanyang Technological
University, Singapore 639798, Singapore [email protected],
[email protected]
2 Nan
Xiao is with
search and Technology
the Singapore-MIT
Centre, Singapore
[email protected]
Alliance for Re138602, Singapore
capacity introduced in this paper characterizes the stabilizability conditions. Essentially, to keep the η-moment of the
state of an unstable scalar plant bounded, it is necessary
and sufficient for the feedback channel’s anytime capacity
corresponding to anytime-reliability α = ηlog2 |λ| to be
greater than log2 |λ|, where λ is the unstable eigenvalue of the
plant. The anytime capacity has a more stringent reliability
requirement than the Shannon capacity. However, it is worthy
noting that there exist no systematic method to calculate the
anytime capacities of channels.
In control community, the anytime capacity is usually
studied under the mean square stability requirement, for
which the anytime capacity is commonly named as the
mean square capacity. For example, [5] characterizes the
mean square capacity of a fading channel. [6] studies the
mean square stabilization problem over a power constrained
AWGN channel and characterizes the critical capacity to
ensure mean square stabilizability. They further show that
the extension from linear encoders/decoders to more general
causal encoders/decoders cannot provide additional benefits
of increasing the channel capacity [7]. Specifically, the
results stated above deal with fading channels or AWGN
channels separately. While in wireless communications, it
is practical to consider them as a whole. In this paper,
we are interested in a power constrained fading channel
which is corrupted by both fading and AWGN. We aim
to find the critical condition on the channel to ensure the
mean square stabilizability of the system. Note that [8] has
derived the necessary and sufficient condition for such kind
of channel to ensure mean square stabilizability under a
linear encoder/decoder. It is still unknown whether we can
achieve a higher channel capacity with more general causal
strategies. This paper provides a positive answer to this
question.
This paper is organized as follows. Problem formulation
and some preliminaries are given in Section 2. Section 3
provides the results for scalar systems. Section 4 discusses
the extension to vector systems. Section 5 provides numerical illustrations and this paper ends with some concluding
remarks in Section 6.
II. P ROBLEM F ORMULATION AND P RELIMINARIES
This paper studies the following single-input discrete-time
linear system
xt+1 = Axt + But
(1)
where x ∈ Rn is the system state and u ∈ R is the control
input. Without loss of generality, we assume that all the
eigenvalues of A are unstable, i.e., |λi (A)| ≥ 1 for all i =
1, 2, . . . , n [7]. The initial value x0 is randomly generated
from a Gaussian distribution with zero mean and bounded
covariance Σx0 . The system state xt is observed by a sensor
and then encoded and transmitted to the controller through
a power constrained fading channel. The communication
channel is modeled as
rt = gt st + nt
(2)
in which st denotes the channel input; rt represents the
channel output; {gt } is an i.i.d. stochastic process modeling
the fading effects and {nt } is the additive white Gaussian
noise with zero-mean and known variance σn2 . The channel
input st must satisfy an average power constraint, i.e.,
E{s2t } ≤ P . We also assume that x0 , g0 , n0 , g1 , n1 , . . . are
independent. In the paper, it is assumed that after each
transmission, the instantaneous value of the fading factor
gt is known to the decoder, which is a reasonable assumption for slowly varying channels with channel estimation
[9]. The instantaneous Shannon channel capacity is ct =
gt2 P
1
with ct being measured in nats/transmission.
2
2 ln 1 + σn
The feedback configuration among the plant, the sensor and
the controller, and the channel encoder/decoder structure are
depicted in Fig. 1. In this paper, we try to find requirements
Plant
Controller/Decoder
Fig. 1.
Sensor/Encoder
Network control structure over power constraint fading channel
on the power constrained fading channel such that there
exists a pair of causal encoder/decoder {ft }, {ht } that can
mean square stabilize the LTI dynamics (1), i.e., to render
limt→∞ E{xt x0t } = 0.
To solve this problem, the following preliminaries are
needed, which are borrowed from [7]. Throughout the paper,
a sequence {χi }ti=0 is denoted by χt ; random variables are
denoted by uppercase letters, and their realizations by lower
case letters. All random variables are assumed to exist on a
common probability space with measure P. The probability
density of a random variable X in Euclidean space with
respect to Lebesgue measure on the space is denoted by pX ,
and the probability density of X conditioned on the σ-field
generated by the event Y = y by pX|y . Let the expectation
operator be denoted by E, and the expectation conditioned on
the event Y = y by Ey . We use log to denote the logarithm
to the base two, and ln to denote the natural logarithm.
The differential entropy of X is defined by H(X) =
−E{lnpX }, provided that the defining integral exists. Denote
the conditional entropy of X given the event Y = y by
Hy (X) = H(X|Y = y) = −Ey {lnpX|y }, and the random
variable associated with Hy (X) by HY (X). The average
conditional entropy of X given the event Y = y and averaged
over Y is defined by H(X|Y ) = E{HY (X)}, and the
average conditional entropy of X given the events Y = y
and Z = z and averaged only over Y by Hz (X|Y ) =
Ez {HY,Z (X)}. The conditional mutual information between
two random variables X and Y given the event Z = z
is defined by Iz (X; Y ) = Hz (X) − Hz (X|Y ). Given a
random variable X ∈ Rn with entropy H(X), the entropy
2
1
e n H(X) . Denote the
power of X is defined by N (X) = 2πe
conditional entropy power of X given the event Y = y by
2
1
e n Hy (X) , and the random variable associated
Ny (X) = 2πe
with Ny (X) by NY (X). The average conditional entropy
power of X given the event Y = y and averaged over
Y is defined by N (X|Y ) = E{NY (X)}, and the average
conditional entropy power of X given the events Y = y
and Z = z and averaged only over Y by Nz (X|Y ) =
Ez {NY,Z (X)}. The following lemma shows that the entropy
power of a random variable provides an estimation of the
lower bound for its variance.
Lemma 1 ([7]): Let X be an n-dimensional random variable. Then Ny (X) ≤ n1 Ey {kXk2 }.
Lemma 2: Let X be an n-dimensional random variable,
f (X) be a function of X, and Y = f (X) + N with N
being a random variable that is independent with X. Then
I(X; Y ) = I(f (X); Y ).
Proof:
Since H(Y |X) = H(Y |X, f (X)) ≤
H(Y |f (X)), we have H(Y ) = I(X; Y ) + H(Y |X) ≤
I(X; Y ) + H(Y |f (X)). Thus H(Y ) − H(Y |f (X)) =
I(Y ; f (X)) ≤ I(X; Y ). Besides, noting that X → f (X) →
Y forms a Markov chain, the data processing inequality [10]
implies that I(X; Y ) ≤ I(f (X); Y ). Combining the two
facts, we have I(X; Y ) = I(f (X); Y ).
Remark 1: Lemma 2 indicates that for the AWGN channel, the amount of information that the channel output contains about the source is equal to the amount of information
that the channel output contains about the channel input.
III. S CALAR S YSTEMS
To better convey our ideas, we start with scalar systems.
Consider the following scalar system
xt+1 = λxt + ut
E{x20 }
σx20 .
(3)
where |λ| ≥ 1 and
=
With the communication
channel given in (2), the stabilizability result is stated in the
following theorem.
Theorem 1: There exists a causal encoder/decoder pair
{ft }, {ht }, such that the system (3) can be stabilized over
the communication channel (2) in mean square sense if and
only if
1
σ2
log|λ| < − logE{ 2 n 2 }
(4)
2
σn + gt P
Theorem 1 indicates that the mean square capacity
of the power constraint fading channel is CMSC =
σ2
− 12 logE{ σ2 +gn 2 P }. In the following, we will prove the
n
t
necessity and sufficiency of Theorem 1, respectively. The
proof essentially follows the same steps as in [11], [7], [12],
however, with some differences due to the channel structure.
A. Proof of Necessity
The proof of necessity follows from the intuition below. In
view of Lemma 1, the entropy power provides a lower bound
for the mean square value of the system state. We thus can
use the average entropy power as a measure of the uncertain
region of the system state and analyze its update. At time t,
the controller maintains a knowledge of the uncertain region
of xt . When it takes action on the plant, the average uncertain
region of xt+1 predicated by the controller is expanded to λ2
times that of xt . This is the iteration we term as dynamics
update, which describes the update of the uncertain region
of x maintained by the controller from time t to t + 1. After
receiving information about xt+1 from the sensor through
the communication channel, the controller can reduce the
predication error of the uncertain region of xt+1 by a factor
σ2
of E{ σ2 +gn 2 P }. This is the iteration we term as communin
t
cation update, which describes the update of the uncertain
region of x maintained by the controller at time t + 1 after
it has received the information about xt+1 from the sensor
through the communication channel. Thus to ensure mean
σ2
square stability, the average expanding factor λ2 E{ σ2 +gn 2 P }
n
t
of the system state’s uncertain region should be smaller than
one, which gives the necessary requirement in Theorem 1.
The formal proof is stated as follows. Here we use the
uppercase letters X, S, R, G to denote the random variables
of the system state, the channel input, the channel output and
the channel fading coefficient. We use the lowercase letters
x, s, r, g to denote their realizations.
1) Communication Update: The average entropy power of
Xt conditioned on (Rt , Gt ) is N (Xt |Rt ,Gt )=E{NRt ,Gt (Xt )}(a)
=
(b)
E{E{NRt ,Gt (Xt )|Rt−1 ,Gt }} =
≥e
2E{HRt ,Gt (Xt )|Rt−1 =r t−1 ,Gt =g t }
(d)
= e2H(Xt |Rt ,R
t−1
(e)
= e2(H(Xt |R
t−1
(f )
t−1
(g)
t−1
= e2(H(Xt |R
≥ e2(H(Xt |R
(h)
=r t−1 ,Gt =g t )
=r t−1 ,Gt =g t )−I(Xt ,Rt |Rt−1 =r t−1 ,Gt =g t ))
=r t−1 ,Gt =g t )−I(St ,Rt |Rt−1 =r t−1 ,Gt =g t ))
=r t−1 ,Gt =g t )−ct )
t−1
t−1
t−1
t−1
= e−2ct e2H(Xt |R =r ,G =g )
where (c) follows from Jensen’s inequality; (d) follows
from the definition of conditional entropy; (e) follows
from the definition of conditional mutual information; (f )
follows from Lemma 2; (g) follows from the definition of
channel capacity, i.e., I(St , Rt |Rt−1 = rt−1 , Gt = g t ) ≤ ct
and (h) follows from the fact that Gt is independent with
2H
(X )
1
E{e−2Ct e Rt−1 ,Gt−1 t }=
Xt , we have N (Xt |Rt ,Gt )≥ 2πe
2
σn
t−1
t−1
E{ 2 2 }N (Xt |R
,G
).
σ +g P
n
t
t
t
t
B. Proof of Sufficiency
To prove the sufficiency, we need to construct a pair of
encoder and decoder. The encoder and decoder are designed
following an ”estimation then control” strategy. The controller consecutively estimates the initial state x0 by using the
received information from the channel and then applies an
equivalent control to the plant. The reason for adopting such
strategy is explained as follows. The response
of the linear
Pt−1
system is xt = λt (x0 − x̂t ) with x̂t = − i=0 λ−1−i ui ,
which means E{x2t } = λ2t E{(x0 − x̂t )2 }. We can treat x̂t
as an estimate of the controller for the initial state x0 . If the
estimation error E{(x0 − x̂t )2 } converges to zero at a speed
that is greater than λ2 , i.e., there exists η > λ2 and α > 0,
such that E{(x0 − x̂t )2 } ≤ ηαt , the mean square value of the
2 t
system state would be bounded by E{x2t } ≤ α λη . Thus
lim E{x2t } = 0, i.e., system (3) is mean square stable. This
t→∞
intuition can be formalized using the following lemma.
Lemma 3 ([12]): If there exists an estimation scheme x̂t
for the initial system state x0 , such that the estimation error
et = x̂t − x0 satisfies the following property,
2H t
(Xt )
1
R ,Gt
|Rt−1 ,Gt }}
2πe E{E{e
where (a) follows from the law of total expectation and
(b) follows from the definition of entropy power. Since
E{e2HRt ,Gt (Xt ) |Rt−1 = rt−1 , Gt = g t }
(c)
t
2) Dynamics Update: Since e2H(Xt+1 |R =r ,G =g ) =
t
t
t
t
t
t
t
t
(i)
(j)
e2H(λXt +Ut |R =r ,G =g ) = e2H(λXt |R =r ,G =g ) =
t
t
t
t
t
t
t
t
e2H(Xt |R =r ,G =g )+2 ln |λ| = λ2 e2H(Xt |R =r ,G =g ) where
(i) follows from the fact that ut = ht (rt , g t ) and (j)
follows
from Theorem
8.6.4 in [10], we have N (Xt+1 |Rt ,Gt )≥
n
o
2 2HRt ,Gt (Xt )
1
E 2πe λ e
=λ2 N (Xt |Rt ,Gt ).
3) Proof of Necessity: Combining the results of
communication update and dynamics update, we have
σ2
N (Xt+1 |Rt ,Gt )≥λ2 E{ 2 n 2 }N (Xt |Rt−1 ,Gt−1 ). In view of
σn +gt P
Lemma 1, N (Xt+1 |Rt , Gt ) should converge to zero
σ2
asymptotically. Thus λ2 E{ σ2 +gn 2 P } < 1, which is (4) and
n
t
this proves the necessity.
lim A
t→∞
t
E{et }
0
E{et et }(A0 )t
=0
(5)
=0
(6)
then the system (1) can
square stabilized
be mean
by the
Pt
t
t−i
controller ut = K A x̂t + i=1 A Bui−1 with K
being selected such that A + BK is stable.
When gt is known at the receiver, channel (2) resembles
an AWGN channel. Shannon shows that when estimating a
Gaussian random variable through an AWGN channel, the
minimal mean square estimation error can be attained by
using linear encoders and decoders, respectively [13]. And
P σ2
the minimal mean square error variance is given by σ2 +gn2 P .
n
t
Thus through one channel use, we can at best decrease the
2
σ
estimation error by a factor of σ2 +gn 2 P . Since gt is i.i.d.,
n
t
we can transmit the estimation error from the decoder to
the encoder and iteratively conduct the minimal mean square
estimation process. Then the estimation error would decrease
σ2
σ2
on average at a speed of E{ σ2 +gn 2 P }. If λ2 E{ σ2 +gn 2 P } <
n
n
t
t
1, in view of Lemma 3, system (3) can be mean square
stabilized. The estimation strategy actually follows the principle of the well-known scheme of Schalkwijk [14], which
utilizes the noiseless feedback link to consecutively refine
the estimation error. The detailed encoder/decoder design and
stability analysis are given as follows.
1) Encoder/Decoder Design: Suppose the estimation of
x0 formed by the decoder is x̂t at time t and the estimation
error is et = x̂t − x0 . The encoder is designed as
s
P
x0
s0 =
σx20
s
(7)
P
st =
(x̂t−1 − x0 ) , t ≥ 1
σe2t−1
The decoder is designed as
r
σx20
x̂0 =
r0
P
E{rt et−1 |gt }
rt , t ≥ 1
x̂t = x̂t−1 −
E{rt2 |gt }
(8)
with σe2t−1 representing the variance of et−1 .
2) Proof of Sufficiency: Since r0 = g0 s0 + n0q
, in view
σ2
x0
of (7) and (8), we have e0 = (g0 − 1)x0 +
P n0 .
Because g0 , x0 , n0 are independent and x0 , n0 follows a zero
mean Gaussian distribution, we know that the conditional
probability distribution of e0 given the event g0 is Gaussian
σ2 σ2
and E{e0 |g0 } = 0, E{e20 |g0 } = (g0 − 1)2 σx20 + xP0 n . Thus
E{e0 } = E{E{e0 |g0 }} = 0 and E{e20 } = E{E{e20 |g0 }} =
σ2
σ2 σ2
IV. V ECTOR S YSTEMS
E{(g0 − 1)2 }σx20 + xP0 n .
For t ≥ 1, in view of (7) and (8), we have
For vector systems, the situation becomes complicated by
the fact that we have n sources xi,0 and only one channel,
where xi,0 denotes the i-th element of x0 . Firstly, we would
analyze the achievable minimal mean square estimation error
for estimating x0 over the channel (2) during one channel
use. Consider the following Markov chain
E{rt et−1 |gt }
et = et−1 −
rt
E{rt2 |gt }
s
E{rt et−1 |gt }
P E{rt et−1 |gt }
et−1 −
nt
= 1 − gt
σe2t−1 E{rt2 |gt }
E{rt2 |gt }
Thus the conditional probability distribution for et given the
event gt is Gaussian. We also have
E{et } = E{E{et |gt }}
s
n
o
P E{rt et−1 |gt }
E{e
|g
}
= E 1 − gt
t−1
t
σe2t−1 E{rt2 |gt }
s
n
P E{rt et−1 |gt } o
(a)
= E 1 − gt
E{et−1 }
2
σet−1 E{rt2 |gt }
where (a) follows from the fact that gt is independent with
et−1 . Since E{e0 } = 0, we further know that E{et } ≡ 0.
The sufficient condition (5) is satisfied.
Since et−1 , gt and nt are independent, we have
E{e2t−1 |gt } = E{e2t−1 } q
and E{n2t |gt } = E{n2t }, which im
2
plies E{rt2 |gt } = E gt σ2P et−1 +nt |gt = σn2 +gt2 P
et−1 q
and E{rt et−1 |gt } = E et−1 gt σ2P et−1 + nt |gt =
et−1
q
2
t et−1 |gt }
gt P σe2t−1 . Since E{e2t |gt } = E{e2t−1 |gt } − E{rE{r
,
2 |g }
t
t
we also have
σ2
E{e2t−1 } σ2 +gn 2 P
n
t
E{e2t |gt }
=
E{e2t−1 }
, which implies
E{e2t }
σ2
E{e2t−1 }E{ σ2 +gn 2 P }. Thus if λ2 E{ σ2 +gn 2 P } < 1, the den
n
t
t
signed encoder/decoder pair can guarantee (6). In view of
Lemma 3, the sufficiency of Theorem 1 is proved.
Remark 2: We can show that CMSC is smaller than the
Shannon capacity, which is CShannon = E{ct } [9]. From
Jensen’s inequality, we know that E{2−2ct } ≥ 2−2E{ct }
and the equality holds if and only if ct is a constant. Thus
1
1
≤ 12 log 2−2E{c
it follows that CMSC = 12 log E{2−2c
=
t}
t}
E{ct } = CShannon and the equality holds if and only if ct
is a constant.
Remark 3: By letting gt in (4) be the Bernoulli distribution with failure probability , and taking the limit σn2 → 0
and P → ∞, we can show that the necessary and sufficient
condition to ensure mean square stabilizability for the real
erasure channel is < λ12 , which recovers the result in [5].
If we let gt be a constant with gt = 1, then the studied
power constrained fading channel degenerates to the AWGN
channel and the (4) degenerates to 21 log(1 + σP2 ) < log|λ|,
n
which recovers the result in [4], [6]. If σn2 = 0 and the event
gt = 0 has zero probability measure, the right hand side of
(4) becomes infinity. Then for any λ, (4) holds automatically.
This is reasonable since we have assumed that gt is known
at the decoder side, thus if there is no additive noise, the
channel resembles a perfect communication link. Since (3)
is controllable, we can always find a pair of encoder and
decoder to stabilize the system.
g 2 P E{e2 }
− tσ2 +g2t−1
n
tP
= E{E{e2t |gt }}
=
=
X0 → St = ft (X0 ) → Rt → X̂t = ht (Rt )
where X0 ∈ Rn denotes the Gaussian initial state with
covariance matrix Σx0 ; ft (·) is a scalar-valued function
denoting the channel encoder for (2); Rt denotes the channel
output and X̂t is the estimation of X0 formed by the decoder
with decoding rule ht (·).
Denote the estimation error as et = X0 − X̂t , in view of
2
1
e n H(et |Rt ) . Since
Lemma 1, we have n1 trE{et e0t } ≥ 2πe
H(et |Rt ) = H(X0 − ht (Rt )|Rt ) = H(X0 |Rt )
= H(X0 ) − I(X0 ; Rt )
(a)
= H(X0 ) − I(ft (X0 ); Rt )
1
1
g2 P
≥ ln((2πe)n det(Σx0 )) − ln(1 + t 2 )
2
2
σn
where (a) follows from Lemma 2, thus we have
trE{et e0t } ≥ n det(Σx0 )
n1
σn2
gt2 P + σn2
From the above inequality, we know that the minimal mean
2
σn
square error is given in terms of g2 P +σ
2 . However, this
t
n
is only for the sum of the estimation errors ei,t with ei,t
being the i-th element of et . There is no indication on the
convergence speed for every single ei,t . Lemma 3 implies
that we should design the encoder/decoder to render that
2
limt→∞ λ2t
i E{ei,t } = 0 for all i, which places separate
requirements for the convergence speed of each ei,t . Thus we
need to optimally allocate channel resources to each unstable
state variable.
The previous analysis also implies that we should treat the
unstable modes of A separately. Here we focus on the real
Jordan canonical form of system (1). Let λ1 , . . . , λd be the
distinct unstable eigenvalues (if λi is complex, we exclude
from this list the complex conjugates λ∗i ) of A in (1), and
let mi be the algebraic multiplicity of each λi . The real
Jordan canonical form J of A then has the block diagonal
structure J = diag(J1 , . . . , Jd ) ∈ Rn×n , where the block
Ji ∈ Rµi ×µi and detJi = λµi i , with
mi if λi ∈ R
µi =
2mi otherwise
It is clear that we can equivalently study the following
dynamical system instead of (1)
xk+1 = Jxk + T Bui
(9)
for some similarity matrix T . Let U = {1, . . . , d} denote the
index set of unstable eigenvalues.
Theorem 2: There exists a causal encoder/decoder pair
{ft }, {ht }, such that the LTI dynamics (1) can be stabilized
over the communication channel (2) in mean square sense if
d
X
σ2
1
µi log|λi | < − logE{ 2 n 2 }
2
σn + gt P
i=1
(10)
and only if (log|λ1 |, . . . , log|λd |) ∈ Rd satisfy that for all
vi ∈ {0, . . . , mi } and i ∈ U
X
v1
v
σn2
(11)
ai vi log|λi | < − logE
2
σn2 + gt2 P
i∈U
P
where v = i∈U ai vi , and ai = 1 if λi ∈ R, and ai = 2
otherwise.
Proof: For the proof of necessity, notice that each
block Ji has an invariant real subspace Avi of dimension
ai vi , for any vi ∈ {0, . . . , mi }. Consider the subspace A
formed by taking the product of the invariant subspaces Avi
for each
total dimension of A is
P real Jordan block. The
V
v =
a
v
.
Denote
by
x
of
the components of x
i∈U i i
belonging to A. Then xV evolves as
V V
xV
k+1 = J xk+1 + QT uk
(12)
Πi∈U λai i vi .
where Q is a transformation matrix and detJ V =
Since Xk is mean square stable, it is necessary that the
subdynamics (12) is mean square stable. Similar to the
necessity proof in Theorem 1, we may derive the necessary
condition (11). And this completes the proof of necessity.
Here we prove the sufficiency using the idea of Time
Division Multiple Access (TDMA). Based on the previous
encoder/decoder design for scalar systems, the following
information transmission strategy is designed for the vector
system. Without loss of generality, here we assume that
λ1 , . . . , λd are real and mi = 1. For other cases, readers can
refer to the analysis discussed in Chapter 2 of [1]. Specifically, under this assumption, J is a diagonal matrix and
d = n. The sensor transmits periodically with a period of τ .
During one channel use, the sensor only transmits the estimation error of the j-th value of x0 using the scheme devised for
scalar systems. The relative transmission frequency for the jth value of x0Pis scheduled to be αj among the τ transmission
n
period with j=1 αj = 1. The receiver maintains an array
that represents the most recent estimation of x0 , which is set
to 0 for t = 0. When the information about the j-th value
of x0 is transmitted, only the estimation of the j-th value of
x0 is updated at the decoder side, and the other estimation
values remain unchanged. After updating the estimation, the
controller takes action as the one designed in Lemma 3. If
the diagonal elements of At E{et e0t }(A0 )t converge to zeros
2
asymptotically, i.e., for i = 1, . . . , n, limt→∞ λ2t
i E{ei,t } = 0
, the conditions in Lemma 3 can be satisfied. Since the transmission is scheduled periodically, we only need to require
that limk→∞ λ2kτ
E{e2i,kτ } = 0, ∀i = 1, . . . , n. Following
i
our designed transmission scheme, we have E{e2i,kτ } =
σ2
σ2
E{ σ2 +gn 2 P }αi kτ E{e2i,0 }. If λ2i E{ σ2 +gn 2 P }αi < 1 for all
n
n
t
t
i = 1, . . . n, the sufficient condition in Lemma 3 can be
satisfied. To complete the proof, we only need to show the
σ2
equivalence between the requirement λ2i E{ σ2 +gn 2 P }αi < 1
n
t
P
n
for all i = 1, . . . n and (10). On one hand, since i=1 αi =
2
σ
1, if λ2i E{ σ2 +gn 2 P }αi < 1 for all i = 1, . . . n, we know that
n
t
(10) holds. On the other hand, if (10) holds, we can simply
log|λ
i|
choose αi = P log|λ
, which satisfies the requirement that
i|
i
2
Pn
σn
2
αi
< 1 for all i = 1, . . . , n.
2 +g 2 P }
i=1 αi = 1 and λi E{ σn
t
The sufficiency is proved.
V. N UMERICAL I LLUSTRATIONS
A. Scalar Systems
The authors in [8] derive the mean square capacity
of a power constrained fading channel with linear encoders/decoders. The necessary and sufficient condition for
µ2 P
scalar systems is 12 log(1 + σ2 Pg+σ2 ) > log|λ| with µg and
g
n
σg2 being the mean and variance of gt . We can similarly
define the mean square capacity of the power constrained
fading channel with linear encoders/decoders as CMSL =
µ2g P
1
2 ). Simply assume that the fading follows
2 log(1 + σg2 P +σn
the Bernoulli distribution with failure probability , then
the Shannon capacity, the mean square capacity achievable
with causal encoders/decoders and the mean square capacity achievable with linear encoders/decoders are given as
2
σn
+P
P
1
,
CShannonBD = 1−
2 , CMSCBD = − 2 log
2 +P
2 log 1+ σn
σn
(1−)2 P
1
2
CMSLBD = 2 log 1 + (1−)P +σ2 . For fixed P and σn ,
n
the channel capacities are functions of . Let P = 1 and
σn2 = 1, the channel capacities in relation to the erasure
probability are plotted in Fig. 2. It is clear that CShannonBD ≥
CMSCBD ≥ CMSLBD at any given erasure probability . This
0.08
0.5
Capacity Channel
0.4
Necessity in Theorem 2
0.3
CShannon
Sufficiency in Theorem 2
0.06
0.2
CMSC
0.1
the Necessary and Sufficient Condition
in Theorem 3.1 in [8]
0.0
0.2
0.4
0.6
0.8
1.0
log|λ2 |
CMSL
0.0
0.04
Erasure Probability ϵ
Fig. 2.
2 =1
Comparison of different channel capacities when P = 1, σn
result is obvious since we have proved that the Shannon
capacity is no smaller than the mean square capacity with
causal encoders/decoders. Besides, we have more freedom in
designing the causal encoders/decoders compared with the
linear encoders/decoders, thus allowing to achieve a higher
capacity. The three kinds of capacity degenerate to the same
when = 0 and = 1, which represent the AWGN channel
case and the disconnected case respectively.
B. Vector Systems
the two dimensional LTI system (9) with J =
λConsider
1 0
,
and
the communication channel is (2) in which
0 λ2
the fading follows the Bernoulli distribution with failure
probability . In view of Theorem 2, a sufficient condition
to ensure mean square stabilizability is that (log|λ1 |, log|λ2 |)
should lie in the region of log|λ1 | + log|λ2 | < CMSCBD . The
necessary requirement is given by the following region in
(log|λ1 |, log|λ2 |) plane
log |λ1 | < CMSCBD , log |λ2 | < CMSCBD
σn2 21
log |λ1 | + log |λ2 | < − log + (1 − )
σn2 + P
The necessary and sufficient condition to ensure mean square
stability using linear encoders/decoders for this system is
given in [8], which states that (log|λ1 |, log|λ2 |) should be
in the region constrained by log|λ1 | + log|λ2 | < CMSLBD .
Selecting P = 1, σn2 = 1 and = 0.8, we can plot the
regions for (log|λ1 |, log|λ2 |) indicated by the sufficiency
and necessity in Theorem 2 and that indicated in Theorem
3.1 in [8] in Fig. 3. We can observe that the region of
(log|λ1 |, log|λ2 |) that can be stabilized with the designed
causal encoders/decoders in Section IV is much larger than
that can be stabilized by linear encoders/decoders in [8].
Thus by extending endocers/decoders from linear settings to
causal requirements, we can tolerate more unstable systems.
VI. C ONCLUSION
This paper characterized the requirement for a power constrained fading channel to allow the existence of a causal encoder/decoder pair that can mean square stabilize a discretetime LTI system. The mean square capacity of the power
constrained fading channel with causal encoders/decoders
was given. It was shown that this mean square capacity is
smaller than the Shannon capacity and they coincide with
each other for some special situations. Throughout the paper,
the capacity was derived with the assumption that there
0.02
0.00
0.00
0.02
0.04
0.06
0.08
log|λ1 |
Fig. 3. Stability region of (log|λ1 |, log|λ2 |) indicated by Theorem 2 for
a vector system
exists a perfect feedback link from the channel output to
the channel input. What would the capacity be for power
constrained fading channels when there is no such feedback
link or there is only a noisy feedback link is still under
investigation.
R EFERENCES
[1] G. Como, B. Bernhardsson, and A. Rantzer, Information and Control
in Networks. New York: Springer, 2014.
[2] G. N. Nair and R. J. Evans, “Stabilizability of stochastic linear
systems with finite feedback data rates,” SIAM Journal on Control
and Optimization, vol. 43, no. 2, pp. 413–436, 2004.
[3] A. Matveev and A. Savkin, “An analogue of shannon information
theory for detection and stabilization via noisy discrete communication
channels,” SIAM Journal on Control and Optimization, vol. 46, no. 4,
pp. 1323–1367, 2007.
[4] A. Sahai and S. Mitter, “The necessity and sufficiency of anytime
capacity for stabilization of a linear system over a noisy communication link - part i: Scalar systems,” IEEE Transactions on Information
Theory, vol. 52, no. 8, pp. 3369–3395, 2006.
[5] N. Elia, “Remote stabilization over fading channels,” Systems &
Control Letters, vol. 54, no. 3, pp. 237–249, 2005.
[6] J. H. Braslavsky, R. H. Middleton, and J. S. Freudenberg, “Feedback
stabilization over signal-to-noise ratio constrained channels,” IEEE
Transactions on Automatic Control, vol. 52, no. 8, pp. 1391–1403,
2007.
[7] J. S. Freudenberg, R. H. Middleton, and V. Solo, “Stabilization and
disturbance attenuation over a gaussian communication channel,” IEEE
Transactions on Automatic Control, vol. 55, no. 3, pp. 795–799, 2010.
[8] N. Xiao and L. Xie, “Analysis and design of discrete-time networked
systems over fading channels,” in Proceedings of the 30th Chinese
Control Conference, (Yantai, China), pp. 6562–6567, 2011.
[9] A. J. Goldsmith and P. P. Varaiya, “Capacity of fading channels with
channel side information,” IEEE Transactions on Information Theory,
vol. 43, no. 6, pp. 1986–1992, 1997.
[10] T. M. Cover and J. A. Thomas, Elements of information theory.
Hoboken, N.J. : Wiley-Interscience, 2006.
[11] P. Minero, M. Franceschetti, S. Dey, and G. N. Nair, “Data rate
theorem for stabilization over time-varying feedback channels,” IEEE
Transactions on Automatic Control, vol. 54, no. 2, pp. 243–255, 2009.
[12] U. Kumar, J. Liu, V. Gupta, and J. Laneman, “Stabilizability across a
gaussian product channel: Necessary and sufficient conditions,” IEEE
Transactions on Automatic Control, vol. 59, pp. 2530–2535, Sept 2014.
[13] A. Gattami, “Kalman meets shannon,” in Proceedings of the 19th IFAC
World Congress, (Cape Town, South Africa), pp. 2376–2381, 2014.
[14] J. Schalkwijk and T. Kailath, “A coding scheme for additive noise
channels with feedback–i: No bandwidth constraint,” IEEE Transactions on Information Theory, vol. 12, no. 2, pp. 172–182, 1966.
| 3cs.SY
|
Extending programs with debug-related features,
with application to hardware development
Nik Sultana[ Salvator Galea[ David Greaves[ Marcin Wójcik[ Noa Zilberman[ Richard Clegg]
Luo Mai\ Richard Mortier[ Peter Pietzuch\ Jon Crowcroft[ Andrew W Moore[
arXiv:1705.09902v1 [cs.PL] 28 May 2017
[
Cambridge University
\
Imperial College, London
Abstract
The capacity and programmability of reconfigurable hardware such as FPGAs has improved steadily over the years,
but they do not readily provide any mechanisms for monitoring or debugging running programs. Such mechanisms
need to be written into the program itself. This is done using ad hoc methods and primitive tools when compared to
CPU programming. This complicates the programming and
debugging of reconfigurable hardware.
We introduce Program-hosted Directability (PhD), the
extension of programs to interpret direction commands at
runtime to enable debugging, monitoring and profiling. Normally in hardware development such features are fixed at
compile time. We present a language of directing commands,
specify its semantics in terms of a simple controller that is
embedded with programs, and implement a prototype for
directing network programs running in hardware. We show
that this approach affords significant flexibility with low impact on hardware utilisation and performance.
Keywords debugging, FPGA, program directing, profiling,
high-level synthesis, aspect-oriented programming
1.
Introduction
When debugging and monitoring programs running on microprocessors we usually benefit from hardware support that
is leveraged by an Operating System to inspect and modify
running processes (Gagnon et al. 2007). But when a program
is run on reconfigurable hardware platforms, one does not
usually have an operating system, a notion of process, nor
any hardware support for debugging. Programs must contain
additional logic to enable debugging, monitoring, and profiling during their execution, because the environment does not
provide visibility into running programs by default.
Field-Programmable Gate Array devices (FPGAs) are a
form of programmable hardware consisting of a grid of logic
blocks whose function and wiring can be flexibly reconfigured. FPGAs are used to perform functions for which a
full-featured general-purpose CPU is not appropriate. For
such functions, FPGAs can operate at a higher throughput
and consume much less electricity than CPUs (Mittal 2014,
]
Queen Mary University, London
§4.4). This makes FPGAs especially appealing for some
problems and environments, a recent example being datacentres (Putnam et al. 2014).
Despite their appeal as a computing device, the programmability of FPGAs has been hampered by the need
for low-level hardware-description languages traditionally
used to program them, such as Verilog and VHDL, which
“requires programmers to assume the role of hardware designers” (So et al. 2002). Research has yielded various approaches for high-level synthesis (HLS): this tends to consist
of using (fragments of) existing languages, such as C or Java,
to describe hardware.
In addition to a programmability gap between FPGAs
and CPUs, there is a debuggability gap that has received
far less attention. The programmability of FPGAs has improved over the years, but they are not debuggable by default (Potkonjak et al. 1995). FPGAs provide no visibility
into the running program, and their standard tooling provides
very limited support for this. For full visibility one could
simulate the program, (e.g., on a workstation) but the simulation can be slower by a factor of 106 (Camera et al. 2005)
because of the sheer amount of detail that must be simulated.
In this paper we propose to improve the debuggability
of programs running on FPGAs by using a domain-specific
control language inspired by program directing (Sosič 1992)
to generate in-program support for debugging. In addition to
debugging, this can be used to monitor and profile programs.
We call our approach Program-hosted Directability (PhD).
It involves extending the user’s program to service direction commands at runtime. Extending the program involves
inserting (i) named extension points which can contain
runtime-modifiable code in a computationally weak language (no recursion), and (ii) state to be used for bookkeeping by that code, to implement direction features.
For example, the direction command “trace X
max_trace_idx” (where X is a variable in the program’s
source code) logs updates made to a variable, and appends
these updates to a buffer (of size “max_trace_idx” for later
inspection). We translate this command into (i) a snippet that
gets injected into the program (Figure 1) and (ii) a modification of the program to allocate space for this snippet,
if V_trace_idx < max_trace_idx then
V_trace_buf [ V_trace_idx ] := V;
inc V_trace_idx ;
continue
else
inc V_trace_overflow ;
break
Figure 1. Code that implements the direction command
“trace X max_trace_idx”. If the buffer is not full then the
new value of X is logged, the index incremented, and control
is handed back to the program that hosts this code. Otherwise indicate depletion of the associated buffer resource and
break the program’s execution.
and for state require by this snippet: such as the variable
V_trace_idx, and array V_trace_buf. More examples are
given in §3, and the overall scheme is shown in Figure 2.
Idiomatic direction features (such as tracing, breakpoints,
etc) are compiled down to a weak language and executed
by an interpreter embedded in the program. For consistency
with the description of the interpreter’s behaviour as hardware, we refer to it as a controller. The controller is invoked
by extension points, which are inserted into the program
through transformation. In the above example, the extension
points would be added after each update to a variable, to ensure that the update is considered for logging.
Extension points are added at compile time depending
on what direction commands we want to support at runtime. For example, starting with the sequence of statements
. . . ; si ; si+1 ; . . . we insert two extension points @L and
@M , resulting in the program . . . ; si ; @L; si+1 ; @M ; . . ..
When extension point L is reached at runtime, a stored procedure associated with L is executed by the controller. This
procedure ultimately hands back control to the host program, or starts an interactive session with a program director, which can send the controller further commands, and
update its stored procedures.
Our work systematises ad hoc debugging and profiling
extensions to programs, and generalises the facilities currently made available for hardware development. Moreover,
we can include only the features needed, thus improving
the utilisation of the hardware, and its power consumption.
The approach is extensible: one can code additional direction commands, or variations (§3.3). In this paper we show
how to give semantics to program direction commands in
terms of the placement of extension points, the code that is
to be run by the controller, and the interaction between the
controller and the direction tool that manages it.
We prototype PhD using an HLS and obtain a uniform
interface for directing the software and hardware instances
of the same program, allowing us to unify the debugging of
these instances (which otherwise require diverse tools).
We believe that PhD can yield practical benefits in hardware development and deployment. As a technique, PhD is
vendor-neutral and compiler-neutral, and the communica-
Original program behaviour
(Normal interaction
with external world)
Program
Hosted directability
Controller
Director
Figure 2. A controller is embedded into the program, and
acts as the agent of the director. The director and controller
implement a protocol to exchange commands and their outputs.
tion between the director and controller can be adapted to
suit the program. For example, in our prototype we send direction commands via the network. PhD can be implemented
in different ways: in our prototype we implemented it entirely in a high-level language via HLS, but one could also
bolt it onto more finely engineered hardware blocks. Despite
running as hardware, the directability features can be enabled and reconfigured at runtime: this consists of updating
the code stored by the controller. In contrast, existing techniques for FPGAs involve including fixed function circuits in
the design. Finally, PhD extends a program with a direction
mode, thus facilitating the in-field debugging of programs.
Through PhD we hope to contribute to the convergeance
between the debuggability of programs on FPGAs, and those
on CPUs, which enjoy extensive hardware and OS support,
and which in turn benefits sophisticated monitoring systems
such as DTrace (Gregg and Mauro 2011). The ideas described in this paper are not necessarily tied to the languages,
compilers, FPGA or other equipment we used in our prototype. We make the following contributions:
• We describe how to translate familiar high-level idioms
for debugging, profiling and monitoring (§3.1), which
we call direction commands, into a low-level language
for controlling program state at runtime (§3.2), for an
example language §3.4.1.
• We relate the direction commands with our low-level lan-
guage through a specification (§3.4) in which programs
are ordered by the directability features they support.
This provides a basis on which we can reason that one
program is “more debuggable” than another at runtime.
• A prototype implementation and its evaluation (§4),
where we measure the effect of directability on FPGA
utilisation and performance.
2.
FPGA debugging gap
Irrespective of how we write our program as a hardware
description, once a so-called “bitstream” is generated that
configures an FPGA to run our program, the program can
only be tested as a black box, and we cannot understand or
further influence its behaviour at that stage.
The need for better FPGA debugging features is becoming more urgent since:
• FPGA chips are getting larger, which allows them to run
more complex programs. Complex programs are more
likely to be buggy, which necessitates more debugging.
• FPGAs are also used for simulating hardware designs,
since behavioural simulations are very slow. (Wang et al.
2011; Chu et al. 2015)
• FPGAs are being deployed in large production environ-
ments, such as Microsoft’s datacentres (Putnam 2014).
Hardware development has driven the development of
formal methods to establish system correctness (Fix 2008),
which enabled the development for methods for software (Ball
et al. 2006; Godefroid et al. 2012). Unfortunately the verification is done on the Register-Transfer Level (RTL), a
higher-level description of a hardware circuit in languages
such as Verilog, and not on the generated bitstream. Thus
debugging may still be needed.
2.1
Debugging(FPGA) != Debugging(CPU)
Debugging concepts from software do not correspond directly with debugging the same program on an FPGA. A lot
of the core issues involved were discussed in the pioneering
work of Koch et al. (1998) and in many other work on HLS
debugging (Goeders and Wilton 2014; Monson and Hutchings 2015). These are the main points:
• Multiple source lines might be executed concurrently in
hardware. Code is represented at the source, registertransfer, and gate levels. This has important consequences for debugging, described in the following points.
The correspondence between these levels necessitates
keeping metadata from compilation for debugging.
• Depending on the debugged artefact, stepping by (source)
line might be less useful than stepping by cycle.
• Breakpoints become more tricky to interpret, since they
are usually set on a specific line of code. In hardware
there may be several overlapping lines being executed at
a breakpoint (Koch et al. 1998, §4.3). Moreover, the output of operations on previous lines might only be available after some clock cycles have elapsed. Depending on
what the user wishes to do, they might prefer if the breakpoint is triggered after the elapsing of these cycles. Furthermore, part of the next line of source code might have
started executing. This suggests that a strict indication of
sequentiality needs to be communicated to the HLS compiler if the usual breakpoint semantics are desired.
• As mentioned in the introduction, FPGAs do not provide
hardware support such as debug registers to assist with
analysing running programs.
2.2
Current techniques for FPGA debugging
Some existing techniques help narrow the debugging gap on
FPGAs. Co-simulation involves comparing the behavioural
simulation between HLS and RTL. This can be considered
a special case of relative debugging (Sosič and Abramson
1997) but it does not provide visibility into the hardware
instance of a program. Another technique involves in-system
testing: testing a large part of the system, though possibly not
all. This does not provide visibility into the hardware either.
Current practice employs two techniques for FPGA debugging. Trace buffers are the most popular technique for
debugging FPGAs. It requires a programmer to identify signals of interest in the circuit at compile time, then an embedded logic analyser is synthesised that uses on-chip memory to record traces for these signals. This suffers from two
problems: only a limited number of signals may be viewed
(limited by on-chip memory), and traces have a limited window size (for the same reason). Traces may be conditional, to
avoid using up buffer space unnecessarily, but this technique
is difficult to use because it involves generating the bitstream
each time. Register scanning allows you to see the values of
all registers on the FPGA, but requires “stopping the world”
to enable reading and sending it off-chip. This slows down
tests, and thus register scanning has been supplanted by trace
buffers.
Both register scanning and trace buffers usually send
recorded data off-chip via the JTAG (Joint Test Action
Group) interface, a standardised instrumentation method (Bennetts and Osseyran 1991). This method is not scalable, since
its transfer rate is far less than the FPGA throughput.
In summary, existing techniques consist of including
“fixed function” modules as part of your hardware description at compile time. This has the advantage of being
lightweight since these circuits are specialised to perform a
single function, but it has the disadvantage of being inflexible. Generating a hardware bitstream can take hours, and
the added overhead costs for runtime-reconfigurable debuggability and monitoring features might not be affordable in
some use-cases. Furthermore, these techniques cannot be
used in production environments.
2.3
High-Level Synthesis (HLS)
HLS involves the use of a high-level language, such as Java,
C, C++, and OCaml, to write a hardware description. This
takes advantage of the features of, and tooling available
for, the high-level language. An RTL description is then
generated from the high-level description.
Using an HLS to describe hardware enables one to run
the HLS description as a software program, and to debug it
as such, by using standard tools to compile and debug Java
programs for instance. This prunes bugs from the eventual
bitstream and avoids regenerating the bitstream. Irrespective
of whether testing is directed at software or hardware, it can
take many tests to find a fault. Software-based testing could
help detect logic errors in our code, but it could not help
us find some important classes of problems: ‘[Testing in] the
silicon mode permits the analysis of bugs that are “invisible”
at the RTL level’ (Calagar et al. 2014). We outline the main
cases below:
1. Interface mismatch. We need to understand whether a
problem occurs because of a mismatch between one module and the rest of the circuit. Recall that behavioural simulation might not be applied to the whole design, and incorrect assumptions about the enclosing circuit can result
in the simulation test succeeding but the hardware tests
failing.
2. Reproducability. Some faults are triggered during highthroughput tests, and are difficult to find when testing other instances of the program. Other faults result
from features of the hardware and transient environmental states – such as “Single Event Effects” manifested
through the flipped or stuck bits from the interaction of
charged particles with semiconductors (Sari et al. 2014;
Krishnaswamy et al. 2008).
3. Toolchain problems. Diagnosing bugs in the compiler
toolchain becomes easier if we can see into the compiled
program’s operation. Bugs are not unusual in both HLS
and RTL toolchains.
2.4
Current research on FPGA debugging
Various improvements have been explored for the techniques
described above. For example, one could write summaries
to the trace-buffer, rather than the explicit trace (Goeders
and Wilton 2014). Another idea consists of multiplexing all
signals and choosing which to observe at debug time rather
than at compile time (Hung and Wilton 2013), enabling
“observation without recompilation”. An additional idea is
to buffer into a fast external memory (Panjkov et al. 2015).
Table 1 surveys closely related work, classifying it according to the features they provide. Related work is contrasted in more detail in §5. In Table 1 extension points
means being able to extend the program at certain points
at runtime. Interruption refers to asynchronous interruption
of a program by a debugger. Fine granularity means that
we can look at arbitrary parts of a program; for instance
McKechnie et al. (2009) only allow inspecting at the module
boundary. Software instance means that the program can be
run on the CPU as a process, independent from the FPGA.
Network/Control refers to how a technique is implemented:
as a control loop or as a network that feeds signals to a
logic analyser. Use leftover resource means that the debug
circuitry does not compete with the program’s circuit. And
embed at Source/HDL refers to whether the source code or
its HDL image is updated to include the debug circuitry.
2.5
PhD design features
PhD came about after we found ourselves extending our ad
hoc debugging and monitoring code to support additional
features, instigating us to study the problem more rigorously.
Table 1 shows the features supported by PhD. Initially
it supported state inspection and update. Updating state enabled us to influence data-dependant control flow. Extending
the controller to include branching and extension points en-
abled us to support more features, such as assertion checking, which involves breaking if a condition is satisfied.
As an idea, PhD is neither committed to HLS nor HDL
description. In our prototype we implemented it in HLS for
convenience, but often better performance can be gained by
relying partly or fully on modules written in HDL.
3.
A language and model for program
directing
In this section we describe D, a language of direction commands that we extracted by analysing the commands commonly given to profilers and debuggers such as gdb.1 In Figure 1 we described one such command and how we code
the high-level direction command as a program that will be
executed by the controller.
3.1
Direction language D
Our direction language D consists of the commands listed in
Table 2. These commands can have three kinds of parameters: (i) symbols relating to the program, (such as variable or
function names, or labels), (ii) relations over program variables, which we denote by the symbol hBi, indicating the
conditions on which statements might apply, and (iii) measures of resource, which we symbolise by h$i, indicating a
finite resource that is allocated for the execution of a command. Symbol X is a metavariable ranging over variable
identifiers in the source program, and L is a metavariable
ranging over labels in the source program. A label is associated with a single position in the program, (e.g., line 5 in
function “main”), but a single position might be associated
with multiple labels. below.
Let DhBi be the set of possible conditions that can be
used. We assume that at least true ∈ DhBi. We can also
allow additional truth conditions, and in this model we will
have (V1 = V2 ) ∈ DhBi for arbitrary V1 , V2 ranging over
program values or variables. For example, “watch v (v =
5)” would instruct the controller to watch a variable v, and
switch to interactive directing when v = 5.
Dh$i ∈ N describes the maximum quantity of some resource when carrying out a direction command. This value
must be less than the compile-time allocation of the resource, to ensure the provision of sufficient resource for
the command at runtime. This is needed to size the buffers
used for tracing. For example, “count reads v true 5000”
will count the number of reads of v, and break after 5000
reads have been made. This could be done to avoid overflow, or to capture some behaviour of interest. Similarly,
“trace start v true 500” breaks after 500 instances of v
have been recorded in the trace. The trace buffer must accommodate at least 500 entries.
1 https://www.gnu.org/software/gdb/
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
C
C
C
N
C
C
C
C
X
embed at Source/HDL
use leftover resource
X
Network/Control
X
X
HLS (vs HDL)
X
X
runtime reconfigurable
X
X
X
software instance
(see §2.5) PhD
X
X
X
X
timing checks
X
X
X
hang detection
X
X
X
X
assertion checking
X
X
X
X
fine granularity
extension points
X
interruption
state updating
X
stepping
trace recording
X
X
break points
state inspection
X
X
X
X
X
X
X
Features
(Sosič 1992) Dynascope
(Goeders and Wilton 2014) HLS-Scope
(Calagar et al. 2014) Inspect
(Panjkov et al. 2015)
(Hung and Wilton 2014) QuickTrace
(Koch et al. 1998) SLE/CADDY
(Monson and Hutchings 2015)
(Curreri et al. 2011)
(Camera et al. 2005) BORPH
System
S
H
S
H
H
S
H
H
C
Table 1. Survey of features provided by debugging systems. Blacked-out boxes mean “not applicable”.
Command
print X
break L hBi
unbreak L
backtrace h$i
watch X hBi
unwatch
X
reads X hBi h$i
writes X hBi h$i
count
calls fname hBi h$i
start X hBi h$i
stop X
clear X
trace
print X
full X
Behaviour
Print the value of variable X from the source program.
Activate a (conditional) breakpoint at the position of label L.
Deactivate a breakpoint.
Print the “function call stack”.
Break when X is updated and satisfies a given condition.
Cancel the effect of the “watch” command.
Count the reads or writes to a variable X, or the calls to a function fname.
Trace a variable, subject to a condition being satisfied, and up to trace some length.
Stop tracing a variable.
Clear a variable’s trace buffer.
Print the contents of a variable’s trace buffer.
Check if a variable’s trace buffer is full.
Table 2. Directing commands making up language D. Note that count has similar subcommands to those of trace, to clear
the counters, get their current value, and find out if a maximum value has been reached.
3.2
Controller
High-level direction commands such as those in Table 2 are
ultimately translated into programs that run on a simple controller embedded in the program. We model the controller
as a CASP machine (for “Counters, Arrays, and Stored Procedures”, the constituents of the machine’s memory). CASP
machines are very weak. They are more structurally complex
than register machines (Shepherdson and Sturgis 1963) since
they have separate memories for storing arrays and registers, but CASP machines are computationally much weaker
than register machines, unable to encode partial computable
functions. A limited form of memory indirection is permitted through a collection of arrays. The language lacks any
means for defining recursive functions, or branching to arbitrary addresses. Any more complex computation must be
done by the director; the controller simply provides a controlled access to the program’s memory.
We describe the language of CASP programs in Figure 3.
We rely on the following meta-variables and syntax categories: P programs, E expressions, I indices, U updatable
values, V values, N numerals (corresponding to Z), X variable identifiers, and R the array identifiers, where the names
for variables and those for arrays are disjoint, X ∩ R = ∅.
@L : {P } is a placement command: it updates the code at
extension point having label L to be P . Note that placement
commands may not be nested in our model: for instance, this
is not a valid program:
P
::= E
|
U := E
|
op U
op ∈ {inc, dec}
|
P1 ; P 2
|
if E then P1
else P2
|
break
|
continue
|
@L : {P 0 }
E
I
U
V
::= V
|
−V
|
V1 op V2
op ∈ {=, <}
::= N
|
X
::= X
|
R[I]
::= I
|
R[I]
Figure 3. Syntax for CASP programs.
@L : {if x = 1 then @M : {break}
else @M : {continue}}
We do not want to allow programs to be self-modifying in
this way, since it complicates reasoning about them.
Ending programs. Both continue and break indicate
the end of a CASP program, but differ in what happens
before resuming the host program (in which the controller
is embedded). continue simply resumes where the host
program left off, whereas break switches into an interactive
direction mode. In this mode, the controller may receive
commands from the director, execute them, and send an
acknowledgement back.
The remaining syntax forms and constants used above are
standard and intuitive. Owing to its simplicity the semantics
of this language are straightforward, and are given in §C.1.
3.3
Examples
CASP programs are to D what microprograms are to an Instruction Set Architecture (Smith and van Leeuwen 1974).
We give some examples of coding program direction commands below, before describing the behaviour of program
direction commands in more detail in the next two sections.
Conditional tracing. Let BASIC represent the program
that codes the behaviour of “trace V...” from Figure 1.
We can code the conditional variant of this command,
for example when V is less than some threshold value
V_trace_threshold (and where >= is syntactic sugaring):
if V >= V_trace_threshold then
continue
else
BASIC
Sampled tracing. The sampled variant involves allocating
an additional variable to count the interval between samples,
and storing the desired sample interval:
if V_trace_samp = 0 then
V_trace_samp := V_trace_samp_interval ;
BASIC
else
dec V_trace_samp ;
continue
Profiling. The command “count writes v” causes a counter
to be incremented each time a variable is updated:
if V_count_writes > V_count_writes_max then
break
else
inc V_count_writes ;
continue
Watchpoints. The command “watch v hBi” causes the
program to break (for interactive guidance) when a variable’s value satisfies some predicate. Let B be the CASPlevel value to the D-level hBi parameter (we give such a
function in §3.5.1). The code is similar to that in the profiling example above, except that B is user-provided.
if B then break else continue
Breakpoints. “break L hBi” causes the program to break
when it reaches a specific label, and if condition B is satisfied. The coding is identical to that in the watchpoint
example, but they differ in their placement: breakpoints
are placed at programmer-specified positions in the code
whereas watchpoints are associated with labels where variables are updated. This difference cannot be seen from the
snippet, but will become evident in the formalisation of the
program direction commands, which we start next.
3.4
Directability ordering
In this section we define a relation x @ x0 to mean “x0
is more directable than x”, where x, x0 are triples (D, C, p)
and (D0 , C 0 , p0 ), each representing three interdependent parties: the director, controller and program. The user (or their
agent) issues direction commands to the director, which interacts with the program’s state via its agent, the controller,
embedded in the program. We use this relation to give semantics to the direction commands in terms of interaction
with CASP machines.
Our directability relation gathers information about the
three parties involved, and describes how their interdependence is revealed by the directing commands: for example,
the director would not be able to execute trace X if the program did not have a variable called X, or if the controller
had not been allocated a trace buffer.
In our notation, D represents the director’s state (a set of
facts representing its knowledge about the controller’s state,
such as which breakpoints exist, and whether they are active
or not). C is the controller’s state, consisting of a (C, A, SP )
machine, cf §3.2. p is a program. We will define an example
language in §3.4.1 to aid our formalisation.
We also include in the relation some information about
why one triple is less directable than the other. We therefore
index the relation by (i) C ⊆ D the direction commands
(§3.1) supported by (D, C, p), (ii) c ∈ D the additional
command supported by (D0 , C 0 , p0 ), and Dc ∈ (D0 → D0 )
the semantics of this command. Note that C denotes the set
of direction commands that is supported simultaneously by
(D, C, p), i.e., these commands are allocated separate state.
Written out in full, we obtain this relation:
c
(D, C, p) @C (D0 , C 0 , p0 ) : Dc
In §3.5 we will instantiate such a relation by formalising
commands from D in terms of CASP machines. Note that
this describes how the directing commands are translated
into CASP programs, but we do not fully formalise the
director: Dc is written in an ML-like pseudocode.
Our formalisation is devised in away that avoids the mutual interference of direction commands. That is, the same
program can be subjected to any combination of direction
commands. To make this non-interference more precise, we
introduce some definitions. Let D̆ = D0 \D and C˘ = C 0 \C.
We say that Dc is relevant to D0 \D if it only manipulates state or elements introduced in D0 and C 0 . Furthermore
commands are disjoint if they introduce non-overlapping
state. That is, for any two commands c1 and c2 , for any
prior states C and D their respective new states are disjoint:
(D1 \D) ∩ (D2 \D) = ∅ = (C1 \C) ∩ (C2 \C).
3.4.1
Program language
In this section we specify a first-order imperative language to
support our formalisation of program direction commands.
Unlike CASP machines (§3.2) this language is computationally strong: recursive functions over the integers can be
encoded. The language’s simplicity enables the relation of
program direction with CASP machines, while avoiding excessive formal complexity. Formalising transformations for
realistic languages—even simple transformations (Schäfer
et al. 2008)—is usually fraught with complex definitions,
and we avoid that here.
The language grammar is given next. Note that for simplicity we deliberately overlay the meta-variables for variables and numerals (X and N ) over those for CASP machines. This simplifies the interface with CASP programs,
which will be executed at extension points within host programs. p ranges over programs, s over statements, e over
expressions, τ over types, and vdecl and fdecl over the declaration of variables and functions respectively.
p
vdecl
fdecl
s
e
::=
::=
::=
::=
|
|
|
::=
|
|
|
vdecl fdecl return fname(ē)
τX
τ fname(x̄){s; return e}
skip
X := e
if e then s; s
extend{L1 , . . . , Ln }
N
X
fname(ē)
e1 op e2
op ∈ {+, −, ==, <}
In this language the only data type in τ is the integer type. The only unusual construct in this language is
extend{L1 , . . . , Ln }. This indicates an extension point,
where control is passed to the CASP machine (§3.2). As
before, L is a metavariable ranging over labels drawn from
a denumerable set.
The semantics of this language are straightforward, and
are given in §C.2. Intuitively, the behaviour of extend{L1 ,
. . . , Ln } is as follows. If n = 0 then the command has no
effect. Otherwise, the stored procedure associated with each
Li is called, in any order, and run to completion, noting the
last instruction of each Li . The last instruction of any CASP
program is either ‘continue’ or ‘break’ (§3.2). If all Li end
in ‘continue’, then the behaviour of extend{L1 , . . . , Ln }
is to continue executing the next statement in the host program. Otherwise, if at least one Li ends in ‘break’, then the
controller switches into interactive mode. In this mode, control remains with the controller, until the director sends it a
‘continue’ command, at which point control is returned to
the program.
We make the simplifying assumption that all the statements in the user’s program are interspersed with ‘extend’:
that is, if the user writes s0 ; . . . ; sn then this is translated
into extend{}; s0 ; extend{}; . . . ; extend{}; sn ; extend{}.
In the next section we populate these extension points with
labels, to extend the directability of a program.
Semantics for D
3.5
In this section we encode program direction commands (§3.1)
into interactions between the director and controller (§3.2).
Note that this describes how the directing commands are
translated into CASP programs, but we do not fully formalise the director: Dc is written in an ML-like pseudocode.
3.5.1
Break
We start by formalising the meaning of the “break” command. Intuitively, this command adds a breakpoint to a program: an extension point is at the position of the breakpoint
is labelled with L, and the associated state is set up in the
director and controller. To support this command:
• A program p is extended to include a label L at the posi-
tion where the breakpoint is to be placed. This extension
is formalised by the premise p <1L p0 , which means that p
is identical to p0 except for the label L occurring at some
extension point. This is defined formally in §B along with
related definitions, such as that of a position in the program.
• The label L must not appear in the original program. We
write this as L 6∈ p using an abbreviation defined in §B.
• The controller’s state is extended to store the procedure
associated with L. Furthermore, the breakpoint is activated by default. We use the abbreviated notation C˘ =
{SP [L 7→ break]} to indicate this extension, where SP
is the stored-procedure memory in the CASP.
• The director’s state is extended to encode whether the
breakpoint is currently active or not. It is activated by
default, thus: D̆ = {(«bp», L, 1)}, where “«bp»” is a
unique token we use for breakpoints, and 1 is a token
we use to indicate that the breakpoint is active. Soon we
will formalise the “unbreak” command, which flips this
value to 0.
• The director’s behaviour for the “break” command, Dc ,
involves activating the breakpoint unless already active.
In the definition of Dc below we use the notation P
N , which we use to mean that the director sent the CASP
program P to the controller, and received the reply N . Thus
“@L : {break}
pLq” means that the director instructed
the controller to store the program break at L, and that it
expects to get pLq back (which is a code indicating where
the program is stored, as formalised in §C.1).
In Dc we use the notation “(«bp», L, 0 7→ 1) :∈ D0 ” to
abbreviate {(«bp», L, 1)} ∪ (D0 \{(«bp», L, 0)}).
The formalisation of the “break” command follows. We
can use a simplified notation since our commands will be
both relevant to D0 \D and disjoint:
p
break L hBi
@C
L 6∈ p
p <1L p0
D̆ = {(«bp», L, 1)}
C˘ = {SP [L 7→ Jbreak L hBiKSP ]}
0
0
0
D
0
c = λD . if («bp», L, 1) ∈ D then D
p
else
Jbreak L hBiK
pLq;
(«bp», L, 0 7→ 1) :∈ D0
where
Jbreak L hBiKSP = conditional hBi break
Jbreak L hBiK = @L : {Jbreak L hBiKSP }
We use Jbreak L hBiK to denote the meaning of
“break L hBi” to the director, as a CASP program.
Jbreak L hBiKSP is the value used to initialise the stored
program associated with L.
The meaning of hBi is translated for inclusion in the
CASP program by the following function:
conditional
hBi t =
t
if I1 == I2 then t
else continue
if hBi = true
if hBi = (I1 = I2 )
We now turn to the “unbreak” direction. To be able to
issue this direction, the breakpoint needs to exist—thus we
have a dependency on the “break” direction earlier:
p
unbreak L
@C
(break L hBi) ∈ C
D̆ = {}
C˘ = {}
p
Dc = λD0 . if («bp», L, 0) ∈ D0 then D0
else Junbreak LK
pLq;
(«bp», L, 1 7→ 0) :∈ D0
where
Junbreak LK = @L : {continue}
Note that the program, and the states of the controller and
director are not changed. The only extension is made to the
behaviour of the director, which is extended with a function
to unset the breakpoint. This time we didn’t set anything
in the controller since there is no default behaviour and
additional state required for unsetting a breakpoint. This
is because unbreaking a breakpoint doesn’t establish the
breakpoint, whereas breakpoint does.
To print the value of a variable X, that variable needs to
exist in the program (X ∈ Varp ), and we need to have at
least one extension point (through which we can send the
print command).
(break L hBi) ∈ C
X ∈ Varp
D̆ = {}
˘
print X
C = {}
p @C p
N;
D = λD0 . X
c
print(N )
This time we didn’t use a placement command to update the
behaviour of the controller; we simply ran a query. print(N )
is pseudocode that uses a print function in the director. Recall that we formalise directions in terms of CASP machines,
and don’t formalise the director’s behaviour. X ∈ Varp ensures that X is a variable in p. Varp is defined in §B.
3.5.2
Trace
The most important command related to tracing is “trace
start”; the other commands depend on it.
X ∈ Varp
∀L ∈ XL . L 6∈ p
Positionsp0 (XL ) = PostUpdatep0 (X)
p <XL p0
p
trace start X hBi h$i
@C
p0
D̆ =
{(«t», X, 1)}
C [Xi ] = 0, C [Xof ] = 0, A[Xa [h$i]],
SP [L 7→ Jtrace start X hBi h$iKSP ]
C˘ =
for each L ∈ XL
Dc = λD0 . if («t», X, 1) ∈ D0 then D0
else
for each L ∈ XL :
@L : {Jtrace start X hBi h$iKSP }
pLq;
(«t», X, 0 7→ 1) :∈ D0
Jtrace start X hBi h$iKSP =
conditional
hBi
if Xi < h$i then
Xa [Xi ] := X;
inc Xi ;
continue
else
inc Xof ;
break
pLq;
where Jtrace stop XKSP = continue.
The remaining commands are formalised in §A, supported by program-level predicates and functions defined
in §B.
4.
Implementation and Evaluation
We prototyped the ideas described in the previous section
by extending network programs with program-hosted directability, and compiled them to run on an FPGA. We then
evaluated the effect of these directability features on the program in which they are embedded. Our approach enables us
to make fine-grained modifications to directability, and we
evaluate the overhead from supporting different CASP machine instructions.
4.1
ut
to
Controller
Figure 4. Transformation of the program to include a controller. Normal packets are handled as normal, but direction
packets are passed to the controller. Pink dots represent extension points, one of which is added within the control flow
of the original program in this illustration.
Focus. From its description, the PhD idea is not constrained to a specific kind of program. In our prototype we
focussed on using it to work with network programs however, for two reasons:
1. It allows us to test remote directing over standard network equipment. In our survey of tools and techniques
(Table 1) only Dynascope has network access, but it does
not work for FPGAs.
2. It allows us to use industrial high-precision network measuring equipment to see the effects on the program hosting a controller.
p
D̆ = {}
C˘ = {}
Dc = λD0 . if («t», X, 0) ∈ D0 then D0
else
for each L ∈ XL :
@L : {Jtrace stop XKSP }
(«t», X, 1 7→ 0) :∈ D0
Program
ke
(trace start X hBi h$i) ∈ C
Positionsp (XL ) = PostUpdatep (X)
trace stop X
@C
Program
c
Pa
Applying this rule depends on a set of labels XL in p0 that
don’t exist in p. p0 is the least extension of p that includes
these labels. Furthermore, these labels coincide with the
positions in the program occurring after X has been updated.
The controller’s state is extended with the buffer index,
Xi , initialised to 0; the overflow indicator Xof , initialised to
0 (for false); and Xa is an array that can hold h$i elements.
Each L ∈ XL labels the positions immediately after the
variable has been updated.
p
t in
ke
c
Pa
Prototype
In our prototype we manually transformed programs to include the controller. This transformation was straightforward: we wrote the controller (implementing a CASP machine) and added extension points to the program (consisting
of calls to the controller to execute a stored procedure).
Use-cases. As programs we used implementations of DNS
and Memcached that we had written previously to run on FPGAs, as part of earlier research. DNS (Domain Name System) is a ubiquitous name-resolution system used on private and public packet-switched networks such as the Internet (Mockapetris 1987). This implementation was around
700 lines of C# . Memcached (Fitzpatrick 2004) is a wellknown, in-memory key/value store that caches read results
in memory to quickly respond to queries. The protocol uses a
number of basic commands such as GET (retrieve a value associated with the provided key), SET (store a key/value pair)
and DELETE (remove a key/value pair) and has both ASCII
and binary protocols. In this work as proof-of-concept we
have implemented a limited version of Memcached supporting GET/SET/DELETE using the binary protocol over UDP
and supporting 6 byte keys and 8 byte values. This implementation was almost 900 lines of C# .
Method. We transformed the DNS and Memcached implementations in two ways: (i) adding code to check whether a
received packet is a direction packet intended for the controller (see Figure 4), in which case the controller (and not
the original program) processes the packet; (ii) adding an extension point in the body of the (DNS or Memcached) main
loop, allowing us to influence and observe the program from
that point. We form an enumerated type that corresponds to
the program variables whose values the controller may access and change at runtime. The code for each value of the
enumerated type is used to refer to the program value, to
10G Port
10G Port
10G Port
10G Port
this module, the packets are passed to the physical interfaces
and are transmitted.
10G Port
Input
Arbiter
Main Logical Core
Output
Queues
10G Port
10G Port
10G Port
PCIe & DMA
PCIe & DMA
Figure 5. NetFPGA Reference Pipeline. Each of the implementations in our prototype consists of a separate Main Logical Core.
instruct the controller to increment it, for example. Building
tool support to automate parts of this process seems feasible.
Direction packets. Direction packets are network packets
in a custom and simple packet format, whose payload consists of (i) code to be executed by the controller, or (ii) status replies from the controller to the director. It enables us
to remotely direct a running program, similar to gdb’s ‘remote serial protocol’.2 Our design uses a simple direction
language, and works for instances of a program that run both
as software and hardware (on FPGAs), whereas gdb requires
special backends for each architecture.
Controller. Our controller follows the description of CASP
machines very closely (§3.2). It has two features. First,
memory is organised into Counters, Arrays and Stored procedures. Counters include variables in the original program,
as well as extra registers used for program directing. In this
prototype we only support numeric datatypes; structured
datatypes could be encoded in principle. Secondly, a function that interprets the language of CASP machines. This is
used to branch to stored procedures when their corresponding extension points are reached.
Tools and equipment. We wrote our programs in C# , and
used the Kiwi high-level synthesis (HLS) system (Singh and
Greaves 2008) that statically recompiles .NET bytecode into
Verilog. The Verilog code generated by Kiwi was slotted
into open-source reference code from the NetFPGA project,3
and compiled to run on the NetFPGA SUME board (Zilberman et al. 2014), a low-cost, PCIe host adapter card able
to support 40Gb/s and 100Gb/s applications. At the core
of the board is a Xilinx Virtex-7 690T FPGA device. The
NetFPGA reference designs share the same FPGA architecture, illustrated in Figure 5, of multiple physical interfaces
surrounding a logical data-path. Traffic to the FPGA enters
through one of four 10Gb/s ports or from a PCIe interface,
and is passed into the main data-path by an input arbiter,
which arbitrates between all ingress interfaces. Packets are
then processed within the main logical core, and are diverted
to their destination through an output queues module. From
2 http://www.embecosm.com/appnotes/ean4/
embecosm-howto-rsp-server-ean4-issue-2.html
3 http://netfpga.org/
4.2
Evaluation
We evaluate our prototype by carrying out a quantitative
analysis of the impact that the controller has on the program
in which it is embedded. This impact is measured in terms of
utilisation of resources on the FPGA, and the performance
of the host program.
An FPGA consists of an interconnected grid of logic
blocks, which in turn contain resources such as memory and
logic functions (in the form of so-called look-up tables).
Flip-flops are primitive storage circuits.
Table 3 shows the utilisation and performance for DNS
and Memcached, extended with different controller features.
These features show a fine-grained decomposition of the
instructions supported by the controller: reading a variable,
writing a variable, and incrementing a variable. We see that
the impact on utilisation and performance is minimal.
Performance is analysed in three ways. The duration is
the number of clock cycles that are needed by the program
to process a packet within the main logical core, as extracted
from simulation. In our hardware, each clock cycle takes
10ns. Latency is the time taken for a program to service
requests from the network. Throughput is the rate of requests
that can be serviced by the program (before packets start to
be dropped).
Using the controller we can read and change the program’s state at the packet rate. Using JTAG and the Virtex-7
FPGA we can read data at up to 66Mbps, three orders of
magnitude less than the maximum throughput that the NetFPGA can sustain over high speed serial interface –e.g. the
PCIe channel. In principle the direction controller could use
any slice of that, subject to not interfering with the hosting
program too much.
In Table 4 we compare DNS extended with one extension
point, against DNS extended with an embedded logic analyser (ELA). An ELA is a standard technique in hardware
development, and consists of a circuit that passively monitors the program, creating a trace of the program’s execution. DNS+2e is the DNS extended with one extension point
that only contains a NOP. In (Count) the extension point’s
stored procedure is changed into a counter, and in (Trace) it
is changed to emulate the behaviour of the ELA.
These results confirm that the resources overhead is minimal, making PhD a feasible solution. We note that in the
use-cases detailed below the FPGA resources were never exhausted, and consumed less than 25% of the logic resources,
even for complex services.
In addition to the quantitative evaluation, we note that
PhD is vendor-neutral and runtime-reconfigurable, and can
be used for remote in-field debugging, whereas standard
techniques for FPGA debugging do not provide this.
Artefact
Utilisation (%)
Logic
Performance
Duration Latency QueriesFlip-flops (#cycles) (µs)
per-sec
(KQPS)
DNS
+R
+W
+I
100.00
103.40
115.05
109.79
100.00
102.76
106.04
106.12
57
-
1.85
1.85
1.84
1.84
1176
1176
1176
1176
Memcached
+R
+W
+I
100.00
99.17
99.80
100.63
100.00
100.29
100.74
100.69
64
-
2.03
2.03
2.04
2.03
952
952
952
952
Table 3. Profile of utilisation and performance. Read,
Write, and Increment are instructions supported by the controller. Latency is indicated at the 99th percentile. The hardware generation process involves an optimisation step to
place and route components, and on occasion this results in
more utilisation-efficient allocations. The duration for extensions to DNS (clock cycles) did not change since the critical
path of the circuit was not affected.
Artefact
Utilisation (%)
Logic
Performance
Duration Latency QueriesFlip-flops (#cycles) (µs)
per-sec
(KQPS)
DNS+ELA 99.74
100.40
57
1.83
1176
DNS+2e
234.61
(Count) 234.46
(Trace) 218.30
151.06
151.81
151.84
57
62
70
1.86
1.94
1.99
1176
1064
1010
Table 4. Utilisation and performance profile of the
DNS+ELA against the DNS having one extension point,
where the extension point is NOP, packet counting, or variable tracing. Latency is indicated at the 99th percentile.
Test setup. We use a host running Ubuntu server 14.04LTS,
kernel version 4.4.0-45-generic. The host hardware is a single 3.5GHz Intel Xeon E5-2637 v4 on a SuperMicro X10DRG-Q motherboard. An Endace 9.2SX2 DAG card (7.5ns
time-stamping resolution) and a NetOptics passive-optical
tap are used to intercept client-server traffic and permit independent measurement of client & server latency. For the
throughput tests, OSNT (Antichi et al. 2014) is used to control the rate of the generated requests.
5.
Related work
Portable debugging, program directing, and debugging
languages. We were inspired by previous work on portable
debugging (Ramsey and Hanson 1992; Hood 1996; Hanson and Raghavachari 1996) and program directing (Sosič
1995). That work usually makes the assumption that software is compiled to run on a general-purpose CPU however,
whereas we also target reconfigurable hardware. The study
of languages for debugging is a decades old subject (Balzer
1969) and includes sophisticated languages for high-level
querying of programs (Johnson 1977; Golan and Hanson
1993; Winterbottom 1994). Compared to that work, our
work separates more starkly between the role (and language)
of the director, and that of the controller, which have separate
languages. This separation guided our modelling in §3.
Dynascope. Dynascope (Sosič 1992) provides an extremely
fine grained execution stream of events–at the level of machine instructions–providing a complete description of a
program’s runtime behaviour. We provide selective (and
programmable) visibility by default, in the interest of performance. A more detailed stream could be produced if wished.
Dynascope is generative by default (since you need to set filters to ignore events) whereas we are generative by direction
(you only get events that you inserted code to generate).
High-level synthesis and runtime debugging. The closest
work is that by McKechnie et al. (2009) who adapt a debugging paradigm first developed for network-on-chip devices
(NoC). They provide transaction-level granularity, consisting of domain-specific (high-level) events–less fine grained
than Dynascope’s execution stream. Compared to our work
they offer a less flexible interface, but they take a more detailed account of different sorts of interconnects between
components–formats such as LocalLink and PLB.
Many other systems inject code to emit and observe
events to aid with debugging of complex designs, such as
“System on Chip” designs (Lee and Lysecky 2015). SeaCucumber (Tripp et al. 2002) was the first to support sourcelevel (HLS) debugging, both during behavioural simulation and during hardware execution. Monson and Hutchings
(2015) take a different approach: source-code transformation
(at the HLS level) to introduce “event observability ports” to
enable runtime visibility of variables’ values –but note that
it’s not always possible to observe an expression an interest. We took the approach of Monson and Hutchings (2015),
relying on source transformation rather than on the HLS
system.
Overhead can be reduced by being selective about what to
monitor, rather than monitoring everything by default. This
was studied by Goeders and Wilton (2016) who compared
the amount of visibility afforded by different schemes when
recording events such as reads and updates. We left the
choice of what to observe to the programmer.
At present the LegUp HLS system (Calagar et al. 2014)
appears to provide the best support for debugging. In a contribution to that system, Hung and Wilton (2014) take a
different approach to us: they use a two-step incremental
compilation, the second step of which compiles the monitoring system by reclaiming unused FPGA resources. This
approach is less likely to interfere with the timing behaviour
of the observed circuit. Their model is more specialised to
trace-buffers, and it would be interesting to generalise it to
support our directing controller.
Testing of hardware is traditionally concentrated on the
RTL description, where tools and techniques have been developed to verify designs prior to synthesis (Foster 2008; Fix
2008) . The difficulty of checking hardware entirely prior
to synthesis has led to research into the inclusion of runtime monitors in hardware (Todman et al. 2015). Curreri
et al. (2011) describe a system that translates source-level
assertions into monitors. This technique is also used to detect hangs and possible timing overruns. We currently do not
support such monitoring, since it requires a source-level notion of time that we currently do not provide, and “watchdog
threads” that we currently do not include.
Finally, Camera et al. (2005) described a hardware OS,
the Berkeley Operating system for ReProgrammable Hardware (BORPH), on which user programs were run. Their
stitcher extend the user program to support debugging, providing a rich system. Like BORPH we require the designer
to provide “hints to the system of what aspects of the design
may need to be explored at runtime.”
Formalising program directors. Since ‘debugging’ is such
a vague term (compared to ‘compiling’, which has a clearer
functional behaviour), its verification objective is hard to formalise. Perhaps as a consequence of this there has been little work on formalising and verifying debuggers, and usually entirely theoretical (Zhu 2001). Kishon et al. (1991) describe transformations over functional programs to include
monitoring behaviour. Compared to this work, our transformations are not based on continuations (and we deliberately
avoid needing first-order functions to avoid departing too
much from the conventional hardware programming mindset). We use an approach based on operational semantics,
similar to that used by da Silva (1992), but different in several ways: (i) we support a different set of directing commands, (ii) we don’t insert commands into the bytecode for
the debugger to keep track of separations between subexpressions, and to keep track of the path through the expression (program), (iii) da Silva (1992), introduces a language for specifying debuggers whereas we introduce an
operational language for inspecting and changing the program’s runtime state, (iv) we do not consider the equivalence between debuggers. Sansom and Peyton Jones (1997)
describe the profiling of functional programs by using “cost
centres”, a paradigm to identify locations in a program at
an expression-level (rather than functional-level) granularity.
Others have continued that to make it more practical (Faddegon and Chitil 2015). All the techniques described above
rely on program transformation. They all use functional languages as examples whereas we use a simpler imperative
language, to model the paradigm of register-level hardware
programming more directly. Because of the weaving of debugging code into the program all these techniques can (retrospectively) be seen as special-cases of aspect-oriented programming, described next.
Aspect-oriented programming (AOP). AOP involves the
inclusion (“weaving”) of code (called “advice”) during
compile-time or run-time, depending on whether certain
compile- or run-time conditions are satisfied. AOP is a linchpin paradigm for tracing and monitoring (Avgustinov et al.
2006; Hamlen and Jones 2008) but advice can be arbitrary
functions, and as a consequence they might have undesirable
effects on runtime (Avgustinov et al. 2007). This is an important characteristic to our work, where we wanted to reduce
the power of added code. Djoko Djoko et al. (2012) have
categorised advice based on the degree of influence they can
have on the observable behaviour of a program; and Dantas
and Walker (2006) characterise less intrusive advice in their
work on “harmless advice”.
6.
Conclusion
Having poor programming and debugging support hinders
the potential of computing architectures such as FPGAs,
which are gaining importance in modern datacentres. By
using a program directing approach we subsume several
activities that can involve the interactive runtime analyses
of programs (Sosič 1992). Our language-based approach has
two extensibility benefits: (i) more efficient controllers can
be implemented without changing the source language, and
(ii) third-parties could extend or customise the language of
direction commands without changing the controller that
these commands compile to.
Using a program director brings security risks, since it
may alter the control-flow of a program, and this may be exploited (Abadi et al. 2005). We sought to mitigate this risk
by making the presence of the debug mode very apparent,
to reduce the chances of excessive debug functionality being
included in deployment (Pauli 2016). Dynascope was used
to diagnose errors in the Dynascope compiler itself (Sosič
1992, §3.1), but bugs could render such a task impossible,
for reasons similar to the hereditary potential of vulnerabilities in compilers (Thompson 1984). This highlights the
preference to have the code for the director be “correct by
construction”, preferably automatically generated.
In work such as this we invariable come across the observer effect (Mytkowicz et al. 2008) that monitoring code
has on the monitored code. This is known to affect the visibility of timing bugs in software (Neville-Neil 2014), but
even “invisible” such bugs can leak important information
to an adversary (Cock et al. 2014).
The overhead due to monitoring is a very important consideration when evaluating system measurements, and monitoring systems try to do their utmost to reduce it (Gregg and
Mauro 2011; Anderson et al. 2014). A key weakness of our
prototype is that it does not use low-level (RTL-level) techniques to reduce overhead, but we mitigate this by allowing
extension points to be specialised, e.g., these may only be
a breakpoint, a watchpoint, or both, etc. This simplifies the
circuitry we get. A more sophisticated approach would in-
volve organising the director to operate in a different clock
domain if possible, following McKechnie et al. (2009). Not
having to sustain a large clock-distribution network leads to
power saving, since unused logic blocks don’t need to be
switched, and less heat is dissipated from leakage. It also
makes placement and routing easier, which usually reduces
the compilation time.
The main strength of our approach is that it provides a
uniform interface for the flexible directing of software and
hardware instances of programs at runtime. To our knowledge no other system provides this. We used an HLS system
that allows us to run the resulting code both on software and
on hardware, but the debugging methods used for either were
hitherto separate.
P. Avgustinov, J. Tibble, and O. de Moor. Making Trace Monitors
Feasible. SIGPLAN Not., 42(10):589–608, Oct. 2007. ISSN
0362-1340. doi: 10.1145/1297105.1297070. URL http://doi.
acm.org/10.1145/1297105.1297070.
Acknowledgments
R. G. Bennetts and A. Osseyran. IEEE standard 1149.1-1990
on boundary scan: History, literature survey, and current status.
Journal of Electronic Testing, 2(1):11–25, 1991. ISSN 15730727. doi: 10.1007/BF00134941. URL http://dx.doi.org/
10.1007/BF00134941.
We thank the many people who contributed to this paper.
Matthew Grosvenor helped us with evaluation ideas and
reusing the QJump infrastructure. Olaf Chitil, Paolo Costa,
Klaus Gleissenthall, Tim Harris, Simon Moore, and Robert
Soulé helped improve the paper through their feedback.
This work has received funding from the EPSRC NaaS
grant EP/K034723/1, European Union’s Horizon 2020 research and innovation programme 2014-2018 under the SSICLOPS (grant agreement No. 644866), and the Leverhulme
Trust Early Career Fellowship ECF-2016-289.
References
M. Abadi, M. Budiu, U. Erlingsson, and J. Ligatti. Control-flow
Integrity. In Proceedings of the 12th ACM Conference on Computer and Communications Security, CCS ’05, pages 340–353,
New York, NY, USA, 2005. ACM. ISBN 1-59593-226-7. doi:
10.1145/1102120.1102165. URL http://doi.acm.org/10.
1145/1102120.1102165.
T. Ball, E. Bounimova, B. Cook, V. Levin, J. Lichtenberg, C. McGarvey, B. Ondrusek, S. K. Rajamani, and A. Ustuner. Thorough Static Analysis of Device Drivers.
SIGOPS Oper.
Syst. Rev., 40(4):73–85, Apr. 2006. ISSN 0163-5980. doi:
10.1145/1218063.1217943. URL http://doi.acm.org/10.
1145/1218063.1217943.
R. M. Balzer. EXDAMS: EXtendable Debugging and Monitoring System. In Proceedings of the May 14-16, 1969, Spring
Joint Computer Conference, AFIPS ’69 (Spring), pages 567–
580, New York, NY, USA, 1969. ACM. doi: 10.1145/1476793.
1476881.
URL http://doi.acm.org/10.1145/1476793.
1476881.
N. Calagar, S. D. Brown, and J. H. Anderson. Source-level debugging for FPGA high-level synthesis. In 2014 24th International
Conference on Field Programmable Logic and Applications
(FPL), pages 1–8, Sept 2014. doi: 10.1109/FPL.2014.6927496.
K. Camera, H. K.-H. So, and R. W. Brodersen. An Integrated Debugging Environment for Reprogrammble Hardware Systems.
In Proceedings of the Sixth International Symposium on Automated Analysis-driven Debugging, AADEBUG’05, pages 111–
116, New York, NY, USA, 2005. ACM. ISBN 1-59593-050-7.
doi: 10.1145/1085130.1085145. URL http://doi.acm.org/
10.1145/1085130.1085145.
T. V. Chu, S. Sato, and K. Kise. Ultra-fast NoC emulation on a
single FPGA. In 2015 25th International Conference on Field
Programmable Logic and Applications (FPL), pages 1–8, Sept
2015. doi: 10.1109/FPL.2015.7294021.
J. Anderson, R. N. M. Watson, D. Chisnall, K. Gudka, I. Marinos,
and B. Davis. TESLA: Temporally Enhanced System Logic
Assertions. In Proceedings of the Ninth European Conference
on Computer Systems, EuroSys ’14, pages 19:1–19:14, New
York, NY, USA, 2014. ACM. ISBN 978-1-4503-2704-6. doi:
10.1145/2592798.2592801. URL http://doi.acm.org/10.
1145/2592798.2592801.
D. Cock, Q. Ge, T. Murray, and G. Heiser. The Last Mile:
An Empirical Study of Timing Channels on seL4. In Proceedings of the 2014 ACM SIGSAC Conference on Computer
and Communications Security, CCS ’14, pages 570–581, New
York, NY, USA, 2014. ACM. ISBN 978-1-4503-2957-6. doi:
10.1145/2660267.2660294. URL http://doi.acm.org/10.
1145/2660267.2660294.
G. Antichi, M. Shahbaz, Y. Geng, N. Zilberman, A. Covington,
M. Bruyere, N. McKeown, N. Feamster, B. Felderman, M. Blott,
et al. OSNT: Open source network tester. IEEE Network, 28(5):
6–12, 2014.
J. Curreri, G. Stitt, and A. D. George. High-level Synthesis of
In-circuit Assertions for Verification, Debugging, and Timing
Analysis. Int. J. Reconfig. Comput., 2011:1:1–1:17, Jan. 2011.
ISSN 1687-7195. doi: 10.1155/2011/406857. URL http:
//dx.doi.org/10.1155/2011/406857.
P. Avgustinov, E. Bodden, E. Hajiyev, L. Hendren, O. Lhoták,
O. de Moor, N. Ongkingco, D. Sereni, G. Sittampalam, J. Tibble, and M. Verbaere. Aspects for Trace Monitoring. In
Proceedings of the First Combined International Conference
on Formal Approaches to Software Testing and Runtime Verification, FATES’06/RV’06, pages 20–39, Berlin, Heidelberg,
2006. Springer-Verlag. ISBN 3-540-49699-8, 978-3-540-496991. doi: 10.1007/11940197_2. URL http://dx.doi.org/10.
1007/11940197_2.
F. Q. B. da Silva. Correctness Proofs of Compilers and Debuggers:
an Approach Based on Structural Operational Semantics. PhD
thesis, School of Informatics, University of Edinburgh, Oct.
1992.
D. S. Dantas and D. Walker. Harmless Advice. In Conference Record of the 33rd ACM SIGPLAN-SIGACT Symposium on
Principles of Programming Languages, POPL ’06, pages 383–
396, New York, NY, USA, 2006. ACM. ISBN 1-59593-027-2.
doi: 10.1145/1111037.1111071. URL http://doi.acm.org/
10.1145/1111037.1111071.
S. Djoko Djoko, R. Douence, and P. Fradet. Aspects Preserving
Properties. Sci. Comput. Program., 77(3):393–422, Mar. 2012.
ISSN 0167-6423. doi: 10.1016/j.scico.2011.10.010. URL http:
//dx.doi.org/10.1016/j.scico.2011.10.010.
M. Faddegon and O. Chitil. Algorithmic Debugging of Real-world
Haskell Programs: Deriving Dependencies from the Cost Centre
Stack. In Proceedings of the 36th ACM SIGPLAN Conference on
Programming Language Design and Implementation, PLDI ’15,
pages 33–42, New York, NY, USA, 2015. ACM. ISBN 9781-4503-3468-6. doi: 10.1145/2737924.2737985. URL http:
//doi.acm.org/10.1145/2737924.2737985.
B. Fitzpatrick. Distributed caching with memcached. Linux Journal, 2004(124), 2004.
L. Fix. Fifteen Years of Formal Property Verification in Intel, pages
139–144. Springer Berlin Heidelberg, Berlin, Heidelberg, 2008.
ISBN 978-3-540-69850-0. doi: 10.1007/978-3-540-69850-0_8.
URL http://dx.doi.org/10.1007/978-3-540-69850-0_8.
H. Foster. Assertion-Based Verification: Industry Myths to Realities, pages 5–10. Springer Berlin Heidelberg, Berlin, Heidelberg, 2008. ISBN 978-3-540-70545-1. doi: 10.1007/
978-3-540-70545-1_3. URL http://dx.doi.org/10.1007/
978-3-540-70545-1_3.
M. N. Gagnon, S. Taylor, and A. K. Ghosh. Software protection
through anti-debugging. IEEE Security Privacy, 5(3):82–84,
May 2007. ISSN 1540-7993. doi: 10.1109/MSP.2007.71.
P. Godefroid, M. Y. Levin, and D. Molnar. SAGE: Whitebox
Fuzzing for Security Testing. Commun. ACM, 55(3):40–44, Mar.
2012. ISSN 0001-0782. doi: 10.1145/2093548.2093564. URL
http://doi.acm.org/10.1145/2093548.2093564.
J. Goeders and S. J. E. Wilton. Effective FPGA Debug for HighLevel Synthesis Generated Circuits. In 2014 24th International
Conference on Field Programmable Logic and Applications
(FPL), pages 1–8, Sept 2014. doi: 10.1109/FPL.2014.6927498.
J. Goeders and S. J. E. Wilton. Quantifying observability for insystem debug of high-level synthesis circuits. In 2016 26th
International Conference on Field Programmable Logic and
Applications (FPL), pages 1–11, Aug 2016. doi: 10.1109/FPL.
2016.7577371.
M. Golan and D. R. Hanson. Duel – A Very High-Level Debugging Language. In USENIX Winter Conference, pages 107–117,
1993.
B. Gregg and J. Mauro. DTrace: Dynamic Tracing in Oracle
Solaris, Mac OS X, and FreeBSD. Prentice Hall Professional,
2011.
K. W. Hamlen and M. Jones. Aspect-oriented In-lined Reference
Monitors. In Proceedings of the Third ACM SIGPLAN Workshop
on Programming Languages and Analysis for Security, PLAS
’08, pages 11–20, New York, NY, USA, 2008. ACM. ISBN 9781-59593-936-4. doi: 10.1145/1375696.1375699. URL http:
//doi.acm.org/10.1145/1375696.1375699.
D. R. Hanson and M. Raghavachari. A machine-independent debugger. Softw., Pract. Exper., 26(11):1277–1299, 1996.
R. Hood. The P2D2 Project: Building a Portable Distributed Debugger. In Proceedings of the SIGMETRICS Symposium on Par-
allel and Distributed Tools, SPDT ’96, pages 127–136, New
York, NY, USA, 1996. ACM. ISBN 0-89791-846-0. doi: 10.
1145/238020.238058. URL http://doi.acm.org/10.1145/
238020.238058.
E. Hung and S. J. Wilton. Towards Simulator-like Observability
for FPGAs: A Virtual Overlay Network for Trace-buffers. In
Proceedings of the ACM/SIGDA International Symposium on
Field Programmable Gate Arrays, FPGA ’13, pages 19–28, New
York, NY, USA, 2013. ACM. ISBN 978-1-4503-1887-7. doi:
10.1145/2435264.2435272. URL http://doi.acm.org/10.
1145/2435264.2435272.
E. Hung and S. J. E. Wilton. Accelerating FPGA Debug: Increasing Visibility Using a Runtime Reconfigurable Observation and
Triggering Network. ACM Trans. Des. Autom. Electron. Syst.,
19(2):14:1–14:23, Mar. 2014. ISSN 1084-4309. doi: 10.1145/
2566668. URL http://doi.acm.org/10.1145/2566668.
M. S. Johnson. The Design of a High-level, Language-independent
Symbolic Debugging System. In Proceedings of the 1977
Annual Conference, ACM ’77, pages 315–322, New York,
NY, USA, 1977. ACM. ISBN 978-1-4503-3921-6. doi: 10.
1145/800179.810221. URL http://doi.acm.org/10.1145/
800179.810221.
A. Kishon, P. Hudak, and C. Consel. Monitoring Semantics:
A Formal Framework for Specifying, Implementing, and Reasoning About Execution Monitors. In Proceedings of the
ACM SIGPLAN 1991 Conference on Programming Language
Design and Implementation, PLDI ’91, pages 338–352, New
York, NY, USA, 1991. ACM. ISBN 0-89791-428-7. doi: 10.
1145/113445.113474. URL http://doi.acm.org/10.1145/
113445.113474.
G. H. Koch, W. Rosenstiel, and U. Kebschull. Breakpoints and
Breakpoint Detection in Source-level Emulation. ACM Trans.
Des. Autom. Electron. Syst., 3(2):209–230, Apr. 1998. ISSN
1084-4309. doi: 10.1145/290833.290843. URL http://doi.
acm.org/10.1145/290833.290843.
S. Krishnaswamy, I. L. Markov, and J. P. Hayes. On the Role of
Timing Masking in Reliable Logic Circuit Design. In Proceedings of the 45th Annual Design Automation Conference, DAC
’08, pages 924–929, New York, NY, USA, 2008. ACM. ISBN
978-1-60558-115-6. doi: 10.1145/1391469.1391703. URL
http://doi.acm.org/10.1145/1391469.1391703.
J. C. Lee and R. Lysecky. System-Level Observation Framework
for Non-Intrusive Runtime Monitoring of Embedded Systems.
ACM Trans. Des. Autom. Electron. Syst., 20(3):42:1–42:27, June
2015. ISSN 1084-4309. doi: 10.1145/2717310. URL http:
//doi.acm.org/10.1145/2717310.
P. E. McKechnie, M. Blott, and W. A. Vanderbauwhede. Debugging FPGA-based Packet Processing Systems Through
Transaction-level Communication-centric Monitoring. SIGPLAN Not., 44(7):129–136, June 2009. ISSN 0362-1340. doi:
10.1145/1543136.1542470. URL http://doi.acm.org/10.
1145/1543136.1542470.
R. Milner, M. Tofte, and R. Harper. The Definition of Standard ML.
MIT Press, 1990.
S. Mittal. A survey of techniques for improving energy efficiency
in embedded computing systems. International Journal of Computer Aided Engineering and Technology, 6(4):440–459, 2014.
P. Mockapetris. Domain names – concepts and facilities, Nov.
1987. https://tools.ietf.org/html/rfc1034.
York, NY, USA, 2014. ACM. ISBN 978-1-4503-2671-1. doi:
10.1145/2554688.2554767. URL http://doi.acm.org/10.
1145/2554688.2554767.
J. S. Monson and B. L. Hutchings. Using Source-Level Transformations to Improve High-Level Synthesis Debug and Validation
on FPGAs. In Proceedings of the 2015 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays, FPGA
’15, pages 5–8, New York, NY, USA, 2015. ACM. ISBN 9781-4503-3315-3. doi: 10.1145/2684746.2689087. URL http:
//doi.acm.org/10.1145/2684746.2689087.
M. Schäfer, T. Ekman, and O. de Moor. Challenge Proposal:
Verification of Refactorings. In Proceedings of the 3rd Workshop
on Programming Languages Meets Program Verification, PLPV
’09, pages 67–72, New York, NY, USA, 2008. ACM. ISBN 9781-60558-330-3. doi: 10.1145/1481848.1481859. URL http:
//doi.acm.org/10.1145/1481848.1481859.
T. Mytkowicz, P. F. Sweeney, M. Hauswirth, and A. Diwan. Observer Effect and Measurement Bias in Performance Analysis.
Technical Report 1042-08, University of Colorado at Boulder, 6
2008.
J. C. Shepherdson and H. E. Sturgis. Computability of Recursive
Functions. J. ACM, 10(2):217–255, Apr. 1963. ISSN 0004-5411.
doi: 10.1145/321160.321170. URL http://doi.acm.org/10.
1145/321160.321170.
G. V. Neville-Neil. Outsourcing Responsibility. Commun. ACM,
57(10):28–29, Sept. 2014. ISSN 0001-0782. doi: 10.1145/
2661051. URL http://doi.acm.org/10.1145/2661051.
S. Singh and D. J. Greaves. Kiwi: Synthesis of FPGA Circuits from
Parallel Programs. In Field-Programmable Custom Computing
Machines, pages 3–12, 2008.
Z. Panjkov, A. Wasserbauer, T. Ostermann, and R. Hagelauer. Hybrid FPGA debug approach. In 2015 25th International Conference on Field Programmable Logic and Applications (FPL),
pages 1–8, Sept 2015. doi: 10.1109/FPL.2015.7294023.
C. H. Smith and J. van Leeuwen. Microprogrammed Random
Access Stored Program Machines. SIGACT News, 6(3):23–32,
July 1974. ISSN 0163-5700. doi: 10.1145/1008311.1008315.
URL http://doi.acm.org/10.1145/1008311.1008315.
D. Pauli. ‘Pork Explosion’ flaw splatters Foxconn’s Android
phones. Oct 2016. http://www.theregister.co.uk/2016/
10/14/pork_explosion_foxconn_flaw/.
B. So, M. W. Hall, and P. C. Diniz. A Compiler Approach to
Fast Hardware Design Space Exploration in FPGA-based Systems. In Proceedings of the ACM SIGPLAN 2002 Conference
on Programming Language Design and Implementation, PLDI
’02, pages 165–176, New York, NY, USA, 2002. ACM. ISBN
1-58113-463-0. doi: 10.1145/512529.512550. URL http:
//doi.acm.org/10.1145/512529.512550.
M. Potkonjak, S. Dey, and K. Wakabayashi. Design-for-debugging
of Application Specific Designs. In Proceedings of the 1995
IEEE/ACM International Conference on Computer-aided Design, ICCAD ’95, pages 295–301, Washington, DC, USA, 1995.
IEEE Computer Society. ISBN 0-8186-7213-7. URL http:
//dl.acm.org/citation.cfm?id=224841.225054.
R. Sosič and D. Abramson. Guard: A relative debugger. Software:
Practice and Experience, 27(2):185–206, 1997.
A. Putnam. Large-Scale Reconfigurable Computing in a Microsoft
Datacenter. In Proceedings of the 26th IEEE HotChips Symposium on High-Performance Chips (HotChips 2014). IEEE, August 2014.
R. Sosič. Dynascope: A Tool for Program Directing. SIGPLAN
Not., 27(7):12–21, July 1992. ISSN 0362-1340. doi: 10.
1145/143103.143110. URL http://doi.acm.org/10.1145/
143103.143110.
A. Putnam, A. M. Caulfield, E. S. Chung, D. Chiou, K. Constantinides, J. Demme, H. Esmaeilzadeh, J. Fowers, G. P. Gopal,
J. Gray, M. Haselman, S. Hauck, S. Heil, A. Hormati, J. Y. Kim,
S. Lanka, J. Larus, E. Peterson, S. Pope, A. Smith, J. Thong,
P. Y. Xiao, and D. Burger. A reconfigurable fabric for accelerating large-scale datacenter services. In 2014 ACM/IEEE 41st International Symposium on Computer Architecture (ISCA), pages
13–24, June 2014. doi: 10.1109/ISCA.2014.6853195.
R. Sosič. A Procedural Interface for Program Directing. Software:
Practice and Experience, 25(7):767–787, 1995. ISSN 1097024X. doi: 10.1002/spe.4380250704. URL http://dx.doi.
org/10.1002/spe.4380250704.
N. Ramsey and D. R. Hanson. A Retargetable Debugger. In Proceedings of the ACM SIGPLAN 1992 Conference on Programming Language Design and Implementation, PLDI ’92, pages
22–31, New York, NY, USA, 1992. ACM. ISBN 0-89791-4759. doi: 10.1145/143095.143112. URL http://doi.acm.org/
10.1145/143095.143112.
T. Todman, S. Stilkerich, and W. Luk. In-circuit Temporal Monitors
for Runtime Verification of Reconfigurable Designs. In Proceedings of the 52Nd Annual Design Automation Conference, DAC
’15, pages 50:1–50:6, New York, NY, USA, 2015. ACM. ISBN
978-1-4503-3520-1. doi: 10.1145/2744769.2744856. URL
http://doi.acm.org/10.1145/2744769.2744856.
P. M. Sansom and S. L. Peyton Jones. Formally Based Profiling for Higher-order Functional Languages. ACM Trans. Program. Lang. Syst., 19(2):334–385, Mar. 1997. ISSN 0164-0925.
doi: 10.1145/244795.244802. URL http://doi.acm.org/10.
1145/244795.244802.
J. L. Tripp, P. A. Jackson, and B. L. Hutchings. Sea Cucumber:
A Synthesizing Compiler for FPGAs, pages 875–885. Springer
Berlin Heidelberg, Berlin, Heidelberg, 2002. ISBN 978-3-54046117-3. doi: 10.1007/3-540-46117-5_90. URL http://dx.
doi.org/10.1007/3-540-46117-5_90.
A. Sari, D. Agiakatsikas, and M. Psarakis. A Soft Error Vulnerability Analysis Framework for Xilinx FPGAs. In Proceedings
of the 2014 ACM/SIGDA International Symposium on Fieldprogrammable Gate Arrays, FPGA ’14, pages 237–240, New
D. Wang, N. E. Jerger, and J. G. Steffan. DART: A Programmable Architecture for NoC Simulation on FPGAs. In
Proceedings of the Fifth ACM/IEEE International Symposium
on Networks-on-Chip, NOCS ’11, pages 145–152, New York,
K. Thompson. Reflections on Trusting Trust. Commun. ACM,
27(8):761–763, Aug. 1984.
ISSN 0001-0782.
doi: 10.
1145/358198.358210. URL http://doi.acm.org/10.1145/
358198.358210.
NY, USA, 2011. ACM. ISBN 978-1-4503-0720-8. doi:
10.1145/1999946.1999970. URL http://doi.acm.org/10.
1145/1999946.1999970.
P. Winterbottom. ACID: A Debugger Built From A Language. In
USENIX Winter Conference, pages 211–222, 1994.
M.-Y. Zhu. Formal Specifications of Debuggers. SIGPLAN Not.,
36(9):54–63, Sept. 2001. doi: 10.1145/609769.609778. URL
http://doi.acm.org/10.1145/609769.609778.
N. Zilberman, Y. Audzevich, G. A. Covington, and A. W. Moore.
NetFPGA SUME: Toward 100 Gbps as Research Commodity.
IEEE Micro, 34(5):32–41, 2014.
A.
More program direction commands
Program-level predicates and functions mentioned in these
definitions are given in §B.
A.1
Tracing
A.2
D̆ = {}
C˘ = {}
cur _idx ;
Dc = λD0 . Xi
for i = cur _idx to 0:
Xa [i]
N;
print(N );
D0
Watching
This bears some similarity to the rule to start tracing (in §3.5.2)
since we inspect the variable after it’s updated, and carry out
some action. When tracing, this action consists of writing to
the trace buffer. When watching, it consists of breaking.
X ∈ Varp
∀L ∈ XL . L 6∈ p
Positionsp0 (XL ) = PostUpdatep0 (X)
p <XL p0
We continue from §3.5.2, where the symbols we use here are
introduced (such as Xi and Xof ).
(trace start X hBi h$i) ∈ C
p
trace clear X
@C
p
D̆ = {}
˘
C = {}
0
Dc = λD . Jtrace clear XK
D0
0;
p
watch X hBi
@C
p0
D̆ =
{(«w», X, 1)}
SP [L 7→ Jwatch X hBiKSP ]
˘
C=
for each L ∈ XL
Dc = λD0 . if («w», X, 1) ∈ D0 then D0
else
for each L ∈ XL :
@L : {Jwatch X hBiKSP }
(«w», X, 0 7→ 1) :∈ D0
pLq;
where
where
Jtrace clear XK =
Xi := 0;
Xof := 0
Jwatch X hBiKSP = conditional hBi break
Next we add the command to check whether the trace
buffer is full.
Unwatching a variable follows a pattern we’ve encountered before, when stopping tracing for instance.
(watch X hBi) ∈ C
Positionsp (XL ) = PostUpdatep (X)
(trace start X hBi h$i) ∈ C
p
trace full X
@C
D̆ = {}
C˘ = {}
Dc = λD0 . Jtrace full XK
print(N );
D0
p
p
N;
where
Jtrace full XK = Xof
(trace start X hBi h$i) ∈ C
trace print X
@C
p
D̆ = {}
C˘ = {}
Dc = λD0 . if («w», X, 0) ∈ D0 then D0
else
for each L ∈ XL :
@L : {Junwatch XKSP }
(«w», X, 1 7→ 0) :∈ D0
pLq;
where
Finally the command to retrieve the contents of the trace
buffer. Here we first find out the current index value for
the trace buffer, then we work back and extract the buffer’s
contents.
p
unwatch X
@C
p
Junwatch XKSP = continue
A.3
Profiling
We could profile programs to count the number of writes
and reads of variables, or function calls. We show the rule
for counting the number of writes; counting variable reads
and function calls are similar, differing in the positioning of
labels. That is, to count writes we position labels immediately after each line that updates that variable–this is similar
to tracing (§3.5.2) and watching (§A.2). To count variable
reads we position after lines in which those variables appear.
To count function calls we position just before the first line
in the function.
X ∈ Varp
∀L ∈ XL . L 6∈ p
Positionsp0 (XL ) = PostUpdatep0 (X)
p <XL p0
p
count write start X hBi h$i
@C
p0
D̆ =
{(«c», X, 1)}
C [Xcount ] = 0, C [Xof ] = 0, A[Xa [h$i]],
˘=
SP
[L
→
7
Jcount
write
start
X
hBi
h$iK
]
C
SP
for each L ∈ XL
Dc = λD0 . if («c», X, 1) ∈ D0 then D0
else
for each L ∈ XL :
@L : {Jcount write start X hBi h$iKSP }
pLq;
(«c», X, 0 7→ 1) :∈ D0
where
Jcount write start X hBi h$iKSP =
conditional
hBi
if Xcount < h$i then
inc Xcount ;
continue
else
inc Xof ;
break
Stopping the count, clearing it, printing the current count,
and checking whether the maximum value has been reached,
is done in a very similar way to the rules for tracing (§3.5.2).
B.
Program analyses and transformations
A position of statement s in program p is a vector π ∈ Nd
where 1 < d < ω. Intuitively, the first position of the vector
refers to the function in which the statement is positioned.
The second until the last positions indicate statements (and
substatements, in the case of if-then statements) in which
the statement is positioned, moving from the outermost to
the innermost containing statement. Within the innermost
containing statement, if the statement number is i, then the
last component of the vector is either i+0 or i+1, depending
on whether it points to just before or just after the statement.
Definition B.1 (Positions). Let
p = vdecl fdecl return fname(ē)
Then
def
Πp =
[
0≤i≤n
{i} × Πfdecl i
where n is |fdecl |, the number of function declarations in p.
Let fdecl = τ fname(x̄){s; return e} and s = s0 ; . . . ; sn .
Then
[
def
Πfdecl =
({i} × Πsi ) ∪ {n + 1}
0≤i≤n
Finally,
S
0≤i≤n ({i} × Πri ) ∪ {n + 1}
def
Πs =
if s = if e then r0 ; . . . ; rn
∅ otherwise
Next we define a function Posp maps a position π ∈
Πp in p to the corresponding statement s, in symbols:
Posp (π) = s.
Definition B.2 (Statement at a position). Let
p = vdecl fdecl return fname(ē)
Then we define
def
Posp ((i, π)) = FPosfdecl i (π)
Let fdecl = τ fname(x̄){s; return e} and s = s0 ; . . . ; sn .
Then
def
FPosfdecl ((i, π)) = SPossi (π)
Finally,
0
0
SPosri (π ) if π = (i, π )
def
SPoss (π) =
and s = if e then r0 ; . . . ; rn
s
if π = ∅
Definition B.3 (Program p contains label L). We write L ∈
p iff ∃π, S. Posp (π) = extend(S) ∧ L ∈ S.
Note that a label L may only be used at most once in a
program. We expect the following to hold for all programs.
We do not include this as a premise to any of the rules, to
reduce clutter.
∀π1 , π2 , S1 , S2 . (Posp (π1 ) = extend(S1 ) ∧ L ∈ S1 )
∧ (Posp (π2 ) = extend(S2 ) ∧ L ∈ S2 )
−→ (S1 = S2 ∧ π1 = π2 )
We now formalise the predicate that programs p and p0 are
identical save for the addition of L in one of the extension
points.
Definition B.4 (Program identity modulo label L).
def
p <1L p0 = ∀π, s, S.
(Posp (π) = s −→
((s 6= extend(S) −→ Posp0 (π) = s)
∧ (s = extend(S) −→
Posp0 (π) = extend(S 0 ) ∧ (S 0 = S ∨ S 0 \S ⊆ {L}))))
∧ (Posp0 (π) = s −→
((s 6= extend(S) −→ Posp (π) = s)
∧ (s = extend(S 0 ) −→
Posp (π) = extend(S) ∧ (S 0 = S ∨ S 0 \S ⊆ {L}))))
We define a related predicate that formalises whether
two programs are related by the above predicate through an
‘interpolation’ of labels drawn from a set.
Definition B.5 (Program identity modulo set of labels). Let
S = {L0 , . . . , Ln } is a set of labels. Then,
def
p <S p0 = ∃p1 , . . . , pn−1 . p <1L0 p1 <1L1 p2 <1L2 . . . <1Ln p0
Definition B.6 (Set of variables in program p). Let p =
vdecl fdecl return fname(ē) and vdecl = vdecl 0 . . . vdecl n .
Then
[
def
Varp =
{Xi | vdecl i = τi Xi }
0≤i≤n
Definition B.7 (Set of variables in program p).
def
PostUpdatep (X) = {π + 1 | ∃E. Posp (π) = (X := E)}
where if π = (m0 , . . . , mn ) then π+1 = (m0 , . . . , mn +1).
The positions in which a label L occurs. This should be a
singleton set.
Definition B.8 (Positions of label L).
def
Positionsp (L) = {π | ∃S. Posp (π) = extend(S) ∧ L ∈ S}
We lift this to work on sets of labels and overload the notation: if SL is a set of labels then
def [
Positionsp (SL ) =
(Positionsp (L))
L∈SL
C.
Language semantics
C.1
CASP machines
The specification of CASP machines was given in §3.2.
The machine’s configuration consists of the triple (S, ia, P ),
where ia ∈ {◦, •} indicates whether the machine is operating in batch or interactive modes respectively, and S =
(C, A, SP ) indicates the machine’s memory: C ∈ (X * Z)
are the counters, A ∈ (R * Z * Z) the arrays, and
SP ∈ (L * P ) the stored procedures.
The machine’s big-step operational semantics are described using the notation L ` (S, ia, P ) =⇒ (S 0 , ia 0 , N )
to indicate that the machine operates in the context of a specific label L to evaluate program P into numeral N , and
possibly updating the other components of its configuration. We use the notation L ` (S, ia, P ) ⇓ N to abbreviate
L ` (S, ia, P ) =⇒ (S, ia, N ) when S and ia are unaffected
by the evaluation.
The semantic rules are given in Figure 6. We use the
notation pLq to denote a total injective map from labels
to numerals. Thus we can identify which label is being
executed. If in interactive mode the director sends “break”
to the controller, the director learns the label that broke.
Note that the “continue” and “break” commands change
the state of the ia component of the configuration, switching
it to batch (resuming the program) and interactive respectively.
Note also that placement (@L : {P }) is only allowed
during interactive mode (i.e., ia = •) since we do not
want the controller to normally run code that updates other
extension points’ code.
C.2
Example language
The syntax for the example language was given in §3.4.1.
In this language we have a valuation function for identifiers R ∈ (X * Z), and a configuration for the reduction semantics consists of a triple hR, c̄, c̄i where c is either
a statement, a ‘return’ statement, or an expression: c ::=
s | return e | e. c̄ denotes a sequence of 1 or more such
c: c0 , . . . , cn where n > 0. We define the concatenation operator : such that if d¯ = d0 , . . . , dm where m > 0, then
c̄ : d¯ = c0 , . . . , cn , d0 , . . . , dm .
Intuitively, hR, c̄1 , c̄2 i describes a configuration where c̄1
is to be evaluated “now”, and c̄2 is to be executed after c̄1 has
been evaluated. The evaluation rules are given in Figure 7.
To reduce clutter in the rules, we assume two pieces implicit state, that we avoid threading around the configurations. The first is a mapping F from function names to their
bodies, and the second is the state of the CASP machine S,
which includes the valuation R that models the store for program variables.
We use the following evaluation context for syntactic objects in c (statements, ‘return’ statements, and expressions):
E
::=
|
|
|
|
|
[]
X := E
if E then s; s
fname(N0 , . . . , Nn , E, e0 . . . , em )
E op e
N op E
op ∈ {+, −, ==, <}
We handle ‘extend’ by expanding it to a sequence of
singleton extends, and invoking the rule in Figure 8.
Rules for y model the interactive mode between the director and the controller. In these semantics we only see
evaluation from the point of view of the program: the program’s evaluation −→ yields to the controller =⇒ at extension points, which either yields back to the program (in case
of continue) or else switches to interactive mode (in case
of break), leading to an interleaving between y (obtaining
a command from the director) and =⇒ (executing it). Note
that the rules for y are side-effecting, and are formalised as
such similar to streams as used by Milner et al. (1990). In
the rules for y, P is a command obtained from the director,
and N is a result sent back to the director.
L ` (S, ia, continue) =⇒ (S, ◦, pLq)
L ` (S, ia, V ) ⇓ N
L ` (S, ia, −V ) ⇓ −N
L ` (S, ia, break) =⇒ (S, •, pLq)
L ` (S, ia, V1 ) ⇓ N1
L ` (S, ia, V2 ) ⇓ N2
L ` (S, ia, V1 op V2 ) =⇒ (S, ia, N )
L ` (S, ia, U ) ⇓ N
L ` (S, ia, op U ) =⇒ (S[U 7→ M ], ia, M )
(
1
N=
−1
(op = (=) ∧ N1 = N2 ) ∨ (op = (<) ∧ N1 < N2 )
o/w
(
N +1
M=
N −1
L ` (S, ia, E) ⇓ N
L ` (S, ia, U := E) =⇒ (S[U 7→ N ], ia, N )
S(C, X) = N
L ` (S, ia, X) ⇓ N
op = inc
op = dec
L ` (S, ia, I) ⇓ N
S(R[A], N ) = M
L ` (S, ia, A[I]) ⇓ M
L ` (S, ia, E) ⇓ N
L ` (S, ia, P ) =⇒ (S 0 , ia 0 , M )
L ` (S, ia, if E then P1 else P2 ) =⇒ (S 0 , ia 0 , M )
(
P =
P1
P2
N =1
N = −1
L ` (S, ia, P1 ) =⇒ (S 0 , ia 0 , N )
ia 6= ia 0
L ` (S, ia, P1 ; P2 ) =⇒ (S 0 , ia 0 , N )
L ` (S, ia, P1 ) =⇒ (S 0 , ia 0 , M )
ia = ia 0
L ` (S 0 , ia 0 , P2 ) =⇒ (S 00 , ia 00 , N )
L ` (S, ia, P1 ; P2 ) =⇒ (S 00 , ia 00 , N )
L ` (S, •, @L0 : {P }) =⇒ (S[SP [L0 ] 7→ P ], •, pL0 q)
L ` (S, ia, N ) ⇓ N
Figure 6. Dynamic semantics for CASP instructions (§3.2)
hR, skip, s; resti −→ hR, s, resti
hR, N, s; resti −→ hR, s, resti
hR, X := N, resti −→ hR[X 7→ N ], skip, resti
N ≤0
hR, if N then s̄, resti −→ hR, skip, resti
hR, s; r̄, resti −→ hR, s, r̄ : resti
hR, X, resti −→ hR, R(X), resti
N >0
hR, if N then s̄, resti −→ hR, s̄, resti
hR, e, skipi −→ hR0 , N, skipi
hR, E[e], resti −→ hR0 , E[N ], resti
N1 + N2
N1 − N2
1
hR, N1 op N2 , skipi −→ hR, M, skipi M =
0
1
0
op
op
op
op
op
op
=+
=−
= (==) ∧ N1 = N2
= (==) ∧ N1 6= N2
= (<) ∧ N1 < N2
= (<) ∧ N1 ≥ N2
(fname, s̄; return e) ∈ F
hR, s̄; return e, skipi −→ hR0 , return N, skipi
hR, fname(N̄ , resti −→ hR0 , N, resti
Figure 7. Dynamic semantics for the example language (§3.4.1)
L ` (S, ◦, SP [L]) =⇒ (S 0 , ◦, N )
hR, extend{L}, s; resti −→ hR0 , skip, resti
L ` (S, •, P ) =⇒ (S 0 , •, N )
L ` (S 0 , •) y (S 00 , ◦)
L ` (S, •) y (S 00 , ◦)
L ` (S, •) y (S 0 , ◦)
L ` (S, ◦, break) =⇒ (S, ◦, continue)
L ` (S, •, P ) =⇒ (S 0 , ◦, N )
L ` (S, •) y (S 0 , ◦)
Figure 8. Linking of the programming language evaluation with that of CASP controller
| 6cs.PL
|
Pseudorehearsal in actor-critic agents
Marochko Vladimir∗ , Leonard Johard†,Manuel Mazzara‡
arXiv:1704.04912v1 [cs.AI] 17 Apr 2017
Innopolis University Email: ∗ [email protected], † [email protected], ‡ [email protected]
Abstract—Catastrophic forgetting has a serious impact in
reinforcement learning, as the data distribution is generally
sparse and non-stationary over time. The purpose of this study
is to investigate whether pseudorehearsal can increase performance of an actor-critic agent with neural-network based policy
selection and function approximation in a pole balancing task
and compare different pseudorehearsal approaches. We expect
that pseudorehearsal assists learning even in such very simple
problems, given proper initialization of the rehearsal parameters.
I. I NTRODUCTION
Reinforcement learning is a promising and growing area
of machine learning based on training agents on positive
or negative feedback to their actions. The agent gets an
observation of the environment, chooses which action to
perform, receives a reward and modifies its way of choosing
future actions according to this reward. If the agent reached
a state that is clearly better then the previous one in terms
of the given task or successfully completed the task, then the
reward is positive. If the agent reached a state clearly worse
or failed the task, then the reward is negative.
Reinforcement learning algorithms can be divided into two
different classes based on their ways to choose long-term
optimal actions. One is based on generation of a model of
the environment and evaluation of the states in this model
to predict future reward. This is the class of model-based
algorithms. The other - model-free algorithms - are based
on trial-and-error learning based on habitual and conditioned
responses tied to certain stimuli. We will focus on modelbased algorithms.
The best situation is if the environment is fully observable
and the number of states defined by agent is finite and
not too big. In this case agent can just keep all possible
states in the memory and evaluate each state directly. Of
course, in real life this situation is very rare, so different
kinds of approximation are used. One of the possible model
approximations widely used is neural networks, which
constitutes a powerful tool for implementation of memory.
Unfortunately neural networks are vulnerable to a problem
known as catastrophic forgetting. Within reinforcement
learning influence of catastrophic forgetting is usually even
more serious then in supervised learning and pseudorehearsal
is one of possible ways of catastrophic forgetting elimination.
We have shown that in Q-learning algorithms pseudorehearsal
can improve performance significantly. [1] and now want to
test it on more interesting and complex actor-critic algorithm.
Actor-critic methods are one of the types of reinforcement
learning model-based algorithms based on TD-learning.
Actor-critic agents have a separate memory structure to
explicitly represent the policy independent of the value
function. Policy function is used to choose the next action.
A value function evaluates the state reached by agent. Actorcritic agent takes an action chosen by the policy structure
- actor. This actor is a probabilistic function from state to
action. After that the agent receives reward and the critic
consequently evaluates the action done. Finally, the critic’s
evaluation is used to update the actor. As a critic has two
interacting neural networks, an actor-critic’s vulnerability to
catastrophic forgetting may be very serious.
II. T HEORETICAL
ISSUES
A. Actor-critic algorithm
Actor-critic approaches achieves a high performance because they require minimal computation in order to select
actions – if the policy is explicitly stored, no computations
are needed for action selection. Actor-critics can also learn an
explicitly stochastic policy which is very useful in continuous
learning problems which are common for reinforcement learning. Actor-critic algorithms, as stated above, uses the actor to
choose actions based on the current state. The actor is typically
a policy gradient function. Policy gradient actors can take the
shape of a neural network and constructs distribution of actions
probabilities, och which one is then selected at each step.
Policy gradient methods has good convergence properties,
so they can reach its optimum quickly, but it does not
store information about the environment. This means its
performance is limited. Critics learn about and estimates
the policy which is currently being followed by the actor.
It provides a critique which takes the form of a TD error,
which drives all learning in both the actor and the critic.
The critic is typically a state-value function, which means it
keeps information about environment and can predict which
action will lead a better state. On the other hand the agent
needs to explore an environment for a longer time to reach
its optimum.[2] After each action selection critic determines
whether things have gone better or worse than expected. If
TD error is positive - tendency to select the last action done is
a for the future. If TD error is negative – this tendency should
be weakened. Both the actor and the critic can use neural
networks as a function approximation, which allows the
agent to execute tasks in continuous and partially observable
environments.
B. Catastrophic forgetting
Catastrophic forgetting is a common problem in neural
networks. The problem occurs when a neural network that
has properly learnt to execute some tasks meets changing
conditions or learns the new task. This neural network learns
the new task without any problems, but the old information
might be almost fully erased in the process. This might not
be a serious problem in supervised learning tasks like digit
recognition where network was trained ones and never retrained, but in reinforcement learning, where networks learning
repeats regularly, e.g. at each episode or even each time step catastrophic forgetting can seriously damage performance. [3]
During the online learning in continuous space things become
even worse. In this case the networks rarely receives real
feedback and most of its learning iterations are based on its
own assumptions. Even minor noise after one thousand step
can make network to forget everything it has learnt.
The cause of catastrophic forgetting is based on the mathematical nature of neural networks. In linear networks that
occurs because neural networks base their prediction from
input data on the vector orthogonality. Two different sets
with low orthogonality make the same neurons return different
outputs on similar inputs, so that the learning of one set erases
knowledge about the other. In the non-linear case the situation
is not so clear, but tends to be similar - there is significant
catastrophic forgetting when the information is very distributed
and highly overlapping between sets. [4]
C. Pseudorehearsal
One of the methods or solving catastrophic forgetting is
pseudorehearsal. It uses a two-step process: the first step is
construction of the set of pseudopatterns and the second is
training the network on pseudopatterns combined in batches
with real patterns. The common way of pseudopattern construction is creating a set of pseudovectors, feeding them
through the network and saving the outputs at each layer.
This approach saves memory compared to rehearsal-based
approaches because no real examples needed to be kept
for catastrophic forgetting elimination. This approach doesn’t
involve any changes to the network structure as some other
approaches do.
The generated approximations of the real data are sufficiently
accurate in practice to reduce forgetting. Even extremely crude
generative models have proven highly effective. In the original
work in this area by [5], pure noise fed to the network was
able to almost completely eliminate catastrophic interference.
The argument of the authors was that, although the input is
completely random, the activation distributions in deeper levels
of the network will be representative of the learnt input data.
Pseudorehearsal methods have been demonstrated to significantly decrease and almost completely eliminate the catastrophic forgetting in unsupervised learning [5], supervised
learning [6] and reinforcement learning [7]. It is interesting
to note that the results of Baddeley suggest that the widely
studied ill conditioning might not be the main bottleneck of
reinforcement learning after all. Instead, their results indicate
that the catastrophic forgetting is the main bottleneck for
reinforcement learning problems.
We will try two basic pseudorehearsal approaches during our
research. The one is using pseudopatterns for the correction
of learning weights in learning with respect to orthogonality
between the learned example and pseudovectors. The other
is learning the network in classic batch-backpropagation way
with one real example in a batch and others are pseudopatterns.
We will vary size of the pseudopatterns batch and frequency
of their reinitialization to find the ones which improve performance the most.
III. E XPERIMENTAL DESIGN
We apply pseudorehearsal algorithms to an actor-critic agent
executing real reinforcement learning task. The task is a double
pole-balancing problem, well-known reinforcement learning
task mentioned for example by Sutton. [8]. The task of agent
is to balance a two poles installed on a cart for as long
as it possible by pushing the cart left or right. There is no
positive reward in the problem and negative reward is given
if any of poles falls down or the cart reaches the end of
track. The primary result of these experiment is the number
of steps for each agent can balance the pole. The bigger
this number the better. Another secondary output is a time
spent for computations. We want to know how good was the
performance improvement compared to computational cost.
1) Observation: Two different observations are used for
experimental comparison. The first observation type given
to the agents constitute a fully observable Markov decision
process - the agent knows all the information about the current
cart’s state: cart position, velocity and acceleration and each
poles’ falling angle, angular velocity and angular acceleration.
Therefore the best possible behaviour can be easily calculated
through analytical methods, while the complex structures as
neural networks create a huge amount of data redundancy
which leads to a high amounts of catastrophic forgetting. For
this reason this observation type is a very good one for testing
tools of elimination of catastrophic forgetting.
The other observation type is a partially observable Markov
decision process. In this case agent receives only information
about current position of the cart and the poles. The neural
networks are used appropriately - modelling the unknown
environment. With this type of observation we might see how
good is catastrophic forgetting in more realistic settings.
The observation is represented as a real valued vector, where
each ith observed parameter is written into one of two vector
cells: 2 ∗ ith if the parameter is positive or (2 ∗ i + 1)th if
parameter is negative. The second vector entity assigned to
parameter is assigned to zero. After that, the linear parameters are divided by 20 and angular are divided by 60 for
normalization
.
2) Performance metric: Agent tries to complete the task
for 1000 runs, then the number of successful steps and
computation time in each try is saved. After that we can
use different straight comparisons and statistical approaches
to compare the results.The first step of straight comparisons is
just subtraction of the one resulting vector from another, let’s
call the resulting vector difference vector. The next simple
comparison practices are: drawing episode/step graphs for
visual evaluation of agent’s behaviour; drawing similar graphs
but with mean of 10-20 consecutive steps instead of real values
for evaluation of agent’s tendencies, drawing these graphs for
difference vectors. Of course these are just preparation of data
for further statistical processing.
The main measures we use to analyze resulting vectors are the
mean of the all resulting vector entities, the median and the
root mean squared deviation.
The mean of the vector tells us how good was the performance overall, the higher value the better. Root mean squared
deviation tells us how strong was the influence of catastrophic forgetting, because cases of catastrophic forgetting
intervention can be noted by significant change of performance
between steps - usually to the side of decreasing.
The median denotes how good is agent at removing bad
information. Comparison of mean and median can give much
information about the agent, which we can deduce from the
nature of this parameters. If the median is higher than the
mean, then this means that the largest deviations are on the side
of lower than average performance. Therefore, the agent learns
successfully and the catastrophic forgetting occasionaly has a
strong impact, but this influence is then quickly eliminated.
If the mean is higher then median - then while the agent
tries to choose optimal policy, sometimes successfully, most
of its runs are influenced by catastrophic forgetting being
too high. If the mean and median are nearly the same then
learning and forgetting counterbalance each other. Of course
the higher absolute values of mean and median denote the
higher overall performance of the agent and very important
too. However, in some reinforcement learning tasks where the
cost of mistake is high, like drone flight, we might choose the
approach with worse performance, but with better catastrophic
forgetting elimination, because reaching the goal slower is
better then fast breaking of the machine in process. After the
application of all this methods we need to apply significance
test - like student’s t-test to the results to see if our results are
significant and therefore if we can make statement based on
research.
IV. C ONCLUSION
This work will show us if the actor-critic algorithm is
vulnerable to the catastrophic forgetting, and how good is
pseudoreharsal in decreasing of this problem in the case
of two interacting neural networks. We will find the best
possible parameters of pseudorehearsal approaches and try to
find dependency between the pseudorehearsal parameters and
agent’s performance and to explain this dependency.
If the pseudorehearsal well be proved to successfully eliminate
catastrophic forgetting for continuous actor-critic algorithms,
many reinforcement learning tasks will become easier to solve,
and let agents quickly react on the instant changes in the
environment. So this research may widen the sphere of the
applications of reinforcement learning agents.
R EFERENCES
[1] V. Marochko, L. Johard, and M. Mazzara, “Pseudorehearsal in value
function approximation,” in 11th KES International Conference, KESAMSTA 2017 Vilamoura, Algarve, Portugal, June 2017 Proceedings,
2017.
[2] J. Beitelspacher, J. Fager, G. Henriques, and A. McGovern, “Policy
gradient vs. value function approximation: A reinforcement learning
shootout,” School of Computer Science, University of Oklahoma, Tech.
Rep, 2006.
[3] A. Cahill, Catastrophic Forgetting in Reinforcement-Learning Environments. PhD thesis, University of Otago, 2011.
[4] O.-M. Moe-Helgesen and H. Stranden, “Catastophic forgetting in neural
networks,” Dept. Comput. & Information Sci., Norwegian Univ. Science
& Technology (NTNU), Trondheim, Norway, Tech. Rep, vol. 1, p. 22,
2005.
[5] A. Robins, “Catastrophic forgetting, rehearsal and pseudorehearsal,” Connection Science, vol. 7, no. 2, pp. 123–146, 1995.
[6] R. Ratcliff, “Connectionist models of recognition memory: constraints
imposed by learning and forgetting functions.,” Psychological review,
vol. 97, no. 2, p. 285, 1990.
[7] B. Baddeley, “Reinforcement learning in continuous time and space:
Interference and not ill conditioning is the main problem when using
distributed function approximators,” IEEE Transactions on Systems, Man,
and Cybernetics, Part B (Cybernetics), vol. 38, no. 4, pp. 950–956, 2008.
[8] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction.
MIT press Cambridge, 1998.
| 2cs.AI
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.